Datasets:

ArmelR

/

the-pile-splitted

like 21

=6.5in =-0.5in i **CHARM CORRELATION AS A DIAGNOSTIC PROBE** .5cm **OF QUARK MATTER** 1.5cm **Rudolph C. Hwa** .5cm Institute of Theoretical Science and Department of Physics University of Oregon, Eugene, OR 97403 **Abstract** > The use of correlation between two open-charm mesons is suggested to give information about the nature of the medium created in heavy-ion collisions. Insensitivity to the charm production rate is achieved by measuring normalized cumulant. The acollinearity of the $D$ momenta in the transverse plane is a measure of the medium effect. Its dependence on nuclear size or $E_T$ provides a signature for the formation of quark matter. The conventional probes of quark matter in heavy-ion collisions, such as dileptons and $J/\psi$ suppression, have not provided conclusive evidence about the creation or absence of quark-gluon plasma [@bm]. A large part of the difficulties involve the ambiguities arising from competing processes and from uncertainties in the initial normalizations of some key quantities. An effective probe should be free of such ambiguities. In this paper we suggest the possibility that charm correlation may be such a probe. The correlation proposed is between open-charm mesons at nearly opposite directions. The charm quark is used both for tagging and for probing. Heavier quark can be used when appropriate; charm will be used as a generic term for heavy quark in the following discussion. Lighter partons and associated minijets are too copiously produced at RHIC and LHC [@kk] to be useful for our purpose of tagging and probing. The idea is based simply on the dual requirements that the signature should be independent of the production rate but sensitive to the medium through which the probe traverses. Appropriately normalized cumulant can satisfy the first requirement, while the transverse deviation from exact back-to-back correlation meets the second. Briefly stated, it is suggested that one searches for $D\bar{D}$ produced in the transverse plane at $y=0$ with their momenta nearly collinear, but opposite. The acollinearity in the transverse plane is the measure of interest. To enhance the effect, experimental cuts should be made on the magnitudes of the $D$-meson momenta so that they are nearly equal and not too large. We expect that the mean acollinearity is smaller if the medium is deconfined quark matter than if it is not. The proposed measure is similar in spirit to the acoplanarity of jets suggested by Appel [@da], but significantly different in substance. The major differences are: (a) the jet axes cannot be as precisely determined as the $D$-meson momenta, (b) at high energy too many jets are produced resulting in contamination and deterioration of the correlation signal, (c) we emphasize the difference between the propagation of a $c$ quark through a deconfined medium and that of a $D$-meson through a confined medium, and (d) the phenomenology of $D\bar{D}$ correlation can reveal interesting physics even in kinematic regions where perturbative QCD (pQCD) is unreliable. Let and be the momenta of the two detected charge-conjugate $D$ mesons. The cumulant is $$\begin{aligned} c(\vec p_1,\vec p_2) = \rho_2(\vec p_1,\vec p_2) - \rho_1(\vec p_1)\rho_1(\vec p_2), \label{1}\end{aligned}$$ where $\rho_n$ is the $n$-particle distribution function; it is the irreducible part of the two-particle correlation. In a heavy-ion collision the cumulant can in general be expressed in the form $$\begin{aligned} c(\vec p_1,\vec p_2) =\int {d^3k_1 \over k^0_1} {d^3k_2 \over k^0_2} S(\vec k_1,\vec k_2)H(\vec k_1,\vec p_1)H(\vec k_2,\vec p_2), \label{2}\end{aligned}$$ where $S(\vec k_1,\vec k_2)$ is the probability of producing two partons with momenta and , and $H(\vec k_i,\vec p_i)$ is the hadronization function that connects the parton $i$ at the point of creation to the hadron detected with momentum $\vec p_i$. Hereafter we adopt the convention of using the symbol $k\ (p)$ for parton (hadron) momentum. It should be stressed that $H$ is not simply the fragmentation function usually used for jet considerations because firstly the produced parton must traverse a dense medium and suffer momentum degradation before fragmentation, and secondly the hadronization process may be recombination [@kd; @rh] instead of fragmentation. In fact, it has been shown that the data of open charm production in the forward region of hadronic collisions can be well described by recombination [@rh2], but badly by pQCD or fragmentation model [@ga]. In the domain where pQCD is reliable one can write $S(\vec k_1,\vec k_2)$ for $AB$ collision as $$\begin{aligned} S(\vec k_1,\vec k_2) =c \int {d^3k_a \over k^0_a} {d^3k_b \over k^0_b} F_A (k_a) F_B(k_b) \delta^4 (k_a+k_b-k_1-k_2) \left|M(a+b\rightarrow1+2)\right|^2 \label{3}\end{aligned}$$ plus other terms of similar structure, if more than one hard subprocess are important. In (\[3\]) $c$ is a numerical constant, $F_A(k_a)$ is the parton distribution in nucleus $A$, and $M(a+b\rightarrow1+2)$ is the amplitude of the hard subprocess involved. There is a great deal of physics contained in the determination of $F_{A,B}$, which depends on nucleon structure function at small $x$, gluon distribution, nuclear shadowing, initial-state radiation, preequilibrium and possibly thermal interactions, space-time evolution, etc. So much uncertainty is involved in the problem that the study of open-charm production has been suggested as a means to learn more about the parton dynamics in the early phase of nuclear collision [@zl; @pl], i.e., the reverse of using $F_{A,B}$ to predict measureable quantities. While that is certainly a worthwhile project to pursue, our proposal here is to circumvent all that complication and proceed with the use of $S(\vec k_1,\vec k_2)$ independent of the details about $F_{A,B}$. From $S(\vec k_1,\vec k_2)$ not only can two-particle inclusive distribution be determined as in (\[2\]), the one-particle distribution can also be obtained as follows $$\begin{aligned} \rho_1 (\vec p_2) = \int {d^3 k_1 \over k^0_1} {d^3k_2 \over k^0_2} S(\vec k_1,\vec k_2) H(\vec k_2,\vec p_2) . \label{4}\end{aligned}$$ To free our signal from the uncertainties of the primordial parton dynamics, let us define the singly-normalized cumulant function $$\begin{aligned} C(\vec p_1,\vec p_2) = c(\vec p_1,\vec p_2)/ \rho_1 (\vec p_2). \label{5}\end{aligned}$$ It is clear from (\[2\]) and (\[4\]) that $C(\vec p_1,\vec p_2)$ should be insensitive to the rate of charm production. Hereafter we shall regard hadron 2 (with momentum $\vec p_2$) as the trigger particle, against which we study the properties of the probe particle 1. Of course, it is the relative momentum between the trigger and probe that is important, but conceptually it is efficient to identify (arbitrarily) one of the two $D$ mesons as the trigger and define the axes such that the trigger momentum in every event is always aligned along a fixed direction, say $-\hat x$ axis, with $\hat y$ being the axis normal to the scattering plane containing the beams and the trigger. The aim is to study the momentum distribution of the other $D$ meson in the neighborhood of the $+\hat x$ axis. Since the parton momenta and can have large longitudinal imbalance due to unequal momenta, $\vec k_a$ and $\vec k_b$, of the initial colliding partons, but they can have only limited total transverse momentum $\vec K_T = \vec k_{1_T} + \vec k_{2_T}$ due to the small intrinsic transverse momenta of the initial partons and to the initial-state radiation, we can avoid the complexity of the full structure of $S(\vec k_1,\vec k_2)$ if we restrict and to only the near neighborhood of a common transverse plane. That is achieved by requiring that and lie only in the transverse plane at $y=0$. For brevity we shall refer to that plane as . Since we expect the angular differences between $\vec k_i$ and $\vec p_i$ to be small, that requirement therefore forces $\vec k_1$ and $\vec k_2$ to be very close to  also. The aim of this problem is to learn about the medium effect through $H(\vec k_i,\vec p_i)$, on which the measureable $C(\vec p_1,\vec p_2)$ depends. With $\vec k_i$ near  those partons do not participate in the longitudinal expansion of the system, which is another area of large uncertainties. But even in  there are several possible processes leading from the partons $(k_i)$ to the hadrons $(p_i)$, each involving a different hadronization function $H(\vec k_i,\vec p_i)$. So far we have not specified the kinds of partons carrying $k_i$. They can be high-momentum quarks of the $u$ and $d$ types, or lower-momentum gluons, all capable of fragmenting into the $D$ mesons. Since the two fragmentation processes are independent, $\vec p_1$ and $\vec p_2$ are not correlated even though $\vec k_1$ and $\vec k_2$ are. To narrow down the hadronization process we make use of the experimental freedom to require further that the magnitudes $p_1$ and $p_2$ are nearly equal within a narrow range. Moreover, that magnitude should not be too high, say, in the $2-5$ GeV range. In so doing we can maximize the contribution from $c \bar {c}$ pair creation to the formation of $D \bar {D}$. There are several stages of reasoning involved here, which we now describe. The possible parton types are $q$ and $c$, where $q$ denotes $u$, $d$, or $g$ collectively. Let the flavor labeling not be encumbered by concerns about quark or antiquark differences. We postpone our consideration about the $s$ quark until later. The possible hadronization processes are fragmentation $(F)$ and recombination $(R)$, for which $k>p$ in $F$, but $k<p$ in $R$. Thus for the production of $D$ there are four possible processes: $F(q \rightarrow D)$, $F(c \rightarrow D)$, $R(q \rightarrow D)$, and $R(c \rightarrow D)$. At high collision energies there are so many hard subprocesses [@kk] that there are enough transversly moving partons to make recombination competitive with fragmentation in the formation of $D$ in . That is not the case in $pp$ or $pA$ collisions. In fact, for any given momentum $p$ of $D$ in , it is more favored to recombine two lower momenta partons to add up to $p$ than to create a higher momentum parton which subsequently decays to $p$, since the probability of creating high $k_T$ partons falls off as a power $k_T^{-\alpha}$, with $\alpha >4$. For single-particle inclusive distribution a comparison between the two hadronization processes in the production of particles in  has been studied quantitatively [@rh] with the result that for $p_T<6$ GeV/c and for large nuclei the rate of hadronization through recombination is at least an order of magnitude higher than through fragmentation. For a pair of correlated particles the $R/F$ ratio of the rates would be squared. Thus we may ignore $F(q \rightarrow D)$ and $F(c \rightarrow D)$ in the following discussion. Since $m_c \approx 5\,m_q$ for constituent quark masses, the momentum fractions $x_c$ and $x_q$ of $c$ and $q$, respectively, in $D$ are on the average very different, with $x_c \approx 5\,x_q$; hence, the recombination function for $c+q \rightarrow D$ is maximum when $\vec k_c\approx 5\,\vec k_q$ in the same direction [@rh2; @rh3]. For and nearly equal and opposite, the production of $D \bar{D}$ is therefore dominated by the creation first of $c \bar{c}$ pair with $\vec k_1\approx -\vec k_2$ followed by recombination with low-momentum $q$ quarks, rather than by the process where a created $q \bar{q}$ pair dictates the momenta of the $D$ mesons. In short, $R(c \rightarrow D)$ is more important than $R(q \rightarrow D)$. In the following we shall focus on the process where $S(\vec k_1, \vec k_2)$ in (\[2\]) describes the hard production of a $c \bar{c}$ pair, and the two $H$ functions represent $R(c \rightarrow D)$. The only part in the problem that has a firm theoretical footing is the amplitude $M$ for hard scattering in (\[2\]), which is calculable in pQCD. Even there, charm production with $k_T \sim 2$ GeV/c is in the grey area of reliability. As mentioned earlier, the parton distribution $F_A(x_a)$ and $F_B(x_b)$ are quite uncertain, but the normalized cumulant $C(\vec p_1, \vec p_2)$ is insensitive to all of them, including $M$; furthermore, we emphasize the misalignment of and , their magnitudes being selected by experimental cuts. Thus the signature we seek depends mainly on $H(k_i \rightarrow p_i)$, which is sensitive to the medium that stands between the creation of $c \bar{c}$ and the detected hadrons $D \bar{D}$. That is just what a good probe should be. It is unfortunate that $H(k_i \rightarrow p_i)$ cannot at this point be calculated precisely in QCD, perturbative or otherwise. However, the discovery of unambiguous experimental signature is more important than having reliable theoretical calculations at this stage. On the basis of reasonable arguments we indicate below what that signature might look like. The four vectors $\vec k_i$ and $\vec p_i$ $(i=1,2)$ are all very close to . We consider below only their projections $\vec k_{i_T}$ and $\vec p_{i_T}$ on . For brevity we shall omit the subscripts $T$, unless there is confusion. Let the angles among these four vectors be labeled as follows: $\theta_1 (\vec k_1, \vec p_1)$, $\theta_2 (\vec k_2, \vec p_2)$, $\theta_{12} (\vec k_1, -\vec k_2)$, and $\phi (\vec p_1, -\vec p_2)$. For simplicity let the distributions in these angles be represented by Gaussians: ${\rm exp} (-\theta_1^2/ \lambda_1^2)$, ${\rm exp} (-\theta_2/ \lambda_2^2)$, ${\rm exp} (-\theta_{12}^2/ \lambda_{12}^2)$, and ${\rm exp} (-\phi^2/ \lambda^2)$, respectively. The distribution in $\phi$ is a Gaussian also because of the convolution theorem, since (\[2\]) implies $$\begin{aligned} c(\vec p_1, \vec p_2) \propto \int d\theta_1 d\theta_2\,{\rm exp} \left( - {\theta_1^2 \over \lambda_1^2} -{\theta_{12}^2 \over \lambda_{12}^2}- {\theta_2^2 \over \lambda_2^2}\right) , \label{6}\end{aligned}$$ where $\theta_{12}= \phi - \theta_1 - \theta_2$. Thus if $\lambda_1$, $\lambda_2$ and $\lambda_{12}$ are all small compared to $\pi$, then $$\begin{aligned} \lambda^2 = \lambda_1^2 + \lambda_2^2 + \lambda_{12}^2 . \label{7}\end{aligned}$$ One can improve on these integrations with better knowledge on $S(\vec k_1, \vec k_2)$ and $H(\vec k_i, \vec p_i)$, but the result will not differ much from (\[7\]), which encapsules the property that the width in $\vec p_1\cdot$is the rms sum of those in $\vec p_1\cdot$, $\vec k_1\cdot$and $\vec p_2\cdot$. Since we require $|\vec p_1| \approx |\vec p_2| \equiv p$, we may write $C(\vec p_1, \vec p_2)$ as $C(p, \phi)$. Then we have $$\begin{aligned} C(p, \phi) = C(p)\,e^{-\phi^2/ \lambda^2} . \label{8}\end{aligned}$$ Alternatively, one can define $\vec p_1=-\vec p_2+\vec p_t$ (note the small t), where $\vec p_t$ is nearly normal to the trigger axis but still in . Thus experimentally, the data on $C(\vec p_1, \vec p_2)$ can be presented as distributions in $\xi \equiv p_t/p$ for various values of the trigger momentum $p$. Empirical $C(p, \xi)$ should be sharply peaked in $\xi$, though not necessarily Gaussian; the width can be represented by $\lambda$ in (\[8\]) in the following discussion. In order to keep all the quantities in (\[7\]) small, $p$ should not be too small; neither should it be too large so that the rate of producing two correlated $D$ mesons does not become too low. The range $2<p<5$ GeV/c appears to be reasonable. In that range the $c$ and $\bar{c}$ quarks are beams in the dense medium with momenta $k_1$ and $k_2$, both $< p$, insufficient to give any hope to the validity of pQCD in describing their passages through the medium. Nevertheless, we expect charm quarks in that momentum range to be sensitive to the medium effects and can provide us with useful information. We now consider the various contributions to (\[6\]). Nonzero $\theta_{12}$ means that $\vec K \equiv \vec k_1 + \vec k_2 \neq 0$. Nonzero $\vec K$ is due to the intrinsic transverse momenta of the partons plus the recoil from initial-state radiation before the hard subprocess that creates the $c \bar{c}$ pair. We expect $|\vec K|$ to be small, and set its average value at $0.3$ GeV/c for low $k_i$ and allow it to be higher for higher $k_i$. Approximating $k_1=k_2$ and denoting them collectively as $k$, we have cos$\theta_{12} =1-K^2/2k^2$, and $\theta_{12} \simeq K/k$. Thus for $k>2$ GeV/c, we have $\lambda_{12} \leq 0.15$. More importantly, $\lambda_{12}$ is independent of the medium in which the created $c$ quarks will traverse, and will therefore not affect our signature of the medium effect. Next, we consider the last two terms of (\[6\]) due to the $H$ functions in (\[2\]), which are the heart of the problem. The two $H$ functions independently describe the hadronization processes from $\vec k_i$ to $\vec p_i$, as they proceed along the paths $l_1$ and $l_2$, through the medium in essentially opposite directions in  from the point of creation of $c \bar{c}$. Since the result we seek concerns the sum $\lambda_1^2 + \lambda_2^2$ in (\[7\]), it is equivalent to that due to one $c$ quark traversing the entire path $L= l_1+ l_2$ from one end of the medium to the other in , passing through the point of $c \bar{c}$ creation. The transverse expansion of the medium is not rapid. If $R$ is the average radius of the system that is relevant to this study, then when averaged over all possible points of the $c \bar{c}$ creation in  and over all orientations of the path, the mean path length $L$ of crossing a disc of radius $R$ is $L= 4R/3$. Our consideration is thus reduced to the hadronization of a $c$ quark with initial momentum $\vec k$ passing through a slab of the medium of thickness $L$ and emerging as a $D$ meson with momentum $\vec p$. There are two scenarios to examine. One is that the medium consists of deconfined quarks and gluons, while the other is of high-density hadrons with quarks confined. Let them be referred to as quark matter (QM) and hadron matter (HM), respectively. Of course, they represent two extreme cases, and other scenarios that stand between them are possible. If we know the nature of the signatures for the extreme cases, what happens in the intermediate cases can be estimated by interpolation. Thus for now it is sensible to consider just QM and HM. .3cm 1. [*Quark Matter*]{}.Being deconfined, the medium cannot support the formation of any hadrons nor the existence of any color flux tubes in it. Thus the $c$ quark that traverses the QM medium remains as a $c$ quark. It may lose momentum and deviate from straightline path, but the formation of $D$ can occur only after passing through the medium by recombining with a $\bar{q}$ at the exit point. If pQCD were applicable, one could study in detail the effect of multiple scattering as in [@xw; @mt], and determine the degree of energy loss and $k_t$ gain. Hereafter we use $k_t$ to denote parton momentum transverse to $\vec k$. However, for $k$ as low as $2$ GeV/c, the reliability of pQCD is questionable. Qualitatively, one expects the radiative energy loss to be reduced for heavy quarks compared to light quarks. We assume that the loss of longitudinal momentum is roughly compensated by the gain in momentum due to recombination so that $p \approx k$. The more important aspect of the problem is the $k_t$ gain. If one regards the result of [@xw] that takes the LPM effect into account as being valid, then $k_t$ due to gluon radiation is of the order of the color screening mass $\mu$, independent of the number of multiple scatterings. The cumulative effect on $k_t$ due to elastic scatterings depends on whether the random-walk model is valid or the quantum coherence effect is important. In the former case the process is Markovian and $k_t$ would increase with $L$, while in the latter case it would not. The coherent LPM effect in the longitudinal component is non-Markovian. It has been known that the former is more relevant to conventional large-$p_T$ processes, while the latter is for low-$p_T$ processes. Our $H(c\rightarrow D)$ in the present problem belongs to the latter category. The following experimental facts support the latter. Exhaustive studies of $h_1 A \rightarrow h_2X$ inclusive reactions at high energies and low $p_T$ have revealed that, for $h_2$ in the beam fragmentation region of $h_1$, the $p_T$ distribution of $h_2$ is essentially independent of $A$ [@ps; @db]. Because of the flavor dependence of the $h_2$ distribution, the processes can be well interpreted in the parton picture by considering the valence quark that is common in $h_1$ and $h_2$ [@kd; @rh3; @lg]. Take $h_1=\pi^+$ and $h_2=K^+$ to be specific. For $K^+$ with high $x_F$, it is the high momentum $u$ quark in $\pi^+$ that leads to $K^+$ by recombination. The $p_T$ of $K^+$ reflects the $k_T$ of the $u$ quark after traversing the target. The independence of $\langle p_T\rangle$ on $A$ implies that there is no $k_T$ broadening due to multiple-scattering effect on the $u$ quark. Specifically, the data of [@db] indicate that $\sigma (p_T=0.5)/\sigma(p_T=0.3)$ stays essentially uncharged at 0.5 and 0.48 for $A=Cu$ and $Pb$, respectively, for $p_{\pi^+}=100$ GeV/c and $p_{K^+}=80$ GeV/c. Although it is a $u$ quark traversing normal nuclei, whereas our problem involves a $c$ quark traversing dense QM, the independence of $k_t$ on the path length is likely to be due to a common origin: quantum coherence in low-$k_t$ processes is non-Markovian, so $k_t$ does not increase with $L$. The same lack of $A$ dependence is found in [@ps] for $h_1=p$ and $h_2=\Lambda$. We emphasize that it is true only at low $p_T$. Significant $A$ dependence at high $p_T$ is not excluded, such as in the production of massive dileptons in $pA$ collisions [@bor; @ald], where no $A$ dependence is seen until $p_T$ exceeds 2 GeV/c. On the basis of these arguments we adopt the following position. Firstly, the magnitude of $k_t$ is small, since heavy quark suffers less deflection; we take it to be of order $\mu$, which according to $\mu^2= 4\pi \alpha_s T^2$ gives $\mu \approx 0.4$ GeV for $\alpha_s=0.3$ and $T= 200$ MeV. Secondly, $k_t$ is expected to be independent of $L$, although a gentle increase with $L$ cannot be ruled out on firm theoretical ground. Since the recombination of $c$ with a $\bar {q}$ to form a $D$ after passing through the QM does not increase $p_t$ beyond the $k_t$ gained, we arrive at the result that $\lambda_1^2 + \lambda_2^2 \approx \lambda_{12}^2$. Consequently, we have $\lambda \approx 0.5/p$, where $p$ is in units of GeV/c. It is clear that if we want $\lambda$ to be small, $p$ should not be small. That is why we have set $p>2$ GeV/c. .3cm 2. [*Hadron Matter*]{}.Consider now the scenario where the created $c \bar{c}$ pair find themselves in a densly packed hadronic medium. The recombination of $c$ with $\bar{q}$ and $\bar{c}$ with $q$ take place rapidly, and it is the $D$ and $\bar{D}$ that traverse the HM in opposite directions. As before, the combined effect on and can be represented by a $D$ of momentum $p$ going through a slab of HM of thickness $L$. Being a low-momentum hadron ($2<p<5$ GeV/c) the $D$ interacts strongly with the mainly pionic medium. The number of multiple collisions in $L$ is $\nu =n_\pi \sigma_{D\pi} L$, where $n_\pi$ is the pion density. The energy dependence of $\sigma_{D\pi}$ is not known, but its magnitude (in the few mb range) is definitely much greater than partonic cross section. Since $\lambda_1^2 + \lambda_2^2$ is proportional to $\nu$, we therefore expect $p_t$ to be larger (compared to the QM case) and to increase significantly with $L$. Herein lies the major difference between the two medium effects on $\xi =p_t/p$. Since $\lambda_{12}$ is the same for the two media, the difference does not depend on it. .3cm Putting together the above considerations leads to our suggestion for the signature of QM vs. HM. Measure $C(p,\xi)$. Plot the $\xi$ dependence for fixed $p$ and determine the mean $\bar{\xi}$. Examine how $\bar{\xi}$ depends on $E_T$ and $A$. For HM, $\bar{\xi}$ should be large and increase with $E_T$ and $A$. If, at high enough $s$, $E_T$ and $A$, quark matter is created, then $\bar{\xi}$ should drop down to a low value and become essentially independent of $E_T$ and $A$. This transition from high and increasing $\bar{\xi}$ to low and roughly constant $\bar{\xi}$ is the signature of QM formation. In short, quarks are smaller than mesons; their difference should be revealed in the measurement of $\bar\xi$. It would also be of interest to study the $p$ dependence of $\bar{\xi} (p)$ . If $\sigma_{D\pi}$ increase with $p$, then $\bar{\xi}_{\rm HM}$ would be constant in $p$, in contradistinction from $\bar{\xi}_{\rm QM}(p)$ which decreases with $p$. Another interesting problem arises when the recombination of $c\ (\bar{c})$ with $\bar{s}\ (s)$ is considered. Since in QM the hadronization occurs outside the quark phase, the width $\bar{\xi}_{D_s \bar{D}_s}$ for $D_s \bar{D}_s$ correlation should not differ significantly from $\bar{\xi}_{D \bar{D}}$. However, in HM the formation of $D_s$ and $\bar{D}_s$ occurs inside the medium, and because of the suppressed $D_s \pi$ scattering due to the OZI rule at low energy, we expect $\bar{\xi}_{D_s \bar{D}_s}<\bar{\xi}_{D \bar{D}}$. The observation of these differences would add to our understanding of what occurs in these systems. What is described in this paper is for idealized systems. In reality the system may be much more complicated and the transition of the behavior of $\bar{\xi}$ may be very gradual. If so, it is unlikely that any other signature would be clear-cut, since it is the system itself that is not sufficiently distinctive. Our proposal deals with the nature of the matter probed, independent of other inessential complications, such as the absolute normalization of the charm production rate, the precise value of the initial temperature, or the validity of pQCD. To have a proper theoretical treatment of the problem is essential ultimately, but for now the need for a distinctive experimental signature seems to be more urgent. Even if the suggested signature turns out to be ineffective because of the complexity not considered in this initial investigation, charm correlation should nevertheless reveal much information about heavy-ion collisions not available so far. I am grateful to X.N. Wang for helpful discussions. This work was supported, in part, by the U.S. Department of Energy under grant No. DE-FG06-91ER40637. B. Müller, Nucl. Phys, A[**590**]{}, 3c (1995); M. Gyulassy, [*ibid*]{}. A[**590**]{}, 431c (1995). K. Kajantie, P.V. Landshoff, and J. Lindfors, Phys. Rev. Lett. [**59**]{}, 2527 (1987); K.J. Eskola, K. Kajantie, and J. Lindfors, Nucl. Phys. B[**323**]{}, 37 (1989); X.N. Wang and M. Gyulassy, Phys. Rev, D[**44**]{}, 3501 (1991). D.A. Appel, Phys. Rev. D[**33**]{}, 717 (1986). K.D. Das and R.C. Hwa, Phys. Lett. [**68**]{}B, 459 (1977); R.C. Hwa, Phys. Rev. D[**22**]{}, 1593 (1980). R.C. Hwa, Phys. Lett. B[**276**]{}, 497 (1992). R.C. Hwa, Phys. Rev. D[**51**]{}, 85 (1995). G.A. Alves [*et al*]{}., Phys. Rev. Lett. [**72**]{}, 812 (1994). Z. Lin and M. Gyulassy, Phys. Rev. C[**51**]{}, 2177 (1995). P. Lévai, B. Müller, and X.N. Wang, Phys. Rev. C[**51**]{}, 3326 (1995). R.C. Hwa, Phys. Rev. D[**27**]{}, 653 (1983). X.N. Wang, M. Gyulassy, and M. Plümer, Phys. Rev. D[**51**]{}, 3436 (1995). M. Thoma, in [*Quark-Gluon Plasma 2*]{}, edited by R.C. Hwa (World Scientific, Singapore, 1995). P. Skubic [*et al*]{}., Phys. Rev. D[**18**]{}, 3115 (1978). D.S. Barton [*et al*]{}., Phys. Rev. D[**27**]{}, 2580 (1983). L. Gatignon [*et al*]{}., Phys. Lett. [**115B**]{}, 329 (1982). P. Bordalo [*et al*]{}., Phys. Lett. [**B193**]{}, 373 (1987). D.M. Alde [*et al*]{}., Phys. Rev. Lett. [**66**]{}, 133 (1991).

ArXiv

The colossal magnetoresistance (CMR) effect [@Kusters] in doped manganese oxides has attracted a lot of attention. The interplay of charge, orbital and magnetic order results in a very rich phase diagram [@giant]. The parent compound LaMnO$_3$ has orthorhombic symmetry at low temperature. The Mn$^{+3}$ ion has $d^4$ ($t_{2g}^{3}$,$e_g^{1}$) configuration with an “inert” $t_{2g}$ core (spin 3/2) and a half-filled doubly degenerate $e_g$-type $d$ orbital which is Jahn-Teller (JT) unstable. Ignoring rotation of the MnO$_6$ octahedra, which occurs below 1010 K, the JT symmetry breaking is cubic to tetragonal [@Rodriguez] at ${\rm T}_{\rm JT}=$750 K. The corresponding orbital order [@Murakami] has $x$- and $y$-oriented E$_g$ orbitals alternating in the $x-y$ plane with wave vector $\vec{Q}=(\pi,\pi,0)$. This in turn causes layered antiferromagnetic (AFA) order to set in at T$_{\rm N}=$140 K. The electronic structure of LaMnO$_3$ has been studied, for example, by photoemission [@Park; @Dessau] and by first principles calculations [@Pickett]. Still there is controversy about the nature of the low energy excitations, arising from the interplay between strong on-site Coulomb repulsion (which leads to magnetic order) and strong electron-phonon (e-p) interactions [@Kanamori] (which lead to orbital order). When an electron is removed from the JT-ordered ground state, e-p coupling causes the hole to self-localize in an “anti-JT” small polaron state. In a previous paper [@Allen1] we have described the localized polaron in adiabatic approximation. Residual non-adiabatic coupling allows the hole to disperse with band width narrowed by Huang-Rhys factor $e^{-3\Delta/4\hbar\omega}\approx 10^{-4}$. The photoemission process is sudden. The emitted electron with wavevector $\vec{k}$ leaves a hole in a lattice “frozen” in the unrelaxed JT state. Ignoring lattice relaxation, this hole would disperse with band width $2t\approx$1 eV ($t$ is the hopping parameter), as shown on Fig. \[fig1\]. However, this is not a stationary state and must be regarded as a superposition of exponentially narrowed small polaron bands. Such bands have anti-JT oxygen distortions at each site, but a sufficient number of vibrational quanta are also excited such that the anti-JT distortion at time zero is “undone”. This is “Franck-Condon principle”. The measured spectrum at wavevector $\vec{k}$ will consist of a central $\delta$-function at the energy of the frozen lattice (dispersive) band $\varepsilon_{1,2}(\vec{k})$, plus multiple vibrational side-bands at energy $\varepsilon_{1,2}(\vec{k})\pm n\hbar\omega$, with an overall Gaussian envelope whose width is approximately the polaron binding energy. Franck-Condon broadening has been seen in photoemission spectra of solid nitrogen and oxygen [@Himpsel]. We find at each wavevector $\vec{k}$ the photoemission spectrum has an intrinsic Franck-Condon broadening indicated by error bars in Fig. \[fig1\]. The position of the maximum disperses with $\vec{k}$-vector close to the “frozen” lattice spectrum. A qualitative picture of this process has been given by Sawatzky [@Sawatzky] in the context of high temperature superconductors and by Dessau and Shen [@Dessau] for the manganites. The present paper gives an exact algebraic prediction for the Angle-Resolved-Photoemission-Spectra (ARPES) of a model Hamiltonian for LaMnO$_3$. Our model Hamiltonian [@Allen1], first introduced by Millis [@Millis], has hopping ${\cal H}_{\rm el}$, electron-phonon ${\cal H}_{\rm ep}$, and lattice ${\cal H}_{\rm L}$ energies: $$\begin{aligned} {\cal H}_{\rm el}&=&t \sum_{\ell,\pm} \left\{[c^{\dagger}_x(\ell) c_x(\ell \pm \hat{x})] + [x\rightarrow y] + [y \rightarrow z] \right\} \nonumber\\ {\cal H}_{\rm ep}&=&-g\sum_{\ell,\alpha} \hat{n}_{\ell,\alpha} (u_{\ell,\alpha}-u_{\ell,-\alpha}) \nonumber\\ {\cal H}_{\rm L}&=&\sum_{\ell,\alpha}(P^2_{\ell,\alpha}/2M +Ku^2_{\ell,\alpha}/2). \label{hmodel}\end{aligned}$$ In these formulas $c^{\dagger}_{\alpha}(\ell)$ creates a state with orbital $\psi_{\alpha}=|3\alpha^2-r^2>$, where $\alpha=x,y,z$. These three orbitals span the two dimensional $e_g$ subspace and can be expressed in terms of the conventional orthogonal basis $\Psi_2=d_{x^2-y^2}$, $\Psi_3=d_{3z^2-r^2}=\Psi_z$; specifically $\Psi_{x,y}=\pm\sqrt{3}/2\Psi_2-\Psi_3/2$. The resulting ${\cal H}_{\rm el}$ coincides with the nearest-neighbor two-center Slater-Koster [@Slater] hopping Hamiltonian with overlap integral $t=(dd\sigma)$ and $(dd\delta)=0$. The hopping parameter $t=0.5$ eV is chosen to agree with an [*ab initio*]{} $e_g$ band width of 1 eV [@Pickett]. The e-p interaction ${\cal H}_{\rm ep}$ is modeled by a linear energy reduction of an occupied $\psi_x$ orbital ($\hat{n}_{\ell,x}=c^{\dagger}_{x}(\ell)c_{x}(\ell)$) if the corresponding two oxygens in the $\pm \hat{x}$ direction expand outwards, and similarly for $\hat{y}$ and $\hat{z}$ oxygens if $\psi_y$ or $\psi_z$ orbitals are occupied. The strength of the e-p coupling $g$ determines the JT splitting $2\Delta=1.9$ eV, which is fitted to agree with the lowest optical conductivity peak [@Jung; @Allen2]. Static oxygen distortions $2u=\sqrt{2\Delta/M\omega^2}$=0.296 [Å]{} given by our model agree well with neutron diffraction data 0.271 [Å]{} [@Rodriguez]. For the lattice term ${\cal H}_{\rm L}$ we use a simplified model where oxygen vibrations along Mn-O-Mn bonds are local Einstein oscillators. The displacement $u_{\ell,\alpha}$ is measured from cubic perovskite position of the nearest oxygen in the $\hat{\alpha}$-direction to the Mn atom at $\ell$. The oxygen vibrational energy $\hbar\omega=\hbar\sqrt{K/M}=0.075$ eV is taken from Raman data [@Raman]. In addition there is a large on-site Coulomb repulsion U and a large Hund energy. These terms inhibit hopping except to empty sites where the t$_{2g}$ core spins are aligned correctly. In adiabatic approximation one can solve this problem for U=0 or U=$\infty$. Both cases give a good description of the observed cooperative JT order. When U=0, the ground state wavefunction is: $$|{\rm GS},0>=\prod_{\vec{k}}c_{\vec{k}1}^{\dagger} c_{\vec{k}2}^{\dagger}|{\rm vac}>. \label{gs0}$$ A JT gap $\approx 2\Delta$ opens and the lower two bands of energy $\lambda_{\vec{k}1}$, $\lambda_{\vec{k}2}$ are filled. The photohole as initially created has energy: $$\begin{aligned} \lambda_{1,2}^2=\Delta^2+t^2(2C_x^2+C_xC_y+2C_y^2)\pm t|C_x+C_y| \nonumber\\ \sqrt{\Delta^2+4t^2(C_x^2-C_xC_y+C_y^2)}, \label{bs}\end{aligned}$$ where $C_{x,y}=\cos k_{x,y}$ and $C_z$ not entering at T=0 K. These bands are shown in Fig. \[fig1\](a) as dashed lines. At this point, at least in principle, could proceed numerically to find the polaronic energy lowering. However, it is both easier and more realistic to switch to U=$\infty$. The ground state wavefunction: $$|{\rm GS}, \infty>=\prod_{\ell}^A c_X^{\dagger}(\ell) \prod_{\ell^{\prime}}^B c_Y^{\dagger}(\ell^{\prime})|{\rm vac}>, \label{gs1}$$ has orbitals $\Psi_{X,Y}=-(\Psi_3\mp\Psi_2)/\sqrt{2}$ occupied singly on interpenetrating $A$ and $B$ sublattices. This is a fully correlated state with zero double occupancy, while Eq. (\[gs0\]) is a band wavefunction in which two electrons are found on the same Mn atom with non-zero probability. For U$\geq$ 6t $\approx 3$ eV, the state (\[gs1\]) has lower energy than state (\[gs0\]) [@Allen1]. Neglecting creation of orbital defects with energy $\approx 2\Delta$ the “frozen” lattice approximation predicts photohole energies $\Delta+\varepsilon_{1,2}(\vec{k})$, where $\varepsilon_{1,2}(\vec{k})=\pm(t/2)(C_x+C_y)$. We need to add a non-adiabatic treatment of the e-p coupling. The effective Hamiltonian ${\cal H}_{\rm eff}={\cal H}_{\rm el}+{\cal H}_{\rm ep}+{\cal H}_{\rm L}$ for the single hole is: $$\begin{aligned} {\cal H}^{\rm A}_{\rm el}&=&\sum_{\ell \in A}\frac{t}{4} \bigl(d_Y^{\dagger}(\ell \pm x)d_X(\ell)+ d_Y^{\dagger}(\ell \pm y)d_X(\ell)\bigr) \nonumber \\ {\cal H}^{\rm A}_{\rm ep}+{\cal H}^{\rm A}_{\rm L}&=&\sum_{\ell \in A} d_X^{\dagger}(\ell)d_X(\ell) \biggl[\Delta+\sum_{\alpha}\kappa_{\alpha} \Bigl(a_{\alpha}(\ell)+ \nonumber\\ a_{\alpha}^{\dagger}(\ell)-b_{\alpha}(&\ell-&\alpha) -b_{\alpha}^{\dagger}(\ell-\alpha)\Bigr)\biggr] + \sum_{\alpha}a_{\alpha}^{\dagger}(\ell)a_{\alpha}(\ell). \label{eff}\end{aligned}$$ Here the operator $d^{\dagger}_X(\ell)=c_X(\ell)$ creates a hole in the JT ground state by destroying an electron on orbital $X$ at site $\ell$ (if $\ell \in A$ sublattice), and the operator $d^{\dagger}_Y(\ell)=c_Y(\ell)$ creates a hole on $B$ sublattice (if $\ell \in B$). The phonon operators $a_{\alpha}^{\dagger}(\ell)$ or $b_{\alpha}^{\dagger}(\ell)$ create vibrational quanta on the $\ell+\hat{\alpha}/2$ oxygen, if $\ell \in A$ or $\ell \in B$ respectively. The e-p coupling constants are $\kappa_{x,y,z}=\sqrt{\Delta/12}(1+\sqrt{3}/2 ; 1-\sqrt{3}/2 ; 1)$. The Hamiltonian and all other energy parameters $\Delta$ and $t$ in Eq. \[eff\] are in units of $\hbar\omega$. The total Hamiltonian has an additional term ${\cal H}_{\rm el}^{\rm B}+{\cal H}_{\rm ep}^{\rm B}+{\cal H}_{\rm L}^{\rm B}$ which is obtained from Eq. (\[eff\]) by interchanging operators: $$d_{Y}\leftrightarrow d_{X}, a_{x}\leftrightarrow b_{y}, a_{y}\leftrightarrow b_{x}, a_{z}\leftrightarrow b_{z} \label{change}$$ and summing over the $B$ sublattice. Following Cho and Toyozawa [@Cho] we are able to diagonalize Hamiltonian (\[eff\]) in a very large truncated basis of functions with a hole present on site $\ell$ and an arbitrary number of vibrational quanta ${\rm p}_{\pm x},{\rm p}_{\pm y},{\rm p}_{\pm z}$ on the six displaced neighboring oxygens: $$\begin{aligned} |\Psi^{\rm A}(\ell,\{p\})>=d_{X}^{\dagger}(\ell) \prod_{\alpha} {\rm U}_{\ell}^{a_{\alpha}}(-\kappa_{\alpha}) \frac{(a^{\dagger}_{\alpha}(\ell))^{p_{+\alpha}}}{\sqrt{p_{+\alpha}!}} \nonumber \\ {\rm U}_{\ell-\alpha}^{b_{\alpha}}(\kappa_{\alpha}) \frac{(b^{\dagger}_{\alpha}(\ell-\alpha))^{p_{-\alpha}}}{\sqrt{p_{-\alpha}!}} |{\rm GS},\infty>. \label{trial}\end{aligned}$$ The displacement operator ${\rm U}_{\ell}^{a}(\kappa)=\exp{[-\kappa(a_{\ell}-a^{\dagger}_{\ell})]}$ makes the ${\cal H}_{\rm ep}+{\cal H}_{\rm L}$ part of the Hamiltonian diagonal. To get basis functions $|\Psi^{\rm B}(\ell,\{p\})>$ for holes on the $B$ sublattice, the operators in Eq. (\[trial\]) should be interchanged according to Eq. (\[change\]). The next step is to build Bloch wavefunctions by Fourier transformation of the basis functions Eq. (\[trial\]). Then the Hamiltonian (\[eff\]) will be diagonal with respect to $\vec{k}$-vector. The hopping term of the Hamiltonian ${\cal H}_{\rm el}$ couples the $|\Psi^{\rm A}(\ell,\{p\})>$ and $|\Psi^{\rm B}(\ell',\{p\})>$-wavefunctions on the neighboring sites. The vibrational wavefunctions give a product of Huang-Rhys factors, but a shared oxygen contributes a non-factorizable overlap integral. However if one treats this shared oxygen as two independent atoms, one coupled to each site, then the Hamiltonian has a simple form: $$\begin{aligned} {\cal H}^{\rm AA}_{pp'}(\vec{k})&=&{\cal H}^{\rm BB}_{pp'}(\vec{k})= \delta_{\{p\}\{p'\}} \biggl[\frac{\Delta}{4}+\sum_{\alpha}(p_{+\alpha}+p_{-\alpha})\biggr] \nonumber\\ {\cal H}^{\rm AB}_{pp'}(\vec{k})&=&{\cal H}^{\rm BA}_{pp'}(\vec{k})= \varepsilon(\vec{k}) \prod_{\alpha}(-1)^{p_{-\alpha}+p'_{-\alpha}} \nonumber\\ \biggl[ P(p_{+\alpha}&,\kappa_{\alpha}&)P(p_{-\alpha},\kappa_{\alpha}) P(p'_{+\alpha},\kappa_{\alpha})P(p'_{-\alpha},\kappa_{\alpha}) \biggr]^{1/2}, \label{kspace}\end{aligned}$$ where $P(p,\kappa)=\exp(-\kappa^2)\kappa^{2p}/p!$ is a Poisson distribution. Since off-diagonal terms factorize, the analytical solution is available in this approximation: $$\begin{aligned} &&\Psi_{\lambda}^{1,2}(\vec{k})=\sum_{\{p\}=0}^{\infty} \prod_{\alpha}(-1)^{p_{-\alpha}} \biggl[ P(p_{+\alpha},\kappa_{\alpha})P(p_{-\alpha},\kappa_{\alpha}) \biggr]^{1/2} \nonumber\\ &&\frac{\Psi_{\vec{k}}^{\rm A}(\{p\})\pm\Psi_{\vec{k}}^{\rm B}(\{p\})} {\sqrt{2G'(x_{\lambda})} \bigl(\sum_{\alpha^{\prime}}(p_{+\alpha^{\prime}}+p_{-\alpha^{\prime}}) %+p_{+y}+p_{-y}+p_{+z}+p_{-z} -x_{\lambda}\bigr)}. \label{funsol}\end{aligned}$$ The corresponding eigenvalues are: $$\begin{aligned} &&E_{\lambda}^{1,2}(\vec{k})=\frac{\Delta}{4}+x_{\lambda}^{1,2}(\vec{k}) , \ \ \ 1+\varepsilon_{1,2}(\vec{k})G(x_{\lambda}^{1,2})=0 \nonumber\\ &&G(x_{\lambda})=e^{-3\Delta/4}\sum_{p=0}^{\infty}\frac{(3\Delta/4)^p}{p!} \frac{1}{p-x_{\lambda}}. \label{eigsol}\end{aligned}$$ The $G'(x_{\lambda})$ function in Eq. (\[funsol\]) is a first derivative of $G(x_{\lambda})$ and makes wavefunctions normalized. The correct solution needs a numerical diagonalization of the Hamiltonian which explicitly includes vibrational states of the four shared oxygens. As can be seen on Fig. \[fig2\] the difference is negligible between a typical spectrum obtained using approximation (\[eigsol\]) and correct numerical treatment. The ground state of the Hamiltonian (\[kspace\]), with energy (from Eq. \[eigsol\]) $E_0(\vec{k})=\Delta/4+x_0(\vec{k})$, corresponds to the anti-JT polaron. Its effective mass, deduced from $d^2x_0(\vec{k})/d\vec{k}^2$, provides a realistic alternative (exact for $\Delta\rightarrow 0$ or $\infty$) to the available variational approaches [@Katja] or exact quantum Monte Carlo simulations [@Pasha]. An ARPES experiment measures the spectral function $A(\vec{k},\omega)=-\frac{1}{\pi}G(\vec{k},\omega)$ with momentum $\vec{k}$ fully resolved, provided there is no dispersion in the direction perpendicular to the surface. Although LaMnO$_3$ is cubic, because of the layered AFA magnetic structure, at low temperatures Mn $e_g$ electrons are two-dimensional and the spectrum can be measured: $$A(\vec{k},\omega)= \sum_{f}|<f|d^{\dagger}(\vec{k})|{\rm GS},\infty>|^2\delta(E-E_f). \label{impart}$$ The operator $d^{\dagger}(\vec{k})$ excites a hole from the JT ground state. Summation over final eigenstates $|f>$ includes summation over branch index $i=1,2$ and number of phonons $\lambda=0,1...\infty$. The sequence of delta functions in Eq. (\[impart\]) should be replaced by convolved local densities of phonon states, which we approximate by a Gaussian, $\delta(E)\rightarrow\exp(-E^2/2\gamma^2)/\sqrt{2\pi}\gamma$. Substituting solution (\[funsol\]), (\[eigsol\]) into equation (\[impart\]), we obtain the spectral function: $$A(\vec{k},\omega)= \sum_{\lambda,i}\frac{G^2(x_{\lambda}^i)}{G'(x_{\lambda})} \delta(E-E_{\lambda}^i). \label{arpes}$$ Equation (\[arpes\]) along with (\[eigsol\]) gives the ARPES spectrum normalized to $\int d\omega A(\vec{k},\omega)=1$. The first energy moment of the spectrum [@Cho] coincides with the free hole energy calculated in the “frozen” lattice approximation $\Delta+\varepsilon_{1,2}(\vec{k})$ shown on Fig. \[fig2\]. The edge of the spectrum corresponds to polaron creation at energy $\approx \Delta/4$. This transition is weaker by 3 orders of magnitude than the peak at $\approx\Delta+\varepsilon_{1,2}(\vec{k})$. At room temperature magnetic order is lost. The paramagnetic state is modeled by a mean field approximation, namely scaling the effective hopping integral by 2/3 and allowing hopping in $\pm\hat{z}$ direction. This modifies the single particle energy band entering Eq. (\[kspace\]) to $\varepsilon_{1,2}(\vec{k})=t/3\bigr(-2C_z\pm(C_x+C_y)\bigl)$. But the JT orbital order is not destroyed at T=300 K and Franck-Condon broadening is still expected. When spins are disordered, $\vec{k}$ is not a good quantum number and additional broadening is expected. Only phonon broadening of the ARPES along with peak positions are shown on Fig. \[fig1\](b). The angle-integrated spectrum, shown on Fig. \[fig3\] for low and high temperatures, has a width of about 1.2 eV and is almost temperature independent. The uncorrelated (U=0) band structure, shown for comparison, is sensitive to magnetic order and therefore temperature dependent. The existing photoemission data [@Park; @Dessau] are consistent with our predictions. Higher resolution experiments are needed to test the theory and to unravel the nature of the lowest energy excitations in the LaMnO$_3$. To make such an experiment possible, a single domain sample (having 2D dispersion at T=0K) is needed, with good control of oxygen concentration [@Dabrowski]. We thank P. D. Johnson for helpful conversations. This work was supported in part by NSF Grant No. DMR-9725037. R. M. Kusters, J. Singelton, D. A. Keen, R. McGreevy, and W. Hayes, Physica B [**155**]{}, 362 (1989); S. Jin, T. H. Tiefel, M. McCormack, R. A. Fastnacht, R. Ramesh, and L. H. Chen, Science [**264**]{}, 413 (1994). , edited by C.N.R. Rao and B. Raveau (World Scientific, Singapore, 1998). J. Rodriguez-Carvajal, M. Hennion, F. Moussa, A. H. Moudden, L. Pinsard, and A. Revcolevschi, Phys. Rev. B [**57**]{}, 3189 (1998). Y. Murakami, J. P. Hill, D. Gibbs, M. Blume, I. Koyama, M. Tanaka, H. Kawata, T. Arima, Y. Tokura, K. Hirota, and Y. Endoh, Phys. Rev. Lett. [**81**]{}, 582 (1998). J. H. Park, C. T. Chen, S. W. Cheong, W. Bao, G. Meigs, V. Chakarian and Y. U. Idzerda, Phys. Rev. Lett. [**76**]{}, 4215 (1996); T. Saitoh A. E. Bocquet, T. Mizokawa, H. Namatame and A. Fujimori, Phys. Rev. B [**51**]{}, 13942 (1995). D. S. Dessau and Z. X. Shen, in [*Colossal Magnetoresistive Oxides*]{}, edited by Y. Tokura (World Scientific, Singapore, 1998). W. E. Pickett and D. Singh, Phys. Rev. B [**53**]{}, 1146 (1996); S. Satpathy, Z. S. Popović and F. R. Vukajlović, Phys. Rev. Lett. [**76**]{}, 960 (1996); I. Solovyev, N. Hamada and K. Terakura, Rhys. Rev. Lett. [**76**]{}, 4825 (1996); Y.-S. Su, T. A. Kaplan, S. D. Mahanti and J. F. Harrison, Phys. Rev. B [**61**]{}, 1324 (2000). J. Kanamori, J. Appl. Phys. [**31**]{}, 14S (1960); K. I. Kugel and D. I. Khomskii, Sov. Phys. Usp. [**25**]{}, 231 (1982). P. B. Allen and V. Perebeinos, Phys. Rev. B [**60**]{}, 10747 (1999). F. J. Himpsel, N. Schwentner, E. E. Koch, Phys. Status Solidi B [**71**]{}, 615 (1975). G. A. Sawatzky, Nature [**342**]{}, 480 (1989). A. J. Millis, Phys. Rev. B [**53**]{}, 8434 (1996). J. C. Slater and G. F. Koster, Phys. Rev. [**94**]{}, 1498 (1954). J. H. Jung, K. H. Kim, D. J. Eom, T. W. Noh, E. J. Choi, J. Yu, Y. S. Kwon, and Y. Chung, Phys. Rev. B [**55**]{}, 15489 (1997); J. H. Jung, K. H. Kim, T. W. Noh, E. J. Choi, and J. Yu,Phys. Rev. B [**57**]{}, 11043 (1998). P. B. Allen and V. Perebeinos, Phys. Rev. Lett. [**83**]{}, 4828 (1999). M. N. Iliev, M. V. Abrashev, H.-G. Lee, V. N. Popov, Y. Y. Sun, C. Thomsen, R. L. Meng, and C. W. Chu, Phys. Rev. B [**57**]{}, 2872 (1998); V. B. Podobedov, A. Weber, D. Romero, J. P. Rice and H. D. Drew, Phys. Rev. B [**58**]{}, 43 (1998). K. Cho and Y. Toyozawa, J. Phys. Soc. Jpn. [**30**]{}, 1555 (1971). A. H. Romero, D. W. Brown and K. Lindenberg, Phys. Rev. B [**59**]{}, 13728 (1999). P. E. Kornilovitch, Phys. Rev. Lett. [**84**]{}, 1551 (2000). B. Dabrowski, R. Dybzinski, Z. Bukowski, O. Chmaissem, J. D. Jorgensen, J. Solid State Chem. [**146**]{}, 448 (1999).

ArXiv

--- abstract: 'We present the discovery observations of the optical counterpart of the $\gamma$-ray burster GRB990712 taken 4.16 hours after the outburst and discuss its light curve observed in the V, R and I bands during the first $\sim$35 days after the outburst. The observed light curves were fitted with a power-law decay for the optical transient (OT), plus an additional component which was treated in two different ways. First, the additional component was assumed to be an underlying galaxy of constant brightness. The resulting slope of the decay is $0.97\pm^{0.05}_{0.02}$ and the magnitudes of the underlying galaxy are: ${\rm V} = 22.3 \pm 0.05$, ${\rm R} = 21.75 \pm 0.05$ and ${\rm I} = 21.35 \pm 0.05$. Second, the additional component was assumed to be a galaxy plus an underlying supernova with a time-variable brightness identical to that of GRB980425, appropriately scaled to the redshift of GRB990712. The resulting slope of the decay is similar, but the goodness-of-fit is worse which would imply that either this GRB is not associated with an underlying supernova or the underlying supernova is much fainter than the supernova associated with GRB980425. The galaxy in this case is fainter: ${\rm V} = 22.7 \pm 0.05$, ${\rm R} = 22.25 \pm 0.05$ and ${\rm I} = 22.15 \pm 0.05$; and the OT plus the underlying supernova at a given time is brighter. Measurements of the brightnesses of the OT and the galaxy by late-time HST observation and ground-based observations can thus assess the presence of an underlying supernova.' author: - 'K.C. Sahu, P. Vreeswijk, G. Bakos, J.W. Menzies, A. Bragaglia, F. Frontera, L. Piro, M. D. Albrow, I. A. Bond, R. Bower, J. A. R. Caldwell, A. J. Castro-Tirado$^{12,13}$, F. Courbin, M. Dominik, J.U. Fynbo, T. Galama, K. Glazebrook, J. Greenhill, J. Gorosabel, J. Hearnshaw, K. Hill, J. Hjorth, S. Kane, P. M. Kilmartin, C. Kouveliotou R. Martin, N. Masetti P. Maxted, D. Minniti, P. M[ø]{}ller, Y. Muraki, T. Nakamura, S. Noda, K. Ohnishi, E. Palazzi, J. van Paradijs, E. Pian, K. R. Pollard, N.J. Rattenbury, M. Reid, E. Rol, T. Saito, P. D. Sackett, P. Saizar, C. Tinney, P. Vermaak, R. Watson, A. Williams, P. Yock A. Dar' title: 'Discovery of the optical counterpart and early optical observations of GRB990712[^1]' --- Introduction ============ The past two years have witnessed tremendous progress in our understanding of Gamma-Ray Burst (GRB) phenomena. Thanks to the ability of Beppo-Sax to provide arcminute sized error boxes for the bursts, the first optical counterpart was detected in February 1997 for GRB970228 (van Paradijs et al. 1997; Costa et al. 1997). The HST observations of this GRB provided the first clear indication that the GRB is associated with an external galaxy and is unrelated to its nuclear activity (Sahu et al. 1997a). Shortly thereafter, the redshift measurement for GRB 970508 proved beyond doubt that GRBs are extragalactic in nature (Metzger et al. 1997; Djorgovski et al. 1997). More than a dozen GRB optical counterparts have been detected since then, with redshifts as high as 3.42 in the case of GRB971214 (Kulkarni et al. 1998). The extensive observations of GRB afterglows in X-ray, optical and radio wavelengths have been shown to be consistent with cosmological fire-ball models (e.g. Paczyński & Rhoads 1993; Mészáros & Rees 1997; Wijers et al. 1997). However, the exact cause of the GRBs has remained elusive, and progress in determining the actual mechanism causing the bursts has been slow. The two leading models for GRBs involve the collapse of a massive star (e.g. Woosley 1997; Paczyński 1998), and the merging of a neutron star with either another neutron star or a black hole (e.g. Eichler et al. 1989; Narayan, Paczyński, & Piran 1992; Sahu et al. 1997b). The “isotropic equivalent energy" of some GRBs is as high as 10$^{54}$ ergs, which exceeds the energy equivalent of the total mass involved in the latter model, and hence is thought to favor the former. However, the beaming factor of the emission of the GRB in different wavebands can be high (in both models), so the energetics alone may not be conclusive in favoring one of the models over the other. Since the lifetime of a very massive star is of the order of a million years or shorter, the GRBs are expected to be within or close to star-forming regions in the massive-star collapse model. On the other hand, since the kick-velocity of a newly formed neutron star is of the order of 200 [$\mbox{km\,s}^{-1}$]{}, and since the neutron star, on average, is already about 10$^8$ years old at the time of the burst, GRBs are generally expected to be far from star-forming regions in the merging neutron-star model (Bloom, Sigurdsson & Pols 1999). Thus the location of the GRB with respect to the star-forming regions in the host galaxy could be a distinguishing feature which can help in settling the question of the cause of the GRBs. However, if the presence of dense interstellar material is prerequisite for the onset of the optical afterglow, then OTs would be observed [*only*]{} in star-forming regions regardless of whether GRBs result from neutron-star mergers or from the collapse of massive stars. In any case, if some OTs are found far away from star-forming regions, neutron-star merger model would be favored. If all OTs are found in star-forming regions, the situation is less clear, and depending on the model, one may need to investigate other aspects to find distinguishing features. For example, one consequence of the massive-star collapse model is that the GRB should be accompanied by an underlying supernova, whose brightness variation is distinct from the power-law decay behavior of the OT and may be detectable. There is a clear bimodality in the observed burst durations: long bursts have timescales of about 10 to 200 seconds, and short bursts have timescales of 0.1 to 1 second (Kouveliotou et al. 1995). All the afterglows that have been discovered so far belong to the subgroup with long bursts (Fishman 1999). All the OTs for which the host galaxies have been observed – although this sample is small – are found to be in galaxies whose spectral indicators suggest star-forming activity (e.g. Fruchter et al. 1999). So there is an interesting possibility that the long-duration GRBs are caused by the collapse of rapidly-rotating massive stars, and hence they occur in star-forming regions. The short-duration bursts may be of a different origin, and may be caused by the coalescence of two neutron stars. The detection of optical counterparts of short duration bursts, and their observations would be important to understand this scenario better. GRB990712, being at $z=0.434$ (Galama et al. 1999a; Hjorth et al. 2000), is one of the closest GRBs observed so far and hence provides a good opportunity to determine the possible presence of an underlying supernova, the location of the GRB within the galaxy, and the luminosity of the galaxy in different wavebands. In this paper we present the optical observations of GRB990712 leading to the discovery of its optical counterpart, and subsequent optical imaging observations taken in different filters during the first $\sim$35 days after the burst. Observations ============ GRB990712 was simultaneously detected on 1999 July 12.69655 with the Gamma-Ray Burst Monitor and the WFC unit 2 onboard the BeppoSAX satellite. The burst lasted for about 30 seconds in both the $\gamma$-ray (40 - 700 keV) and X-ray (2-26 keV) energy ranges and had a double-peaked structure. While its intensity in $\gamma$-rays is moderate, it exhibited the strongest X-ray prompt emission observed to date (Heise et al. 1999). The initial Beppo-SAX position, with an error circle of 5 arcmin radius, was revised to one with an error radius of only 2 arcmin. Our first observations were made at this latter position with the SAAO 1m telescope at Sutherland, South Africa on July 12.87 UT, about 4.16 hours after the burst during time generally dedicated to PLANET microlensing observations (Albrow et al. 1998). The optical image was taken through an R filter, with an integration time of 900 sec. The detector was a $1024 \times 1024$ SITe CCD with an image scale of 0.309 arcsec per pixel and a total field of view of $5.3 \times 5.3$ square arcmin. Comparison of the image with the Digitized Sky Survey (DSS) image for the same field showed the presence of a new source well above the sky background, that was absent in the DSS image. This new source had a brightness of R = 19.4 $\pm$ 0.1 mag, which is about 2 magnitudes brighter than the limiting magnitude of the DSS image, indicating that the new source was the optical counterpart of the GRB (Bakos et al. 1999a,b). Fig. 1 shows the discovery image, along with the DSS image taken of the same region. Also marked in the figure are the reference stars used for photometric calibration of the OT. A spectrum of this source obtained a few hours later revealed emission and absorption lines which were used to derive redshift of 0.434 for this source (Galama et al. 1999a). Subsequent observations by various groups showed the decaying nature of this source further confirming the identification of the OT with the GRB. We continued the observations with a multi-wavelength optical follow-up campaign using the telescopes at SAAO, ESO and AAO. A log of all the observations used in this analysis along with the derived magnitudes of the OT are listed in Table 1. The measurements include the acquisition images taken for the spectroscopic observations with the ESO VLT shortly after the discovery of the OT (Vreeswijk et al. 1999, in prep.). Astrometry and Photometry of the Optical Counterpart ---------------------------------------------------- An astrometric solution of the field was carried out using 10 reference stars in the image which are marked in Fig. 1. Their coordinates, as taken from the USNO 2.0 catalog which uses the Hipparcos frame of reference, are listed in Table 2. The pixel centroids of the reference stars were determined using two-dimensional Gaussian fits. These centroids were combined with the USNO coordinates to determine an astrometric solution of the field. The resulting position of the optical counterpart is RA(2000) = 22$^h$ 31$^m$ 53.$^s 061 \pm 0^s.011$, Dec(2000)= $-73^d \ 24^\prime \ 28.^{\prime\prime}58\, \pm\, 0^{\prime\prime}.05$. Since sky conditions deteriorated due to clouds just after the discovery observations of the OT, no standard star observations could be taken on that night. The images were calibrated through observations of the standard stars F203 and F209 (Menzies et al. 1989) taken on 13th July at SAAO. All the photometric measurements were carried out using the IRAF aperture photometry task PHOT. The photometry of a few secondary standards in the field (Fig. 1) was carried out using the standard star observations, and the photometry of all GRB observations was then performed via these secondary standards. Since only two standard stars were observed, the extinction correction due to differential airmass between the GRB and the standard star observations were not applied in the preliminary analysis (Bakos et al. 1999a,b). Instead, the photometric measurements taken with the adjacent 50cm telescope were used to get the extinction coefficients, which resulted in a correction of –0.075 mag in V, –0.06 mag in R, and –0.02 mag in I. (This is consistent with the discrepancy pointed out by Kemp & Halpern 1999). Our final magnitudes of the secondary photometric standards are listed in Table 2, and Table 1 lists the derived magnitudes of the OT in different bands at different epochs including one extra measurement reported by Kemp and Halpern (1999). Analysis ======== The Light Curve in Different Wavebands -------------------------------------- All the photometric measurements listed in Table 1 are also shown in Fig. 2. The observed light curves in different bands clearly show a continuously-changing slope suggesting a additional component to the power-law decay of the OT as first noted by Hjorth et al. (1999a). In an approach similar to the one followed by Hjorth et al. (1999b), we have fitted this additional component in two ways: (i) with an underlying galaxy of constant brightness, and (ii) an underlying galaxy plus a supernova similar to GRB980425, appropriately scaled to the redshift of GRB990712. First, a power law decline of the form $f(t) \propto t^{-\alpha}$ for the OT and a constant contribution from the background galaxy was used to fit the observed light curves simultaneously in different bands, with the same slope for all the bands. The resultant slope is $0.97$, and the total $\chi^2$ is 47 for 39 d.o.f. (Relaxing the condition of the same slope in different bands does not alter the fit parameters significantly). The total $\chi^2$ is, however, fairly flat (between 44 and 52) for any value of $\alpha$ between 0.95 and 1.02 beyond which the total $\chi^2$ increases rapidly. Fig. 2 shows the best fit to the light curves. The derived magnitude of the underlying galaxy is not very sensitive to the exact value of $\alpha$ and are found to be ${\rm V} = 22.3 \pm 0.05$, ${\rm R} = 21.75 \pm 0.05$ and ${\rm I} = 21.35 \pm 0.05$ where the uncertainties represent the variation of the derived magnitudes as a result of changing $\alpha$ between 0.95 and 1.02. One possible consequence of the massive-star model for the GRB is that one should observe an underlying supernova in the lightcurve of an OT. The physical connection of a GRB with a supernova (SN), first suggested by the discovery of a peculiar type Ic SN in the error box of GRB980425 (Galama et al. 1998; Iwamoto et al. 1998), has been strengthened further by recent observations of other GRBs. Castro-Tirado and Gorosabel (1999) suggested that the light curve of GRB980326 resembled that of a SN, and indeed Bloom et al. (1999) showed that the late time light curve of GRB980326 can be explained by an underlying SN1998bw type SN at a redshift of around unity. For the afterglow of GRB970228, Reichart (1999) and Galama et al. (1999b) find that a power-law decay plus SN1998bw light curve redshifted to the distance of the burst, $z = 0.695$ (Djorgovski et al. 1999), fits the observed light curve very well. In light of these new findings, we have also fitted the observed light curve of GRB 990712 assuming the presence of an underlying supernova. In order to determine the contribution of the underlying supernova, we first calculate the expected V, R, I magnitudes of SN1998bw, placing it at the redshift of GRB990712, $z = 0.434$. This includes wavelength shifting and time profile stretching (both by a factor of 1+z), and rescaling the magnitudes for a distance corresponding to $z=0.434$, assuming $\Omega_0$ = 0.2. To account for the wavelength shift correctly, we interpolate the redshifted UBVRI broadband flux spectrum of the SN for each bin in time with a spline fit, obtain the V, R, and I fluxes at their effective wavelengths, and convert these back to obtain magnitudes in the observer’s frame (Fukugita et al. 1995). The resultant SN light curves in different bands are shown at the bottom of Fig. 3. In order to fit the GRB light curve, the SN flux is subtracted from the observed GRB flux, and the residual fluxes are assumed to be due to the OT and the underlying galaxy. The procedure outlined earlier is then used to fit the light curves. The $\chi^2$ in this case is 52 for 42 d.o.f., which is clearly higher than the model without the supernova. The resulting slope of the decay is $0.96$ which is quite similar but the magnitudes of the underlying galaxy are fainter: ${\rm V} = 22.7 \pm 0.05$, ${\rm R} = 22.25 \pm 0.05$ and ${\rm I} = 22.15 \pm 0.05$. The brightness of the OT of the GRB at a given time is also slightly lower, but the combined brightness of the OT and the underlying supernova (which is the quantity that can be measured if the OT can be resolved from the galaxy) is higher than the brightness of the OT in the absence of an underlying supernova. It is important to note, however, that there is no evidence that the SNe possibly underlying the GRB afterglows have the same brightness or the same decay behavior. Our assumption that the underlying supernova is identical to SN1998bw is dictated by two reasons. First, this is the only SN associated with a GRB whose light curve has been monitored extensively. Second, the number of data points in our light curves does not allow us to vary the characteristics of the underlying SN light curve. So the fact that the $\chi^2$ is higher in this case can be misleading since the true underlying supernova may be different from that of SN1998bw, making the resultant $\chi^2$ higher for our simplified model. This chi-square analysis indicates that that either GRB990712 is not associated with an underlying supernova or the underlying supernova is much fainter than the supernova associated with GRB980425. However, the qualitative conclusion that the galaxy is expected to be fainter and the OT plus the SN is expected to be brighter in the presence of an underlying supernova is unlikely to change. Hence late time HST observations, in which the OT is well resolved so that the brightness of the OT and the galaxy can be estimated separately, or late time ground-based observations when the brightness of the OT is negligible, would greatly help in determining the presence of an underlying supernova. In both the above scenarios (i.e. with and without an underlying supernova), the power-law index of the decay $\alpha$ is about $1$, making it one of the slowest decline rates of all the OTs observed so far. Since $\alpha \le$1 would lead to a divergence of the total energy integrated over time, the slope must steepen at later times. This has been observed for GRB990510 which declined with $\alpha = 0.82$ at early times, later steepening with time (Harrison et al. 1999). Power-law decays are thought to arise from electrons shocked by the relativistic expansion of the debris into the ambient medium. In such a case, the information on the change of slope can be used to derive information on the cooling rate of the electrons in the post-shock region (see, e.g. Wijers et al. 1997; Sari, Piran, & Narayan 1998; Livio & Waxman 1999). Spectral Energy Distribution of the OT -------------------------------------- If the OT emission is due to the synchrotron radiation from the swept-up electrons in the post-shock region, then its energy distribution is expected to be of the general form $f_\nu \propto t^{-\alpha} \nu^\beta$. In such a case, one expects a relationship between the power law index of the energy distribution of the electrons $p$, the spectral slope $\beta$, and the decay constant $\alpha$ (Sari et al. 1998). One must distinguish two cases: (i) both the peak frequency $\nu_{\rm m}$ and the cooling frequency $\nu_{\rm c}$ are below the optical/IR waveband, in which case $p = (4\alpha +2)/3 = 1.97 \pm 0.04$ and $\beta = -p/2 = -0.99 \pm 0.02$, and (ii) $\nu_{\rm m}$ has passed the optical/IR waveband, but $\nu_{\rm c}$ has not yet done so, in which case $p = (4\alpha +3)/3 = 2.31 \pm 0.04$ and $\beta = -(p-1)/2 = -0.66 \pm 0.02$. We can now compare the theoretically expected spectral slope ($\beta$) with the observed one. Although perhaps best determined by the spectroscopic measurements. our light curve analysis provides magnitudes of the OT in different bands, which can be directly used to derive $\beta$. We should note here that the brightness of the underlying galaxy and the possible contamination by an underlying supernova makes the determination of the magnitudes of the OT less reliable at later times, and our value of $\alpha$ is most likely dominated by the early part of the light curve when the OT was bright. Furthermore, $\alpha$ was kept fixed in $all$ the bands for the entire period of our observations, which implies that we cannot detect any spectral [*evolution*]{} of the OT. The $synphot/calcphot$ task in IRAF was used to determine the V$-$R and R$-$I values for a range of $\beta$ values, which were then compared with the colors derived from the light curve analysis to determine the spectral slopes. The V$-$R color implies a spectral slope of $-$1.8 $\pm$ 0.5, and the R$-$I color implies a spectral slope of $-$0.7 $\pm$ 0.1. Thus the slope derived from the V and R bands is closer to the case (i) mentioned above, and the slope derived from the R and I bands is closer to the case (ii). This is roughly consistent with the theoretical expectations and implies that, in the early part of the light curve, $\nu_m$ has passed the V and R and I bands, but $\nu_c$ has not passed the I band. The Background Galaxy --------------------- The inferred magnitudes for an underlying galaxy suggest that the galaxy is relatively bright compared to the OT, and it may be possible to see the host galaxy directly in the images. The NTT images taken about 8 days after the burst had the best seeing ($\sim$ 1 arcsec) and the R images indeed show some hint of a slight extension. To further investigate this, all the images taken in B,V,R and I bands were co-added, giving rise to a single deep image of a total integration time of 40 min. From this combined image, we constructed a model point-spread-function (PSF) from a few isolated bright stars in the field using IRAF/DAOPHOT. This model PSF was used to subtract the point source contribution from the OT. The PSF-subtracted image shows some residual distribution elongated in the east-ward direction of the OT over a total of about 2 arcsec, which is probably the contribution of the underlying galaxy. Thus HST observations would show the detailed structure of the galaxy (Fruchter et al. in prep). The derived magnitudes of the underlying galaxy were used to calculate its luminosity. The GRB is at a Galactic latitude and longitude of –40 and 315 degrees, respectively. The extinction models of Burstein & Heiles (1982) and Schlegel et al. (1998) yield E(B–V) of 0.015 and 0.033, respectively, for the position of the GRB. This extinction is small and similar to SN1998bw, which is neglected in our analysis. To derive the luminosity of the host galaxy, we use a redshift of z=0.434 which, for H$_0$ = 70 [$\mbox{km\,s}^{-1}$]{}and $\Omega$=0.2, corresponds to a luminosity distance of 2160 Mpc, and a distance modulus of 41.67 mag. Applying the K-correction for z=0.434, the inferred luminosity of the galaxy is of the order of $L_\star$ (depending on the presence of an underlying supernova), where $L_\star$ corresponds to the luminosity of a typical galaxy at the ‘knee’ of the observed galaxy luminosity function (see, e.g., Lilly et al. 1999). If we exclude SN1998bw (which is not only ‘nearby’ ($z = 0.008$), but for which the energy released in $\gamma$-rays is much smaller than any other GRB), then GRB990712 is the closest ‘cosmological’ GRB, and the apparent magnitude of its host galaxy is the brightest among all the GRB host galaxies observed so far. Compared to the host galaxy of GRB990713, which is the next brightest, the host galaxy of GRB990712 is about 0.25 magnitudes brighter in R if there is an underlying supernova, and more than a magnitude brighter if there is no underlying supernova. Akerlof, C., et al. 1999, Nature, 398, 400 Bakos, G., et al. 1999a, GCN 387 Bakos, G., et al. 1999b, IAUC 7255 Bloom, J. S., et al. 1999, Nature, 401, 453 Bloom, J. S., Sigurdsson, S., & Pols, O. R. 1999, MNRAS, 305, 763 Burstein, D., & Heiles, C. 1982, AJ, 87, 1165 Castro-Tirado, A. J., & Gorosabel, J. 1999, A&AS, 138, 449 Costa, E., et al. 1998, Nature, 387, 783 Djorgovski, S. G., et al. 1997, Nature, 387, 876 Eichler, D., Livio, M., Piran, T., & Schramm, D. N. 1989, Nature, 340, 126 Fishman, G. J. 1999, in Proc. STScI Symp. on “Supernovae and Gamma-Ray Bursts", ed. M. Livio, K. C. Sahu & N. Panagia (in press) Frail, D. A., Kulkarni, S. R., & Bloom, J. S. 1999, ApJ, 525, L81 Fruchter, A., et al. 1999, ApJ, 519, L13 Fukugita, M., Shimasaku, K., & Ichikawa, R. 1995, PASP, 107, 945 Galama, T. J., et al. 1997, Nature, 387, 479 Galama, T. J., et al. 1998, , 395, 670 Galama, T. J., et al. 1999a, GCN 388 Galama, T. J., et al. 1999b, ApJ, submitted (astro-ph/9907264) Harrison, F. A., et al. 1999, ApJ, 523, 121 Heisse, J., et al. 1999, IAUC 7221 Hjorth, J., et al. 1999a, GCN 389 Hjorth, J., et al. 1999b, GCN 403 Hjorth, J., et al. 2000, to appear in ApJ (astro-ph/0003383) Iwamoto, K., et al. 1998, , 395, 672 Kemp, J., & Halpern, J. 1999, GCN 402 Kouveliotou, C., et al. 1995, in AIP Conf. Proc. 384, Gamma Ray Bursts – 3rd Huntsville Symposium, ed. C. Kouveliotou, M. F. Briggs, & G. J. Fishman (New York: AIP), 42 Kulkarni, S. R., et al. 1998, Nature, 393, 35 Lilly, S. J., et al. 1999, ApJ, 518, 641 Livio, M., & Waxman, E. 1999, ApJ, submitted (astro-ph/9911160) Menzies, J. W., Cousins, A. W. J., Banfield, R. M., & Laing, J. D. 1989, SAAO Circular 13, 1 Mészáros, P., & Rees, M. J. 1997, ApJ, 476, 232 Metzger, M. R., et al. 1997, Nature, 387, 879 Narayan, R., Paczyński, B., & Piran, T. 1992, ApJ, 395, L83 Paczyński, B. 1998, ApJ, 494, L45 Paczyński, B., & Rhoads, J. E. 1993, ApJ, 418, L5 Reichart, D. E. 1999, ApJ, 521, L111 van Paradijs, J., et al. 1997, Nature, 386, 686 Sahu, K. C., et al. 1997a, Nature, 387, 476 Sahu, K. C., et al. 1997b, ApJ, 489, L127 Sari, R., Piran, T., & Narayan, R. 1998, ApJ, 497, L17 Schlegel, D., Finkbeiner, D., & Davis, M., 1998, ApJ, 500, 525 Waxman, E. 1997, ApJ, 485, L5 Wijers, R. A. M. J., Rees, M. J., & Mészáros, P. 1997, MNRAS, 288, L51 Woosley, S. E, & Wallace, R. K. 1982, ApJ, 258, 716 [lcll]{}\ Jul 12.873& 19.349 $\pm$ 0.019&SAAO 1m&\ Jul 13.129 & 20.175 $\pm$ 0.011&VLT 8m\ Jul 13.151 & 20.183 $\pm$ 0.009&VLT 8m\ Jul 13.174 & 20.251 $\pm$ 0.085&ESO 1.5m\ Jul 13.328 & 20.470 $\pm$ 0.10&MJUO 61cm\ Jul 13.383 & 20.533 $\pm$ 0.016&VLT 8m\ Jul 13.395 & 20.460 $\pm$ 0.057&CASLEO 2.15m\ Jul 13.750 & 20.857 $\pm$ 0.088&SAAO 1m&\ Jul 14.127 & 20.968 $\pm$ 0.025&VLT 8m\ Jul 14.287 & 20.991 $\pm$ 0.040&NTT 3.5m\ Jul 14.683 & 21.490 $\pm$ 0.27&MJUO 61cm\ Jul 14.764 & 21.201 $\pm$ 0.123&SAAO 1m&\ Jul 16.403 & 21.420 $\pm$ 0.050&NTT 3.5m&\ Jul 20.421 & 21.550 $\pm$ 0.050& NTT 3.5m\ Aug 02.533 & 21.584 $\pm$ 0.036&AAT 3.9m\ Aug 12.232 & 21.650 $\pm$ 0.030&VLT 8m&\ Aug 16.320 & 21.750 $\pm$ 0.060&CTIO 0.9m$^{\dag}$\ Aug 16.445 & 21.779 $\pm$ 0.041&AAT 3.9m\ \ Jul 13.156& 20.523 $\pm$ 0.017 &SAAO 1m\ Jul 13.158& 20.516 $\pm$ 0.015 &SAAO 1m\ Jul 13.405& 20.852 $\pm$ 0.023 &VLT 8m\ Jul 13.406& 20.866 $\pm$ 0.028 &VLT 8m\ Jul 14.157& 21.290 $\pm$ 0.041 &VLT 8m\ Jul 14.298& 21.458 $\pm$ 0.053 &VLT 8m\ Jul 14.787& 21.795 $\pm$ 0.089&SAAO 1m&\ Jul 20.433& 22.065 $\pm$ 0.100& NTT 3.5m\ Aug 02.511& 22.136 $\pm$ 0.042& AAT 3.9m\ Aug 16.471& 22.305 $\pm$ 0.054& AAT 3.9m\ \ \ Jul 13.774& 20.266 $\pm$ 0.082&SAAO 1m&\ Jul 14.271& 20.603 $\pm$ 0.078&NTT 3.5m\ Jul 14.810& 20.933 $\pm$ 0.258&SAAO 1m\ Jul 20.439& 20.950 $\pm$ 0.100&NTT 3.5m\ Aug 02.554& 21.164 $\pm$ 0.165&AAT 3.9m\ Aug 16.414& 21.420 $\pm$ 0.100&AAT 3.9m\ [lccccccc]{} OT & 22:31:53.0614 & -73:24:28.576 & 579.40 & 620.02\ 1 & 22:31:50.8933 & -73:24:25.280 & 609.98 & 630.95& 17.185 & 16.40 & 15.64\ 2 & 22:31:40.1507 & -73:25:09.400 & 758.34 & 487.54& 16.975 & 16.65 & 16.29\ 3 & 22:31:49.7413 & -73:26:09.250 & 624.22 & 294.44& 16.455 & 15.97 & 15.50\ 4 & 22:32:18.6200 & -73:24:09.920 & 224.05 & 680.45& 15.905 & 15.27 & 14.65\ \ \ 5 & 22:31:49.3380 & -73:24:53.130 & 631.68 & 537.68\ 6 & 22:32:02.5307 & -73:25:27.980 & 446.48 & 427.78\ 7 & 22:32:00.6387 & -73:23:20.810 & 474.52 & 839.78\ 8 & 22:31:37.5760 & -73:23:34.160 & 794.42 & 796.29\ 9 & 22:32:21.9453 & -73:24:27.680 & 179.47 & 623.97\ 10 & 22:32:16.1573 & -73:24:55.230 & 259.72 & 535.90 [^1]: Based on observations collected at SAAO, Sutherland; ESO, Paranal and La Silla (ESO Programs 63.O-0618 and 63-O-0567); and AAT, Australia.

ArXiv

--- abstract: 'High redshift sources suffer from magnification or demagnification due to weak gravitational lensing by large scale structure. One consequence of this is that the distance-redshift relation, in wide use for cosmological tests, suffers lensing-induced scatter which can be quantified by the magnification probability distribution. Predicting this distribution generally requires a method for ray-tracing through cosmological N-body simulations. However, standard methods tend to apply the multiple thin-lens approximation. In an effort to quantify the accuracy of these methods, we develop an innovative code that performs ray-tracing without the use of this approximation. The efficiency and accuracy of this computationally challenging approach can be improved by careful choices of numerical parameters; therefore, the results are analysed for the behaviour of the ray-tracing code in the vicinity of Schwarzschild and Navarro-Frenk-White lenses. Preliminary comparisons are drawn with the multiple lens-plane ray-bundle method in the context of cosmological mass distributions for a source redshift of $z_{s}=0.5$.' author: - | M. Killedar$^{1,2}$[^1], P. D. Lasky$^{3,4}$, G. F. Lewis$^{1}$ & C. J. Fluke$^{4}$\ $^{1}$Sydney Institute for Astronomy, School of Physics, The University of Sydney, NSW, 2006, Australia\ $^{2}$Dipartimento di Fisica dell’Università di Trieste, Sezione di Astronomia, Via Tiepolo 11, I-34131 Trieste, Italy\ $^{3}$Theoretical Astrophysics, Eberhard Karls University of Tübingen, Tübingen 72076, Germany\ $^{4}$Centre for Astrophysics and Supercomputing, Swinburne University of Technology, PO Box 218, Hawthorn, 3122, Victoria, Australia bibliography: - 'MKpaper02.bib' date: 'Accepted 2011 October 15. Received 2011 October 04' title: 'Gravitational Lensing with Three-Dimensional Ray Tracing' --- \[firstpage\] cosmology: theory – gravitational lensing – methods: numerical – large scale structure of Universe Introduction ============ Weak gravitational lensing of distant galaxies by large-scale structure leads to the distortion of their images. In an inhomogeneous universe, each line of sight probes a slightly different integrated mass and is sheared differently; overdense regions result in magnification and underdense regions cause demagnification. The result is that for a given redshift $z$, a source is magnified (or demagnified) by some amount $\mu$, [*relative to the case of a perfectly homogeneous universe on all scales*]{}. This magnification has an associated probability distribution that reflects the existence of structure on a range of scales. Large surveys for galaxies, quasars and supernovae suffer a two-fold [*magnification bias*]{} as a result of the same phenomenon. When a source is magnified, its total surface area on the sky is increased; consequently, for a given area of sky observed, the region of the source plane being sampled is decreased. On the other hand, magnification will push otherwise too-faint sources above the observational threshold in flux-limited surveys [@C81; @TOG84]; this is particularly important for optically-selected quasar surveys [@P82; @BL96; @W11]. Together, these two effects have counteracting influences on source counts, but generally do not cancel each other out. Additionally, for sources that are demagnified the effect is reversed; indeed, the majority of sources are de-magnified relative to the case of a perfectly homogeneous universe. The net effect on source counts, observed luminosity functions, source redshift distributions and any resulting bias depends on selection procedures, intrinsic source properties and the probabilistic lensing effect [@DR74; @P82; @LB00]. The intrinsic source luminosity function and magnification probability distribution function (hereafter $\mu$PDF) are unknowns; the former is generally modelled as a Schechter function [@S76]. An appropriate model for the latter — which may be produced by ray-tracing through N-body simulations — is the subject of the present study (see Sec. \[mpdf\]). @HMF00 and @JL11 have demonstrated that a power-law tail in the $\mu$PDF significantly changes the shape of the bright end of the luminosity function and generates a considerable magnification bias for high-redshift sources. Gravitational lensing causes magnification or demagnification of Type Ia supernovae (SNIa), resulting in a scatter in the inferred distance-redshift distribution, particularly for high-redshift sources [@J08]. The lensing-induced scatter introduces a bias and uncertainty in the inferred values of the matter density parameter, ${\Omega_{\mathrm{M},0}}$, the deceleration parameter, $q_{0}$ and the dark energy equation of state, $w$ [@W97; @A06; @J08]. For example, in the case of the $z=1.7$ supernova SN1997ff, the estimated magnification would easily bridge the gap between evidence supporting a [$\Lambda$CDM]{} universe and that supporting an empty universe. The lensing effect can be ‘averaged out’, for supernovae samples, by combining measurements from a sufficiently large number of sources at similar redshifts [@HW98; @HL05]. However, a similar scatter will be induced by lensing in the Hubble diagram constructed from measurements of gravitational waves, the so-called standard sirens [@M93; @HH05; @JGM07]. Efforts are now underway to provide an appropriate method for correcting the observed brightnesses of individual objects for magnification [e.g. @G06; @JMS09; @S10; @HGK11]. By accounting for the non-Gaussian shape of the $\mu$PDF, one is best able to correct for the effects of lensing statistically [@HHC10; @SH11]. There are many approaches to determining the $\mu$PDF for a given cosmological model. The optical scalar equations [@S61], which describe the evolution of the cross-section of a small beam of light, have been applied to various mass distributions leading to useful redshift-distance relations in the limiting case when the line of sight is far away for inhomogeneities [@K69; @DR74; @D77]. The equations have also been applied to an infinitesimal beam transported through a generalized inhomogeneous universe that, on average, satisfies the Friedmann-Lemaître-Robertson-Walker (FLRW) geometry in order to re-derive the maximal angular diameter distance [@SSE94]. A simple integration of the optical scalar equations would be possible in the case of a known metric, but that is the crux of the problem - the local metric for an inhomogeneous universe has no general (analytic) solution. Crude model universes incorporate inhomogeneities and describe their effects on light propagation [e.g. @DR72; @L88; @FS89]. However, when the cosmological structure being probed is highly non-linear and lines of sight have small impact parameters, studying light deflection requires the use of numerical techniques. The non-linear and hierarchical growth of the large-scale structure is generally modelled by cosmological N-body simulations, with the propagation of light and its subsequent lensing computed by ray-tracing through these simulations. Some methods consider only one dominant lens in each line of sight; the lensing object is thin compared to the distances between observer, lens and source - this is known as the thin-lens approximation. However, multiple lenses coincidental along the line of sight may be responsible for a lensing event [@WBO05; @DO06]. The existence of multiple lens plane also accounts for the findings of @P01: that the magnification probability, $P(\mu>1)$ is mostly independent of the parameter that quantifies the normalisation of the matter power spectrum, $\sigma_{8}$. @R92 identified a feature of the $\mu$PDF: a ‘bump’ that was only evident when considering two-dimensional projections of matter; they speculated that changes to the caustic structure resulting from the projection was somewhat responsible. The presence of this caustic-induced bump was confirmed by @L97, who derived a semi-analytic expression for this feature in the limit of low-optical depth. Therefore, in the context of cosmological structure, the thin-lens approximation should be replaced, at the very least, by the multiple-lens plane approach (see Sec. \[multilensplane\]), where large volumes of matter are projected onto a series of lens planes [@BN86; @K87] . The ray-shooting method, developed by @KRS86 [@SW88; @W90; @WCO98], embodies this approach. Subsequent introduction of tree-methods to measure the deflection angle at each lens-plane have produced efficient algorithms. Early ray-shooting techniques were applied to a random distribution of point masses, which therefore did not incorporate the intricacies of clustering properties within various cosmological models [@SW88; @PW89; @LP90]. Later studies coupled the multiple lens-plane approach with mass distributions taken from N-body simulations to study the effect and magnitude of gravitational lensing [e.g. @J90]. While it is tempting to continue using the previous methods because of their simplicity, one must first quantify the effect that the various approximations have on the resulting predictions. Galaxy redshift surveys such as the CfA survey [@D82] and the 2dFGRS [@Cole05] have revealed that elongated structures exist in the form of filaments, which stretch across large voids between galaxy clusters. These are also evident in cosmological simulations. If such filaments are projected onto lens planes, the resulting magnification would be overestimated. Though earlier computational limitations, such as memory and processing, warranted the need for these simplifications, we are now entering an era where more accurate methods are within our reach. Only a few studies have numerically integrated the null geodesic equations from observer to source. The earliest of these studies assumed metrics that were approximated via a simplified model for inhomogeneities and derived distance-redshift relations that were compared to the Dyer-Roeder approximations [@FS89; @WT90]. @KFT90, drew attention to the spread in angular diameter distances given a small enough beam-size and the fact that the average value was consistently lower than the solution for a homogeneous universe. @T98 numerically integrated the null geodesic equations through N-body simulations, albeit at low-resolution, finding that various cosmologies exhibited differences in angular diameter distances. @CBT99 developed a three-dimensional algorithm for modelling weak gravitational lensing; comparing the two- and three-dimensional shear, they found that the projection of structure on 100 h$^{-1}$Mpc scales led to errors of up to 9 per cent depending on the redshift of the lens box. @VW03 performed three-dimensional ray-tracing by computing the deflection angle for rays many times along their path, advancing them in step-sizes of $L_{\mathrm{Box}}/N_{g}$ along the line of sight. In their study, they used a relatively small number of grid points along the line-of-sight $N_{g}=32$, such that the step-size was approximately 10 h$^{-1}$Mpc, but boxes of half that size were used for numerical tests. They found that their convergence maps were unchanged to 0.1 per cent. @WH00 run simulations of structure formation while simultaneously propagating photons through said simulation; they were able to produce convergence fields from the light rays they follow. What has been missing thus far is a one-to-one comparison of results that take a multiple-lens approach and a direct numerical integration of the null geodesic equations; this comparison is carried out in the present work. In Section \[gravlensing\], the basic theory and notation of gravitational lensing and the relevant statistics are presented. We introduce the method of numerically integrating the null geodesic equations in Section \[method\] with specific reference to its multiple lens plane counterpart; in Section \[models\], we establish the accuracy and limitations of the method. We describe the simulations used for modelling the cosmological mass distribution and compare statistical predictions determined by the two methods in Section \[results\]. Finally, a summary of our findings is presented in Section \[summary\]. Gravitational Lensing {#gravlensing} ===================== Here, we introduce the conventional notation for the components that describe the effect of gravitational lensing on the shape of a beam to first order. Each background source experiences lensing in the form of convergence, $\kappa$, and shear, $\boldsymbol{\gamma}$. Convergence is the isotropic Ricci focusing of a beam due to enclosed matter, while shear is the tidal stretching of the beam along a particular axis due to asymmetric matter distribution; together, they serve to increase the area of the source on the sky, resulting in its magnification, due to the conservation of surface brightness. The two-dimensional effective lensing potential, $\psi_{2D}$, is given by $$\label{potential} \psi_{2D}(\mathbf{x}) = \frac{1}{\pi}\int_{\mathbb{R}^{2}} d^2x'\kappa(\mathbf{x}')\mathrm{ln}\left|\mathbf{x}-\mathbf{x'}\right| .$$ Here $\mathbf{x}$ is a dimensionless vector formed by scaling the image position, $\boldsymbol{\xi}$: $$\label{scaleimagepos} \mathbf{x} = \frac{\boldsymbol{\xi}}{\xi_{0}}.$$ For a single thin lens, $\kappa$ is equivalent to the scaled surface density of the lens, and is related to the gravitational potential via the Poisson Equation $$\label{poisson} \frac{1}{2}\nabla^{2}\psi_{2D} = \kappa = \frac{\Sigma}{\Sigma_{\mathrm{crit}}},$$ where the critical surface density for gravitational lensing is given by $$\label{sigmacrit} \Sigma _{\mathrm{crit}} = \frac{c^{2}}{4\pi G}\frac{D_{s}}{D_{d}D_{ds}}.$$ Thus lensing is characterised by $D_{s}$, $D_{d}$ and $D_{ds}$, the angular diameter distances from the observer to the source, from the observer to the (thin) lens, and from the lens to the source respectively. Total shear can be written in complex notation; its dependency on the two orthogonal components are given by: $$\label{shearcomps} \boldsymbol{\gamma} = \gamma_{1} + \mathit{i}\gamma_{2},$$ where the two components are linearly related to the second derivatives of the (projected) gravitational potential along two orthogonal directions by $$\label{shearvpot} \gamma_{1} = \frac{1}{2} (\psi_{,11}-\psi_{,22}) \qquad \mathrm{and} \qquad \gamma_{2} = \psi_{,12},$$ where the indices after the comma denote partial differentation and we temporarily drop the subscript $2D$ for clarity. The deformation of the beam is described as a mapping from the source plane to the image (observed) plane. The Jacobian, $\pmb{A}$, of the lens mapping is a real and symmetric $2\times2$ matrix given by $$\label{lensmap} \pmb{A} \equiv \begin{pmatrix} 1-\kappa - \gamma_{1} & - \gamma_{2} \\ - \gamma_{2} & 1-\kappa + \gamma_{1} \\ \end{pmatrix}.$$ The flux magnification of an image is given by the inverse of the determinant of the Jacobian, so $$\label{magnification} \mu = \frac{1}{(1-\kappa)^{2} - \lvert\gamma\rvert^{2}}.$$ One may now consider the effect of lensing on the apparent position of the image of the source of light. The scaled deflection angle is the gradient of the lensing potential: $$\begin{aligned} \boldsymbol{\alpha_{2D}} & = \nabla\psi_{2D} \label{defangleA}\\ & = \frac{1}{\pi}\int_{\mathbb{R}^{2}} d^2x'\kappa(\mathbf{x}') \frac{\mathbf{x}-\mathbf{x'}}{\left|\mathbf{x}-\mathbf{x'}\right|^2} .\label{defangleB}\end{aligned}$$ The gravitational lens equation is therefore given by $$\label{lenseqn} \boldsymbol{\beta} = \boldsymbol{\theta} - \boldsymbol{\alpha_{2D}}(\boldsymbol{\theta}),$$ where $\bm{\beta}$ is the angular source position and $\bm{\theta}$ is the angular position of the image on the sky. The deflection angle, $\boldsymbol{\hat{\alpha}}$, is related to its scaled counterpart by: $$\label{defangle} \boldsymbol{\hat{\alpha_{2D}}} = \frac{\xi_{0}D_{s}}{D_{d}D_{ds}} \boldsymbol{\alpha_{2D}}.$$ The magnification probability distribution {#mpdf} ------------------------------------------ The probability that a source at redshift $z$ would be magnified by an amount within the interval \[$\mu$, $\mu+\mathrm{d}\mu$\] is $p(\mu,z)d\mu$. It satisfies : $$\label{normaliseprob} \int^{\infty}_{0} p(\mu,z)d\mu = 1.$$ When the $\mu$PDF is convolved with intrinsic luminosity distributions for standard candles, the result describes the observed spread in magnitudes. Flux conservation [@W76] demands that $p(\mu,z)$ satisfies $$\label{fluxconserve} \langle\mu\rangle \equiv \int^{\infty}_{0}\mu p(\mu,z)d\mu = 1.$$ There exists a minimum magnification, indeed a minimum convergence, which corresponds to a line of sight that encounters no matter between observer and source; this is also referred to as an empty beam [@DR72]. The distribution function peaks at values below $\mu=1$ and is highly skewed towards high magnifications [@HMF00]. For low redshift sources and/or small lensing optical depths, the high magnification tail will exhibit a power law trend $p(\mu)\propto\mu^{-3}$, which results in a formally divergent standard deviation [@VO83]. This result was originally derived analytically for a random distribution of compact lenses where only one lens dominates; however, @P93 have noted that for higher optical depths, the slope of tail may become shallower. The precise shape of the distribution depends on the assumed density profile of the lenses [@Y08]. The spread in this distribution increases with source redshift [@BL91]. For a fixed source redshift, the shape of the distribution, particularly for low magnifications, depends mostly on the cosmological parameters $\sigma_{8}$ and the vacuum density parameter, ${\Omega_{\mathrm{\Lambda},0}}$ [@P01]. The high magnification region suffers from low number statistics and since it represents the effects of strong lensing, cannot be probed by ray-bundle methods, which we now discuss. Ray Tracing {#method} =========== Multiple lens plane methods {#multilensplane} --------------------------- The multiple lens-plane method requires the total three-dimensional mass distribution to be sliced up into contiguous boxes; each box is then projected onto a plane perpendicular to the line-of-sight, usually placed at the centre of the box. In backward ray-tracing methods, light rays are propagated from the observer to the source with deflections [*only*]{} occurring in successive lens planes. The deflection at a given plane is the result of matter within that plane only. The formalism for the resulting deflection can be written in terms of the multiple lens-plane equation, derived and developed by @BN86, @SW88 and @J90; the source position, $\eta$, after deflection by $N$ lens planes: $$\boldsymbol{\eta} = \frac{D_{os}}{D_{o1}}\boldsymbol{\xi}_{1} - \sum^{N}_{i=1}D_{is}\boldsymbol{\hat{\alpha}_{i}}(\boldsymbol{\xi_{i}}),$$ where $\xi_{i}$ is the position of the ray in the $i^{\mathrm{th}}$ lens-plane; $\xi_{1}$, therefore, is the position of the image. $D_{os}$, $D_{o1}$ and $D_{is}$ are the angular diameter distances from the observer to the source, from the observer to the first lens-plane, and from the $i^{\mathrm{th}}$ lens to the source respectively. Notice that high density regions in a plane may not necessarily be dense in real space. One approach to studying the statistics of gravitational lensing is the ray-bundle method (RBM), developed by @FWM99, and see also @F99. Here, the multiple lens-plane method is used to propagate light rays from the observer to a given source plane. Instead of employing a grid-based technique for calculating magnifications across the source plane, the approach favoured by other backward ray-tracing codes [e.g. @JSW00; @P01], the RBM models each individual line of sight as an ‘infinitesimal bundle’ of light rays around a central ray and follows this bundle back to the source plane. An initially circular image is distorted by the time it reaches the source plane as a result of convergence and shear. These quantities can be determined for this specific line of sight as the image-source association is maintained. In the RBM, N-rays in a regular polygon represent a circular image and their positions at the source plane are fitted with an ellipse; the Jacobian matrix (Eqn. \[lensmap\]) is determined for each bundle and solved to determine $\mu$, $\kappa$ and $\gamma$. The ray-tracing method presented in this work uses the RBM design of a ray-bundle with 8 rays, but does away with the multiple lens-plane treatment of the lensing mass. Instead, the evolution of the cross section of the bundle is determined by integrating the null geodesic using a numerical gravitational potential obtained from N-body simulations. The weak field metric {#metric} --------------------- The Friedmann-Lemaître-Robertson-Walker (FLRW) geometry describes a universe that is homogeneous on all scales The FLRW line-element for a flat geometry is: $$\label{flrwmet} ds^2=-dt^2+R^2(t)\left[d\chi^2 + \chi^{2}\left(d\theta^2 + \textrm{sin}^2\theta \, d\phi^2\right)\right],$$ where $\chi$, $\theta$ and $\phi$ are comoving coordinates. The metric allows an evolving scale factor, $R(t)$. From here onwards, we quantify the scale factor relative to its current value, i.e. at $t=0$: $$\label{scalefac} a(t) \equiv \frac{R(t)}{R(0)} .$$ We assume that the inhomogeneities present in large scale structure are small enough to be represented as a perturbation to the background FLRW metric, and do not falsify the large-scale predictions made under FLRW geometry. The resulting line element is: $$\label{expandingmetric} ds^{2} = -\big[1 + 2\psi(t,\boldsymbol{x}) \big] dt^{2} + a^{2}(t) \big[1 - 2\psi(t,\boldsymbol{x}) \big] d\boldsymbol{x}^{2}$$ where $i$ denotes the three spatial dimensions. The weak-field metric is applicable where $\psi(t,\boldsymbol{x}) \ll a(t)$. The gravitational potential, which is defined with respect to the local perturbation from a smooth background (see Sec. \[phi\]), can be decoupled $$\label{phicotophys} \psi(t,\boldsymbol{x}) = \frac{\phi(\boldsymbol{x})}{a(t)}$$ such that $\psi$ is defined in physical units and $\phi$ in co-moving. The Christoffel symbols for this metric, presented in Appendix \[AppChristoffel\], are dependent, not only on the gradients of the gravitational potential $\phi$, but $\phi$ [*itself*]{}, which is defined based on the mass perturbation from a smooth background. The Geodesic Equations (see Eqn. \[geodesiceqn\] below), which are the second order differential equations for the four coordinates, are then constructed. The gravitational potential {#phi} --------------------------- The perturbation field: $$\label{perturbation} \delta(t,\boldsymbol{x}) \equiv \frac{\rho(t,\boldsymbol{x}) - \bar{\rho}(t)}{\bar{\rho}(t)}$$ relates the local density, $\rho(t,\boldsymbol{x})$, to the mean matter density, $\bar{\rho}(t)$, the latter of which is given by $$\label{meanmatterdens} \bar{\rho}(t) = {\Omega_{\mathrm{M}}}\rho_{c}(t).$$ The critical density, $\rho_{c}(t)$, is given by $$\label{critdens} \rho_{c}(t) = \frac{3H^{2}(t)}{8\pi G},$$ where $H(t)$ is the Hubble parameter: $$\label{hubbparam} H(t) = \frac{\dot{R}(t)}{R(t)}.$$ The perturbation is related to the gravitational potential $\psi(t,\boldsymbol{x})$, by $$\label{psi} \psi(t,\boldsymbol{x}) = -\frac{4\pi G}{c^{2}}\bar{\rho}(t)\int_{\mathbb{R}^{3}} d\boldsymbol{x}'^{3} \frac{\delta(t,\boldsymbol{x})}{\left|\mathbf{x}-\mathbf{x'}\right|}.$$ The derivatives of the gravitational potential are then $$\label{derivs} \frac{d\psi(t,\boldsymbol{x})}{d\boldsymbol{x}^{j}} = \frac{4\pi G}{c^{2}}\bar{\rho}(t)\int_{\mathbb{R}^{3}} d\boldsymbol{x}'^{3} \delta(t,\boldsymbol{x})\frac{\mathbf{x}-\mathbf{x'}}{\left|\mathbf{x}-\mathbf{x'}\right|^{3}}.$$ One can relate the numerically determined $\phi$ to the gravitational potential via Eqn. \[phicotophys\] and derivatives in a similar fashion: $$\label{phiderivscotophys} \frac{d\psi(t,\boldsymbol{x})}{d\boldsymbol{x}^{j}} = \frac{1}{a(t)} \frac{d\phi(\boldsymbol{x})}{d\boldsymbol{x}^{j}}.$$ Constructing the grids that represent the perturbation field and its derivative requires a few steps. Firstly, if the lensing mass is discretized into particles, then the mass within a cube of co-moving side length $L_{box}$ is assigned to the nodes of a regular grid using the Cloud-in-Cell algorithm [CIC; @HE88] in three dimensions; the mean matter density is then subtracted off. A Fast Fourier Transform is applied to convolve the mass distribution with the appropriate kernels to determine the gravitational field and its derivatives according to Eqns. \[psi\] and \[derivs\], using the popular software package ‘Fastest Fourier Transform in the West’ ([FFTW]{}[^2]). Depending on the lens, a periodic mass distribution may be implied, or not. If not, the density grid is zero-padded before performing the FFT convolution. At many stages during the integration, the local values of the field and its derivatives needs to be evaluated, so an interpolation scheme was written for this purpose; it is described in detail in Appendix \[AppInterp\]. There are two numerically intensive parts of the ray-tracing method: the first is the set of Fast Fourier Transforms required to calculate the gravitational potential and its derivatives at every grid point on a 3D mesh; the second is the interpolation required at each time-step to determine the values of the same quantities at the exact position of the light ray. Note also, that the Fourier-grid resolution sets a lower limit to the scale of structure probed, as the mass is smoothed over this grid scale. This is assuming that this scale is reasonably larger than the softening length employed in the N-body simulations used to model the cosmological mass distribution. Three-dimensional ray-tracing {#newcode} ----------------------------- Our approach is based on avoiding the approximation of the multiple lens plane method. Instead, the path of a photon is numerically integrated from the Geodesic Equation: $$\label{geodesiceqn} \frac{d^{2}x^{\alpha}}{d\lambda^{2}} = - \Gamma^{\alpha}_{\beta\gamma} \frac{dx^{\beta}}{d\lambda} \frac{dx^{\gamma}}{d\lambda}$$ Here $ \Gamma^{\alpha}_{\beta\gamma} $ denotes the Christoffel symbols associated with the metric, $x^{i}$ represents any of the four coordinates specified by the superscript $i$, and $\lambda$ is the affine parameter. The four second-order differential equations are reduced to eight coupled first-order differential equations, which are integrated with the affine parameter $\lambda$ as the dependent variable. A classical fourth order Runge-Kutta integration scheme with fixed step-size was written and used to perform the integration. The subtlety is that it is the [*$x^{3}$-coordinate*]{}, rather than the affine parameter, that determines the boundaries of the integration. The exact evolution of the affine parameter will be different for each ray-bundle, and is not known in advance. Therefore, the integral is repeated in small steps of $\lambda$, $\Delta\lambda_{\mathrm{RK}}$,until the $x^{3}$ reaches the value required at the source plane $D_{C}(z_{s})$. The step-size is chosen such that the estimated resultant step in $x^{3}$ is a fixed fraction, $f_{\mathrm{RK}}$, of the Fourier-grid resolution: $$\label{RKstepsize} \Delta\lambda_{\mathrm{RK}} = \frac{f_{\mathrm{RK}} L_{box}}{n_{\mathrm{FFT}}\dot{x}^{3}_{0}} ,$$ where the overdot denotes the derivative with respect to $\lambda$, the extra subscript $0$ denotes values at $t=0$, and $n_{\mathrm{FFT}}$ denotes the number of points along one side of the Fourier-grid (see Sec. \[phi\]). If the ray overshoots the source plane, a simple linear interpolation, using the current and previous positions, is used to find the final source position. The use of a fixed step-size is deemed appropriate as the values of the Christoffel symbols are interpolated from gridded values, as described in Sec. \[phi\]. There would be little information gain from step-sizes much smaller than the grid resolution. Nevertheless, the effect of the choice of step-size is analysed, along with other numerical parameters, with results presented in Sections \[schwarzschild\] and \[fakenfw\]. Evolution of the scale factor {#evolvescalefac} ----------------------------- For certain cosmologies, the Friedmann Equations can be solved to find the specific time dependence of the scale factor; for a spatially flat, radiation-free Lemaitre model (${\Omega_{\mathrm{k}}}=0$, ${\Omega_{\mathrm{M}}}+ {\Omega_{\mathrm{\Lambda}}}= 1$), with ${\Omega_{\mathrm{\Lambda}}}>0$, this dependence is given by : $$\label{scalefac2} a(t) = \left[ \frac{{\Omega_{\mathrm{M},0}}}{{\Omega_{\mathrm{\Lambda},0}}} \sinh^{2} \left( \frac{3}{2}\sqrt{{\Omega_{\mathrm{\Lambda},0}}}H_{0}(t_{0}-t) \right) \right]^{1/3},$$ where $t_{0}$ is given by $$\label{lookbacksingularity} t_{0} = \frac{2}{3H_{0}\sqrt{{\Omega_{\mathrm{\Lambda},0}}}} \sinh^{-1} \frac{{\Omega_{\mathrm{\Lambda},0}}}{{\Omega_{\mathrm{M},0}}}.$$ Eqns. \[scalefac2\] and \[lookbacksingularity\] were derived from Eqn. 15.36 in @Hobson06, the solution for a spatially flat, matter-only Lemaitre model, but with $t$ used to denote [*lookback time*]{} instead. As the photon traverses a single lens box, the scale factor will evolve (i.e. decrease) although the co-moving scale of the structure does not change appreciably over this time-scale. The value of the scale factor is required to evaluate the Christoffel symbols, so we evolve the lookback time for the photon as well as the spatial coordinates and use this to determine the scale factor at each position throughout the box. Compact Lens Models {#models} =================== Here, we present results of ray-tracing in the vicinity of simple lenses. Both of the lens models used are fairly compact, and so the thin lens approximation applied in the analytic solution is suitable. The analytic solutions should be read as magnification for the [*image*]{} at that location, rather than the total magnification of the source. This is an important distinction. The ray-bundle approach only follows the congruence corresponding to a single image, and therefore cannot account for the total magnification of multiply imaged sources; this is one of the reasons why it should not be used to model strong lensing. Yet for these simple lenses, we are still able to test our results against the known magnification of a single image for an otherwise strongly lensed source. Schwarzschild lenses {#schwarzschild} -------------------- The Schwarzschild lens, a singular density point, is the simplest lens one may study. The analytic solution for the magnification for an image at any location is well known and is presented in Appendix \[AppSchw\]. Since the gravitational potential here is equivalent to a kernel (see Eqn. \[psi\]) multiplied by the mass of the lens, we directly compute the gravitational field at each grid point, before performing the ray-tracing. The solution for this lens is scale-invariant, so by increasing the mass of the lens, we may essentially increase the resolution of the grid, and make it possible to study the behaviour of ray-bundles that pass near (or inside) the Einstein radius. In cosmological simulations, the RBM explicitly avoids this as the images near the Einstein ring are part of multiple set of images, and the images inside the ring are the demagnified images which contribute only a small portion of the magnification of the associated source. However, for testing purposes, we allow lines of sight that approach the Einstein radius since the analytic solution is known for each individual image. In the fiducial test case, the Schwarzschild lens has a mass of $10^{15}{\mathrm{M}_{\odot}}$. It is placed at a redshift of $z_{L}=0.35$, giving it an Einstein radius of $65.\mspace{-5.0mu}"6$ for a source redshift of $z_{s}=0.8$, which corresponds to $\sim12$ mesh-points on the $256^{3}$ Fourier-grid. The ray-tracing code developed here is able to reproduce the desired lensing distortion very well. The plots on the left hand side of Fig. \[schwtests\] show that even high magnification events, which occur when the image is near the Einstein radius, are recovered by the numerical method. {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} The numerical error, shown on the right hand side, increases for images that are closer to this radius where the higher-order lensing effects, like flexion, are expected to play a role and an eight-ray bundle has limited application regardless of the choices of other numerical parameters. Note that in each of the tests shown in the Figures \[schwtests\] – \[nfwtests3\], the results break down in the very centre of the lens. We choose 3 per cent as an error threshold to mark the sharp upturn in the average error at small impact parameters. The error surpasses 3 per cent when the impact parameter is smaller than a limiting radius, hereafter referred to as the ‘minimum reliable radius’, or MRR. The MRR can be used as a measure of numerical accuracy and the effect of numerical parameters. For example, the top row of Fig. \[schwtests\] shows the result of decreasing the number of interpolation points so that just one Fourier-grid node [*on either side*]{} of the current position is used to find the local values of the gravitational field and its derivatives. In this case, the MRR expands out to a larger impact parameter where only low magnification regions ($\mu\lesssim1.5$) are recovered with $<3$ per cent accuracy. However, once the number of interpolation points is increased beyond this, this parameter has little influence on the accuracy of the results. Likewise, on the bottom row of Fig. \[schwtests\] we show the result of reducing the Runge-Kutta step-size; a step that is approximately half the size of the Fourier-grid resolution ($f_{\mathrm{RK}}=0.5$) is sufficient for accuracy at moderate magnifications ($1.5\lesssim\mu\lesssim6$). Decreasing this parameter has negligible effect on the results. NFW lenses {#fakenfw} ---------- Various studies of cosmological simulations [@NFW95] have found that dark matter haloes on galactic and cluster scales have mass distributions that are well described by a Navarro-Frenk-White (NFW) profile: $$\label{nfwprof} \rho(r) = \frac{\delta_{c}\rho_{c}}{(r/r_{s})(1+r/r_{s})^{2}}.$$ Here, $\rho_{c}$ is the critical density (see Eqn. \[critdens\]) [*at the halo redshift*]{}; $r_{200}$ is the radius within which the mean density of the halo is $200\rho_{c}$; $r_{s}=r_{200}/{c_{\Delta}}$ is the characteristic scale radius, which marks the transition in the the slope of the profile; and ${c_{\Delta}}$ is the dimensionless concentration parameter. Finally, the characteristic overdensity is given by: $$\label{overdens} \delta_{c} = \frac{200}{3}\frac{{c_{\Delta}}^{3}}{\ln(1+{c_{\Delta}})-{c_{\Delta}}/(1+{c_{\Delta}})}.$$ We use this profile to compare how the multiple-lens plane approach and the three-dimensional approach model lensing around galaxy and cluster haloes, where most of the high magnification events will occur. Lenses with the NFW profile were modelled by discretising the mass contained in one virial radius into a number of particles of equal masses and constructing a fake simulation output in the [GADGET]{} format. The numerically determined gravitational potential is therefore [*not*]{} equivalent to the analytic solution, but suffers from discretisation effects just as a simulated halo would. The spherically symmetric lens is divided into a fixed number of radial bins of equal width, extending from the centre to the virial radius. Each particle is randomly placed in one of the radial bins with a probability proportional to the mass within that bin; the bin mass is found by appropriately integrating the density profile, given in Eqn. \[nfwprof\]. The lenses modelled for this purpose have a virial mass, and total mass, of $M_{vir}/{\mathrm{M}_{\odot}}= 10^{14}$ and concentration parameter ${c_{\Delta}}=7.2$. At a (lens) redshift of $z_{L}=0.35$ and a virial overdensity of $\Delta_{c}= 200$, their virial radii are $0.55$ h$^{-1}$Mpc. The results presented here are for a source redshift of $z_{S}=0.8$. Figures \[nfwtests1\] – \[nfwtests3\] show the magnifications determined by ray-tracing over a range of impact parameters changing one numerical parameter relative to a fiducial choice. The ray-bundles are allowed to encounter the lens at a distance of up to twice the virial radius, which corresponds to an angular separation of $343"$. The analytic solutions for image magnification and shear by a lens with the NFW profile are presented in Appendix \[AppNFW\]. Since the model lens is truncated at one virial radius, the metric on the exterior is given by the Schwarzschild metric, by Birkhoff’s theorem. Thus the analytic solution shown for $\mu(r)$ switches to that appropriate for a Schwarzschild lens outside this impact parameter. {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} Between the MRR and the truncation radius, the error remains fairly constant at 1 per cent. Outside the truncation radius, the error drops sharply by an order of magnitude, which implies that the removal of mass outside this radius is primarily responsible for the aforementioned 1 per cent error. The size of the ray-bundle, or equivalently, the radius of the image or beam, has negligible effect on the ray-tracing results (see top row of Fig. \[nfwtests1\]). By varying the grid size (see top row of Fig. \[nfwtests2\]), we notice that the MRR has a very strong dependence on the Fourier-grid resolution. Our interpretation is that the interpolation scheme used to measure the local gravitational field would smooth over the central cusp, causing the large errors here. We conclude that if small-scale structure is responsible for a caustic (high magnification) then this will be washed out. The MRR has a strong dependence on the number of points used to interpolate the values of the gravitational potential and its derivatives from the Fourier-grid on which they are calculated, as shown in the bottom row of Fig. \[nfwtests1\]. The use of more interpolation points allows for more robust ray-tracing, with a more accurate mean magnification value. However, the computation time rises approximately linearly with the number of interpolation points used, and the interpolation scheme suffers in the vicinity of density peaks. The latter issue should not pose too much of a problem when the lensing mass distribution is large-scale structure, but computational efficiency presents a significant hurdle for ray-tracing procedures such as this. The parameters that most govern the accuracy of the ray-tracing scheme are the Fourier-grid size and mass smoothing length, which have essentially equivalent effects; assigning mass to an Fourier-grid acts to spread it out over a fixed number of grid-points. A CIC mass assigning scheme is different to a Gaussian filter smoothing, but in essence, the larger the area over which mass is spread, the more density peaks are suppressed and the larger the error in the tests shown in Fig. \[nfwtests2\]. On the top row of Fig. \[nfwtests3\], we show that halving the Runge-Kutta step-size has a negligible effect, so half the Fourier-grid resolution is deemed sufficient. The model lens used for this test is subject to a few user-defined variables; a few convergence tests were run to ensure that these were not the cause of the observed features. Trialling different mass resolutions from $10^{9}$ to $10^{11} {\mathrm{M}_{\odot}}$ for a fixed halo virial mass, $M=10^{14}{\mathrm{M}_{\odot}}$, the results are shown in the panels of the middle row of Fig. \[nfwtests3\]. We find that the resulting scatter increases from less than 1 per cent up to 10 per cent at a fixed impact parameter. One could equivalently say that the MRR is larger for haloes that are not as well resolved. The cosmological mass distributions in Sec. \[sims\] have mass resolutions better than $10^{10}{\mathrm{M}_{\odot}}$, so one can expect a low error in magnification even within dense regions. The number of bins that the lens is divided into is a choice that has negligible effect on the ray-tracing results, as shown in the bottom row of Fig. \[nfwtests3\]. One is therefore reassured that the results of the ray-tracing on the discretised NFW profile are relevant to the results of ray-tracing through cosmological mass distributions based on the fiducial parameter choices. @KF08 numerically integrate the null geodesic equations (with the adaptive stepsize Runge-Kutta-Fehlberg 4-5 method) in order to test the accuracy of the thin-lens approximation, but only with reference to strong lensing by singular isothermal sphere (SIS) and Navarro-Frenk-White [NFW; @NFW95] mass profiles, both of which are relatively thin. Large-Scale Structure {#results} ===================== Cosmological simulations {#sims} ------------------------ In order to construct the predicted probability distributions for the weak lensing statistics numerically, numerous null geodesic equations are integrated through a cosmological mass distribution that is generated with N-body simulations. The cosmological simulations are carried out with the parallel Tree-PM SPH code `GADGET2` [@S05] using collisionless particles only. We adopt a [$\Lambda$CDM]{} cosmology, with the following values for the cosmological parameters: ${\Omega_{\mathrm{M},0}}$ = 0.27, ${\Omega_{\mathrm{\Lambda},0}}$ = 0.73, $h$=0.71, $\sigma_{8}=0.9$. The dark matter distribution is discretised into $256^{3}$ particles distributed within a periodic box with co-moving length of $L_{\mathrm{box}}=50$ h$^{-1}$Mpc, resulting in a mass resolution of $m_{p}=6.3\times10^{8}{\mathrm{M}_{\odot}}$; the initial displacements are in a ‘glass’ configuration. The simulations begin at an initial redshift of $z_{i}=39$, with a redshift-dependent gravitational force softening length of $\epsilon_{co}=16$ h$^{-1}$kpc (Plummer-equivalent). The space between an observer and source is divided up into individual regions, each modelled by a snapshot of a cosmological simulation at an appropriate redshift. The snapshot cadence, $\Delta z$, is chosen such that the light travel time corresponds to the length of the boxes. The line-of-sight integrated co-moving distance between an observer and source at $z=z_{s}$ is: $$\label{raytracedistance} D_{c} = c\int_{0}^{z_{s}}\frac{dz}{H(z)}.$$ Since this cannot be easily evaluated in general, we instead assume that the differences between the redshift of adjacent boxes are small, and thus: $$\label{boxsizestack} L_{\mathrm{box}} = \frac{c \Delta z}{H(z)},$$ which is easily rearranged to determine the snapshot cadence. A total of 28 boxes are required to fill the space between an observer and a source at $z\approx0.5$. Note that the scale factor at the position of a photon is required to integrate the null geodesic equations. This scale factor is [*not*]{} identified with the ‘average’ redshift of the current snapshot being traced through; instead it is derived from the lookback time (see Eqn. \[evolvescalefac\]) and evolves during the integration through the box. That is to say, it is assumed that in the time it takes for a photon to traverse a box, the co-moving scale of the structure remains constant, however the physical scale length changes. As a sanity check, the numerically derived value of the scale factor at each box interface is compared to the redshift of the appropriate snapshot. The sampling region ------------------- In order to avoid repeated structure, a number of precautions were taken. Nine independent simulations were run, resulting in nine separate realisations. The snapshot for each required redshift was chosen randomly from among the nine realisations. The chosen box was then randomly translated and rotated $90^{\circ}$ about any/all of the axes. From each 50 h$^{-1}$Mpc box we selected the lensing mass formed by particles within a prism 6.25 h$^{-1}$Mpc across the sky and 50 h$^{-1}$Mpc along the line of sight. This is split into 8 cubes each of volume $ V = 6.25^{3}$ (h$^{-1}$Mpc)$^{3} \approx 244$ (h$^{-1}$Mpc)$^{3}$. Each of these has its mass placed on a $128^{3}$ point grid, using the CIC algorithm. We apply an FFT to these boxes with zero-padding so that the density field (and the derivatives of the gravitational potential) due to structure [*within the box only*]{} is determined on a scale of 49 h$^{-1}$kpc. Note here that the simulation is run on a larger box such that the large-scale modes are included in the formation of structure, but we only use a smaller portion of the box for the lensing (see Fig. \[RayTracingDiagram\] for a cartoon diagram). The null geodesic equations of the eight rays and one anchor of each ray-bundle are numerically integrated using the methods described in Sec. \[newcode\]. {width="\linewidth"} \[RayTracingDiagram\] Lines of sight that fall too close to the edge of the boxes will be artificially sheared, so assuming that most of the lensing occurs due to mass within 0.5 h$^{-1}$Mpc, we only send out ‘beams’ that will fall within the central $5.25\times5.25$ (h$^{-1}$Mpc)$^{2}$ region at the source redshift. For a source redshift of z=0.5, with an angular diameter distance of approximately 900 h$^{-1}$Mpc, this means that the area of sky sampled is about $0.33^{\circ}\times0.33^{\circ}$ or $5\times10^{-6}$ of the entire sky. To increase the sampling area, this entire procedure is used to trace the paths of 10 000 bundles, then repeated for another 10 000 bundles using another lensing mass distribution sampled from the same set of simulations. Numerically determining the $\mu$PDFs {#numericalPDF} ------------------------------------- Using ray-tracing procedures to construct a magnification probability distribution is effectively equivalent to calculating the angular diameter distance in different directions out to the same co-moving distance. The distribution of angular diameter distances for a given source redshift has been investigated by @HB97. In the extremal case of a large beam-size, and therefore, large separations between rays, the angular diameter distances would be the same in all directions and the ‘distribution’ would be a single peak at $D_{fb}(z)$, the average FLRW value, implying homogeneity on those scales. However, if the beam-size is small enough, then the inhomogeneities induce a distribution. Each ray-bundle probes a single light of sight. More specifically, it describes the lensing that has affected a single image of a fixed size on the sky. The number of bundles that exhibit a magnification of $\mu$ are thus proportional to the probability that an [*image*]{} is magnified by $\mu$. However, the statistical quantity of interest is the magnification probability distribution for [*sources*]{}. Therefore, the number counts that are used to produce the magnification probability histograms presented in Sec. \[compareRBM\] are weighted by the area covered by the beam at the source plane, or equivalently, the inverse of the magnification. $$\label{weighthisto} P(\mu)d\mu = \frac{F(\mu,\mu+d\mu)}{\mu}$$ where $F(\mu_{1},\mu_{2})$ is the fraction of all ray-bundles for which $\mu_{1}<\mu<\mu_{2}$. Similarly the mean and standard deviation of these properties are also weighted. Comparison of the PDFs produced with and without lens planes {#compareRBM} ------------------------------------------------------------ The predicted probability distributions of magnification and shear are constructed by ray-tracing $2\times10^{4}$ bundles through the three-dimensional mass distribution described above. The results, shown in Figures \[RBMvsMudMag\] and \[RBMvsMudShear\], are compared to the probability distributions constructed with the multiple lens-plane RBM using $5\times10^{4}$ bundles. However, not all ray-bundles are included in the analysis. The RBM is very well suited to the weak lensing limit and provides a computationally efficient alternative to grid-based ray-shooting methods; however, it underestimates the magnification in cases where multiple imaging is expected as only one image contributes to the magnification of each source. For this reason, rays that fall within a minimum distance of a lens (remembering that a distribution of point masses describes the lensing mass) are excluded from the analysis. This affects the high magnification probabilities deduced, but does not significantly affect the weak lensing analysis [@FWM02]. The RBM avoids the danger of artificial shear due to a source pixel collecting light rays that have passed near the edge of a shooting grid. If a ray-bundle were to suffer from neither convergence nor shear, it would obtain the minimal magnification, $\mu_{eb,min}=1$, or equivalently, $\mu_{fb,min}= D_{fb}^{2}(z_{s})/D_{eb}^{2}(z_{s})$ (see Appendix \[AppBeam\]). For a source redshift of $z=0.5$, this minimum magnification is $\approx 0.965$; the relevant angular diameter distances have been determined with the aid of the [Angsiz]{} routine [@KHS97]. Those bundles that produce magnifications below this minimum value represent the demagnified components of a multiple-image system, a possible consequence of strong lensing. These will highly underestimate the total magnification of the associated source, and so, are also excluded in both methods. On the other hand, if a magnified image is traced back to the source plane, but belongs to a multiple-image system, it will not be identified as such. Although it will also underestimate the total magnification, it will generally do so by a negligible amount. The ray-tracing method developed in the present work does not produce any bundles that must be excluded. ![The magnification probability distribution for a source at redshift $z=0.5$ as predicted by the multiple lens-plane RBM (blue dashed line) and by the three-dimensional equivalent (black solid line)[]{data-label="RBMvsMudMag"}](fig_rayt-cosmo-XXX-mag.eps){width="0.95\linewidth"} The average magnification predicted by both methods is close to unity, as required by flux conservation (see Eqn. \[fluxconserve\]). More precisely, the multiple-lens plane approach finds $\bar{\mu}=1.0199$ while the three-dimensional approach finds a mean magnification of $\bar{\mu}=0.9994$. Using the three-dimensional method to analyse lensing by large-scale structure, we find that large magnifications are not produced, as shown in Fig. \[RBMvsMudMag\]. The slope of the differential magnification probability distribution is found to be much steeper than for RBM for the regime $1.1<\mu<1.5$. The difference in the two methods can also quantified by the standard deviation, found to be $\sigma_{\mu}=0.205$ from the multiple-lens plane approach and $\sigma_{\mu}=0.021$ from the three-dimensional approach, and order of magnitude lower. To understand the reason of this difference, we also plot the probability distribution for shear, shown in Fig. \[RBMvsMudShear\]. While the overall shape of the distribution function is similar as derived by the two methods, the multiple-lens plane method measures more high shear values and less low shear values. ![The shear probability distribution for a source at redshift $z=0.5$ as predicted by the multiple lens-plane RBM (blue dashed line) and by the three-dimensional equivalent (black solid line)[]{data-label="RBMvsMudShear"}](fig_rayt-cosmo-XXX-shear.eps){width="0.95\linewidth"} The projection of matter onto multiple lens-planes may have artificially increased the shearing effect of cosmological lenses. However, the effect of shot-noise in the standard RBM method is identified as another possible reason for the discrepancy. When the three-dimensional approach is used, the gravitational potential is calculated by means of Fourier-methods. In contrast, the multiple-lens plane approach determines the potential field at the location of each ray, but adding the potential due to each simulation particle in the plane individually. Each particle in the simulation represents some underlying smooth mass distribution, but by treating them as point masses, the method is susceptible to the effects of shot-noise. Recently, @T11 discussed the effects of shot-noise on the variance of the convergence found by ray-tracing through N-body simulations. Although they did not use the same ray-bundle approach that we do here, they apply a multiple lens-plane method with an FFT for calculating the gravitational potential in each plane; they compare their results for a range of Fourier-grid resolutions (see Figure 2 of their paper). When the grid resolution is large enough to smear out structure, the variance of the convergence falls below theoretical predictions; although our choice of statistic is different, our results agree with this interpretation (see Fig. \[nfwtests2\] in Sec. \[fakenfw\]). Interestingly, a very small grid-resolution ($<5$ h$^{-1}$kpc) leads to a variance that is larger than the theoretical prediction, a result which they attribute to shot-noise. We agree with this interpretation, noting that Fourier methods applied to grid-resolutions that are much smaller than the mean-interparticle distance would resemble a direct summation as employed by the multiple-lens plane approach in the present work. The $\mu$PDF predicted by the three-dimensional method is subject to certain choices of numerical parameters, some of which have been discussed above. Thus, we test the parameters that are most likely to have an impact on the cosmological lensing results presented above. For example, a parameter that was not relevant to the tests presented in Sec. \[models\] is the force-softening length chosen for the cosmological N-body simulation. We re-run the simulation described earlier but with a force-softening length reduced to $\epsilon = 5$ h$^{-1}$kpc. We sample $3\times10^{4}$ lines of sight through the mass distribution determined with this simulation, and in Fig. \[MudMagSoftening\], we compare the $\mu$PDF to the result for the three-dimensional method shown in Fig. \[RBMvsMudMag\]. For each, the Fourier-grid resolution is 49 h$^{-1}$kpc. The similarities in the PDFs reassures us that any smoothing of structure below 16 h$^{-1}$kpc is not responsible for any error in the PDFs measured by the three-dimensional method. ![The magnification probability distribution for a source at redshift $z=0.5$ as predicted by the three-dimensional method. Two different force-softening lengths applied in the N-body simulations are compared: $\epsilon = 5$ h$^{-1}$kpc (red dashed line), and $\epsilon = 16$ h$^{-1}$kpc (black solid line)[]{data-label="MudMagSoftening"}](fig_rayt-cosmo-mud-stack08-nm128-epsXX-mag.eps){width="0.95\linewidth"} In addition, as the Fourier-grid scale is reduced, the numerical solution is expected to be more accurate, with the caveat that it remains large enough to prevent the effects of shot-noise. In Fig. \[MudMagFFTres\] we compare two different Fourier-grid resolutions: 24 h$^{-1}$kpc for which $9\times10^{4}$ lines of sight have been sampled, and 49 h$^{-1}$kpc for which $3\times 10^{4}$ lines of sight have been sampled. Both tests are run on simulations with a force-softening length reduced to $\epsilon = 5$ h$^{-1}$kpc, which are only able to reliably describe structure on scales larger than 20 h$^{-1}$kpc. The differences are negligible, which demonstrate that most of the weak lensing results from structure above 49 h$^{-1}$kpc scales. Structures on scales larger than 24 h$^{-1}$kpc that are able to produce strong or even ‘moderate’ lensing are rare. ![The magnification probability distribution for a source at redshift $z=0.5$ as predicted by the three-dimensional method. Two different Fourier-grid resolutions are compared: 24 h$^{-1}$kpc (green solid line), and 49 h$^{-1}$kpc (red dashed line)[]{data-label="MudMagFFTres"}](fig_rayt-cosmo-mud-fftXXX-eps05-mag.eps){width="0.95\linewidth"} We recognize that there are multiple ways to create the same Fourier-grid resolution. If the number of cubes taken from the simulation boxes are doubled, but the number of Fourier grid-points halved, the resolution does not change. However, only a quarter of the patch of sky would be sampled, and the computational run-time would double. On the upside, the memory usage reduces to an [*eighth*]{}, which is a significant advantage for such a computational demanding approach. In Fig. \[MudMagBoxes\] we show the result of doubling the number of cubes in such a manner; the Fourier-grid resolution is 24 h$^{-1}$kpc. A total of $3\times10^{4}$ lines of sight have been sampled for the method using 8 cubes, while $6\times10^{4}$ lines of sight have been sampled for the method using 16 cubes. Note that this was not tested on the compact lens models in Sec. \[models\] since cubes are not excised from simulations as they are when analysing cosmological lensing. We are satisfied that the choice of 8 cubes for our previous results (see Fig. \[RBMvsMudMag\] and Fig. \[RBMvsMudShear\]) has no bearing on the results. ![The magnification probability distribution for a source at redshift $z=0.5$ as predicted by the three-dimensional method. The same Fourier-grid resolution (24 h$^{-1}$kpc) is created, but excising either 8 cubes (purple dashed line) or 16 cubes (yellow solid line) from each simulation box[]{data-label="MudMagBoxes"}](fig_rayt-cosmo-mud-stackXX-nmXXX-eps05-mag.eps){width="0.95\linewidth"} Summary and Discussion {#summary} ====================== Modelling the magnification probability distribution of background sources due to gravitational lensing relies on the application of ray-tracing methods. This work has taken the first steps towards making a direct comparison between the predictions of the multiple lens-plane ray-bundle method (RBM) and one that does not invoke the thin-lens approximation; instead, the null geodesic equations are integrated. The efficiency and accuracy of this computationally challenging approach can be improved by careful choices of numerical parameters; therefore, the results are analysed for the behaviour of the ray-tracing code in the vicinity of Schwarzschild and NFW lenses. A range of tests were able to pin down the numerical parameters that play a critical role in the predicted statistics. The behaviour of the ray-bundles in the vicinity of a Schwarzschild lens demonstrated that the method can reproduce large magnifications given sufficient spatial resolution. The limitations are dominated by the spatial resolution of the Fourier-grid and the mass resolution of the discretized lens. Comparisons to a multiple lens-plane algorithm are drawn in the context of cosmological mass distribution for a source redshift of $z_{s}=0.5$. The weak lensing statistics predicted by ray-tracing through simulated cosmological mass distributions found significant differences compared to the results from the original RBM approach. Either the use of multiple lens-planes or shot-noise are responsible for the observed differences. To clarify the dominant factor, a multiple lens-plane ray-bundle method that applies a two-dimensional Fourier-method at each lens-plane is proposed as the next step towards the direct comparison. The method developed here presents a computational challenge as the three-dimensional FFTs prove memory-expensive, but is justified by the need to quantify the multiple lens-plane approximation. We are, as yet, unable to sample the large number of lines of sight required to model the intermediate magnification ($2<\mu<20$) region where @R92 had identified the feature present only in two dimensional lens models. We now discuss future applications of the this ray-tracing method that are outside the scope of the present work. For example, proposed large surveys have the measurement of cosmic shear, with particular focus on the determination of the nature of dark energy, as one of their main science drivers, using a combination of a large area of sky coverage and high-precision photometric redshifts. These include the ground-based Panoramic Survey Telescope and Rapid Response System (Pan-STARRS)[^3], the VST-KIlo-Degree Survey (KIDS)[^4], the Dark Energy Survey (DES)[^5], the Large Synoptic Survey Telescope (LSST)[^6] as well as the space-based Euclid [@L09][^7] and the Joint Dark Energy Mission (JDEM)[^8], which has recently been rebranded as the Wide-Field InfraRed Survey Telescope (WFIRST). The method developed in the present work has the potential to quantify the effects of multiple lens-plane techniques on the theoretical predictions for cosmic shear measurements. Another alternative to ray-tracing is to retain the three dimensional mass distribution, but assume that the deflections are small enough to satisfy the Born approximation. Here, the matter that is integrated along the (un-deflected) line of sight is solely responsible for the convergence. Over larger and larger distances, the Born approximation becomes less and less accurate; several studies have found the requirement for corrections in the construction of the shear power spectrum and higher order bispectrum [e.g. @VW01; @SC06]. We envisage that the three-dimensional method could be used to make a one-to-one comparison with the Born approximation, just as we have done here with the multiple lens-plane method. Finally, with a larger number of ray-bundles and greater sky-sampling, we could turn our attention to higher-order statistics. Flexion is the third-order effect in gravitational lensing and is effectively the gradient of the shear component [see @GB05]. It has recently been demonstrated that flexion can be modelled using a ray-bundle method [@FL11]; this presents an exciting future application of the method, but surely it is even more necessary to justify the use of the multiple lens-plane formalism. Acknowledgments {#acknowledgments .unnumbered} =============== M.K. acknowledges support from the University of Sydney Faculty of Science Postgraduate Award, as well as a fellowship from the European Commission’s Framework Programme 7, through the Marie Curie Initial Training Network CosmoComp (PITN-GA-2009-238356) and would like to thank the Centre for Astrophysics and Supercomputing at the Swinburne University of Technology for their hospitality and for the use of the Green Machine. PL is supported by the Alexander von Humboldt Foundation. This work is undertaken as part of the Commonwealth Cosmology Initiative (www.thecci.org), and funded by the Australian Research Council Discovery Project DP0665574. Christoffel Symbols {#AppChristoffel} =================== The expanding weak-field metric presented in Eqn. \[expandingmetric\] has the following non-zero Christoffel symbols: $$\begin{aligned} \label{christoffel} \Gamma_{tt}^{x} & = & \frac{1}{\mathrm{a}^{3}}\phi_{,x} \notag \\ \Gamma_{tt}^{y} & = &\frac{1}{\mathrm{a}^{3}}\phi_{,y} \notag \\ \Gamma_{tt}^{z} & = & \frac{1}{\mathrm{a}^{3}}\phi_{,z} \notag \\ \Gamma_{tt}^{t} & = & -\frac{\dot{\mathrm{a}}}{\mathrm{a}^{2}}\phi \notag \\ \Gamma_{tx}^{x} = \Gamma_{ty}^{y} = \Gamma_{tz}^{z} & = &\frac{\dot{\mathrm{a}}}{\mathrm{a}^{2}}(\mathrm{a}+\phi) \notag \\ \Gamma_{tx}^{t} = \Gamma_{yy}^{x} = \Gamma_{zz}^{x} & = & \frac{1}{\mathrm{a}}\phi_{,x} \notag \\ \Gamma_{xx}^{x} = \Gamma_{xy}^{y} = \Gamma_{xz}^{z} & = &-\frac{1}{\mathrm{a}}\phi_{,x} \notag \\ \Gamma_{ty}^{t} = \Gamma_{xx}^{y} = \Gamma_{zz}^{y} & = & \frac{1}{\mathrm{a}}\phi_{,y} \notag \\ \Gamma_{yy}^{y} = \Gamma_{yx}^{x} = \Gamma_{yz}^{z} & = &-\frac{1}{\mathrm{a}}\phi_{,y} \notag \\ \Gamma_{tz}^{t} = \Gamma_{xx}^{z} = \Gamma_{yy}^{z} & = & \frac{1}{\mathrm{a}}\phi_{,z} \notag \\ \Gamma_{zz}^{z} = \Gamma_{zx}^{x} = \Gamma_{zy}^{y} & = &-\frac{1}{\mathrm{a}}\phi_{,z} \notag \\ \Gamma_{xx}^{t} = \Gamma_{yy}^{t} = \Gamma_{zz}^{t} & = & \dot{\mathrm{a}}(\mathrm{a}+\phi)\end{aligned}$$ Note that $x$ denotes $x^{1}$, $y$ denotes $x^{2}$, and $z$ denotes $x^{3}$. The Christoffel symbols for this metric are dependent, not only on the gradients of the gravitational potential $\phi$, but $\phi$ [*itself*]{}. However, the weak field condition requires that the magnitude of the perturbations satisfy $\phi \ll\mathrm{a}$. The Geodesic Equations - the second order differential equations for the four coordinates - are then constructed. The Interpolation Scheme {#AppInterp} ======================== Here we only describe the scheme for the evaluation of $\phi(x^{1}_{\star},x^{2}_{\star},x^{3}_{\star})$ given the gridded values $\phi(x^{1},x^{2},x^{3})$, but the same method is applied to evaluate the derivatives of the potential. The scheme is not strictly tri-cubic interpolation, but a three-dimensional version of the bicubic spline. Instead, it initially performs a one-dimensional spline to compute the derivative with respect to the line of sight; i.e. it fits a spline to find $d\phi(x^{1},x^{2},x^{3})/dx^{3}$ across the entire grid. This is only done once for each lensing mass distribution. Each time the local values need to be found, the scheme identifies the gridded values that include $n_{int}$ points on either side of the current location for each dimension, i.e. a $(2n_{int})^{3}$ mesh grid. Then, the scheme performs the following steps: 1. The gridded derivative is interpolated across the chosen dimension, in this case the $x^{3}$-dimension, to find $\phi(x^{1},x^{2},x^{3}_{\star})$, i.e. the value with the current $x^{3}$-coordinate of the photon specified, but the other coordinates corresponding to grid points. $(2n_{int})^{2}$ interpolations are required. 2. Another set of $2n_{int}$ one-dimensional splines are performed, this time across the $x^{1}$-axis, to find $d\phi(x^{1},x^{2},x^{3}_{\star})/dx^{1}$. 3. $2n_{int}$ interpolations along the $x^{1}$-axis determine $\phi(x^{1}_{\star},x^{2},x^{3}_{\star})$. 4. A single one-dimensional spline across the $x^{2}$-axis is used to find $d\phi(x^{1}_{\star},x^{2},x^{3}_{\star})/dx^{2}$ 5. Finally, a single interpolation along the $x^{2}$-axis allows one to evaluate $\phi(x^{1}_{\star},x^{2}_{\star},x^{3}_{\star})$ Schwarzschild Lens {#AppSchw} ================== A single point of infinite density is one of the simplest lens models, and is often used to model a spherically symmetric lens for idealised analytical studies. The relevant (angular) scale length for a so-called Schwarzschild lens of mass $M$ is the Einstein radius: $$\label{Eangle} \theta_{\mathrm{Ein}} = \sqrt{\frac{D_{ds}}{D_{d}D_{s}}} \frac{\sqrt{4GM}}{c} .$$ At the lens plane, this corresponds to a physical scale length given by: $$\label{Eradius} r_{\mathrm{Ein}} = \theta_{\mathrm{Ein}} D_{d}.$$ The gravitational lensing quantities of interest are found by combining the derivatives of this potential, as described in Sec. \[gravlensing\]. They are defined in terms of a dimensionless radial distance, $x=r/r_{\mathrm{Ein}}$. All lines of sight probe regions outside the lens, so the convergence is zero. The solution for the shear for images lensed by a Schwarzschild lens, $\gamma_{\mathrm{Schw}}$, is: $$\label{schwgamma} \gamma_{\mathrm{Schw}}(x) = x.$$ The analytic solution for the magnification for an image at any location, $\mu_{\mathrm{Schw}}$, is: $$\label{schwmu} \mu_{\mathrm{Schw}}(x) = \left(1-\frac{1}{x^{4}}\right)^{-1}.$$ NFW lens {#AppNFW} ======== The NFW profile, given in Eqn. \[nfwprof\], can be considered a thin lens for cases where the observer, lens and source are separated by large angular diameter distances. The following analytic solutions, \[nfwkappa\]-\[nfwgmore\], are therefore derived from projected surface densities. A dimensionless radial distance, $x=r/r_{s}$, has been adopted. Here, the quantities $\rho_{c}$, $r_{s}$, $\delta_{c}$ and $\Sigma_{crit}$ are those defined in Eqns. \[critdens\], \[nfwprof\], \[overdens\] and \[sigmacrit\]. Following @WB00 and @C10, the analytic solution for the radial dependence of the convergence is given by $$\label{nfwkappa} \kappa_{\mathrm{NFW}}(x) = \frac{r_{s}\delta_{c}\rho_{c}}{\Sigma_{crit}} K(x) ,$$ where $K(x)$ is given by $$\label{nfwK} K(x) = \left\{ \begin{array}{rll} &\frac{2}{(x^{2}-1)}\left[1-\frac{2}{\sqrt{1-x^{2}}}\mathrm{tanh}^{-1}\sqrt{\frac{1-x}{1+x}}\right] & \text{if } x<1 \\ \\ &\frac{2}{3} & \text{if } x=1 \\ \\ &\frac{2}{(x^{2}-1)}\left[1-\frac{2}{\sqrt{x^{2}-1}}\tan^{-1}\sqrt{\frac{x-1}{1+x}}\right] & \text{if } x>1 \end{array} \right..$$ By integrating Eqn. \[nfwkappa\] over the area within $r$, one finds the mass within a [*cylinder*]{} of radius $r$: $$\label{nfwcylinder} M_{\mathrm{NFW,cyl}}(x) = 4\pi r_{s}^{3}\delta_{c}\rho_{c} C(x) ,$$ where $C(x)$ is given by $$\label{nfwC} C(x) = \ln\frac{x}{2} + \left\{ \begin{array}{rll} &\frac{1}{\sqrt{1-x^{2}}}\mathrm{cosh}^{-1}\frac{1}{x} & \text{if } x<1 \\ \\ &1 & \text{if } x=1 \\ \\ &\frac{1}{\sqrt{x^{2}-1}}\cos^{-1}\frac{1}{x} & \text{if } x>1 \end{array} \right..$$ The radial dependence of shear is thus given by $$\label{nfwgamma} \gamma_{\mathrm{NFW}}(x) = \frac{r_{s}\delta_{c}\rho_{c}}{\Sigma_{crit}} G(x) ,$$ where $G(x)$ is given by $$\label{nfwG} G(x) = \left\{ \begin{array}{rll} & g_{<}(x) & \text{if } x<1 \\ \\ &\left[ \frac{10}{3}+4\ln\left(\frac{1}{2}\right)\right] & \text{if } x=1 \\ \\ & g_{>}(x) & \text{if } x>1 \end{array} \right..$$ The functions $g_{<}(x)$ and $g_{>}(x)$ are given by $$\begin{aligned} \label{nfwgless} g_{<}(x) &= \frac{8\mathrm{~tanh}^{-1}\sqrt{(1-x)/(1+x)}}{x^{2}\sqrt{1-x^{2}}} + \frac{4}{x^{2}}\ln\left(\frac{x}{2}\right) \notag \\ &- \frac{2}{(x^{2}-1)} + \frac{4\mathrm{~tanh}^{-1}\sqrt{(1-x)/(1+x)}}{(x^{2}-1)\sqrt{1-x^{2}}}\end{aligned}$$ and $$\begin{aligned} \label{nfwgmore} g_{>}(x) & = \frac{8\tan^{-1}\sqrt{(x-1)/(1+x)}}{x^{2}\sqrt{x^{2}-1}} + \frac{4}{x^{2}}\ln\left(\frac{x}{2}\right) \notag \\ & - \frac{2}{(x^{2}-1)} + \frac{4\tan^{-1}\sqrt{(x-1)/(1+x)}}{(x^{2}-1)^{3/2}} .\end{aligned}$$ Combining these with Eqns. \[nfwkappa\] and \[nfwgamma\] with Eqns. \[poisson\] and \[magnification\] allows one to determine the analytic value of the magnification, $\mu$, of an image at a given projected distance from the centre of the NFW lens. Full Beam vs Empty Beam {#AppBeam} ======================= Magnification is the relative increase in flux as a result of gravitational lensing. There is a subtlety here. One may compare the flux received from the source to the flux received if the universe was entirely homogeneous, in which case the magnification is referred to as the full-beam magnification, $\mu_{fb}$. This is the magnification referred to in Sec. \[mpdf\]. The so-called empty-beam magnification, $\mu_{eb}$, is defined relative to the case where all the matter in the universe is locked up in compact objects, and all lines of sight are empty. As this is the case of minimal magnification (no convergence and assumed negligible shear), then $\mu_{eb}$ is always greater than unity. To derive the relationship between the two magnifications, we must compare the solid angles subtended by the source: $\Omega_{fb}$ in the full-beam scenario, $\Omega_{eb}$ in the empty-beam scenario and $\Omega_{img}$ the solid angle of the image observed. Given the conservation of surface brightness, the magnifications are defined by: $$\label{fullbeammag} \mu_{fb} = \frac{\Omega_{img}}{\Omega_{fb}} \qquad \mathrm{and} \qquad \mu_{eb} = \frac{\Omega_{img}}{\Omega_{eb}}.$$ The solid angles are related to the physical area of the source via the angular diameter distances: $$\label{solidanglesfull} \Omega_{fb}(z) = \frac{A_{src}}{(1+z)^{2}D_{fb}(z)}$$ is the appropriate relationship for the full-beam case, and $$\label{solidanglesempty} \Omega_{eb}(z) = \frac{A_{src}}{(1+z)^{2}D_{eb}(z)}.$$ is appropriate for the empty-beam case. Combining Equations \[fullbeammag\], \[solidanglesfull\] and \[solidanglesempty\], one is easily able to convert between the two: $$\label{comparemag} \frac{\mu_{fb}}{\mu_{eb}} = \frac{D_{fb}^{2}(z)}{D_{eb}^{2}(z)}.$$ The angular diameter distance in the full-beam case, $D_{fb}(z)$, is equivalent to the Dyer-Roeder distance $D(\tilde{\alpha}=1; z)$ which is also the FLRW solution: $$\label{angularDD12} D_{fb}(z) \equiv \frac{c}{H_{0}(1+z)} \int^{z}_{0}\frac{dz'}{\left[{\Omega_{\mathrm{M},0}}(1+z')^{3}+ {\Omega_{\mathrm{\Lambda},0}}\right]^{1/2}}.$$ The empty-beam angular diameter distance, $D_{eb}(z)$, is found by solving the Dyer-Roeder equation for $\tilde{\alpha}=0$ : $$D_{eb}(z) \equiv D(\tilde{\alpha}=0; z) = c \int^{z}_{0} \frac{1}{(1+z')^{2}H(z')} dz' \label{DRempty}$$ \[lastpage\] [^1]: Research undertaken as part of the Commonwealth Cosmology Initiative (CCI: www.thecci.org), an international collaboration supported by the Australian Research Council (ARC). E-mail: [email protected] (MK); [email protected] (PL); [^2]: http://www.fftw.org/ [^3]: http://pan-starrs.ifa.hawaii.edu [^4]: http://www.astro-wise.org/projects/KIDS/ [^5]: http://www.darkenergysurvey.org/ [^6]: http://www.lsst.org [^7]: http://sci.esa.int/euclid [^8]: http://jdem.lbl.gov/ and http://jdem.gsfc.nasa.gov/

ArXiv

--- abstract: 'In this paper, we give a harmonic analysis proof of the Neumann boundary observability inequality for the wave equation in an arbitrary space dimension. Our proof is elementary in nature and gives a simple, explicit constant. We also extend the method to prove the observability inequality of a visco-elastic wave equation.' address: 'Department of Mathematical Sciences, Clemson University, Clemson, South Carolina 29634' author: - Walton Green - Shitao Liu - Mishko Mitkovski title: 'A Harmonic Analysis Proof of the Boundary Observability Inequality for the Wave Equation and Visco-Elastic Equation' --- Introduction ============ Let $\Omega \subseteq {\mathbb{R}}^d$ be an open, bounded domain with smooth boundary ${{\partial \Omega}}$. Consider the following backwards wave equation generated at final time $T$. $$\label{waveeq}\left\{ \begin{array}{lr} w_{tt}(x,t)-\Delta w(x,t) =0 & \mbox{in }\Omega \times [0,T]\\ w(x,T)=w_0(x) \quad w_t(x,T)=w_1(x) \hspace{5ex}& \mbox{in }\Omega \\ w(x,t) = 0 & \mbox{on }{{\partial \Omega}}\times [0,T]\end{array} \right.$$ where $w_0 \in H^1_0(\Omega)$ and $w_1 \in L^2(\Omega)$. The problem we are interested in is the boundary observability inequality: There exists $c >0$ such that for all $(w_0,w_1) \in H^1_0(\Omega) \times L^2(\Omega)$, $$\label{waveob} c \int_\Omega |\nabla w_0(x)|^2 + |w_1(x)|^2 \, dx \le \int_0^T \int_{{\partial \Omega}}\left|\dfrac{\partial w}{\partial \nu} (x,t)\right|^2\, dS(x) \, dt$$ It is well known (e.g. by the Hilbert Uniqueness Method [@lions88]) that the observability inequality (\[waveob\]) is equivalent to the exact controllability of the dual equation to (\[waveeq\]): $$\label{waveeqcontrol}\left\{ \begin{array}{lr} u_{tt}(x,t)-\Delta u(x,t) =0 & \mbox{in }\Omega \times [0,T]\\ u(x,0)=u_0(x) \quad u_t(x,0)=u_1(x) \hspace{5ex}& \mbox{in }\Omega \\ u(x,t) = f(x,t) & \mbox{on }{{\partial \Omega}}\times [0,T] \end{array} \right.$$ for $f \in L^2({{\partial \Omega}}\times[0,T])$ and $(u_0,u_1) \in L^2(\Omega) \times H^{-1}(\Omega)$. Exact controllability refers to the question: Given initial states $(u_0,u_1)$ and final states $(u_T,u_T')$, does there exist $T>0$ and $f \in L^2({{\partial \Omega}}\times[0,T])$ such that $u(x,T) = u_T(x)$ and $u_t(x,T)=u_T'(x)$? There exists an extensive body of literature about the exact boundary controllability for the wave equation (or other typed hyperbolic equations). The problem has been very well studied and we refer to the books [@komornik95; @L-Tbook; @Lionsbook; @micu02], and the references therein for a literature review of the problem. The typical method is to use the duality and transfer the controllability problem for the wave equation into the observability question of the dual problem. Even though such equivalence between the controllability and observability was long noticed [@dolecki77], it is not until the mid 80s that mathematicians started to develop systematic methods to prove the observability inequality, especially in the general multidimensional setting. It is well known by now that the observability inequality (\[waveob\]) may be proved using microlocal analysis [@b-l-r], the multiplier method [@ho86; @lions88], or Carleman estimates. The latter is especially powerful and has become a major tool to prove observability inequalities since it can deal with the situations where there are lower-order terms or variable principle coefficients appearing in the wave equation. We cite only a few references here [@gllt; @ltz; @zhang10] and refer to the other works quoted in these papers for readers who are interested in the details of this method. The idea of using harmonic analysis to prove the observability inequality originated much earlier. It seems that Russell [@russell78] was the first person who systematically explored the relationship between control problems and harmonic analysis. The moment method of Russell has been extended in different directions [@hansen; @komornik05; @loreti12; @pandolfi], but the common feature of all these results has been the requirement for the space dimension to be equal to one. Probably, the most comprehensive treatment, to date, on the use of complex exponentials and harmonic analysis in control problems is the monograph [@avdonin95] where, in addition, approximate controllability results (even in higher space dimension) are obtained using complex exponentials in concert with standard uniqueness results. Still, to the best of our knowledge, nobody has given a harmonic analysis proof of the exact controllability (and observability) of the wave equation in higher space dimensions. In this note, we complete this gap by providing a proof of the observability inequality in an arbitrary spatial dimension using the harmonic analysis method. We also apply this method to show observability for a wave equation with memory kernel, also known as the visco-elastic wave equation, which is of the form $$\label{wavemem} y_{tt}-\Delta y = \int_0^t M(t-s) \Delta y(s) \, ds.$$ Our motivation is from [@loreti12; @pandolfi15] in which exact controllability of $(y,y_t)$ is achieved but only for dimension $d \le 3$ using the moment method of Russell. In section \[visco\] we extend this to an arbitrary space dimension. Carleman estimates have been applied to the heat equation with memory kernel (same as (\[wavemem\]) but only first-order in time) [@fu09], but we are not aware of the use of Carleman estimates to prove the observability for the visco-elastic equation (\[wavemem\]). A separate noteworthy contribution to this problem is [@kim93] in which exact controllability is established in arbitrary dimension using the classical compactness-uniqueness argument. The paper is organized as follows. In section \[harm\], we reformulate the observability inequality as a Riesz sequence property for a suitably chosen system of functions. This property is then established for the regular wave equation (\[waveeq\]) in section \[main\] and extended to the visco-elastic wave equation (\[wavemem\]) in section \[visco\]. Harmonic Analysis Reformulation {#harm} =============================== The standard one-dimensional moment method focuses on estimates concerning sequences of complex exponentials. In order to extend it to higher spatial dimensions, we need to consider the following system of functions. Let $\{\phi_n\}_{n=1}^\infty$ be an orthonormal basis in $L^2(\Omega)$ of eigenfunctions of the Dirichlet Laplacian. In other words, $$\left\{\begin{array}{rclr} -\Delta \phi_n &=& \lambda_n^2 \phi_n & \mbox{in } \Omega\\[2mm] \phi_n&=& 0 &\mbox{on } {{\partial \Omega}}\end{array} \right.$$ It is well known that $0 < \lambda_1 \le \lambda_2 \le \cdots$ and $\lambda_n \to \infty$. For simplicity, we set $\lambda_n = \text{sgn}(n)\lambda_{|n|}$ and $$\label{psidef} \psi_n = \dfrac{1}{\lambda_n}\dfrac{{\partial}\phi_{|n|}}{{\partial}\nu} \quad \text{on }{{\partial \Omega}}$$ for $n \in {\mathbb{Z}}\backslash \{0\}$ (henceforth ${\mathbb{Z}}_0$), denoting by $\nu(x)$ the outward normal vector to ${{\partial \Omega}}$ at $x$. This system of functions $\{\psi_n\}$ has been investigated as far back as the 1940’s by Rellich [@rellich46]. More recently, Hassel and Tao [@hassell02] reinvigorated the interest in this system from the harmonic analysis perspective. As a consequence, much more precise linear independence results about it were obtained (which extend the independence results that follow from the observability inequality) [@barnett11; @triggiani08; @xu12]. Our treatment of the system $\{\psi_n\}$ is directly influenced by this more recent work. The following two features enter in our proof: 1) the system $\{\psi_n\}$ is linearly independent in certain small spectral windows (for close values of $\lambda_n$); 2) complex exponentials are independent for $\lambda_n$ with large enough gap. Thus, when combined together, we achieve independence without restrictions on $\lambda_n$. The notion of independence we use is stronger than the classical linear independence. It is given precisely by the notion of a Riesz-Fischer sequence (see e.g. [@avdonin95], [@young01]). A sequence $\{e_n\}$ in a Hilbert space ${\mathcal{H}}$ is said to be a Riesz-Fischer sequence if there exists a constant $c>0$ such that $$\label{rfseq} c \sum |a_n|^2 \le \left\| \sum a_ne_n \right\|^2_{\mathcal{H}}$$ for all finite sequences $\{a_n\}$. We now state the relationship between Riesz-Fischer sequences and the observability inequality. This relationship is well-known to the experts in the field, but since we were not able to find a good reference, we decided to include a short proof of it. \[prop1\] The observability inequality (\[waveob\]) holds for all $(w_0,w_1) \in H^1_0(\Omega) \times L^2(\Omega)$ if $\{\psi_ne^{i\lambda_nt}\}_{n \in {\mathbb{Z}}_0}$ is a Riesz-Fischer sequence in $L^2({{\partial \Omega}}\times [0,T])$, i.e. there exists $c >0$ such that $$\label{riesz} c \sum |a_n|^2 \le \int_0^T \int_{{\partial \Omega}}\left| \sum a_n \psi_n(x) e^{i\lambda_n t}\right|^2 \, dS(x) \, dt$$ for all $\{a_n\}_{n \in {\mathbb{Z}}_0} \in \ell^2$. Let $(w_0,w_1) \in H^1_0(\Omega) \times L^2(\Omega)$. We will represent the solution $w$ to (\[waveeq\]) by separation of variables. In the space variable, we expand onto $\{\phi_n\}$. There exist $\{\xi_n\}, \{\eta_n\} \in \ell^2$ such that $$w_0 = \sum_{n=1}^\infty \xi_n \phi_n \quad \quad w_1 = \sum_{n=1}^\infty \eta_n \phi_n$$ Since $w_0 \in H^1_0(\Omega)$, by the orthonormality of $\{\phi_n\}$, $$\int_\Omega |\nabla w_0(x)|^2 \, dx = -\int_\Omega w_0\Delta w_0 \, dx = \int_\Omega \left( \sum \xi_n \phi_n \right)\left( \sum \lambda_n^2\xi_n \phi_n \right) = \sum |\lambda_n \xi_n|^2$$ therefore $\{ \lambda_n \xi_n \}\in \ell^2$. Set $\tilde \xi_n = \lambda_n \xi_n$. Then, $$w_0 = \sum \dfrac{\tilde \xi_n}{\lambda_n} \phi_n$$ Additionally, we consider the following ordinary differential equation to account for the time variable. $$\left\{ \begin{array}{lrl} z_{jn}''(t) + \lambda_n^2 z_{jn}(t) = 0 & t \in [0,T]&j=1,2\\ z_{1n}(T) = 1 \quad z_{1n}'(T) = 0 &t=T &\\ z_{2n}(T) = 0 \quad z_{2n}'(T) = -\lambda_n &t=T& \end{array} \right. \label{ode}$$ Solutions to (\[ode\]) are of the form $z_{1n}(t) = \cos(\lambda_n(T-t))$ and $z_{2n}(t) = \sin(\lambda_n(T-t))$. Thus, we can represent $w$ solving (\[waveeq\]) as $$\label{dualrep} w(x,t) = \sum \left[ \dfrac{\tilde\xi_n}{\lambda_n}\cos(\lambda_n(T-t)) - \dfrac{\eta_n}{\lambda_n}\sin(\lambda_n(T-t)) \right]\phi_n(x)$$ Then the observability inequality (\[waveob\]) takes the following form: $$\begin{aligned} \notag c \sum &|\tilde \xi_n|^2 + |\eta_n|^2 \\ &\le \int_0^T \int_{{\partial \Omega}}\left|\sum \left[ \tilde \xi_n\cos(\lambda_n(T-t)) - \eta_n\sin(\lambda_n(T-t)) \right]\dfrac{1}{\lambda_n}\dfrac{\partial \phi_n}{\partial \nu}(x) \right|^2 \, dS(x)\, dt \label{equiv} \end{aligned}$$ Using the Euler formula and setting $a_n = \tilde \xi_{|n|} + i \operatorname{sgn}(n)\eta_{|n|}$ for $n \in {\mathbb{Z}}_0$, (\[waveob\]) and (\[equiv\]) are equivalent to $$c \sum |a_n|^2 \le \int_0^T \int_{{\partial \Omega}}\left| \sum a_n \psi_n e^{i \lambda_n t} \right|^2 \, dS \, dt$$ Equivalence between observability (\[waveob\]) and the Riesz-Fischer property (\[riesz\]) actually holds when restricted to sequences $\{a_n\}$ with the property $a_n = \bar{a}_{-n}$. In what follows, we show a slighly stronger result by considering all $\ell^2$ sequences of coefficients. Main Result {#main} =========== \[waveriesz\] Let $R>0$ such that $\Omega \subseteq B(x_0,R)$ for some $x_0 \in {\mathbb{R}}^d$. Then, for $T > 2R$, there exists $c >0$ such that $$c \sum |a_n|^2 \le \int_0^T \int_{{\partial \Omega}}\left| \sum a_n \psi_n(x) e^{i \lambda_n t} \right|^2 \, dS(x) \, dt$$ for all $\{a_n\} \in \ell^2$. Moreover, $c = \dfrac{2(T-2R)}{C_\Omega}$ where $C_\Omega$ is a positive constant dependent only on $\Omega$. We first state two preliminary lemmas concerning the functions $\{\psi_n\}$. Define the following operator $A:H^1_0(\Omega) \to L^2(\Omega)$ which connects the boundary terms $\psi_n$ with the interior eigenfunctions $\phi_n$. $$(Au)(x) = m(x) \cdot \nabla u(x) \mbox{ where } m(x) = x-x_0 \label{am}$$ for $u \in H^{1}_0(\Omega)$ and $x \in \Omega$. \[quasio\] Let $A$ and $m$ be defined by (\[am\]). Then, for all $j,k \in {\mathbb{Z}}_0$, $$\int_{{\partial \Omega}}(m \cdot \nu) \psi_j \overline{\psi_k} \, dS = \left\{ \begin{array}{cl} \displaystyle \dfrac{\lambda_j^2-\lambda_k^2}{\lambda_j\lambda_k}\int_{\Omega}A\phi_{|j|}\overline{\phi_{|k|}} \, dx & \text{if } |j|\neq |k|; \\[5mm] 2 & \text{if }j=k; \\[2mm] -2 &\text{if } j=-k \end{array} \right.$$ We use the fact that $$A\phi_j(x) = (m\cdot\nu)\frac{{\partial}\phi_j}{{\partial}\nu}(x), \ \forall x\in{\partial}\Omega, \ j \in {\mathbb{N}}$$ as in [@rellich46] since $\phi_j = 0$ on ${{\partial \Omega}}$. Applying Green’s Theorem and the fact that $\Delta A - A\Delta = 2\Delta$, $$\begin{aligned} \int_{\partial\Omega}(m\cdot\nu)\psi_j(x)&\overline{\psi_k(x)} \ dS =\frac{1}{\lambda_j\lambda_k}\int_{\partial\Omega}A\phi_{|j|}\frac{{\partial}\overline{\phi_{|k|}}}{{\partial}\nu} \ dS \\[2mm] & = \frac{1}{\lambda_j\lambda_k}\int_{\Omega} A\phi_{|j|}\Delta\overline{\phi_{|k|}} - \Delta(A\phi_{|j|})\overline{\phi_{|k|}} \ dx\\[2mm] & = \left\{ \begin{array}{cl}\displaystyle \frac{\lambda_j^2-\lambda_k^2}{\lambda_j\lambda_k}\int_{\Omega}A\phi_{|j|}\overline{\phi_{|k|}} \ dx & \text{if } |j|\neq |k|; \\[3mm] \displaystyle \frac{1}{\lambda_j\lambda_k}\int_{\Omega}2\lambda_j^2|\phi_{|j|}|^2 \ dx = \pm2 & \text{if } | j|=|k|. \end{array} \right.\end{aligned}$$ \[abdd\] The sequence $\{\lambda_j^{-1} A\phi_{|j|}\}_{j \in{\mathbb{Z}_0}}$ is quasi-orthogonal in $L^2(\Omega)$. More precisely, for all $u \in \ell^2({\mathbb{Z}}_0)$, $$\label{abddeq} \int_\Omega \left| \sum_j u_j \dfrac{A \phi_{|j|}}{\lambda_j} \right|^2\le R^2 \sum_j \left(|u_j|^2 - u_j\bar u_{-j}\right)$$ Secondly, $$\label{aadjoint} \int_\Omega A\phi_{|j|} \phi_{|k|} = -\int_\Omega \phi_{|j|} A \phi_{|k|}$$ for $|j| \ne |k|$. Notice that the system $\{\lambda_j^{-1}\nabla \phi_{|j|} \}_{j \in {\mathbb{Z}}_0}$ has some sense of orthogonality. Indeed, for each $j,k \in {\mathbb{Z}}_0$, $$\int_\Omega \dfrac{\nabla \phi_{|j|} \cdot \nabla \overline \phi_{|k|}}{\lambda_j\lambda_k} = -\int_\Omega \dfrac{\phi_{|j|} \Delta \overline{\phi_{|k|}}}{\lambda_j\lambda_k} = \left\{ \begin{array}{rc} 0 & |j| \ne |k| \\ 1 & j=k \\ -1 & j=-k\end{array} \right.$$ Then, using the definition of $A$ in (\[am\]) and the Cauchy-Schwarz Inequality, we obtain for $\{u_j\} \in \ell^2({\mathbb{Z}}_0)$ $$\int_\Omega \left| \sum_j u_j \dfrac{A \phi_j}{\lambda_j}\right|^2 \le R^2\int_\Omega\left|\sum_j u_j \dfrac{\nabla \phi_j}{\lambda_j} \right|^2 = R^2 \left( \sum_j |u_j|^2 - \sum_j u_j\bar u_{-j} \right)$$ Now we proceed to the second statement in the lemma. Recalling $m$ from (\[am\]) and using ${\partial}_i$ to denote $\frac{{\partial}}{{\partial}x_i}$, $$m_i {\partial}_i \phi_j \overline \phi_k = {\partial}_i (m_i \phi_j \overline \phi_k) - ({\partial}_i m_i) \phi_j \phi_k - m_i \phi_j {\partial}_i \overline{\phi_k}$$ Summing over $i=1,\ldots,d$ and integrating over $\Omega$ yields $$\label{asym} \int_\Omega A \phi_j \overline{\phi_k} = \int_\Omega \nabla \cdot (m \phi_j\overline{\phi_k}) - d \int_\Omega \phi_j \overline{\phi_k} - \int_\Omega \phi_j A\overline{\phi_k}$$ which gives the desired identity since $\phi_j=0$ on ${{\partial \Omega}}$ and $\{\phi_j\}$ are orthonormal. Now we complete the proof of Theorem 1. To be concise, all sums are assumed to be taken over ${\mathbb{Z}}_0$ unless otherwise stated. For $C_{\Omega} :=\max_{x\in{\partial}\Omega}[m(x)\cdot\nu(x)]\le R$, we have the following estimate using Lemma \[quasio\]. $$\begin{aligned} C_{\Omega}\int_{{\partial}\Omega}\int_0^T&\left|\sum_{j }a_je^{i\lambda_jt}\psi_j(x)\right|^2 \ dt\, dS \ge \sum_{j}\sum_{k}a_j\bar a_k\int_{0}^{T}e^{i(\lambda_j-\lambda_k)t} \ dt \int_{{\partial}\Omega}(m\cdot\nu)\psi_j\overline{\psi_k} \ dS \nonumber\\[2mm] = 2T\sum_{j}&|a_j|^2 -\sum_j a_j\bar{a}_{-j} \dfrac{e^{i2\lambda_jT}-1}{i\lambda_j} \notag\\ &+\sum_{j}\sum_{k\neq \pm j}a_j\bar{a}_k\left(\frac{1}{i\lambda_j}+\frac{1}{i\lambda_k}\right)\left(e^{i(\lambda_j-\lambda_k)T}-1\right)\int_{\Omega}A\phi_j\overline{\phi_k} \ dx \label{orig}\end{aligned}$$ Notice when $k= j$, the terms in the double summation actually have zero value. Thus we may include them in the summation. Moreover, applying the second statement in Lemma \[abdd\], $$\sum_{j}\sum_{k\neq \pm j}a_j\bar{a}_k\left(\frac{1}{i\lambda_j}+\frac{1}{i\lambda_k}\right)\left(e^{i(\lambda_j-\lambda_k)T}-1\right)\int_{\Omega}A\phi_j\overline{\phi_k} \ dx$$ $$= 2 \operatorname{Re}\left(\sum_{j}\sum_{k\neq - j}a_j\bar{a}_k\left(e^{i(\lambda_j-\lambda_k)T}-1\right)\int_{\Omega}\dfrac{A\phi_j}{i\lambda_j}\overline{\phi_k} \ dx\right)$$ We will now include the terms when $k=-j$. Notice that by (\[asym\]), $ 2\operatorname{Re}\left(\int_\Omega A\phi_{|j|}\bar \phi_{|-j|} \right) = -d$. Thus we can rewrite the second two terms in the original inequality (\[orig\]) as $$\begin{aligned} &(d-1)\operatorname{Re}\left(\sum_j a_j\bar a_{-j} \dfrac{e^{i2\lambda_jT}-1}{i\lambda_j}\right) \\ &\hspace{10ex}+2 \operatorname{Re}\left(\sum_{j}\sum_{k}a_j\bar{a}_k\left(e^{i(\lambda_j-\lambda_k)T}-1\right)\int_{\Omega}\dfrac{A\phi_j}{i\lambda_j}\overline{\phi_k} \, dx\right) \end{aligned}$$ Additionally, we have the following identity for the single sum: $$\begin{aligned} &\sum_j a_j\bar a_{-j}\dfrac{e^{i2\lambda_jT}-1}{i\lambda_j} \\ &\hspace{5ex}= \int_\Omega \left( \sum_j a_j e^{i\lambda_jT} \dfrac{\phi_{|j|}}{i\lambda_j} \right)\overline{\left( \sum_k a_k e^{i\lambda_kT} \phi_{|k|} \right)}- \left( \sum_j a_j \dfrac{\phi_{|j|}}{i\lambda_j} \right)\overline{\left( \sum_k a_k \phi_{|k|} \right)}dx\end{aligned}$$ Then we split the double sum into two terms (one with $e^{i(\lambda_j-\lambda_k)T}$ and one with $-1$) and estimate each with the corresponding portion in the above identity. $$\begin{aligned} \notag &\left| \int_\Omega (d-1)\left( \sum_j a_j e^{i\lambda_jT} \dfrac{\phi_{|j|}}{i\lambda_j} \right)\overline{\left( \sum_k a_k e^{i\lambda_kT} \phi_{|k|} \right)} +2\sum_{j}\sum_{k}a_j\bar a_ke^{i(\lambda_j-\lambda_k)T}\int_{\Omega}\dfrac{A\phi_{|j|}}{i\lambda_j}\overline{\phi_{|k|}} \right|\\ & \hspace{7ex}=\left| \int_\Omega \left((d-1)\sum_j a_j e^{i\lambda_jT} \dfrac{\phi_{|j|}}{\lambda_j} + 2\sum_j a_j e^{i\lambda_jT}\dfrac{A\phi_{|j|}}{\lambda_j}\right)\overline{\left( \sum_k a_k e^{i\lambda_kT} \phi_{|k|} \right)} \right|\notag \\ &\hspace{7ex} \le \dfrac{1}{4R} \int_\Omega \left|(d-1)\sum_j a_j e^{i\lambda_jT} \dfrac{\phi_{|j|}}{\lambda_j} + 2\sum_j a_j e^{i\lambda_jT}\dfrac{A\phi_{|j|}}{\lambda_j}\right|^2 + R \left| \sum_k a_k e^{i\lambda_kT} \phi_{|k|} \right|^2 \label{id2}\end{aligned}$$ Note that by (\[asym\]), for any $u \in H^1_0(\Omega)$, $$\|(d-1)u+2Au\|^2 = (d-1)^2\|u\|^2 + 4(d-1)\operatorname{Re}(u,Au) + 4\|Au\|^2$$ $$= (-1-d)(d-1)\|u\|^2 + 4\|Au\|^2 \le 4\|Au\|^2$$ where $\|\cdot\|$ and $(\cdot,\cdot)$ denote the $L^2(\Omega)$ norm and inner product. Apply this to (\[id2\]) with $u=\sum a_je^{i\lambda_jT}\phi_{|j|}\lambda_j^{-1}$. Then, applying (\[abddeq\]) from Lemma \[abdd\], we have $$\label{id3} \dfrac{1}{R} \int_\Omega \left| \sum_j a_j e^{i\lambda_jT}\dfrac{A\phi_{|j|}}{\lambda_j}\right|^2 + R \int_\Omega \left| \sum_k a_k e^{i\lambda_kT} \phi_{|k|} \right|^2$$ $$\le R \sum_j \left(|a_j|^2 - a_j\bar a_{-j} e^{i2\lambda_jT} \right) +R \sum_k \left(|a_k|^2 + a_k\bar a_{-k} e^{i2\lambda_kT} \right) = 2R \sum_j |a_j|^2$$ The other term (with $e^{i(\lambda_j-\lambda_k)T}$ replaced by $-1$) is estimated in a manner similar to (\[id2\]) and (\[id3\]). Substituting (\[id3\]) and the corresponding estimate for $-1$ into the original inequality gives the desired result: $$\begin{aligned} C_{\Omega}\int_0^T&\int_{{\partial}\Omega}\left|\sum_{j}a_je^{i\lambda_jt}\psi_j(x)\right|^2 \, dS \, dt \\[2mm] &\geq 2T\sum_{j}|a_j|^2 + \sum_j a_j\bar{a}_{-j} \dfrac{e^{i2\lambda_jT}-1}{i\lambda_j} \\ &\hspace{5ex}+ \sum_{j}\sum_{k\neq \pm j}a_j\bar a_k\left(\frac{1}{i\lambda_j}+\frac{1}{i\lambda_k}\right)\left(e^{i(\lambda_j-\lambda_k)T}-1\right)\int_{\Omega}A\phi_{|j|}\overline{\phi_{|k|}} \\[2mm] &\geq 2T\sum_{j}|a_j|^2 - 4R\sum_{j}|a_j|^2 =2\left(T - 2R\right)\sum_j|a_j|^2.\end{aligned}$$ Application to the Visco-Elastic Equation {#visco} ========================================= Let $\Omega \subseteq {\mathbb{R}}^d$ be an open, bounded domain with smooth boundary. Consider the following visco-elastic wave equation: $$\left\{ \begin{array}{lr} y_{tt}(x,t) -\Delta y(x,t)= \displaystyle\int_0^t M(t-s)\Delta y(x,s) \, ds \hspace{5ex}&\text{in } \Omega \times [0,T] \\[2.5mm] y(x,0) = y_0(x) \quad y_t(x,0) = y_1(x) \hspace{5ex} &\text{in } \Omega \\[1.5mm] y(x,t) = g(x,t) &\text{on } {{\partial \Omega}}\times [0,T] \end{array} \right. \label{viscocontrol}$$ for given $M \in H^2(0,T)$, $y_0 \in L^2(\Omega)$, $y_1 \in H^{-1}(\Omega)$, and $g \in L^2(\Gamma)$. By a similar argument to the one presented section \[harm\] of this note, the exact controllability of (\[viscocontrol\]) will be shown if it is the case that $\{z_n(t)\psi_n(x)\} \subseteq {{L^2({{\partial \Omega}}\times [0,T])}}$ is a Riesz sequence where $z_n$ satisfies the following time ODE (compare with (\[ode\])). $$\left\{ \begin{array}{lr} z_n''(t) + \lambda_n^2 z_n(t) = -\lambda_n^2 \displaystyle\int_t^T M(s-t) z_n(s) \, ds & t \in [0,T]\\[3mm] z_{n}(T) = 1 \quad \quad z_{1n}'(T) = i\lambda_n \\[2mm] \end{array} \right. \label{viscoode}$$ $$\left\{ \begin{array}{lr} y_{tt}(x,t) -\Delta y(x,t)= \displaystyle\int_t^T M(s-t)\Delta y(x,s) \, ds \hspace{5ex}&\text{in } \Omega \times [0,T] \\[2.5mm] y(x,T) = y_0(x) \quad y_t(x,T) = y_1(x) \hspace{5ex} &\text{in } \Omega \\[1.5mm] y(x,t) = 0 &\text{on } {{\partial \Omega}}\times [0,T] \end{array} \right. \label{viscodual}$$ for given $M \in H^2(0,T)$, $y_0 \in H^1_0(\Omega)$ and $y_1 \in L^2(\Omega)$. By a similar argument to the one presented section \[harm\], the boundary observability inequality (analogous to (\[waveob\])) for (\[viscodual\]) will be acheived if it can be established that $\{z_n(t)\psi_n(x)\} \subseteq {{L^2({{\partial \Omega}}\times [0,T])}}$ is a Riesz-Fisher sequence where $\psi_n$ is as defined in (\[psidef\]) and $z_n$ satisfies the following time ODE (compare with (\[ode\])) for each $n \in {\mathbb{Z}_0}$. $$\left\{ \begin{array}{lr} z_n''(t) + \lambda_n^2 z_n(t) = -\lambda_n^2 \displaystyle\int_t^T M(s-t) z_n(s) \, ds & t \in [0,T]\\[3mm] z_{n}(T) = 1 \quad \quad z_{n}'(T) = i\lambda_n \\[2mm] \end{array} \right. \label{viscoode}$$ This formulation is thoroughly carried out in [@loreti12] and [@pandolfi15] where the following Riesz sequence property (\[viscors\]) is obtained for space dimension $d=1$ and $d \le 3$ respectively. Here we extend this to the general case $d \ge 1$. \[viscoriesz\] Let $R>0$ such that $\Omega \subseteq B(x_0,R)$ for some $x_0 \in {\mathbb{R}}^d$. If $T>2R$, then for $\{z_n\}$ solving (\[viscoode\]) and $\{\psi_n\}$ as defined in (\[psidef\]), $\{z_n\psi_n\}$ is a Riesz sequence in ${{L^2({{\partial \Omega}}\times [0,T])}}$. In other words, there exists $c,C>0$ such that $$\label{viscors} c\sum |a_n|^2 \le \int_0^T \int_{{\partial \Omega}}\left| \sum a_n z_n(t) \psi_n(x) \right|^2 \, dS(x) \, dt \le C \sum |a_n|^2$$ for all finite sequences $\{a_n\}$. The notion of a Riesz sequence is slightly stronger than that of a Riesz-Fisher sequence (\[rfseq\]), we simply add the upper inequality. Nonetheless, the lower inequality is enough to imply observability of (\[viscodual\]) which gives exact controllability for the dual system. Our approach is similar to [@loreti12] and [@pandolfi15] in the sense that we will argue that $\{z_n\psi_n\}$ is in a certain sense “close” to $\{e^{i\lambda_n t} \psi_n\}$ which we already know to be a Riesz-Fisher sequence (Theorem \[waveriesz\]). In [@loreti12], it is shown that there exists $C_1>0$ such that $$\label{zest} \int_0^T |z_n(t) - e^{(\gamma + i \lambda_n) t}|^2 \, dt \le \dfrac{C_1}{\lambda_n^2} \quad \forall \, t \in [0,T]$$ for some $\gamma \in {\mathbb{C}}$ in the special case where $\lambda_n=n$. However, there is no crucial role played by $n$ in the computations so (\[zest\]) can be easily verified with general $\lambda_n$. The key in [@loreti12] is that when $\lambda_n=n$, $\{z_n\}$ and $\{e^{\gamma + i \lambda_n t} \}$ are quadratically close, which means $$\sum_n \int_0^T |z_n(t)-e^{(\gamma + i\lambda_n) t} |^2 \, dt < \infty$$ In [@pandolfi15], the decay (\[zest\]) is improved to $\lambda_n^{-4}$ so quadratically closeness follows from Weyl’s lemma when $d \le 3$. We do not expect to be able to extend the quadratically close property to arbitrary dimensions. Rather, we incorporate the estimates on $\{\psi_n\}$ given below in Lemma \[psib\] to show that $\{z_n\psi_n\}$ and $\{e^{(\gamma + i \lambda_n) t} \psi_n\}$ are Paley-Weiner close. The Theorem \[viscoriesz\] will then become a consequence of the following variation of the classical Paley-Weiner theorem [@young01]: \[perr\] Let $\{e_n\}$ be a Riesz sequence in a Hilbert space ${\mathcal{H}}$ and $\{f_n\} \subseteq {\mathcal{H}}$ be an $\ell^2$-independent sequence, i.e. any $\{c_n\} \in \ell^2$ such that $\sum c_nf_n=0$ implies $c_n=0$ for all $n$. If there exists $q \in (0,1)$ and a finite set of indices $J$ such that $$\label{percond} \left\| \sum_{n \not\in J} a_{n}(e_{n}-f_{n}) \right\|^2 \le q \left\| \sum_{n} a_{n}e_{n} \right\|^2$$ for all finite sequences $\{a_{n}\}$, then $\{f_n\}$ is also a Riesz sequence. Thus the Theorem \[viscoriesz\] will be established if we can show three conditions hold: - $\{e^{(\gamma + i \lambda_n) t}\psi_n\}$ is a Riesz sequence. - $\{z_n\psi_n\}$ is an $\ell^2$-independent sequence. - There exists $q \in (0,1)$ and a finite set of indices $J$ such that, $$\left\| \sum_{n \not\in J} a_n \psi_n \left(z_n - e^{(\gamma + i\lambda_n) t}\right) \right\|^2 \le q \left\| \sum_n a_n \psi_n e^{(\gamma + i\lambda_n) t} \right\|^2$$ for all finite sequences $\{a_n\}$ ( Here and henceforth $\|\cdot\|$ denotes the ${{L^2({{\partial \Omega}}\times [0,T])}}$ norm). We first claim that when $T>2R$, $\{\psi_ne^{i\lambda_nt}\}$ is actually a Riesz sequence, i.e. in addition to the lower inequality from Theorem \[main\], there exists $C_2>0$ such that $$\int_0^T \int_{{\partial \Omega}}\left| \sum a_n \psi_n(x) e^{i \lambda_nt} \right|^2 \, dS \, dt \le C_2 \sum |a_n|^2$$ for all finite sets of scalars $\{a_n\}$. This can be established in a similar manner to Theorem \[waveriesz\] but with the operator $A$ (\[am\]) replaced by $V$ from the proof of Lemma \[psib\] below. Alternatively, in the setting of Proposition \[prop1\], it is equivalent to the following regularity estimate for $w$ solving (\[waveeq\]) which is well-known [@lasiecka86]: $$\int_0^T \int_{{\partial \Omega}}\left| {{\dfrac{{\partial}w}{{\partial}\nu}}}(x,t) \right|^2 \, dS \, dt \le C_2 \int_\Omega |\nabla w_0(x)|^2 + |w_1(x)|^2 \, dx$$ This is then extended to $\{e^{(\gamma + i \lambda_n) t}\psi_n\}$ by noticing that $$\begin{aligned} \label{4.6} \max\{1,e^{\operatorname{Re}(\gamma)T}\} \left\| \sum a_n \psi_n e^{i\lambda_n t} \right\|^2 &\ge \left\| \sum a_n \psi_n e^{(\gamma + i\lambda_n) t} \right\|^2 \\ \nonumber &\ge \min\{1,e^{\operatorname{Re}(\gamma)T}\} \left\| \sum a_n \psi_n e^{i\lambda_n t} \right\|^2 \end{aligned}$$ for all finite sequences $\{a_n\}$. Consider a solution $y=\sum c_nz_n\phi_n\lambda_n^{-1}$ to the equation (\[viscodual\]). In [@kim93], the following unique continuation property is shown: > Let $y$ be a solution to (\[viscodual\]) such that $$\frac{{\partial}y}{{\partial}\nu} =0 \quad \mbox{on} \quad {{\partial \Omega}}\times [0,T]$$ If $T>2R$, then $y=0$ on $\Omega \times [0,T]$. This, in turn, gives $y(x,T)= y_t(x,T) = 0$ for $x \in \Omega$ so $$0 = \int_\Omega |\nabla y(x,T)|^2 + |y_t(x,T)|^2\, dx =\left(\sum |c_n|^2-c_n\bar c_{-n}\right) + \left(\sum |c_n|^2+c_n\bar c_{-n}\right)$$ Therefore $c_n=0$ for all $n$. We now give the key lemma in establishing (iii). \[psib\]Let $\{\psi_n\}$ be defined as in (\[psidef\]). Then there exists $C_\alpha$ dependent only the domain $\Omega$ such that for any finite sequence of scalars $\{a_n\}$, $$\label{psib2} \int_{{\partial \Omega}}\left| \sum a_n \psi_n(x) \right|^2 \, dS(x) \le C_\alpha \left(\sum |a_n|^2 \right)^{1/2}\left(\sum |\lambda_na_n|^2 \right)^{1/2}$$ The estimate (\[psib2\]) may be viewed as stating some degree of orthogonality for $\{\psi_n\}$. In proving this, we follow the techniques in [@barnett11; @xu12] utilizing the following lemma. Let $\Omega$ be bounded with piecewise smooth boundary. Then, there exists a smooth vector field $\alpha$, defined on a neighborhood of $\overline{\Omega}$ such that $$\alpha(x) \cdot \nu(x) \ge 1$$ for almost every $x \in {{\partial \Omega}}$. Define $V:H^1_0(\Omega) \to L^2(\Omega)$ by $(Vu)(x) = \alpha(x) \cdot {\nabla}u(x)$. First we claim that there exists $C_\alpha>0$ such that $$\left| \int_\Omega u[V, \Delta]\bar u \, dx \right| \le C_\alpha \|{\nabla}u\|^2$$ for any $u \in H^3(\Omega) \cap H^1_0(\Omega)$. Indeed, using Einstein notation summing over $i,j=1,2,\ldots,d$ $$\begin{aligned} \Delta Vu &= \partial_{ii} \left( \alpha_j (\partial_j u) \right) \\ &= (\partial_{ii}\alpha_j)(\partial_ju) + 2 (\partial_i\alpha_j)(\partial_{ij} u) + \alpha_j (\partial_{jii}u) \\ &= V \Delta u + (\partial_{ii}\alpha_j)(\partial_ju) + 2 (\partial_i\alpha_j)(\partial_{ij} u) \end{aligned}$$ Integrating by parts once and applying the Poincaré inequality yields $$\begin{aligned} \left| \int_\Omega uV\Delta \bar u- u \Delta V \bar u \, dx \right|&= \left|\int_\Omega u(\partial_{ii}\alpha_j)(\partial_j\bar u) + 2 u(\partial_i\alpha_j)(\partial_{ij} \bar u) \, dx \right|\\ &= \left| \int_\Omega u(\partial_{ii}\alpha_j)(\partial_j\bar u) - 2 \left[(\partial_i \bar u)(\partial_i\alpha_j)+ u (\partial_{ii}\alpha_j)\right](\partial_{j} \bar u) \, dx \right|\\ &\le C_\alpha \int_\Omega | {\nabla}u |^2\, dx \end{aligned}$$ $$\begin{aligned} \Delta Vu &= \sum_{i=1}^d \partial_{ii} \left( \sum_{j=1}^d \alpha_j (\partial_j u) \right) =\sum_{i=1}^d \partial_i \left( \sum_j (\partial_i\alpha_j) (\partial_ju) + \alpha_j (\partial_{ji} u) \right) \\ &= \sum_{i=1}^d \sum_{j=1}^d (\partial_{ii}\alpha_j)(\partial_ju) + 2 (\partial_i\alpha_j)(\partial_{ij} u) + \alpha_j (\partial_{jii}u) \\ &= V \Delta u + \sum_i \sum_j (\partial_{ii}\alpha_j)(\partial_ju) + 2 (\partial_i\alpha_j)(\partial_{ij} u) \end{aligned}$$ Integrating by parts once yields $$\begin{aligned} \int_\Omega uV\Delta \bar u- u \Delta V \bar u &= \int_\Omega \sum_{i=1}^d \sum_{j=1}^d u(\partial_{ii}\alpha_j)(\partial_j\bar u) + 2 u(\partial_i\alpha_j)(\partial_{ij} \bar u) \, dx \\ &= \int_\Omega \sum_{i,j=1}^d u(\partial_{ii}\alpha_j)(\partial_j\bar u) - 2 \left[(\partial_i \bar u)(\partial_i\alpha_j)+ u (\partial_{ii}\alpha_j)\right](\partial_{j} \bar u) \, dx \\ &\le C_\alpha \int_\Omega | {\nabla}u |^2\, dx \end{aligned}$$ Take $u = \sum a_n\phi_n\lambda_n^{-1}$ for a finite set of scalars $\{a_n\}$. Notice that $\|\nabla u\|^2 \le 2\sum |a_n|^2$ and $\|\Delta u\|^2 \le 2\sum |\lambda_na_n|^2$ (the factor of 2 is due to the negative indices as in Lemma \[abdd\]). Then, using Cauchy-Schwartz and the above estimates on $V$, we have $$\begin{aligned} \int_{{\partial \Omega}}\left|{{\dfrac{{\partial}u}{{\partial}\nu}}}\right|^2 \, dS &\le \int_{{\partial \Omega}}(\alpha \cdot \nu) \left|{{\dfrac{{\partial}u}{{\partial}\nu}}}\right|^2 \, dS = \int_{{\partial \Omega}}{{\dfrac{{\partial}u}{{\partial}\nu}}}V\bar u \, dS \\ &= \int_\Omega \Delta u V\bar u- u \Delta V\bar u \, dx \\ &= \int_\Omega \Delta uV\bar u - uV\Delta \bar u + u[V,\Delta] \bar u \, dx \\ &= \int_\Omega \Delta uV\bar u +(\nabla \cdot \alpha) u\Delta \bar u + Vu \Delta \bar u + u[V,\Delta]\bar u \, dx \\ &\le C_\alpha \left(\sum |a_n|^2 \right)^{1/2} \left( \sum |\lambda_na_n|^2 \right)^{1/2} \end{aligned}$$ Let $c_\gamma=(T-2R)\min\{1,e^{\operatorname{Re}(\gamma)T}\}/C_\Omega$ be the constant from the lower Riesz sequence inequality (\[4.6\]) for $\{e^{(\gamma + i\lambda_n)t}\psi_n\}$. Since $\lambda_n \to \infty$, there exists $k \in {\mathbb{N}}$ such that $$\dfrac{c_\gamma^{-1} C_\alpha C_1}{\lambda_k} < 1$$ Take $J = \{ j : |j| < k \}$. Applying Lemma \[psib\] and then the estimate (\[zest\]), we have $$\begin{aligned} \int_0^T \int_{{\partial \Omega}}&\left| \sum_{|n| \ge k} a_n \psi_n(x) \left(z_n(t)-e^{(\gamma+i\lambda_n )t}\right) \right|^2 \\ &\le C_\alpha \left( \int_0^T \sum_{|n| \ge k} |a_n(z_n-e^{(\gamma+i\lambda_n )t})|^2 \right)^{1/2} \left( \int_0^T\sum_{|n| \ge k} |\lambda_na_n(z_n-e^{(\gamma+i\lambda_n )t})|^2 \right)^{1/2} \\ &\le C_\alpha C_1 \lambda_k^{-1} \sum |a_n|^2 \\ &\le C_\alpha C_1 c_\gamma^{-1} \lambda_k^{-1} \left\| \sum a_n e^{(\gamma+i\lambda_n )t} \psi_n \right\|^2 \end{aligned}$$

ArXiv

--- abstract: 'We study the problem of knowledge graph (KG) embedding. A widely-established assumption to this problem is that similar entities are likely to have similar relational roles. However, existing related methods derive KG embeddings mainly based on triple-level learning, which lack the capability of capturing long-term relational dependencies of entities. Moreover, triple-level learning is insufficient for the propagation of semantic information among entities, especially for the case of cross-KG embedding. In this paper, we propose recurrent skipping networks (RSNs), which employ a skipping mechanism to bridge the gaps between entities. RSNs integrate recurrent neural networks (RNNs) with residual learning to efficiently capture the long-term relational dependencies within and between KGs. We design an end-to-end framework to support RSNs on different tasks. Our experimental results showed that RSNs outperformed state-of-the-art embedding-based methods for entity alignment and achieved competitive performance for KG completion.' bibliography: - 'references.bib' --- Introduction {#sect:intro} ============ Knowledge graphs (KGs) store a wealth of structured facts about the real world. Each fact is structured in the form of $(s,r,o)$, where $s,o$ and $r$ denote the subject entity, object entity and their relation, respectively. KGs have gradually become an important resource for many knowledge-driven applications, such as semantic search, question answering and recommender systems. Oftentimes, a single KG is far from complete and cannot support these applications with sufficient facts. To address this problem, two fundamental KG tasks are proposed: (i) **entity alignment**, a.k.a. entity resolution or matching, which aims at integrating multiple KGs by identifying entities in different KGs referring to the same real-world object; and (ii) **KG completion**, a.k.a. link prediction, which aims to complete the missing facts in a single KG. Conventional methods usually rank candidates by exploiting various features, as well as using crowdsourcing [@SiGMa; @PARIS; @Hike]. However, even for a single KG, it can be developed and maintained by different people using different domain knowledge and natural languages, which inevitably makes it heterogeneous. Recently, several methods leverage KG embedding techniques to tackle this problem [@TransE; @ConvE; @MTransE; @IPTransE; @BootEA]. They have shown effectiveness in learning relational information either in a single KG or across multiple KGs. For KG embedding, existing methods start with the assumption that similar entities are likely to have similar relational roles. Their primary focus, therefore, lies in learning from relational triples of entities. Typically, some of them are inspired by the TransE model [@TransE], which interprets $(s,r,o)$ as $\mathbf{s} + \mathbf{r} \approx \mathbf{o}$, where the boldfaces denote the corresponding embeddings. Under this modeling, the embedding of one entity is learned by aggregating the embeddings of its $1$-hop relational neighbors. We refer to this kind of models as *triple-level learning*. Many KG embedding models belong to this kind, including not only translational models like TransE [@TransE], TransH [@TransH] and TransR [@TransR], but also compositional models like DistMult [@DistMult], ComplEx [@ComplEx] and HolE [@HolE], as well as neural models like ProjE [@ProjE] and ConvE [@ConvE]. Triple-level learning has two major limitations: (i) **low expressiveness**. It learns entity embeddings from a fairly local view (i.e., $1$-hop relational neighbors). On one hand, there are many different entities having common local relational neighbors in KGs, such as entities with multi-mapping relations as discussed in [@TransH]. Exploiting local relational neighbors for KG embedding is insufficient. On the other hand, there are many entities having few relational triples (a.k.a. long-tail entities) in real-world KGs [@Long-tail]. With triple-level learning, long-tail entities would receive limited attention, thus their embeddings have low expressiveness; and (ii) **inefficient information propagation**. For the entity alignment task, existing methods rely on *seed alignment* (i.e., prior entity alignment known ahead of time) to bridge two KGs. As triple-level learning uses relational triples of seed entities (entities in seed alignment) to deliver alignment information across KGs, it would limit alignment propagation, especially for long-tail entities and entities that are far away from seed entities. Although the information of multi-hop neighborhoods can be passed with back propagation in different mini-batches [@Survey], the efficiency would be seriously affected, especially in the case of cross-KG embedding. To deal with the limitations, we propose *recurrent skipping networks* (RSNs). Instead of learning the embeddings in a triple-level view, RSNs concentrate on learning from relational paths. A relational path is defined as an entity-relation chain, such as (*United Kingdom*, *country*$^-$, *Tim Berners-Lee*, *employer*, *W3C*), where *country*$^-$ is a reverse relation that we create additionally to enhance the connectivity. It is clear that paths can provide richer relational dependencies than triples without losing the local relational information of entities. RSNs also overcome the limitations that many existing methods are only designed for one specific task of KG embedding. For example, TransR [@TransR] and ConvE [@ConvE] have competitive performance on the KG completion task, but they fail on the entity alignment task. We explain the reasons in later sections. A conventional choice to model relational paths is recurrent neural networks (RNNs). However, RNNs assume that the next element in a sequence depends on the current input and the previous hidden state only, which is inappropriate for KG path modeling. Take a relational path $(..., s,r,o, ...)$ for example. After being fed with $(..., \mathbf{s},\mathbf{r})$, RNNs use the current input $\mathbf{r}$ and the previous hidden state $\mathbf{h}_s$ to infer $o$. However, $\mathbf{h}_s$ is a mix of context, which overlooks the importance of $\mathbf{s}$. In KGs, subject entities are vital for inferring a specific object entity. The local neighbor information would be broken if we use RNNs to model relational paths. To overcome this weakness, RSNs enable the output hidden states of relations to learn a *residual* [@ResNet] from their direct subject entities when inferring object entities, with only a few more parameters. Furthermore, we present an end-to-end framework to support RSNs on different tasks. Specifically, to obtain desired paths, we use the *biased random walks* to efficiently sample paths from KGs. This sampling method differs from normal random walks in that it can fluently control the depth and cross-KG biases of the generated paths. After sampling the paths, we are capable of using RSNs to model the relational paths. To make the embedding learning more effectively, we design *type-based noise-constrained estimation* (NCE), which optimizes the negative example sampling according to the types of elements in paths. The main contributions of this paper are listed as follows: - We propose the path-level learning for KG embedding and design RSNs to remedy the limitations of using RNNs to model relational paths. (Section \[sect:rsn\]) - We present an end-to-end framework to support different KG embedding tasks. It significantly outperformed several state-of-the-art methods for entity alignment and achieved competitive performance for KG completion. (Sections \[sect:meth\] and \[sect:exp\]) Related Work {#sect:work} ============ Path-level Embedding -------------------- PTransE [@PTransE] is one of the path-based KG embedding models. It improves TransE [@TransE] by incorporating relation inferences into KG embedding. For example, if there exist a path $(e_1, r_1, e_2, r_2, e_3)$ and a triple $(e_1, r_3, e_3)$, PTransE models a relation inference by learning $\mathbf{r_1} \oplus \mathbf{r_2} \approx \mathbf{r_3}$, where $\oplus$ denotes the used operator, e.g., add, to merge $\mathbf{r_1}$ and $\mathbf{r_2}$. However, it is worth noting that PTransE only uses relation sequences to enhance triple-level learning but ignores relational dependencies of entities. Thus, PTransE still belongs to the triple-level learning. There are many similar methods that purely leverage relational paths or employ chunk-based paths [@Traverse; @Path-RNN; @NeuralLP]. Different from them, our approach is the first one to fully exploit the potential of KG paths. In the network embedding area, DeepWalk [@DeepWalk] uses the uniform random walks to sample paths in networks and employs Skip-Gram [@word2vec] to model these paths. Skip-Gram learns node embeddings by maximizing the probabilities of their neighbors. node2vec [@node2vec] introduces the biased random walks to refine the process of path sampling from networks. It smoothly controls the node selection strategy to make the random walks explore neighbors in a breadth-first-search as well as a depth-first-search fashion. In this paper, the proposed biased random walks are inspired by node2vec. However, we concentrate on generating deep and cross-KG paths. There are also many methods for graph embedding, e.g., structure2vec [@struct2vec], SSE [@SSE], and JK-Net[@JK-Net]. Similar to the network embedding models, they usually do not consider the semantics and directions of relations. Their main goal is to discover clusters or communities of related nodes. Therefore, we think that these methods cannot directly model complex and directed relations in KGs. KG Embedding ------------ KG embedding has been widely studied in last few years. TransE [@TransE] presents translational embedding, which models a relational triple $(s,r,o)$ as $\mathbf{s}+\mathbf{r}\approx \mathbf{o}$. TransH [@TransH] and TransR [@TransR] improve TransE on modeling complex relations. There are also many non-translational methods. ComplEx [@ComplEx] embeds KGs into complex spaces to enhance the basic model DistMult [@DistMult]. RotatE [@RotatE] is similar to ComplEx, but it defines each relation as a rotation from the subject entity to the object entity. Recently, there also exist several neural models designed for KG completion. ProjE [@ProjE] adopts a simple but effective shared variable neural network, and achieves competitive performance. ConvE [@ConvE] combines the embeddings of subject entities and relations by a 2D convolutional operation. For more KG completion methods, please see [@Survey]. Recently, several studies [@MTransE; @JAPE; @BootEA; @KDCoE] have found that KG embedding can also improve the performance on the entity alignment task. MTransE [@MTransE] reuses TransE [@TransE] to separately train embeddings of different KGs and learns a transition between the KG embeddings. JAPE [@JAPE] is also based on TransE, but it learns embeddings of different KGs in a unified space. Additionally, it leverages attributes to refine entity embeddings. IPTransE [@IPTransE] employs an iterative alignment process to extend PTransE [@PTransE] for entity alignment. As aforementioned, it still belongs to the triple-level learning. BootEA [@BootEA] bootstraps embedding-based entity alignment by using an elaborate algorithm to update alignment during iterations. KDCoE [@KDCoE] co-trains entity relations and descriptions to derive KG embeddings. It requires extra pre-trained multi-lingual word embeddings and descriptions. All these methods use TransE-like models to learn KG embeddings, thus they are not capable of capturing long-term relational dependencies in KGs and the propagation of alignment information between different KGs is also limited. GCN-Align [@GCN-Align] employs graph convolutional networks (GCNs) to embed entities based on adjacent neighborhoods, but it does not consider relation semantics among entities. Recurrent Skipping Networks {#sect:rsn} =========================== In this section, we start with preliminaries and an introduction to RNNs. Then, we describe RSNs in detail. Finally, we compare RSNs with conventional residual learning. Preliminaries ------------- A KG is a directed multi-relational graph where nodes denote entities and edges have labels indicating that there exist some specific relations between the connected entities. Formally, we define a KG as a 3-tuple $\mathcal{G} =(\mathcal{E},\mathcal{R},\mathcal{T})$, where $\mathcal{E}$ and $\mathcal{R}$ denote the sets of entities and relations, respectively. $\mathcal{T} \subseteq \mathcal{E} \times\mathcal{R}\times\mathcal{E}$ is the set of relational triples. Different from the existing methods that learn from triples, in this paper, we concentrate on learning from relational paths. A relational path is an entity-relation chain, where entities and relations appear alternately. The head and tail of a relational path must be entities. We use $(x_1, x_2, ..., x_T)$ to denote a relational path, where $T$ is an odd number. Elements with odd indices are entities while the remaining is intermediate relations. To enhance the connectivity of KGs, we add reverse relations in KGs. For each triple $(s,r,o)$, we add a reverse triple $(o, r^-, s)$, where $r^-$ is distinct from $r$. KG completion is a prevalent task for KG embedding [@TransE]. Given a KG, it aims to predict the object entity $o$ given $(s,r,?)$ or predict the subject entity $s$ given $(?,r,o)$. Given two KGs $\mathcal{G}_1 =(\mathcal{E}_1,\mathcal{R}_1,\mathcal{T}_1)$ and $\mathcal{G}_2 =(\mathcal{E}_2,\mathcal{R}_2,\mathcal{T}_2)$, entity alignment aims to find aligned entity pairs between them. Typically, a small subset of entity alignment, denoted by $\mathcal{S} \subset \mathcal{E}_1 \times \mathcal{E}_2$, is known as seed alignment. So, the input of entity alignment is $\mathcal{G}_1, \mathcal{G}_2$ and $\mathcal{S}$. Oftentimes, the two KGs are assembled as one *joint* KG by copying relational triples of seed entities to their counterparts. For convenience, we also denote the joint KG by $\mathcal{G} =(\mathcal{E},\mathcal{R},\mathcal{T})$. RNNs ---- RNNs are a popular class of neural networks performing well on sequential data types. Given a relational path $(x_1,$ $x_2, ..., x_T)$ as input, we first convert the entities and relations into fixed $d$-dimensional embeddings. Thus, the relational path turns to an embedding sequence $(\mathbf{x}_1, \mathbf{x}_2, ...,\mathbf{x}_T)$. RNNs sequentially read in elements in this sequence and output a hidden state at each time step. The output hidden state at time step $t$, denoted by $\mathbf{h}_t$, is calculated as follows: $$\begin{aligned} \label{eq:rnn} \mathbf{h}_t = \tanh(\mathbf{W}_h \mathbf{h}_{t-1} + \mathbf{W}_x \mathbf{x}_t + \mathbf{b}),\end{aligned}$$ where $\mathbf{W}_h,\mathbf{W}_x$ are the weight matrices. $\mathbf{b}$ is the bias. RNNs are capable of handling input of any length with a few parameters and have achieved state-of-the-art performance in many areas. However, there still exist a few limitations when using RNNs to model relational paths. First, the elements in a relational path have two different types: “entity" and “relation", which always appear in an alternating order. However, the traditional RNNs treat them as the same type like words or graph nodes, which makes capturing semantic information in relational paths less effective. Second, any relational paths are constituted by triples, but these basic structure units are overlooked by RNNs. Let $x_t$ denote a relation in a relational path and $(x_{t-1}, x_t, x_{t+1})$ denote a triple involving $x_t$. As shown in Eq. (\[eq:rnn\]), to predict $x_{t+1}$, RNNs would combine the hidden state $\mathbf{h}_{t-1}$ and the current input $\mathbf{x}_t$, where $\mathbf{h}_{t-1}$ is a mix of the information of all the previous elements $x_1, ..., x_{t-1}$. In fact, it is expected that the information of $x_{t-1},x_t$ in the current triple can be more emphasized. ![Example of RSNs with a $2$-hop relational path[]{data-label="fig:rsn"}](rsn.pdf){width="\columnwidth"} Semantic Enhancement with Skipping Mechanism -------------------------------------------- To remedy the limitations of conventional RNNs, we propose RSNs, which refine RNNs by a simple but effective skipping mechanism. The basic idea of RSNs is to shortcut the current input entity to let it directly participate in predicting its object entity. In other words, an input element in a relational path whose type is “entity" can not only contribute to predicting its next relation, but also straightly take part in predicting its object entity. Figure \[fig:rsn\] illustrates an RSN example. Given a relational path $(x_1, x_2, ..., x_T)$, the skipping operation for an RSN is formulated as follows: $$\begin{aligned} \label{eq:rsn} \mathbf{h}'_t = \begin{cases} \mathbf{h}_t & x_t\in \mathcal{E} \\ \mathbf{S}_1 \mathbf{h}_t + \mathbf{S}_2 \mathbf{x}_{t-1} & x_t\in\mathcal{R} \end{cases},\end{aligned}$$ where $\mathbf{h}'_t$ denotes the output hidden state of the RSN at time step $t$, and $\mathbf{h}_t$ denotes the corresponding RNN output. $\mathbf{S}_1, \mathbf{S}_2$ are the weight matrices, and we share their parameters at different time steps. In this paper, we choose the weighted sum for the skipping operation, but other combination methods can be employed as well. Insight of RSNs --------------- Intuitively, RSNs explicitly distinguish entities and relations, and allow subject entities to skip their connections for directly participating in predicting object entities. Behind this simple skipping operation, there is an important thought to adopt residual learning. Let $F(\mathbf{x})$ be an original mapping, where $\mathbf{x}$ denotes the input, and $H(\mathbf{x})$ be the expected mapping. Compared to directly optimizing $F(\mathbf{x})$ to fit $H(\mathbf{x})$, conventional residual learning hypothesizes that it can be easier to optimize $F(\mathbf{x})$ to fit the residual part $H(\mathbf{x})-\mathbf{x}$. For an extreme case, if an identity mapping is optimal (i.e., $H(\mathbf{x})=\mathbf{x}$), pushing the residual to 0 would be much easier than fitting an identity mapping by a stack of nonlinear layers [@ResNet]. However, different from ResNet [@ResNet] and recurrent residual networks (RRNs) [@RRN], which are proposed to help train very deep networks, RSNs employ residual learning on “shallow" networks. The skipping connections do not link the previous input to the very deep layers, but only focus on each triple in relational paths. Specifically, given a relational path $(..., x_{t-1}, x_t, x_{t+1}, ...)$, where $(x_{t-1}, x_{t}, x_{t+1})$ forms a triple, RRNs leverage residual learning by regarding the process at each time step as a mini-residual network. Take time step $t$ for example. RRNs take the previous hidden state $\mathbf{h}_{t-1}$ as input and learn the residual $\mathbf{h}_t$ by $H(\mathbf{h}_{t-1}, \mathbf{x}_t) - \mathbf{h}_{t-1}$, where $H(\mathbf{h}_{t-1}, \mathbf{x}_t)$ is the expected mapping for $\mathbf{h}_{t-1},\mathbf{x}_t$. Since the information of $x_{t-1}$ is mixed in $\mathbf{h}_{t-1}$, RRNs still ignore the structure of KGs that $x_{t-1}, x_{t}$ should be more emphasized for predicting $x_{t+1}$. Hence, the local (i.e., 1-hop) relations cannot be appropriately modeled. Differently, RSNs leverage residual learning in a new manner. Instead of choosing $\mathbf{h}_{t-1}$ as subtrahend, RSNs directly pick up the subject entity $\mathbf{x}_{t-1}$ as subtrahend. We can write this residual as follows: $$\mathbf{h}_t := H(\mathbf{h}_{t-1}, \mathbf{x}_t) - \mathbf{x}_{t-1},\quad x_t\in\mathcal{R}.$$ The underlying thought is that making the output hidden state $\mathbf{h}_t$ to fit $\mathbf{x}_{t+1}$ may be hard, but learning the residual of $\mathbf{x}_{t+1}$ and $\mathbf{x}_{t-1}$ may be easier. We think that this is the key characteristic of RSNs. Table \[tab:rsn\_example\] shows the differences of RNNs, RRNs and RSNs by an example. Suppose that we are standing at *employer*, it is obvious that learning the residual between *W3C* and *Tim Berners-Lee* can make the optimization much easier. The skipping operation only increases a few more parameters, but it offers an efficient way to remedy the major problem of leveraging sequence models to learn relational paths. We also empirically demonstrate the strengths of RSNs in the performance and convergence speed in our experiments. Architecture of RSNs {#sect:meth} ==================== In this section, we present an end-to-end framework that leverages RSNs for entity alignment and KG completion. We show the full architecture in Appendix \[app:full\_arch\]. Three main modules in this framework are described as follows: - **Biased random walk sampling** generates deep and cross-KG relational paths. - **Recurrent skipping network** models relational paths to learn KG embeddings. We have introduced it in the previous section. - **Type-based noise contrastive estimation** evaluates the loss of RSNs in an optimized way. Biased Random Walks {#sect:rw} ------------------- Towards KG embedding, the desired relational paths should be relatively deep and, for entity alignment, stretch across two KGs. Deep paths carry more relational dependencies than triples for representing the relational roles of entities. Cross-KG paths serve as the bridges between two KGs to deliver alignment information. Because KGs are often large scale, it is impractical to enumerate all possible paths. Besides, not all paths contribute to KG embeddings. Thus, we propose a path sampling method with biased random walks on a single KG and across two KGs, which can efficiently explore deep and cross-KG relational paths for embedding learning. **Conventional random walks**. Using random walks to sample paths from networks has been widely studied for a long time [@DeepWalk]. When being applied to KGs, the unbiased random walks obtain the probability distribution of next entities by the following equation: $$\begin{aligned} \label{eq:rw} \text{Pr}(e_{i+1}\,|\,e_i) = \begin{cases} \frac{\pi_{e_i\rightarrow e_{i+1}}}{Z} & \exists r \in \mathcal{R}\colon (e_i,r,e_{i+1}) \in \mathcal{T}\\ 0 & \text{otherwise} \end{cases},\end{aligned}$$ where $e_i$ denotes the $i^\text{th}$ entity in this walk. $\pi_{e_i\rightarrow e_{i+1}}$ is the unnormalized transition probability between $e_i$ and $e_{i+1}$. $Z$ is the normalization constant. The unbiased random walks choose next entities in a uniform probability distribution.\ **Biased random walks**. We leverage the idea of second-order random walks [@node2vec] and introduce a *depth bias* to smoothly control the depths of sampled paths. Specifically, suppose that we are standing at entity $e_{i}$ at present and the previous step is at $e_{i-1}$. The 1-hop neighbors of $e_{i}$ are the candidates for the next step. As we prefer deep paths, we are inclined to choose the next entity which is far away from $e_{i-1}$. Formally, let $e_{i+1}$ denote a candidate entity. We calculate the depth bias between $e_{i-1}$ and $e_{i+1}$, denoted by $\mu_d(e_{i-1},e_{i+1})$, as follows: $$\begin{aligned} \label{eq:bias_dpt} \mu_d(e_{i-1},e_{i+1}) = \begin{cases} \alpha & d(e_{i-1},e_{i+1}) = 2\\ 1-\alpha & d(e_{i-1},e_{i+1}) < 2 \end{cases},\end{aligned}$$ where $d(e_{i-1},e_{i+1})$ gains the distance of the shortest path from $e_{i-1}$ to $e_{i+1}$, and its values can only range in $\{0,1,2\}$. $\alpha\in(0,1)$ is a hyper-parameter controlling the depths of random walks. To reflect the favors on deeper paths, we set $\alpha>0.5$. Figure \[fig:walk\](a) illustrates an example of the depth-biased random walks. Candidates for the next step are $e_3, e_4$ and $e_5$. Their depth biases are as follows: $\mu_d(e_1,e_3) = \alpha$, $\mu_d(e_1,e_4) = \alpha$ and $\mu_d(e_1,e_5) = 1-\alpha$. Due to $\alpha > 0.5$, we are more likely to go to $e_3$ or $e_4$. ![Samples of biased random walks. For simplicity, we reduce a KG as an undirected graph by merging relations and their corresponding reversed ones. $e_2$ is the current entity that we now stand on and $e_1$ is the previous one.[]{data-label="fig:walk"}](walk.pdf){width="\columnwidth"} Furthermore, we also encourage walking across two KGs to deliver alignment information for entity alignment. In a similar way, we introduce a *cross-KG bias* to favor paths connecting two KGs. To formalize, the cross-KG bias between $e_{i-1}$ and $e_{i+1}$, denoted by $\mu_c(e_{i-1},e_{i+1})$, is defined as follows: $$\begin{aligned} \label{eq:bias_crs} \mu_c(e_{i-1},e_{i+1}) = \begin{cases} \beta & kg(e_{i-1}) \neq kg(e_{i+1})\\ 1-\beta & \text{otherwise} \end{cases},\end{aligned}$$ where $kg(\cdot)$ denotes the KG to which an entity belongs. $\beta\in(0,1)$ is a hyper-parameter controlling the behavior of random walks across two KGs. To favor cross-KG paths, we set $\beta>0.5$. This bias also avoids walking backwards and forwards between entities in the seed alignment. Let us look at Figure \[fig:walk\](b) as an example of KG-biased random walks. $e_1$ and $e_4$ are two entities in KG$_1$, while $e_2$ and $e_3$ are two entities in KG$_2$. $e_2$ is a seed entity. After walking from $e_1$ to $e_2$, we calculate the cross-KG biases as follows: $\mu_c(e_1, e_3)=\beta$ and $\mu_c(e_1, e_4)=1-\beta$. Due to $\beta>0.5$, we prefer to go to $e_3$. Finally, we combine the depth and cross-KG biases into the following bias: $$\begin{aligned} \label{eq:bias} \mu(e_{i-1},e_{i+1}) =\mu_d(e_{i-1},e_{i+1}) \times \mu_c(e_{i-1},e_{i+1}).\end{aligned}$$ The detailed algorithm of the biased random walk sampling is shown in Appendix \[app:algo\]. Note that, biased random walks aim to sample paths which can properly describe a graph, rather than conditionally rank paths. Thus, it is significantly different from path ranking [@PathRanking], which tends to select the paths with similar features due to their high rewards. In our case, we need *randomness* to ensure that all features of a graph are sampled. Type-based NCE -------------- Each element in a relational path can be optimized by learning to predict the next element. As the number of candidate entities or relations is usually large, directly computing the sigmoid loss of each prediction is time-consuming. Thus, we use the noise-contrastive estimation (NCE) [@NCE] to evaluate each output of RSNs, which only requires a small number of negative samples to approximate the integral distribution. To formalize, given the input $(\mathbf{x}_1, \mathbf{x}_2, ...,\mathbf{x}_T)$, the loss of RSNs is defined as follows: $$\begin{split} \label{eq:loss} \mathcal{L} = - \sum_{t=1}^{T-1} \Big( & \log{\sigma(\mathbf{h}'_t\cdot \mathbf{y}_{t})}\\ &+ \sum_{j=1}^{k} \mathbb{E}_{\tilde{y}_j\sim Q(\tilde{y})} \big[ \log{\sigma(-\mathbf{h}'_t\cdot \tilde{\mathbf{y}}_j)} \big]\Big), \end{split}$$ where $\mathbf{y}_{t}$ is the target at time step $t$, $\sigma (\cdot)$ is the sigmoid function, and $k$ is the number of negative samples. A negative example $\tilde{y}_j$ is drawn from the noise probability distribution: $Q(\tilde{y}) \propto q(\tilde{y})^{\frac{3}{4}}$, where $q(\tilde{y})$ is the frequency of $\tilde{y}$ appearing in KGs. Note that the negative samples can be either negative entities or negative relations based on the inference task (entity or relation prediction) at current step. So, we can separate the computation of noise probability distribution according to the target *types*. Specifically, if the current target is an entity, we draw negative samples from the noise probability distribution of entities. Negative relation sampling is carried out similarly. In this way, the candidate sets for negative sampling are compacted and the inapplicable negative examples can also be avoided. Experiments and Results {#sect:exp} ======================= We evaluated RSNs on two representative KG embedding tasks: entity alignment and KG completion. For each task, we conducted experiments on a set of real-world datasets and reported the results compared with several state-of-the-art methods. Due to lack of space, a part of experiments and results are shown in Appendix \[app:more\]. Dataset Preparation ------------------- **Entity alignment datasets.** Although the existing datasets used by the embedding-based entity alignment methods [@MTransE; @JAPE; @BootEA; @GCN-Align] are sampled from real-world KGs, e.g., DBpedia and Wikidata, their entity distributions are quite different from real ones. We argue that this distortion would prevent us from a comprehensive and accurate evaluation of embedding-based entity alignment. In this paper, we design a *segment-based random PageRank sampling* (SRPRS) method, which can fluently control the degree distributions of entities in the sampled datasets. Here, the degree of an entity is defined as the number of relational triples in which the entity involves. We obtained four couples of datasets for embedding-based entity alignment, and each has a normal entity distribution and a dense one. Please see Appendix \[app:dataset\] for more details. **KG completion datasets.** We considered two benchmark datasets, namely FB15K and WN18, for KG completion [@TransE]. FB15K contains 15,000 entities, while WN18 has 18 types of relations. Furthermore, recent studies [@Node+LinkFeat; @ConvE] argued that the two datasets contain redundant triples between the training and test sets. In Appendix \[app:237\], we also showed the results on a modified version called FB15K-237. Experiment Settings ------------------- We implemented RSNs with TensorFlow. The source code and datasets are accessible online.[^1] Please see Appendix \[app:implementation\] for the implementation details. We chose Hits@1, Hits@10 and mean reciprocal rank (MRR) as the evaluation metrics. For entity alignment, we picked up several state-of-the-art embedding-based methods for comparison: MTransE [@MTransE], IPTransE [@IPTransE], JAPE [@JAPE], BootEA [@BootEA] and GCN-Align [@GCN-Align]. As KDCoE [@KDCoE] did not release its full code and we did not particularly sample entities with textual descriptions, we skipped this method. We also deployed the source code of a few KG completion methods on the joint KGs and considered them as additional baselines: TransR [@TransR], TransD [@TransD], ConvE [@ConvE] and RotatE [@RotatE]. Following the previous works, we used 30% of reference alignment as the seed alignment. We tried our best to tune the hyper-parameters for all the methods. For KG completion, we mainly reused the results reported in literature. Due to a few methods did not report the results of some metrics, we conducted the experiments by using the provided source code. Following the previous works, we used the filtered ranks, which means that we would exclude other correct entities when we rank the current test entity. Entity Alignment Results ------------------------ Tables \[tab:normal\_results\] and \[tab:dense\_results\] depict the entity alignment results on the normal and dense datasets, respectively. It is evident that capturing long-term dependencies by relational paths enabled RSNs to outperform all the existing embedding-based entity alignment methods. Also, RSNs achieved better results than RSNs (w/o biases), which demonstrated the effectiveness of the proposed biased random walks. Intuitively, the heterogeneity among different KGs is more severe than one KG with different languages. Therefore, entity alignment between different KGs is harder for the embedding-based entity alignment methods. By establishing long-term dependencies, RSNs captured richer information of KGs and learned more accurate embeddings, leading to more significant improvement on the DBP-WD and DBP-YG datasets, especially on the normal datasets. Our experimental results also showed that the embedding-based entity alignment methods are sensitive to entity distributions. The performance of all the methods on the normal datasets is significantly lower than that on the dense datasets, because the dense datasets contain richer relational triples for KG embedding. Although the normal datasets are more difficult, RSNs still gained considerable advantages compared with the other methods. This stemed from the fact that RSNs learn from relational paths, which can preserve more semantics than triples. It is worth mentioning that RSNs showed larger superiority in terms of Hits@1 and MRR. Hits@1 only considers the correct results at the first position, while MRR also favors the top-ranked results. As aforementioned, RSNs can capture richer information to help identify aligned entities in different KGs. The better results on these two more important metrics verified this point. KG Completion Results --------------------- We also conducted experiments to assess the performance of RSNs on KG completion, by deactivating the cross-KG bias in random walks. Specifically, subject entity $s$ and relation $r$ are regarded as a sequence of length 2. We fed their embeddings to RSNs to predict the next element (i.e., object entity $o$). The experimental results are shown in Tables \[tab:fb15k\] and \[tab:wn18\]. We can see that RSNs obtained comparable performance on both two datasets. More specifically, RotatE performed best on FB15K, followed by our RSNs, which also showed a clear advantage compared with the others. However, their performance gaps were significantly narrowed on WN18. It is worth noting that RSNs outperformed all the translational models that also aim to learn KG embeddings rather than only complete KGs. Explanations of the Results --------------------------- Entity alignment and KG completion exist significant divergences. Several methods that performed pretty well on KG completion, e.g., ConvE, lost their advantages on entity alignment. We argue that this may be caused by that they were particularly designed for KG completion. In other words, they aim to better model a triple instead of learning the relational dependencies in KGs. For instance, ConvE involves the convolutional operation to better predict the missing entities, but the complex networks may hinder the learning of input embeddings. But for entity alignment, we identify aligned entities by directly comparing the trained embeddings. These methods may not be capable of training high-quality embeddings. We also found that RSNs performed better on entity alignment than KG completion. As aforementioned, the performance of KG completion can largely be improved with a sophisticatedly-designed structure for triples, whereas the main gaol of RSNs is to model the long paths. This limits the performance of RSNs for KG completion, which only needs to predict subject or object entities in triples. Further Experiments {#sect:further_exp} =================== Comparison with Alternative Networks ------------------------------------ To assess the feasibility of RSNs, we conducted experiments to compare with RNNs and RRNs [@RRN]. Both RNNs and RRNs used in this experiment were implemented with the same settings of multi-layer LSTM units, dropout and batch normalization. We depict the comparison results on the DBP-WD dataset in Figure \[fig:epoch\]. Because RNNs and RRNs do not consider the local structures of relational paths, they converged at a very slow speed. Differently, RSNs achieved better performance with only $1/30$ epochs, which indicated that this particular residual structure is vital for learning relational paths in KGs. Furthermore, RRNs only achieved little improvement compared with RNNs. This implied that simply combining residual learning with RNNs did not significantly help. ![Hits@1 results w.r.t. epochs to converge[]{data-label="fig:epoch"}](epoch.pdf){width="\columnwidth"} Sensitivity to Random Walk Length --------------------------------- We also want to observe how the random walk length affects the performance of RSNs. In Figure \[fig:length\], on all the eight entity alignment datasets, the Hits@1 results increase sharply from length 5 to 15, which indicates that modeling longer relational paths can help KG embedding obtain better performance. Also, we saw that the performance approaches to saturation from length 15 to 25, which may mean that RSNs have reached the max-length of capturing dependencies in the relational paths. In consideration of efficiency, the results in Tables \[tab:normal\_results\] and \[tab:dense\_results\] are based on length 15. More sensitivity analyses can be found in Appendix \[app:seed\]. Concluding Remarks {#sect:concl} ================== In this paper, we studied the path-level KG embedding learning and proposed RSNs to remedy the problems of using sequence models to learn relational paths. We presented an end-to-end framework, which uses the biased random walks to sample desired paths and models them with RSNs. Our experiments showed that the proposed method can obtain superior performance for entity alignment and competitive results for KG completion. Future work includes studying a unified sequence model to learn KG embeddings using both relational paths and textual information. Acknowledgements {#acknowledgements .unnumbered} ================ This work is supported by the National Key R&D Program of China (No. 2018YFB1004300), the National Natural Science Foundation of China (No. 61872172), and the Key R&D Program of Jiangsu Science and Technology Department (No. BE2018131). Complementary Details {#app:complement} ===================== In this section, we introduce more details of the proposed framework. We first illustrate the architecture by an entity alignment example, and then give the algorithm of sampling relational paths with the biased random walks. Architecture {#app:full_arch} ------------ Figure \[fig:arch\] shows the architecture of RSNs for the entity alignment task. It accepts two KGs as input and adopts an end-to-end framework to align the entities between them. Specifically, it first assembles the two KGs as a joint KG, and then repeatedly samples relational paths by the biased random walks on this KG. The generated paths are converted to embedding sequences according to the index of each element in the paths. It uses RSNs to model them and optimizes this process with type-based NCE. Finally, new alignment can be found by comparing the entity embeddings. ![Architecture of the proposed method for entity alignment[]{data-label="fig:arch"}](arch.pdf){width="\columnwidth"} Algorithm of Biased Random Walk Sampling {#app:algo} ---------------------------------------- We depict the algorithm of biased random walk sampling in Algorithm \[alg:biased\]. It first precomputes the depth biases and the cross-KG biases to avoid repeated computation. Then, it samples the paths based on each triple instead of each entity, since using each entity for initialization may cause certain triples out of paths. It repeats the sampling process in terms of the sampling times and the maximal path length. **Input:** Triple set $\mathcal{T}$, depth bias $\alpha$, cross-KG bias $\beta$, sampling times $n$, max length $l$ Obtain biased transition probability matrices $M_d,M_c$; $p := s\to r\to o$ Look up $M_d, M_c$ and compute normalized transition probability distribution $p_o$ of $o$; Sample next entity $e$ from $p_o$; Sample a relation $r'$ between $o$ and $e$; $p := p\to r'\to e$; Implementation Details {#app:implementation} ---------------------- We built RSNs based on the multi-layered LSTM [@LSTM] (two layers for both entity alignment and KG completion) with Dropout [@Dropout]. We conducted batch normalization [@BN] for both input and output of RSNs. KG embeddings and parameters of RSNs were initialized with Xavier initializer. We trained RSNs by Adam optimizer [@Adam] with mini-batches. Table \[tab:setting\] lists the hyper-parameter settings used in the experiments. Entity Alignment Datasets {#app:dataset} ========================= Random PageRank sampling is an efficient algorithm for large graph sampling [@Sampling]. It samples entities according to the PageRank weights and assigns higher biases to more valuable entities. However, it also favors high-degree entities. To fulfill our requirements on KG sampling, we first divided the entities in a KG into several groups by their degrees. Then, we separately performed random PageRank sampling for each group. The group number and size might be adjusted for several times to make the sampled datasets satisfying our requirements. To guarantee the distributions of the sampled datasets following the original KGs, we used the Kolmogorov-Smirnov (K-S) test to measure the difference. We set our expectation to $\epsilon=5\%$ for all the datasets. The statistics of four couples of sampled datasets for entity alignment are shown in Table \[tab:dataset\]. For the normal datasets, they follow the degree distributions of the original KGs. For example, Figure \[fig:degree\] shows the degree distributions of DBpedia and Wikidata, as well as the sampled datasets from different methods. We can see that our normal datasets best approximate the original KGs. For the dense datasets, we randomly removed entities with low degrees in the original KGs to make the average degree doubled, and then conducted the sampling. Therefore, the dense datasets are more similar to the datasets used by the existing methods [@MTransE; @JAPE; @BootEA; @GCN-Align]. More Experimental Analysis {#app:more} ========================== KG Completion Results on FB15K-237 {#app:237} ---------------------------------- FB15K-237 [@Node+LinkFeat] removes one side of symmetric relation pairs (e.g., *contains* versus *containedBy*). However, this may cut down the connectivity and cause unbalanced data distribution. For example, many methods achieve about 10% on Hits@1 for subject prediction, which is much lower than object prediction (about 30%). Thus, we argue that this dataset is still questionable. Furthermore, the test examples involving symmetric relations are just easy to be predicted, and we should not remove them due to the easiness. This may lean to the methods over-tailored to KG completion. The experimental results on FB15K-237 are shown in Table \[tab:fb15k237\]. RotatE obtained the best results on this dataset, followed by ConvE and RSNs. It is worth noting that, while predicting the entities given two-thirds of one triple is not our primary goal, RSNs still achieved comparable or better performance than many methods specifically focusing on KG completion. This revealed the potential of leveraging relational paths for learning KG embeddings. Sensitivity to Proportion of Seed Alignment {#app:seed} ------------------------------------------- The proportion of seed alignment may significantly influence the performance of KG embedding methods. However, we may not obtain a large amount of seed alignment in real world. We assessed the performance of RSNs and BootEA (the best published method on the entity alignment task currently) in terms of the proportion of seed alignment from 50% down to 10% with step 10%. We depict the results on the DBP-WD dataset in Figure \[fig:seed\]. The performance of the two methods continually dropped with the decreasing proportion of seed alignment. However, the curves of RSNs are gentler than BootEA. Specifically, on the normal dataset, for the four proportion intervals, RSNs lost 7.4%, 8.2%, 16.5% and 30.2% on Hits@1, respectively, while BootEA lost 11.8%, 12.0%, 22.3% and 49.8%. This demonstrated that RSNs are more stable. Additionally, when the proportion was down to 10%, the Hits@1 result of RSNs on the normal dataset is almost twice higher than that of BootEA, which indicated that modeling paths helps RSNs propagate the identity information across KGs more effectively and alleviates the dependence on the proportion of seed alignment. [^1]: <https://github.com/nju-websoft/RSN>

ArXiv

--- abstract: | This paper considers the problem of solving a special quartic-quadratic optimization problem with a single sphere constraint, namely, finding a global and local minimizer of $\frac{1}{2}\mathbf{z}^{*}A\mathbf{z}+\frac{\beta}{2}\sum_{k=1}^{n}\lvert z_{k}\rvert^{4}$ such that $\lVert\mathbf{z}\rVert_{2}=1$. This problem spans multiple domains including quantum mechanics and chemistry sciences and we investigate the geometric properties of this optimization problem. Fourth-order optimality conditions are derived for characterizing local and global minima. When the matrix in the quadratic term is diagonal, the problem has no spurious local minima and global solutions can be represented explicitly and calculated in $O(n\log{n})$ operations. When $A$ is a rank one matrix, the global minima of the problem are unique under certain phase shift schemes. The strict-saddle property, which can imply polynomial time convergence of second-order-type algorithms, is established when the coefficient $\beta$ of the quartic term is either at least $O(n^{3/2})$ or not larger than $O(1)$. Finally, the Kurdyka-Łojasiewicz exponent of quartic-quadratic problem is estimated and it is shown that the exponent is ${1}/{4}$ for a broad class of stationary points. author: - Haixiang Zhang - Andre Milzarek - Zaiwen Wen - Wotao Yin bibliography: - 'references.bib' date: 'Received: date / Accepted: date' title: 'On the geometric analysis of a quartic-quadratic optimization problem under a spherical constraint [^1] ' --- Introduction ============ In this paper, we analyze the geometric properties of the following nonconvex quartic-quadratic problem under a single spherical constraint, $$\begin{aligned} \min_{\mathbf{z}\in\mathbb{C}^{n}}~f(\mathbf{z})=\frac{1}{2}\mathbf{z}^{*}A\mathbf{z}+\frac{\beta}{2}\sum_{k\in[n]}\lvert z_{k}\rvert^{4}\quad\mathrm{s.t.}\quad\lVert\mathbf{z}\rVert_{2}=1, \label{eqn:obj}\end{aligned}$$ where $\beta > 0$ is a fixed interaction coefficient and $A \in {\mathbb C}^{n \times n}$ is a given Hermitian matrix. An important class of applications of this type is the so-called Bose-Einstein condensation (BEC) problem, which has attracted great interests in the atomic, molecule and optical physics community and in the condense matter community. Utilizing a proper non-dimensionalization and discretization, the BEC problem can be rewritten as a quartic-quadratic minimization problem of the form , where the matrix $A$ corresponds to the sum of the discretized Laplace operator and a diagonal matrix. If a non-rotating BEC problem is considered, then the variable $\mathbf{z}$ can be restricted to the real space $\mathbb{R}^{n}$ and problem becomes a real optimization problem. For a more detailed setup of the BEC problem and its specific mathematical formulation, we refer to [@griffin1996bose; @bao2012mathematical; @pethick2002bose]. Our interest in problem and its geometric properties is primarily triggered by related numerical results and observations with Bose-Einstein condensates and Kohn-Sham density functional calculations, see, e.g., [@WenYin13; @hu2017adaptive; @wu2017regularized; @GaoLiuCheYua18], and is motivated by recent landscape results for matrix completion [@ge2016matrix; @sun2016guaranteed; @ge2017no], phase retrieval [@sun2016geometric; @chen2018gradient], phase synchronization [@bandeira2016low; @boumal2016nonconvex; @liu2017estimation], and quadratic programs with spherical constraints [@gao2016ojasiewicz; @liu2017quadratic]. Understanding the geometric landscape of the nonconvex optimization problem is a fundamental step towards understanding and explaining the global and local behavior of the problem and the performance of associated algorithms. Despite recent progress on the geometric properties of nonconvex minimization problems and due to the complex interaction of the quadratic and quartic terms, the landscape of is still elusive. We further note that in [@hu2016note], Hu et al. have shown that the minimization problem can be interpreted as a special instance of the partition problem and thus, it is generally NP-hard to solve . Related Work and Geometric Concepts ----------------------------------- Although nonconvex optimization problems are generally NP-hard, [@murty1987some], direct and traditional minimization approaches, such as basic gradient and trust region schemes, can still be applied to solve certain and important classes of nonconvex problems – with astonishing success – and they remain the methods of choice for the practitioner [@boumal2016nonconvex]. A recent and steadily growing area of research concentrates on the identification of such classes of problems and tries to close the discrepancy between theoretical results and numerical performances, see, e.g., [@sun2016nonconvex; @jain2017non; @chi2018nonconvex] for an overview. Herein geometric observations and techniques play a major role in understanding the landscape and the global and local behavior of a nonconvex problem and of associated algorithms. Specifically, we are interested in the following geometric properties: - All local minimizers are also global solutions, i.e., there are no spurious local minimizers. - The objective function possesses negative curvature directions at all saddle points and local maximizers which allows to effectively escape those points. Condition $(\mathcal P_2)$ is the basis of the so-called *strict-saddle property* and was introduced in [@ge2015escaping; @sun2016geometric; @ge2017no]. The strict-saddle and other related conditions can be used in the convergence analysis and in the design of algorithms to efficiently avoid saddle points. For instance, Sun, Qu, and Wright [@SunQuWri18] established a polynomial-time convergence rate of a Riemannian trust region method that is tailored to solve phase retrieval problems which satisfy the strict-saddle property. Furthermore, in [@lee2016gradient; @panageas2016gradient; @lee2017first-order] it is shown that certain randomly initialized first-order methods can converge to local minimizers and escape saddle points almost surely if the strict-saddle property holds. In the following, we briefly review recent classes of nonconvex optimization problems for which the conditions $(\mathcal P_1)$, $(\mathcal P_2)$, or other desirable geometric properties are satisfied. The generalized phase retrieval (GPR) problem is a popular nonconvex problem which has seen remarkable progress these years, see, e.g., [@jaganathan2015phase; @shechtman2015phase] for an overview. Classical methods that transform the GPR problem into a convex program include convex relaxation techniques [@candes2014solving; @candes2013phaselift; @chen2015exact] and Wirtinger flow algorithms with carefully-designed initialization [@candes2015phase]. Phase retrieval problems are typically formulated as a quartic and unconstrained least squares problem depending on $m$ measurements ${\bf y}_k = |{\bf a}_k^*{\bf z}|$, $k = 1,...,m$. Traditional GPR methods can recover the true signal ${\bf z}$ from the measurements as long as the sample size $m$ satisfies $m \gtrapprox n$ or $m \gtrapprox n \log n$ where $n$ is the dimension of the signal. A provable convergence rate for a randomly initialized trust region-type algorithm is given in [@sun2016geometric] as long as $m \gtrapprox n\log^{3}{n}$ via showing that all the local minima are global and the strict-saddle property holds. When the signal and observations are real, the convergence rate of the vanilla gradient descent method is established by Chen et al. [@chen2018gradient] under the assumption $m \gtrapprox n\log^{13}{n}$. Another interesting class of amenable nonconvex problems are low-rank matrix factorization problems. Classical methods for matrix factorization are based on nuclear norm minimization [@candes2010power; @recht2011simpler] and are usually memory intensive or require long running times. In [@keshavan2010matrix1; @keshavan2010matrix2], Keshavan, Montanari, and Oh showed that the well-initialized gradient descent method can recover the ground truth of those problems. A strong convexity-type property is proved to hold around the optimal solution by Sun and Luo in [@sun2016guaranteed] and the objective function is shown to be sharp and weakly convex in nonsmooth settings by Li et al., [@li2018nonconvex]. Further, the strict-saddle property for the low-rank matrix factorization problem is established in [@ge2017no; @ge2016matrix], as well as for other low rank problems such as robust PCA and matrix sensing. Other classes of nonconvex optimization problems with provable convergence or geometric properties comprise orthogonal tensor decomposition [@ge2015escaping; @ge2017optimization], complete dictionary learning [@arora2015simple; @sun2017complete1; @sun2017complete2], phase synchronization and community detection [@bandeira2016low; @boumal2016nonconvex; @liu2017estimation] and shallow neural networks [@liang2018understanding]. There are also several numerical methods that work well in practice for solving the BEC problem (e.g., tools for numerical partial differential equation [@adhikari2000numerical; @edwards1995numerical] or optimization methods [@garcia2001optimizing; @hu2016note; @wu2017regularized]), but their geometric properties are not known. So far the mentioned concepts allow to cover global structures and landscapes. Instead, local properties and the local behavior of can be captured by the so-called Kurdyka-[Ł]{}ojasiewicz (KL), [@Kur98], or [Ł]{}ojasiewicz inequality, [@lojasiewicz1963propriete]. The [Ł]{}ojasiewicz inequality is a useful tool to estimate the convergence rate of first-order iterative methods in the nonconvex setting [@absil2005convergence; @merlet2013convergence; @schneider2015convergence]. Moreover, the convergence rate of first-order methods satisfying a certain line-search criterion and descent condition can be derived via the KL inequality, [@attouch2009convergence; @bolte2014proximal; @schneider2015convergence], where the rate depends on the KL exponent $\theta$. However, there is no general method to determine or estimate the KL exponent of specific optimization problems, though the existence of the KL exponent is guaranteed in many situations. For optimizing a real analytic function over a compact real analytic manifold (such as problem ), the existence of the KL exponent is established by [Ł]{}ojasiewicz in [@lojasiewicz1963propriete]. There are also several few works that derive explicit estimates of the KL exponent for certain structured problems, such as general polynomials [@Gwo99; @d2005explicit; @Yan08], convex problems [@li2018calculus], non-convex quadratic optimization problems with simple convex constraints [@forti2006convergence; @li2018calculus; @luo1994error; @luo2000error], and quadratic optimization problems with single spherical constraint [@gao2016ojasiewicz; @liu2017quadratic]. Obviously, the above four cases do not cover our constrained quartic-quadratic optimization problem . Contributions ------------- In this work, we investigate different geometric concepts for the quadratic-quartic optimizations problem and give theoretical explanations why first- and second-order methods can perform well on it. In section \[sec:notation\], we first derive several new second- and fourth-order optimality conditions for problem that can be utilized to characterize local and global solutions. These conditions capture fundamental geometric properties of stationary points and local minima and form the basis of our geometric analysis. We then investigate problem (\[eqn:obj\]) in the special case where $A$ is a diagonal matrix. In this situation, we show that a complete characterization of the landscape can be obtained and that problem (\[eqn:obj\]) does not possess any spurious local minima. Furthermore, global solutions can be computed explicitly using a closed-form expression that involves the projection onto an $n$-simplex which requires $O(n \log{n})$ operations. These results can be partially extended to the case where $A$ is a rank-one matrix and we can prove uniqueness of global minima up to a certain phase shift. In general, the complex interplay between the quartic and quadratic terms impedes the derivation of explicit expressions for stationary points and local minima and complicates the landscape analysis of $f$ significantly. However, if either the quartic or the quadratic term dominates the objective function, we can establish the strict-saddle property $(\mathcal P_2)$ and identify and calculate the location and number of local minima. Our methodology is based on a careful discussion of the quartic and quadratic terms for large and small interaction coefficients that is applicable for general deterministic and arbitrary choices of $A$. We note that previous works and results only cover fourth-order unconstrained optimization problems (e.g., phase retrieval), quadratic constrained optimization problems (e.g., matrix completion and phase synchronization), or fourth-order constrained optimization problems without quadratic terms (e.g., fourth-order tensor decomposition). In particular, there is no interaction between quartic and quadratic terms and between their Riemannian derivatives. Different from most nonconvex problems discussed in the literature, our problem does not have a natural probabilistic framework and thus, probabilistic techniques such as concentration inequalities can not be directly applied. In addition, we estimate the KL exponent and establish a Riemannian Łojasiewicz-type inequality for problem (\[eqn:obj\]). Again, the presence of the quartic term considerably complicates the theoretical analysis. In order to deal with the high-order terms appearing in the Taylor expansion, we first separate the nonzero and zero components of a stationary point in order to facilitate the discussion of the leading terms. Then we divide the proof into several cases corresponding to different leading terms. The appearance of the quartic term requires the third-order and the fourth-order terms in the Taylor expansion to fully describe the local behavior, rather than merely the second-order terms. Due to the additional terms, the number of possible leading terms is significantly increased and we carefully analyze the relationship between those different terms. If the matrix $A$ is diagonal, we show that the Łojasiewicz inequality holds at every stationary point of with exponent $\theta = \frac14$. Moreover, this result can be extended to more general choices of $A$, if the problem is restricted to the real space and positive semi-definiteness of the stationary certification matrix is assumed. The proof is based on the diagonal case and on estimates of the local behavior of the objective function and the Riemannian gradient in different subspaces. The positive semi-definiteness assumption is utilized throughout the proof to handle the non-isolated case and can not be easily removed. Although this additional condition represents a stronger notion of global optimality, a wide range of global minima in the real case satisfy this condition. To the best of the authors’ knowledge, our work is the first to estimate and analyze these properties for quadratic-quartic optimization problems over a single sphere. Organization and notations -------------------------- This paper is organized as follows. In section \[sec:notation\], we present second- and fourth-order optimality conditions and characterize global minimizer of problem . Next, in section \[sec:diagonal\] and section \[sec:rank1\], we consider two special cases and investigate geometric properties of problem when $A$ is either diagonal or has rank one. General landscape results for the real case are discussed in section \[sec:strict\_saddle\]. Finally, in section \[sec:KLExp\], we estimate the KL exponent of problem . For $n \in {\mathbb{N}}$, we define $[n] := \{1,...,n\}$ and for ${\bf z} \in {\mathbb{C}^n}$, we set $\|{\bf z}\| = \|{\bf z}\|_2 = \sqrt{{\bf z}^*{\bf z}}$. Let ${\mathbb{S}^{n-1}}$ and $\mathbb{CS}^{n-1}$ denote the $n$-dimensional real and complex sphere, respectively. In the following sections, we will use the notation $\mathcal{M} = \mathbb{S}^{n-1}$ or $\mathcal{M} = \mathbb{CS}^{n-1}$ depending on whether we consider the real or the complex case. The tangent space of $\mathbb{CS}^{n-1}$ at a point $\mathbf{z} \in \mathbb{CS}^{n-1}$ is given by $\mathcal{T}_{\mathbf{z}}\mathcal{M}:=\{\mathbf{v}\in\mathbb{C}^{n}:\Re({\bf v}^{*}{\bf z})=0\}$. For ${\bf z} \in {\mathbb{C}^n}$, $\operatorname*{diag}(\mathbf{z})$ is a diagonal matrix with diagonal entries $z_{1},...,z_{n}$ and we use $|{\bf z}|^2$ to denote the component-wise absolute value, $|{\bf z}|^2 = {\bf z} \odot \bar {\bf z}$, of ${\bf z}$. We use $I$ to denote the ($n \times n$) identity matrix. The Euclidean and corresponding Riemannian gradient of $f$ at $\mathbf{z}$ on $\mathcal M$ are denoted by $\nabla f(\mathbf{z})$ and ${\mathrm{grad\!\;}}{f(\mathbf{z})}$. Similarly, $\nabla^{2}f(\mathbf{z})$ and ${\mathrm{Hess\!\;}}{f(\mathbf{z})}$ represent the Euclidean and Riemannian Hessian, respectively. Throughout this paper and without loss of generality we will assume that the matrix $A$ is positive definite. Furthermore, $A = P \Lambda P^*$ is an associated eigenvalue decomposition of the Hermitian matrix $A$ with $\lambda_1 \geq \lambda_2 \geq ... \geq \lambda_n > 0$, $\Lambda = \operatorname*{diag}(\lambda_1,...,\lambda_n)$, $P = (\mathbf{p}_{1},...,\mathbf{p}_{n}) \in {\mathbb{C}}^{n \times n}$, and $P^*P = I$. Wirtinger Calculus and Optimality Conditions {#sec:notation} ============================================ Since the real-valued objective function $f$ is nonanalytic in ${\bf z}$, we utilize the Wirtinger calculus [@kreutz2009complex; @sorber2012unconstrained] to express the complex derivatives of $f$. Specifically, the Wirtinger gradient and Hessian of $f$ are defined as $$\begin{aligned} &\nabla f(\mathbf{z}):= \begin{bmatrix} \nabla_{\mathbf{z}}f\\ \nabla_{\bar{\mathbf{z}}f}\\ \end{bmatrix},\quad\nabla^{2}f(\mathbf{z}):= \begin{bmatrix} \frac{\partial}{\partial\mathbf{z}}(\frac{\partial f}{\partial\mathbf{z}})^{*} &\frac{\partial}{\partial\bar{\mathbf{z}}}(\frac{\partial f}{\partial \mathbf{z}})^{*}\\ & \\ \frac{\partial}{\partial\bar{\mathbf{z}}}(\frac{\partial f}{\partial\bar{\mathbf{z}}})^{*} &\frac{\partial}{\partial\bar{\mathbf{z}}}(\frac{\partial f}{\partial\bar{\mathbf{z}}})^{*} \end{bmatrix},\end{aligned}$$ where $\nabla_{\mathbf{z}}f(\mathbf{z}):=({\partial f}/{\partial \mathbf{z}})^{*},\nabla_{\bar{\mathbf{z}}}f(\bar{\mathbf{z}}):=({\partial f}/{\partial \bar{\mathbf{z}}})^{*}$ and following [@wu2017regularized], we obtain $\nabla f_{\bf z}({\bf z}) = \frac12 A{\bf z} + \beta \operatorname*{diag}(|{\bf z}|^2){\bf z}$, $\nabla_{\bar{\bf z}} f(\bar {\bf z}) = \overline{\nabla f_{\bf z}({\bf z})}$, and $$\nabla^{2}f(\mathbf{z}) = \begin{bmatrix} \frac{1}{2}A+2\beta\operatorname*{diag}(|{\bf z}|^2) & \beta \operatorname*{diag}(z_{1}^{2},...,z_{n}^{2}) \\ \beta \operatorname*{diag}(\bar{z}_{1}^{2},...,\bar{z}_{n}^{2}) & \frac{1}{2}\bar{A}+2\beta \operatorname*{diag}(|{\bf z}|^2) \end{bmatrix}.$$ Furthermore, using the identification $\mathcal{T}_{\mathbf{z}}\mathcal{M} \equiv \{ {\bf v}, \bar {\bf v} \in {\mathbb{C}^n}: {\bf z}^*{\bf v} + \bar {\bf z}^T \bar {\bf v} = 0 \}$, the Riemannian gradient and Hessian of $f$ are given by $$\begin{aligned} \label{eq:lambda}{\mathrm{grad\!\;}}{f(\mathbf{z})}= \nabla f({\bf z}) -\lambda \begin{bmatrix} \mathbf{z} \\ \bar{\mathbf{z}} \end{bmatrix} \quad \text{and} \quad {\mathrm{Hess\!\;}}{f(\mathbf{z})} = \nabla^2 f({\bf z}) -\lambda I_{2n}, \end{aligned}$$ where $\lambda = {\bf z}^* \nabla_{\bf z} f({\bf z}) = \frac{1}{2}\mathbf{z}^{*}A\mathbf{z}+\beta\lVert\mathbf{z}\rVert_{4}^{4} \in {\mathbb{R}}$, see, e.g., [@Absil2009Optimization Section 3.6 and 5.5]. Let us notice that ${\mathrm{grad\!\;}}{f}$ and ${\mathrm{Hess\!\;}}{f}$ coincide with the standard gradient and Hessian of the Lagrangian $L({\bf z},\mu) = f({\bf z},\bar{\bf z}) - \frac{\mu}{2} (\bar {\bf z}^T {\bf z} - 1)$ when choosing $\mu = \lambda$. Exploiting the symmetry in ${\mathrm{grad\!\;}}{f}$, the associated first-order optimality conditions for now take the form: $$\label{eq:stat} {\mathrm{grad\!\;}}{f({\bf z})} = 0 \quad \iff \quad [A + 2\beta \operatorname*{diag}(|{\bf z}|^2)] {\bf z} = 2\lambda {\bf z}.$$ A point ${\bf z} \in {\mathbb{C}^n}$ satisfying the conditions will be called *stationary point* of problem . We define the curvature of $f$ at ${\bf z}$ along a direction $\mathbf{v} \in {\mathbb{C}^n}$ via $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}] & := \begin{bmatrix} \mathbf{v}^* & {\mathbf{v}}^T \end{bmatrix} {\mathrm{Hess\!\;}}{f(\mathbf{z})} \begin{bmatrix} \mathbf{v}\\ \bar{\mathbf{v}} \end{bmatrix} \\ & = \mathbf{v}^{*} [ A+4\beta \operatorname*{diag}(|{\bf z}|^2)]{\bf v} +2\beta \cdot {\sum}_{k = 1}^n \Re(v_{k}^{2}{\bar z_{k}}^{2})-2\lambda\lVert\mathbf{v}\rVert^{2} \\ & = \mathbf{v}^{*}[ A + 2\beta \operatorname*{diag}(|{\bf z}|^2) - 2\lambda I] \mathbf{v} + 4\beta \cdot {\sum}_{k=1}^n \Re(v_k\bar z_k)^2$$ In the real case, the latter formulae reduce to ${\mathrm{grad\!\;}}{f(\mathbf{z})} = [A+2\beta\operatorname*{diag}(|{\bf z}|^2)]{\bf z}-2\lambda \mathbf{z}$, ${\mathrm{Hess\!\;}}{f(\mathbf{z})}=A+6\beta\operatorname*{diag}(|{\bf z}|^2)-2\lambda I_{n}$, and $H_{f}(\mathbf{z})[\mathbf{v}] :=\mathbf{v}^{T}{\mathrm{Hess\!\;}}{f(\mathbf{z})}\mathbf{v}$. Second-Order Optimality Conditions ---------------------------------- Due to the analogy of the Riemannian expressions and the Lagrangian formalism, we can apply classical optimality results to describe the second-order optimality conditions of problem . In particular, by [@NoceWrig06 Theorem 12.5 and 12.6] we have: \[lemma:son-sos\] Suppose that $\mathbf{z}$ is a local solution of problem . Then, it holds that ${\mathrm{grad\!\;}}{f(\mathbf{z})}=0$ and we have $H_{f}(\mathbf{z})[\mathbf{v}]\geq0$ for all $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}$. Conversely, if $\mathbf{z}$ is a stationary point satisfying ${\mathrm{grad\!\;}}f(\mathbf{z})=0$ and $H_{f}(\mathbf{z})[\mathbf{v}]>0$ for all $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\backslash\{0\}$, then $\mathbf{z}$ is an isolated local minimum of problem . Next, for some ${\bf z} \in {\mathbb{C}^n}$ we define the equivalence class $$\begin{aligned} \label{eq:equi-class} \llbracket {\bf z} \rrbracket := \{ {\bf y} \in {\mathbb{C}^n}: |y_k| = |z_k|, \;\; \forall~k \in [n] \}. \end{aligned}$$ The following theorem gives a general sufficient condition for a stationary point to be a global minimum of problem . \[theorem:glob-suff\] Let ${\bf z} \in {\mathbb{C}^n}$ be a stationary point of problem with corresponding multiplier $\lambda$ and suppose that the matrix $$\begin{aligned} \label{eq:glob-suff} H := A + 2\beta \operatorname*{diag}(|{\bf z}|^2) - 2\lambda I \succeq 0 \end{aligned}$$ is positive semidefinite. Then ${\bf z}$ is a global minimum and all global minima of problem belong to the equivalence class $\llbracket {\bf z} \rrbracket$. Let ${\bf y} \in {\mathbb{C}^n}$ be an arbitrary point with $\|{\bf y}\| = 1$ and let us introduce the polar coordinates $z_i = r_i e^{i \theta_i}$, $y_i = t_i e^{i \phi_i}$ for $r_i, t_i \geq 0$, $\theta_i, \phi_i \in [0,2\pi]$ and all $i \in[n]$. Using the stationarity condition ${\mathrm{grad\!\;}}{f({\bf z})} = 0$ and $\|{\bf z}\| = \|{\bf y} \| = 1$, it holds that $$\begin{aligned} \nonumber f({\bf y}) - f({\bf z}) & = {\frac{1}{2}}{\bf y}^* A {\bf y} - {\frac{1}{2}}{\bf z}^* [ 2\lambda {\bf z} - 2\beta \operatorname*{diag}(|{\bf z}|^2) {\bf z}] + \frac{\beta}{2} (\|{\bf y}\|_4^4 - \|{\bf z}\|_4^4) \\ \nonumber & \hspace{-0ex} = {\frac{1}{2}}{\bf y}^* H {\bf y} - \beta {\bf y}^* \operatorname*{diag}(|{\bf z}|^2){\bf y} + \frac{\beta}{2}(\|{\bf y}\|_4^4 + \|{\bf z}\|_4^4) \\ & \hspace{-0ex} = {\frac{1}{2}}{\bf y}^* H {\bf y} + \frac{\beta}{2} \; {\sum}_{k=1}^n [t_k^2 - r_k^2]^2. \label{eq:objective_value}\end{aligned}$$ Consequently, the positive semidefiniteness of $H$ yields $f({\bf y}) - f({\bf z}) \geq 0$ for all ${\bf y} \in {\mathbb{C}^n}$ with $\|{\bf y}\| = 1$. Suppose now that ${\bf y}$ is a global minimum with ${\bf y} \notin \llbracket {\bf z} \rrbracket$. In this case the last sum in the above expression is strictly positive which, together with the positive semi-definiteness of $H$ yields a contradiction. If problem has two different global minimizers ${\bf y}$ and ${\bf z}$ with $\llbracket {\bf y} \rrbracket \cap \llbracket {\bf z} \rrbracket = \emptyset$, Theorem \[theorem:glob-suff\] implies that $H$ can not be positive semidefinite. Moreover, if condition holds at a stationary point ${\bf z}$, it automatically has to hold at all global minimizers in $\llbracket {\bf z} \rrbracket$. The definiteness condition in Theorem \[theorem:glob-suff\] can be equivalently rephrased as follows: The multiplier $\lambda$ associated with ${\bf z}$ is the minimum eigenvalue of the matrix $A + 2\beta \operatorname*{diag}(|{\bf z}|^2)$ and ${\bf z}$ is the corresponding eigenvector. Characterizations of this type are also known for (quadratic) trust-region subproblems and for general quadratic programs with quadratic constraints, see [@Mor93; @VanThoai2005]. Furthermore, utilizing [@cai2018on Theorem 3.1], it can be shown that such an eigenvector ${\bf z}$ with the stated properties exists under the assumption $0 < \beta \leq (\lambda_{n-1}-\lambda_n)/8$. In this case, the condition is necessary sufficient for global optimality. Fourth-order optimality conditions {#section:foc} ---------------------------------- In the following section, we derive several fourth-order optimality conditions based on a special and finer expansion of the objective function $f$. In contrast to the sufficient conditions in Theorem \[theorem:glob-suff\], this allows us to fully characterize global optima. Let ${\bf z}\in{\mathbb{CS}^{n-1}}$ be an arbitrary stationary point. For ${\bf v} \in \mathcal T_{\bf z}\mathcal M \cap {\mathbb{CS}^{n-1}}$ and $\theta \in {\mathbb{R}}$, we consider the point ${\bf y}=\cos(\theta){\bf z}+\sin(\theta){\bf v} \in {\mathbb{CS}^{n-1}}$. Using this decomposition in , we obtain $$\begin{aligned} f({\bf y}) - f({\bf z}) & = {\frac{1}{2}}{\bf y}^{*}H{\bf y}+\frac{\beta}{2}\sum_{k\in[n]}\left[|y_{k}|^{2}-|z_{k}|^{2}\right]^{2} \\ & = \frac{\sin^2(\theta)}{2} \left[ H_f({\bf z})[{\bf v}] + 2 \beta \sin(2\theta) \cdot {\sum}_{k \in [n]} (|v_k|^2 - |z_k|^2) \Re(\bar z_k v_k) \right] \\ & \hspace{4ex} + \frac{\beta\sin^{4}(\theta)}{2} \left[ {\sum}_{k \in [n]} [(|v_{k}|^{2}-|z_{k}|^{2})^2 - 4\Re(\bar{z}_{k}v_{k})^2] \right]. \end{aligned}$$ Defining $H_{3}({\bf v}) := \beta \sum_{k \in [n]}(|v_k|^2 - |z_k|^2)\Re(\bar z_k v_k)$, $H_4({\bf v}) := \beta {\sum}_{k \in [n]} [(|v_{k}|^{2}-|z_{k}|^{2})^2 - 4\Re(\bar{z}_{k}v_{k})^2]$ and dividing the latter equation by $\sin^4(\theta)$, this yields $$\begin{aligned} G({\bf v},\cot(\theta)) & = \frac{2[f({\bf y}) - f({\bf z})]}{\sin^4(\theta)} \\ & = H_f({\bf z})[{\bf v}] \cdot \cot^2(\theta) + 4 H_3({\bf v}) \cdot \cot(\theta) + 2[f({\bf v}) - f({\bf z})] ,\end{aligned}$$ for $\theta \neq k\pi$, $k \in {\mathbb Z}$. We first propose necessary and sufficient optimality conditions for ${\bf z}$ being a local minimum. \[thm:4order-local\] Let ${\bf z}$ be a stationary point of problem . Then ${\bf z}$ is locally optimal if and only if there exists a constant $M_{{\bf z}}>0$ such that $G({\bf v},t)\geq0$ for all ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and all $t$ such that $|t|\geq M_{{\bf z}}$. By definition, ${\bf z}$ is a local minimum if and only if there exists a constant $\delta_{{\bf z}}>0$ such that $f({\bf y})\geq f({\bf z})$ for all ${\bf y}\in B({\bf z},\delta_{{\bf z}})\cap{\mathbb{CS}^{n-1}}$. Due to $\|{\bf y}-{\bf z}\|^{2}=2-2\cos(\theta)$, the condition ${\bf y}\in B({\bf z},\delta_{{\bf z}})\cap{\mathbb{CS}^{n-1}}$ is equivalent to ${\bf y} = \cos(\theta) {\bf z} + \sin(\theta) {\bf v}$ with ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and $\cos(\theta)\geq1-{\delta_{{\bf z}}^{2}}/{2}$. Moreover, $\cos(\theta)\geq1-{\delta_{{\bf z}}^{2}}/{2}$ implies $$\begin{aligned} |\sin(\theta)|\leq\delta_{{\bf z}}\sqrt{1-{\delta_{{\bf z}}^{2}}/{4}},\quad|\cot(\theta)|\geq\frac{1-\delta_{{\bf z}}^{2}}{\delta_{{\bf z}}\sqrt{1-\delta_{{\bf z}}^{2}/4}} =: M_{\bf z}.\end{aligned}$$ Consequently, it holds that ${\bf y}\in B({\bf z},\delta_{{\bf z}})\cap{\mathbb{CS}^{n-1}}$ if and only if ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and $|\cot(\theta)|\geq M_{{\bf z}}$. Further, by definition of $G$, the necessary and sufficient conditions for ${\bf z}$ being a local minima are equivalent to: there exists a constant $M_{{\bf z}}>0$ such that $G({\bf v},t)\geq0$ for all ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and all $t$ such that $|t|\geq M_{{\bf z}}$. Similarly, we can derive the fourth-order global optimality conditions. \[theorem:glob-opt-full\] Let ${\bf z}$ be a stationary point of problem . Then ${\bf z}$ is a global solution if and only if $$\begin{aligned} & H_f({\bf z})[{\bf v}] \geq 0,~2 H_{3}({\bf v})^2 \leq H_f({\bf z})[{\bf v}] \cdot [f({\bf v}) - f({\bf z})], \quad \forall~{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}},\\ &H_{3}({\bf v}) = 0,~H_{4}({\bf v})\geq0, \quad \forall~{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}~\text{with}~H_f({\bf z})[{\bf v}] =0.\end{aligned}$$ A stationary point ${\bf z}$ is a global minimum if and only if $f({\bf y})\geq f({\bf z})$ for all ${\bf y}\in{\mathbb{CS}^{n-1}}$, which is equivalent to $G({\bf v},t)\geq0$ for all $t\in\mathbb{R}$ and all ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$. However, nonnegativity of the (degenerated) quadratic function $t \mapsto G({\bf v},t)=H_f({\bf z})[{\bf v}] t^{2}+ 4 H_{3}({\bf v})t+ 2 [f({\bf v}) - f({\bf z})]$ on ${\mathbb{R}}$ is equivalent to the conditions stated in Theorem \[theorem:glob-opt-full\]. Finally, we establish fourth-order necessary conditions for local optimality. \[theorem:foc-nes\] Let ${\bf z}$ be a local minima of problem . Then it holds that $$\left[ \begin{array}{l} H_f({\bf z})[{\bf v}]\geq0, \quad \forall~{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}},\\[0.5ex] H_{3}({\bf v})=0,~H_{4}({\bf v})\geq0, \quad \forall~{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}~\text{with}~H_f({\bf z})[{\bf v}]=0,\\[0.5ex] 4H_{3}^{2}({\bf v},{\bf w})\leq H_f({\bf z})[{\bf w}]H_{4}({\bf v}),\\[0.5ex] \quad \forall~{\bf v},{\bf w}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}~\text{with}~H_f({\bf z})[{\bf v}]=0,~H_f({\bf z})[{\bf w}]>0,~{\bf v}^{*}{\bf w}=0, \end{array} \right.$$ where $H_{3}({\bf v},{\bf w})=H_3({\bf v}) + 2\beta\sum_{k\in[n]}\Re(\bar{v}_{k}w_{k})\Re(\bar{v}_{k}z_{k})$. Theorem \[thm:4order-local\] implies that there is $M_{{\bf z}} > 0$ such that $G({\bf v},t)\geq0$ for all ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and $|t|\geq M_{{\bf z}}$. Hence, for fixed ${\bf v}$, we have $G({\bf v},t) = H_f({\bf z})[{\bf v}] t^{2}+ 4 H_{3}({\bf v})t+ 2 [f({\bf v}) - f({\bf z})] \geq 0$ for $t\rightarrow\pm\infty$. Thus, it follows $H_f({\bf z})[{\bf v}] \geq 0$ and if $H_f({\bf z})[{\bf v}] = 0$, it must hold $H_{3}({\bf v}) = 0$ and $2[f({\bf v}) - f({\bf z})] = H_f({\bf z})[{\bf v}] + H_4({\bf v}) = H_4({\bf v})\geq0$. Now we prove the last condition. Suppose ${\bf v}$ and ${\bf w}$ are two vectors in $\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ satisfying $H_f({\bf z})[{\bf v}]=0$, $H_f({\bf z})[{\bf w}]>0$ and ${\bf v}^{*}{\bf w}=0$. Let ${\bf y}=\cos(\theta){\bf v}+\sin(\theta){\bf w}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$, where $\theta\in[-\pi,\pi]\backslash\{0\}$. We consider the limiting process $\theta\rightarrow0$. Then, due to $H_f({\bf z})[{\bf v}] \geq 0$ for all ${\bf v} \in \mathcal{T}_{{\bf z}}\mathcal{M}$, it follows that $H_f({\bf z})[{\bf y}]=H_f({\bf z})[{\bf w}]\cdot\sin^{2}(\theta)$ and $$H_{3}({\bf y})=H_{3}({\bf v},{\bf w})\cdot\sin(\theta)+O(\sin^{2}(\theta)), \quad H_{4}({\bf y})=H_{4}({\bf v})+O(\sin(\theta)). \label{eqn:foc-nes}$$ We first discuss the case $H_{4}({\bf v})=0$. The discriminant of $G({\bf y},t)$ – as a quadratic function of $t$ – is given by $$\begin{aligned} 16H_{3}^{2}({\bf y})-4H_f({\bf z})[{\bf y}](H_f({\bf z})[{\bf y}]+H_{4}({\bf y})) & =16H_{3}^{2}({\bf v},{\bf w})\sin^{2}(\theta)+O(\sin^{3}(\theta)).\end{aligned}$$ If $H_{3}({\bf v},{\bf w})\neq0$, this term is positive for all sufficiently small $\theta$ and hence, $G({\bf y},t)$ has two real roots. The absolute value of the larger root is $$\begin{aligned} \frac{4|H_{3}({\bf y})|+\sqrt{16H_{3}^{2}({\bf y})-4H_f({\bf z})[{\bf y}](H_f({\bf z})[{\bf y}]+H_{4}({\bf y}))}}{2H_f({\bf z})[{\bf y}]}=\Theta\left(|\sin(\theta)|^{-1}\right),\end{aligned}$$ which implies that there does not exist a constant $M_{{\bf z}}>0$ such that $G({\bf y},t)\geq0$ for all ${\bf y}$ and $t$ such that ${\bf y}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and $|t|\geq M_{{\bf z}}$. Thus, we have $H_{3}({\bf v},{\bf w})=0$ in this case. Next, we consider the case $H_{4}({\bf v})>0$ and let us suppose $4H_{3}^{2}({\bf v},{\bf w})>H_f({\bf z})[{\bf y}]H_{4}({\bf v})=\Theta(\sin^{2}(\theta))$. The discriminant of $G({\bf y},t)$ now satisfies $$\begin{aligned} &16H_{3}^{2}({\bf y})-4H_f({\bf z})[{\bf y}](H_f({\bf z})[{\bf y}]+H_{4}({\bf y}))\\ & \hspace{6ex}=4\left[4H_{3}^{2}({\bf v},{\bf w})\sin^{2}(\theta)-H_f({\bf z})[{\bf w}]H_{4}({\bf v})\right]\sin^{2}(\theta)+O(\sin^{3}(\theta))>0,\end{aligned}$$ for all sufficiently small $\theta \neq 0$. As in the last case, the absolute value of the larger root of $t \mapsto G({\bf y},t)$ converges to $+\infty$ as $\theta \to 0$ which yields the same contradiction. Consequently, we have $4H_{3}^{2}({\bf v},{\bf w})\leq H_f({\bf z})[{\bf w}]H_{4}({\bf v})$ by combining the two cases. The fourth-order optimality conditions in Theorem \[theorem:glob-opt-full\] and \[theorem:foc-nes\] resemble other known fourth order conditions, see, e.g., [@Ded95; @Pen17; @CarGouToi18], and might be hard to verify in practice. However, in the real case, the inequality $H_4({\bf v}) \geq 0$ is equivalent to checking $\||{\bf v}|^2 - |{\bf z}|\| \geq 2\sqrt{2} \|{\bf z}\|_4^2$. In this situation, the framework presented in [@AnaGe16] can be used to verify the first two conditions in Theorem \[theorem:foc-nes\] in polynomial time. \[theorem:foc-suf\] Suppose that the point ${\bf z}\in{\mathbb{CS}^{n-1}}$ satisfies the conditions $$\left[ \begin{array}{l} H_f({\bf z})[{\bf v}]\geq0, \quad \forall~{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}},\\[0.5ex] H_{3}({\bf v})=0,~H_{4}({\bf v})\geq0, \quad \forall~{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}~\text{with}~H_f({\bf z})[{\bf v}]=0,\\[0.5ex] 4H_{3}^{2}({\bf v},{\bf w})\leq H_f({\bf z})[{\bf w}]H_{4}({\bf v}),\\[0.5ex] \quad \forall~{\bf v},{\bf w}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}~\text{with}~H_f({\bf z})[{\bf v}]=0,~H_f({\bf z})[{\bf w}]>0,~{\bf v}^{*}{\bf w}=0, \end{array} \right.$$ where $H_{3}({\bf v},{\bf w})=H_3({\bf v}) + 2\beta\sum_{k\in[n]} \Re(\bar{v}_{k}w_{k})\Re(\bar{v}_{k}z_{k})$ and equality in the last inequality holds if and only if $H_{4}({\bf v})=0$. Then ${\bf z}$ is a local minimum of . We prove that there exists a constant $M_{{\bf z}}>0$ such that $G({\bf v},t) \geq 0$ for all ${\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and $t$ with $|t|\geq M_{{\bf z}}$. If this condition holds, then by Theorem \[thm:4order-local\], we know that ${\bf z}$ is a local minima of problem . If $H_{f}({\bf z})[{\bf v}]=0$, then it follows $G({\bf v},t)=H_{4}({\bf v})\geq0$ for all $t\in\mathbb{R}$. Next, let $\epsilon>0$ be a small constant. If $H_{f}({\bf z})[{\bf v}]>\epsilon$, then the roots of the quadratic function $t \mapsto G({\bf v},t)$ are bounded by $$\begin{aligned} \frac{4|H_{3}({\bf v})|+\sqrt{|16H_{3}^{2}({\bf v})-4H_{f}({\bf z})[{\bf v}](H_{f}({\bf z})[{\bf v}]+H_{4}({\bf v}))|}}{2H_{f}({\bf z})[{\bf v}]}\leq\frac{(4+2\sqrt{6})M}{2\epsilon},\end{aligned}$$ where $M :=\max_{{\bf v}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}}\max\left\{H_{f}({\bf z})[{\bf v}],|H_{3}({\bf v})|,|H_{4}({\bf v})|\right\}$. The continuity of the functions $H_{f}({\bf z})[\cdot]$, $|H_{3}(\cdot)|$, and $|H_{4}(\cdot)|$ implies $M<+\infty$. Hence, we have $G({\bf v},t)\geq0$ for all $t\in\mathbb{R}$ such that $|t|\geq (2+\sqrt{6})M\epsilon^{-1}$. We now consider the case $H_{f}({\bf z})[{\bf v}]\in(0,\epsilon]$. Let us define the decomposition $$\begin{aligned} {\bf v}=\cos(\theta)\cdot{\bf u}+\sin(\theta)\cdot{\bf w},~H_{f}({\bf z})[{\bf u}]=0,~H_{f}({\bf z})[{\bf w}]>0,~{\bf u}^{*}{\bf w}=0,\end{aligned}$$ where $\theta\in[-\pi,\pi]$ and ${\bf u},{\bf w}\in\mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$. Specifically, introducing the sets $\mathcal N := \{{\bf v} \in {\mathbb{C}^n}: H_f({\bf z})[{\bf v}] = 0\}$ and $\mathcal W := \mathcal{T}_{{\bf z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}\cap \mathcal N^\bot$, there exists $\bar \sigma > 0$ such that $H_f({\bf z})[{\bf w}] \geq \bar \sigma$ for all ${\bf w} \in \mathcal W$. As before, we have $H_{f}({\bf z})[{\bf v}]=H_{f}({\bf z})[{\bf w}]\cdot\sin^{2}(\theta)$ and in the case $H_{4}({\bf u})=0$, the condition $4H_{3}^{2}({\bf u},{\bf w})\leq H_f({\bf z})[{\bf w}]H_{4}({\bf u})$ implies $H_{3}({\bf u},{\bf w})=0$. Utilizing , this yields $|H_{3}({\bf v})|\leq\eta\sin^{2}(\theta)$ for some universal constant $\eta>0$ and $H_{4}({\bf v})=O(\sin(\theta))$. If the discriminant of the quadratic function $t \mapsto G({\bf v},t)$ is negative, it follows $G({\bf v},t)\geq0$ for all $t\in\mathbb{R}$. Otherwise, if the discriminant is non-negative, then the absolute values of the roots are bounded by $$\begin{aligned} &\frac{4|H_{3}({\bf v})|+\sqrt{16H_{3}^{2}({\bf v})-4H_{f}({\bf z})[{\bf v}](H_{f}({\bf z})[{\bf v}]+H_{4}({\bf v}))}}{2H_{f}({\bf z})[{\bf w}]\cdot\sin^{2}(\theta)}\\ & \hspace{10ex}\leq\frac{4\eta\sin^{2}(\theta)+\sqrt{16\eta^{2}-4H_{f}^{2}({\bf z})[{\bf w}]+O(\sin(\theta))}\sin^{2}(\theta)}{2H_{f}({\bf z})[{\bf w}]\cdot\sin^{2}(\theta)}\\ & \hspace{10ex}\leq\frac{4\eta+\sqrt{16\eta^{2}+1}}{2H_{f}({\bf z})[{\bf w}]}\leq\frac{4\eta+1}{\bar \sigma}.\end{aligned}$$ Consequently, it holds $G({\bf v},t)\geq0$ for all $t\in\mathbb{R}$ such that $|t|\geq {(4\eta+1)}{\bar\sigma}^{-1}$. Otherwise, if $H_{4}({\bf u})>0$, then the last condition of this theorem implies $4H_{3}^{2}({\bf u},{\bf w})<H_{f}({\bf z})[{\bf w}]H_{4}({\bf u})$. In this case, the discriminant of $G({\bf v},\cdot)$ satisfies $$\begin{aligned} &16H_{3}^{2}({\bf v})-4H_{f}({\bf z})[{\bf v}](H_{f}({\bf z})[{\bf v}]+H_{4}({\bf v}))\\ & \hspace{10ex}=4[4H_{3}^{2}({\bf u},{\bf w})-H_{f}({\bf z})[{\bf w}]H_{4}({\bf u})]\cdot\sin^{2}(\theta)+O(\sin^{3}(\theta))<0,\end{aligned}$$ if $\theta$ is chosen sufficiently small and thus, we obtain $G({\bf v},t)\geq0$ for all $t\in\mathbb{R}$. Overall, we can set $M_{{\bf z}} :=\max\{(2+\sqrt{6})M\epsilon^{-1},(4\eta+1){\bar\sigma}^{-1}\}$ and ${\bf z}$ is a local minima of problem . Geometric Analysis of the Diagonal Case {#sec:diagonal} ======================================= In this section, we investigate the geometric properties of problem (\[eqn:obj\]) under the assumption that $A$ is a diagonal matrix, i.e., $A=\operatorname*{diag}(\mathbf{a}) = \operatorname*{diag}(a_{1},a_{2},...,a_{n})$. By setting $u_{k}=\lvert z_{k}\rvert^{2}$, we can reformulate problem as a convex problem $$\begin{aligned} \min_{\mathbf{u}\in\mathbb{R}^{n}}~\frac{1}{2}\mathbf{a}^{T}\mathbf{u}+\frac{\beta}{2}\lVert\mathbf{u}\rVert^{2} \quad\mathrm{s.t.}\quad \mathbf{u}\in\Delta_{n}. \label{eqn:obj_1}\end{aligned}$$ where $\Delta_{n}=\{\mathbf{u}\in\mathbb{R}^{n}:u _{k}\in[0,1], k\in[n], \sum_{k\in[n]} u_{k}=1\}$ is the $n$-simplex. We will use this connection later to show that there are no spurious local minima in the diagonal case and that the global solutions can be characterized via the unique solution of the strongly convex problem . We first derive an explicit representation of critical points of problem . Suppose that $A$ is a diagonal and let $\mathbf{z}\in\mathbb{CS}^{n-1}$ be given. Let us set $ \mathcal I := \{k \in [n]: z_k \neq 0\}$ and $$u_k := 0, \quad \forall~k \in \mathcal I^{\sf C}, \quad u_k := \frac{1}{|\mathcal I|} + \frac{1}{2\beta}\left[ \frac{1}{|\mathcal I|} {\sum}_{i \in \mathcal I} a_i - a_k \right], \quad \forall~k \in \mathcal I.$$ Then, ${\bf z}$ is a stationary point if and only if there exist $\theta_{k}\in[0,2\pi)$, $k \in [n]$, such that $u_k \in (0,1]$ for all $k \in \mathcal I$ and $z_{k} = \sqrt{u_k}e^{i\theta_{k}}$ for all $k$. \[tho:diag\_1\] In the diagonal case, introducing the polar form $\mathbf{z}=(r_{1}e^{i\theta_{1}},\dots,r_{n}e^{i\theta_{n}})^{T}$, the first-order optimality conditions reduce to $$\begin{aligned} (a_{k}+2\beta r_k^{2} - 2\lambda) r_k e^{i \theta_k} = 0, \quad \forall~k\in[n] $$ where $\lambda$ is the associated Lagrange multiplier. Specifically, for all $k \in \mathcal I$, we have $a_{k}+2\beta r_k^{2} = 2\lambda$ and summing these equations, we obtain $$\lambda = \frac{2 \beta + \sum_{k \in \mathcal I} a_k}{2|\mathcal I|}, \quad r_k^2 = \frac{2\lambda - a_k}{2\beta},$$ and $2\lambda - a_k \in (0,2\beta]$ for all $k \in \mathcal I$. The claimed result in Lemma \[tho:diag\_1\] now follows immediately by setting $u_k = r_k^2$, $k \in [n]$. Next, we discuss the local minimizer of problem . By combining Theorems \[theorem:glob-suff\] and \[theorem:char-glob-loc\], we see that there are no spurious local minimizer in the diagonal case, i.e., all local solutions are automatically global solutions of problem . \[theorem:char-glob-loc\] Let $A$ be a diagonal matrix. A point ${\bf z} \in {\mathbb{C}^n}$ is a local minimizer of problem if and only if $$\begin{aligned} {\mathrm{grad\!\;}}{f(\bf z)} = 0 \quad \text{and} \quad H = A + 2\beta \operatorname*{diag}(|{\bf z}|^2) - 2\lambda I \succeq 0, \label{eqn:3-7}\end{aligned}$$ where $\lambda \in {\mathbb{R}}$ is the associated Lagrange multiplier. In addition, every local solution ${\bf z}$ can be represented explicitly and has to satisfy $$z_k = \sqrt{u_k} e^{i \theta_k}, \quad \theta_k \in [0,2\pi), \quad {\bf u} = \mathcal P_{\Delta_n}(-{\bf a}/2\beta), \quad \forall~k\in[n],$$ where $\mathcal P_{\Delta_n}$ denotes the Euclidean projection onto the $n$-simplex $\Delta_n$. According to Theorem \[theorem:glob-suff\], a point satisfying the conditions is a global minimum of problem and hence, it also a local minimum. Let ${\bf z} \in {\mathbb{CS}^{n-1}}$ now be an arbitrary local minimum. Then, the first- and second-order necessary optimality conditions hold at ${\bf z}$, i.e., we have ${\mathrm{grad\!\;}}{f({\bf z})} = 0$ and $$\label{eq:soc-in-theo} H_f({\bf z})[{\bf v}] = {\bf v}^*H{\bf v} + 4\beta \cdot {\sum}_{k=1}^n r_k^2 t_k^2 \cos^2(\theta_k - \phi_k) \geq 0$$ for all ${\bf v} \in \mathcal T_{\bf z}\mathcal M$, where $(r_{1}e^{i\theta_{1}},\dots,r_{n}e^{i\theta_{n}})^{T}$ and $(t_{1}e^{i\phi_{1}},\dots,t_{n}e^{i\phi_{n}})^{T}$ are the corresponding polar coordinates of ${\bf z}$ and ${\bf v}$, respectively. As shown in Lemma \[tho:diag\_1\] and using the stationarity condition ${\mathrm{grad\!\;}}{f({\bf z})} = 0$, it follows $H_{kk} = a_k + 2\beta |z_k|^2 - 2\lambda = 0$ for all $k \in \mathcal I = \{k: z_k \neq 0\}$. Next, for $k \in \mathcal I^{\sf C}$, we define ${\bf v} := e_k$, where $e_k$ denotes the $k$-th unit vector. This choice of ${\bf v}$ obviously fulfills $\Re({\bf v}^* {\bf z}) = 0$ and thus, the optimality condition implies $H_{kk} = H_f({\bf z})[{\bf v}] \geq 0$. Since $H$ is diagonal, this yields $H \succeq 0$. In order to verify the explicit characterization of local minimizers, we notice that ${\bf u} = \mathcal P_{\Delta_n}(-{\bf a}/2\beta)$ is the unique solution of the strongly convex problem . Moreover, using the identity $u_{k} \equiv \lvert z_{k}\rvert^{2}$, $k \in [n]$, every global solution of corresponds to a global minimizer of the problem and vice versa. Since problem does not possess spurious local minimizers, this finishes the proof of Theorem \[theorem:char-glob-loc\]. The latter theorem shows that we can identify and explicitly compute the unique equivalence class $\llbracket {\bf z} \rrbracket$ of global minimizer by a projection onto the $n$-simplex. This can be realized numerically in $O(n\log{n})$ operations, see [@finlayson1987numerical]. Inspired by the analysis of phase synchronization problems in [@bandeira2017tightness], we now study the behavior of global minimizer when the diagonal matrix $A$ is perturbed by a random noise matrix $W$. \[theorem:perturb\] Let $A$ be a given diagonal matrix and let $W \in {\mathbb{C}}^{n\times n}$ be a Hermitian noise matrix with noise level $\sigma > 0$. Suppose that ${\bf z}_0$ is a global minimizer of and that the point ${\bf y} \in {\mathbb{CS}^{n-1}}$ satisfies $f_{\sigma}(\mathbf{y})\leq \min_{{\bf z} \in \llbracket {\bf z}_0 \rrbracket} f_{\sigma}(\mathbf{z})$, where $f_\sigma({\bf z}) := f({\bf z}) + \frac{\sigma}{2} {\bf z}^*W{\bf z}$. Then, it holds that $$\begin{aligned} \min_{{\bf z} \in \llbracket {\bf z}_0 \rrbracket}~\lVert\mathbf{y}-\mathbf{z}\rVert_{4} \leq \sqrt[3]{{2\sigma\beta^{-1}\lVert W\rVert_{2}n^{1/4}}}.\end{aligned}$$ As usual, we introduce the polar coordinates $\mathbf{z}_{0}=(r_{1}e^{i\theta_{1}},...,r_{n}e^{i\theta_{n}})^{T}$ and $\mathbf{y}=(t_{1}e^{i\phi_{1}},...,t_{n}e^{i\phi_{n}})^{T}$. Due to Theorem \[theorem:char-glob-loc\], we can assume $\theta_{k}=\phi_{k}$ and we have $a_k + 2\beta r_k^2 - 2\lambda \geq 0$ for all $k \in [n]$, where $a_k = A_{kk}$ and $\lambda$ is the associated multiplier of ${\bf z}_0$. Thus, using $f_{\sigma}(\mathbf{y})\leq f_{\sigma}(\mathbf{z}_{0})$, this implies $$\begin{aligned} \frac{\sigma}{2}(\mathbf{z}_{0}^{*}W\mathbf{z}_{0}-\mathbf{y}^{*}W\mathbf{y}) & \geq \frac{1}{2}(\mathbf{y}^{*}A\mathbf{y}-\mathbf{z}_{0}^{*}A\mathbf{z}_{0})+\frac{\beta}{2} \cdot {\sum}_{k\in[n]}(t_{k}^{4}-r_{k}^{4})\\ &=\frac{1}{2}\sum_{k\in[n]}(a_{k}+\beta(t_{k}^{2}+r_{k}^{2}))(t_{k}^{2}-r_{k}^{2})\\ & \geq \frac{1}{2}\sum_{k\in[n]}(2\lambda +\beta(t_{k}^{2}-r_{k}^{2}))(t_{k}^{2}-r_{k}^{2}) \\ &=\frac{\beta}{2}\sum_{k\in[n]}(t_{k}^{2}-r_{k}^{2})^{2}\geq\frac{\beta}{2}\sum_{k\in[n]}(t_{k}-r_{k})^{4}=\frac{\beta}{2}\lVert\mathbf{y}-\mathbf{z}_{0}\rVert_{4}^{4}.\end{aligned}$$ Note that the last inequality follows from $(t_{k}+r_{k})^{2}\geq(t_{k}-r_{k})^{2}$. Furthermore, by Hölder’s inequality and by $\|{\bf x}\|_{4/3} \leq n^{1/4} \|{\bf x}\|$, we have $$\begin{aligned} \mathbf{z}_{0}^{*}W\mathbf{z}_{0}-\mathbf{y}^{*}W\mathbf{y} & =\Re((\mathbf{z}_{0}-\mathbf{y})^{*}W(\mathbf{z}_{0}+\mathbf{y})) \leq \lVert\mathbf{z}_{0}-\mathbf{y}\rVert_{4}\lVert W(\mathbf{z}_{0}+\mathbf{y})\rVert_{4/3} \\ & \leq n^{1/4} \lVert\mathbf{z}_{0}-\mathbf{y}\rVert_{4}\lVert W(\mathbf{z}_{0}+\mathbf{y})\rVert \leq 2n^{1/4} \lVert W\rVert_{2}\lVert\mathbf{z}_{0}-\mathbf{y}\rVert_{4}.\end{aligned}$$ Combining the above two inequalities, concludes the proof. If $W \in {\mathbb{C}}^{n \times n}$ is a Hermitian random matrix with i.i.d. off-diagonal entries following a standard complex normal distribution and with zero diagonal entries, then Bandeira, Boumal, and Singer, [@bandeira2017tightness], have shown that the bound $\lVert W\rVert_{2}\leq3\sqrt{n}$ holds with probability at least $1-2n^{-5/4}-e^{-n/2}$. Combing this observation with Theorem \[theorem:perturb\], we can obtain $$\begin{aligned} \min_{{\bf z} \in \llbracket {\bf z}_0 \rrbracket}~\lVert\mathbf{y}-\mathbf{z}_{0}\rVert_{4}\leq\sqrt[3]{6\sigma\beta^{-1}}\cdot n^{1/4}\end{aligned}$$ with probability at least $1-2n^{-5/4}-e^{-n/2}$. Geometric Analysis of the Rank-One Case {#sec:rank1} ======================================= In this section, we investigate the case when $A$ is rank-one and positive semidefinite, i.e., we can write $A=\mathbf{a}\mathbf{a}^{*}$ for some $\mathbf{a}\in\mathbb{C}^{n}$ and the quartic-quadratic problem reduces to $$\begin{aligned} \min_{\mathbf{z}\in\mathbb{C}^{n}}~f(\mathbf{z})=\frac{1}{2}\lvert\mathbf{a}^{*}\mathbf{z}\rvert^{2}+\frac{\beta}{2}\lVert\mathbf{z}\rVert_{4}^{4} \quad \mathrm{s.t.} \quad \lVert\mathbf{z}\rVert_{2}=1. \label{eqn:obj-rank1}\end{aligned}$$ The associated first- and second-order necessary optimality conditions are given by $${\bf a}^*{\bf z} \cdot {\bf a} + 2\beta \operatorname*{diag}(|{\bf z}|^2) {\bf z} = 2\lambda {\bf z}, \quad 2\lambda = \lvert\mathbf{a}^{*}\mathbf{z}\rvert^{2}+2\beta\lVert\mathbf{z}\rVert_{4}^{4}$$ and $H_f({\bf z})[{\bf v}] = {\bf v}^*[{\bf a}{\bf a}^* + 2\beta \operatorname*{diag}(|{\bf z}|^2) - 2\lambda I]{\bf v} + 4\beta \sum_{k=1}^n \Re(v_k \bar z_k)^2 \geq 0$ for all $\mathbf{v} \in {\mathbb{C}^n}$ with $\Re(\mathbf{v}^{*}\mathbf{z})=0$. We now present a first structural and preparatory property of local and global minima. Suppose that $\mathbf{z}$ is a local minimizer of . Then, for all $k\in[n]$ with $a_{k}=0$ it holds that $\lvert z_{k}\rvert^{2}=\frac{\lambda}{\beta}$. \[lem:4-1\] If $a_{k}=0$, the first-order optimality conditions imply $\beta\lvert z_{k}\rvert^{2}z_{k}=\lambda z_{k}$. Let us assume $z_{k}=0$ and let us choose ${\bf v} \in {\mathbb{C}^n}$ with $v_{k}=1$ and $v_{j}=0$ for all $j\neq k$. Due to $\lambda \geq \beta \|{\bf z}\|_4^4 \geq \frac{\beta}{n} > 0$, we obtain $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}]=-2\lambda<0,\end{aligned}$$ which contradicts the second-order necessary optimality conditions. Hence, we have $\lvert z_{k}\rvert^{2}=\frac{\lambda}{\beta}$. In the following sections, we discuss two different classes of local minima, which are characterized by the orthogonality to the vector $\mathbf{a}$. Orthogonal local minima ----------------------- We first analyze the case where the local minimizer $\mathbf{z}$ satisfies $\mathbf{a}^{*}\mathbf{z}=0$. Suppose that $\mathbf{z}$ is a local minimizer satisfying $\mathbf{a}^{*}\mathbf{z}=0$. Then, $\mathbf{z}$ has at most one zero component and all of its nonzero components must have the same modulus. \[thm:4-3\] By the first-order optimality conditions, it follows $z_{k}=0$ or $\lvert z_{k}\rvert^{2}=\frac{\lambda}{\beta}$ for all $k$. Hence, all nonzero components of $\mathbf{z}$ have the same modulus. Without loss of generality we now assume that $\lvert z_{k}\rvert^{2}=\frac{\lambda}{\beta}$ for $1\leq k\leq \tau$ and $z_{k}=0$ for $\tau+1\leq k\leq n$. Due to Lemma \[lem:4-1\], we have $a_{n-1},a_{n}\neq0$ if $\tau \leq n-2$. Let us set $$\begin{aligned} v_{k} = 0, \quad k\in[n-2], \quad v_{n-1}=\frac{a_{n}}{\sqrt{\lvert a_{n-1}\rvert^{2}+\lvert a_{n}\rvert^{2}}},\quad v_{n}=\frac{-a_{n-1}}{\sqrt{\lvert a_{n-1}\rvert^{2}+\lvert a_{n}\rvert^{2}}}.\end{aligned}$$ Then, it holds that $H_{f}(\mathbf{z})[\mathbf{v}]=-2\lambda<0$, which is a contradiction. Thus, we have $\tau = n-1$ or $\tau = n$ (which means that all components $z_k$ are nonzero). Next, we derive conditions under which the existence of such local minima can be ensured. Before we present the formal statement and proof of the main theorem, we discuss a result that is used later in Theorem \[theorem:exi-rank-1\]. If $\|{\bf a}\|_\infty \leq\frac{1}{2}\|{\bf a}\|_1$, there exist phases $\{\theta_{k}\}_{k\in[n]}$, $\theta_k \in [0,2\pi]$, such that $\sum_{k\in[n]}e^{i\theta_{k}}a_{k}=0$. \[lem:4-3\] Let us assume $a_{k}\neq0$ for all $k$. If $n=1$, the condition $\|{\bf a}\|_\infty \leq\frac{1}{2}\|{\bf a}\|_1$ is never satisfied and hence, the statement in Lemma \[lem:4-3\] holds automatically. In the case $n=2$, we have $2\max \{ \lvert a_{1}\rvert, |a_2|\} \leq \lvert a_{1}\rvert+\lvert a_{2}\rvert$. This implies $\lvert a_{1}\rvert=\lvert a_{2}\rvert$ and thus, we can choose $\theta_1$ and $\theta_2$ such that $e^{i\theta_{1}}a_{1}=\lvert a_{1}\rvert$ and $e^{i\theta_{2}}a_{2}=-\lvert a_{1}\rvert$. Otherwise assume $n\geq3$ and $\lvert a_{1}\rvert=\|{\bf a}\|_\infty$. Let $b_{1}=\sum_{k\geq 3}\lvert a_{k}\rvert$, $b_{2}=\lvert a_{2}\rvert$ and $b_{3}=\lvert a_{1}\rvert$. It holds that $b_{1}+b_{2}=\|{\bf a}\|_1-\lvert a_{1}\rvert\geq b_{3}$, which means that the numbers $b_{1},b_{2},b_{3}$ can be interpreted as sides of a (degenerated) triangle. Let $ABC$ be such a triangle embedded into the complex space, where $A,B,C\in\mathbb{C}$ denote the nodes of $ABC$ with $\lvert B-C\rvert=b_{1},\lvert C-A\rvert=b_{2},\lvert A-B\rvert=b_{3}$. Consequently, there exist $\phi_{1},\phi_{2},\phi_{3}\in[0,2\pi)$ such that $B-C=e^{i\phi_{1}}b_{1},C-A=e^{i\phi_{2}}b_{2},A-B=e^{i\phi_{3}}b_{3}$. But then we have $\sum_{k=1}^{3}e^{i\phi_{k}}b_{k}=0$, which completes the proof of the lemma. \[theorem:exi-rank-1\] There exists a local minimizer $\mathbf{z}$ of such that $\mathbf{a}^{*}\mathbf{z}=0$ if and only if $\|{\bf a}\|_\infty \leq\frac{1}{2}\lVert\mathbf{a}\rVert_{1}$, or $\mathbf{a}$ has only one nonzero component and we have $\lVert\mathbf{a}\rVert^{2}\geq2\beta/(n-1)$. Further, if $\mathbf{a}$ satisfies such conditions, all local minima with $\mathbf{a}^{*}\mathbf{z}=0$ are the only global minima of . Let $\mathbf{z}=(r_{1}e^{i\theta_{1}},\dots,r_{n}e^{i\theta_{n}})^{T}$ be a local minimizer of problem (\[eqn:obj-rank1\]) such that $\mathbf{a}^{*}\mathbf{z}=0$. By Theorem \[thm:4-3\], we only need to consider the cases when the local minimizer has no zero component or exactly one zero component. *Case 1.* If $\mathbf{z}$ does not have any zero component, then it follows $\lvert z_{k}\rvert^{2}=\frac{1}{n}$ for all $k$ and we have $$\begin{aligned} \mathbf{a}^{*}\mathbf{z}=\frac{1}{\sqrt{n}}\sum_{k\in[n]}e^{i\theta_{k}}a_{k}=0,\end{aligned}$$ which implies $\lvert a_{j}\rvert=\lvert\sum_{k\neq j}e^{i\theta_{k}}a_{k}\rvert\leq\sum_{k\neq j}\lvert a_{k}\rvert=\lVert\mathbf{a}\rVert_{1}-\lvert a_{j}\rvert$ for all $j \in[n]$. Choosing $a_{j}$ to be the element with maximal modulus, we get $ \|{\bf a}\|_\infty \leq\frac{1}{2}\lVert\mathbf{a}\rVert_{1}$. *Case 2.* Let us suppose $z_{n} = 0$. Then, due to Lemma \[lem:4-1\], we obtain $a_{n}\neq0$. Let us assume that there exists another component $a_{k}\neq0$ for some $k \in [n-1]$. Setting $$\begin{aligned} v_{j}=iz_{j}, \quad j\neq k,n,\quad v_{k}=(1-\lvert a_{n}\rvert)iz_{k},\quad v_{n}= \frac{a_{n}}{\lvert a_{n}\rvert}\bar{a}_{k}\cdot iz_{k},\end{aligned}$$ and normalizing $\mathbf{v}$, we have $\mathbf{a}^{*}\mathbf{v} = \Re({\bf z}^*{\bf v})=0$ and the curvature is given by $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}]=2\beta\sum_{k\in[n-1]} |z_k|^2 \lvert v_{k}\rvert^{2}-2\lambda=-2\lambda\lvert v_{n}\rvert^{2}<0,\end{aligned}$$ which contradicts with the second-order optimality conditions. Hence, we can infer $a_{k}=0$ for all $k\in[n-1]$, which implies that $\mathbf{a}$ has only one nonzero component. By Theorem \[thm:4-3\], we have $\lvert z_{k}\rvert^{2}=\frac{1}{n-1}$ for all $k\in[n-1]$. The second-order necessary optimality conditions yields $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}]=\lvert a_{n}v_{n}\rvert^{2}+\frac{2\beta}{n-1}(1-\lvert v_{n}\rvert^{2})-\frac{2\beta}{n-1}=\left(\lvert a_{n}\rvert^{2}-\frac{2\beta}{n-1}\right)\lvert v_{n}\rvert^{2}\geq0,\end{aligned}$$ for $|v_n| \in [0,1]$ and it follows $\lVert\mathbf{a}\rVert^{2}=\lvert a_{n}\rvert^{2}\geq\frac{2\beta}{n-1}$. We continue with the proof of the second direction. In particular, suppose that $\mathbf{a}$ satisfies the conditions stated in Theorem \[theorem:exi-rank-1\]. We again discuss two cases. *Case 1.* By Lemma \[lem:4-3\], if $\|{\bf a}\|_\infty \leq\frac{1}{2}\lVert\mathbf{a}\rVert_{1}$, we can choose phases $\{\theta_{k}\}_{k\in[n]}$ such that $\sum_{k\in[n]}e^{i\theta_{k}}a_{k}=0$. Let us set $z_{k}= e^{i\theta_{k}} / {\sqrt{n}}$ for all $k\in[n]$. Then, we have $\mathbf{a}^{*}\mathbf{z}=0$ and $f(\mathbf{z})= {\beta}/{(2n)}$, which is the lower bound of the objective function $f$. Thus, in this case $\mathbf{z}$ is a global minimizer of (\[eqn:obj-rank1\]). Moreover, the objective function attains its optimal value if and only if $\mathbf{a}^{*}\mathbf{z}=0$ and $\|{\bf z}\|_{4}^4={1}/{n}$. *Case 2.* If $\mathbf{a}$ has only one nonzero component $a_{n}$ with $\lVert\mathbf{a}\rVert^{2}\geq\frac{2\beta}{n-1}$, the matrix $A$ is diagonal and we can apply the results derived in section 3. Specifically, by Theorem \[theorem:char-glob-loc\], it can be shown that all local minimizer are global minimizer and satisfy $$\lvert z_{k}\rvert=\frac{1}{\sqrt{n-1}}, \quad \forall~k\in[n-1], \quad \text{and} \quad z_{n}=0.$$ This finishes the proof of Theorem \[theorem:exi-rank-1\]. Non-orthogonal local minima --------------------------- We discuss the case when there is no local minimizer $\mathbf{z}$ such that $\mathbf{a}^{*}\mathbf{z}=0$, or, equivalently, $\mathbf{a}$ does not satisfy the conditions in Theorem \[thm:4-3\]. By the first-order optimality conditions, all $z_{k}$ with $a_{k}\neq0$ satisfy $$\begin{aligned} \mathrm{arg}(z_{k})-\mathrm{arg}(a_{k})=\mathrm{arg}(\mathbf{a}^{*}\mathbf{z})=\mathrm{const},\end{aligned}$$ where $\mathrm{arg}(z)$ is the principal angle of the complex number $z$ modulo $\pi$. Since a global shift of the phase will not change the objective function value and the first-order optimality conditions, we can shift $\mathbf{z}$ by a global phase such that the principal angles of the nonzero components are the same as $a_{k}$. In the case $a_{k}=0$, the phase of $z_{k}$ does not influence the objective function value and the first-order optimality conditions and we can adjust ${\bf z}$ to be a real number. Consequently, for every stationary point of problem , we can find a corresponding stationary point which has the same objective function value and satisfies the following ‘consistency’ property. A stationary point $\mathbf{z}$ of problem is called consistent, if it satisfies $$\begin{aligned} \bar a_{k} z_{k}\in\mathbb{R}, \quad \forall~k \; \text{with} \; a_{k}\neq0 \quad \text{and} \quad z_{k} \in\mathbb{R}, \quad \forall~k \; \text{with} \; a_{k}=0.\end{aligned}$$ Note that the corresponding consistent stationary point of a local minimizer of problem does not need to be a local minimizer. On the other hand, shifted consistent stationary points of global minimizer remain global minimizer. In this subsection, we focus on structural properties of consistent stationary points of . Suppose $\mathbf{a}$ does not satisfy the conditions in Theorem \[thm:4-3\] and there exists a local minimizer $\mathbf{z}$ such that $\mathbf{a}^{*}\mathbf{z}\neq0$. If we have $z_{k}=0$ for some $k\in[n]$, then the first-order optimality conditions imply $a_{k}=0$ which contradicts Lemma \[lem:4-1\]. Hence, for any local minima $\mathbf{z}$ with $\mathbf{a}^{*}\mathbf{z}\neq0$, we have $z_{k}\neq0$ for all $k\in[n]$. In the following result, we show that consistent local minima must belong to the same equivalence class defined in . Suppose that $\mathbf{z},\mathbf{y}\in\mathbb{CS}^{n-1}$ are two consistent local minima of problem with $\mathbf{a}^{*}\mathbf{z}\neq0$ and $\mathbf{a}^{*}\mathbf{y}\neq0$. Then, we have ${\bf y} \in \llbracket {\bf z} \rrbracket$. The consistency of the stationary points ${\bf y}$ and ${\bf z}$ implies ${\bf y}^*{\bf z} \in {\mathbb{R}}$ and thus, it holds that $i\mathbf{y} \in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{CS}^{n-1}$ and $i\mathbf{z}\in\mathcal{T}_{\mathbf{y}}\mathcal{M}\cap\mathbb{CS}^{n-1}$. By the second-order necessary optimality conditions, we have $$\begin{aligned} \nonumber H_{f}(\mathbf{z})[i\mathbf{y}]&=\lvert\mathbf{a}^{*}\mathbf{y}\rvert^{2}+2\beta\sum_{k\in[n]}\lvert z_{k} y_{k}\rvert^{2}-2\lambda_\mathbf{z} \geq 0,\\ H_{f}(\mathbf{y})[i\mathbf{z}]&=\lvert\mathbf{a}^{*}\mathbf{z}\rvert^{2}+2\beta\sum_{k\in[n]}\lvert z_{k} y_{k}\rvert^{2}-2\lambda_\mathbf{y} \geq0, \label{eqn:4-10}\end{aligned}$$ where $2\lambda_{\bf z} = |{\bf a}^*{\bf z}|^2 + 2\beta \|{\bf z}\|_4^4$ and $2\lambda_{\bf y} = |{\bf a}^*{\bf y}|^2 + 2\beta \|{\bf y}\|_4^4$. Summing those two inequalities yields $$\begin{aligned} 0 & \leq - 2\beta \left[ \|{\bf z}\|_4^4 - 2 \sum_{k\in[n]}\lvert z_{k} y_{k}\rvert^{2} + \|{\bf y}\|_4^4 \right] = -2\beta \sum_{k \in [n]} [|z_k|^2 - |y_k|^2]^2. \end{aligned}$$ Hence, we have $\lvert z_{k}\rvert =\lvert y_{k}\rvert$ for all $k\in[n]$. Similarly, if $\mathbf{z},\mathbf{y}\in\mathbb{CS}^{n-1}$ are two local minima of problem such that $z_{k}$ and $y_{k}$ have the same phases for all $k\in[n]$, then ${\bf y} \in \llbracket {\bf z} \rrbracket$. We now prove that there are no spurious consistent local minima. If $\mathbf{a}$ does not satisfy the conditions in Theorem \[thm:4-3\], then all the consistent local minima of problem are global minima. Suppose $\mathbf{z},\mathbf{y}\in\mathbb{CS}^{n-1}$ are two consistent local minima. Using the inequalities in and $\lvert z_{k}\rvert = \lvert y_{k}\rvert$, we obtain $\lvert\mathbf{a}^{*}\mathbf{z}\rvert^{2}=\lvert\mathbf{a}^{*}\mathbf{y}\rvert^{2}$. Hence, all consistent local minima have the same objective function value. Since $\mathbf{a}$ does not satisfy the conditions in Theorem $\ref{thm:4-3}$, there exists a consistent global minimizer satisfying $\mathbf{a}^{*}\mathbf{z}\neq0$. This shows that all consistent local minima are global minima. Combining the results of the last two subsections, we see that global minima of problem are unique up to certain shifts in the phase. In particular, we can shift the phases of components with $(\mathbf{a}^{*}\mathbf{z})a_{k}=0$ arbitrarily and shift all the other components by the same angle. Analyzing the Geometric Landscape – the Real Case {#sec:strict_saddle} ================================================= We now investigate a variant of the so-called *strict-saddle property* introduced by Ge, Jin, and Zheng in [@ge2017no Definition 2]. More specifically, as in [@sun2016geometric Theorem 2.2], we strengthen the first condition in [@ge2017no Definition 2] to uniform positive definiteness of the Riemannian Hessian. \[def:ssp\] Let $\xi,\epsilon,\zeta>0$ be given constants. A function $f$ is called $(\xi,\epsilon,\zeta)$-**strict-saddle** if for all $\mathbf{z}\in\mathcal{M}$ one of the following conditions holds: - (Strong convexity). For all $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ we have $H_{f}(\mathbf{z})[\mathbf{v}]\geq\xi$. - (Large gradient). It holds that $\lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert\geq\epsilon$. - (Negative curvature). There exists $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ with $H_{f}(\mathbf{z})[\mathbf{v}]\leq-\zeta$. The strict-saddle property can be utilized to establish polynomial time convergence rates of second-order optimization algorithms, such as the Riemannian trust region method [@sun2016geometric], and almost sure convergence to local minimizer of Riemannian gradient descent methods with random initialization, see, e.g., [@lee2016gradient; @lee2017first-order]. In the following sections, we analyze the geometric landscape of problem and show that the strict-saddle property is satisfied in the real case when the interaction coefficient $\beta$ is sufficiently small or large. In general, due to the intricate relation between the quadratic and quartic terms, we can not expect that the conditions in Definition \[def:ssp\] do hold for all choices of $\beta$ and $A$. For instance, let us consider the example $A := \alpha {\mathds 1}{\mathds 1}^T$, $\alpha \geq 0$, and ${\bf z} := {\mathds 1}/\sqrt{n} \in {\mathbb{S}^{n-1}}$. Then, the associated multiplier $\lambda$ is given by $2 \lambda = \alpha n + 2\beta /n$ and it can be shown that ${\bf z}$ is a stationary point of for all $\alpha$. Furthermore, for all ${\bf v} \in \mathcal{T}_{\mathbf{z}}\mathcal{M}$, it follows $$H_f({\bf z})[{\bf v}] = {\bf v}^T[A + 6\beta \operatorname*{diag}(|{\bf z}|^2) - 2\lambda I]{\bf v} = [ 4\beta n^{-1} - \alpha n ] \|{\bf v}\|^2.$$ Hence, in the case $\alpha = 4\beta / n^2$, the strict-saddle property can not hold. Let us further note that the strong convexity condition in Definition \[def:ssp\] is never satisfied at stationary points in the complex case. In particular, if ${\bf z} \in {\mathbb{CS}^{n-1}}$ is a critical point, then we have $i{\bf z} \in \mathcal{T}_{\mathbf{z}}\mathcal{M}\cap{\mathbb{CS}^{n-1}}$ and $H_f({\bf z})[i {\bf z}] = 0$ which contradicts condition 1. In Figure \[fig:land\], we illustrate different landscapes of the mapping $f$ when the problem is real and three-dimensional and the parameter $\beta$ changes. We consider the setup $$\label{eq:example-A} A = \begin{pmatrix} 1 & 0 & 1 \\ 0 & 1 & 0 \\ 1 & 0 & 1 \end{pmatrix}, \quad \beta \in \{0.25,0.75,1.25,3.25\},$$ and the eigenvalues of $A$ are given by $2$, $1$, and $0$, respectively. Moreover, we use spherical coordinates $(\phi,\theta) \mapsto (\cos(\phi),\sin(\phi)\cos(\theta),\sin(\phi)\sin(\theta))^{T}$, $\phi \in [0,2\pi]$, $\theta \in [0,\pi]$, to plot the objective function $f$ on the sphere. -- -- -- -- Figure \[fig:land\] demonstrates that the landscape of the objective function varies a lot when the interaction coefficient $\beta$ changes. Specifically, it shows that the number of stationary points and local minima increases from $6$ to $26$ and from $2$ to $8$ as $\beta$ increases from $0.25$ to $3.25$. In section \[sec:large\] and section \[sec:small\], we investigate basic geometric features and the strict-saddle property for large and small choices of $\beta$ while keeping the matrix $A$ fixed. In particular, our results will allow us to characterize and describe the geometric landscape of $f$ in the sub-figures (a) and (d) of Figure \[fig:land\]. In the following, we assume $n\geq2$ since the landscape of $f$ is trivial in the case $n=1$. Large interaction coefficient {#sec:large} ----------------------------- In this section, we prove that the objective function possesses the strict-saddle property if the coefficient $\beta$ is chosen sufficiently large and satisfies $\beta \approx C\rho n^{3/2}$ for some constant $C > 0$. Here, $\rho := \lambda_1 - \lambda_n$ denotes the difference between the largest and the smallest eigenvalue of $A$. We analyze the geometric properties and behavior of $f$ on the following sets: - (Strong convexity) $\mathcal{R}_{1} :=\{\mathbf{z}\in{\mathbb{S}^{n-1}}: \| |{\bf z}|^2-{\mathds 1}/n \|_\infty \leq 1/2n\}$, - (Large gradient) $\mathcal{R}_{2} := \{\mathbf{z}\in{\mathbb{S}^{n-1}}: \| |{\bf z}|^2-{\mathds 1}/n \|_\infty \geq 1/2n,\\\min_{k\in[n]} z_{k}^{2}\geq(n-1)/4n^{2}\}$, - (Negative curvature) $\mathcal{R}_{3} :=\{\mathbf{z}\in{\mathbb{S}^{n-1}}:\min_{k\in[n]}z_{k}^{2}\leq(n-1)/4n^{2}\}$. Obviously, these three regions cover the sphere $\mathbb{S}^{n-1}$. An illustration of the regions $\mathcal R_1$–$\mathcal R_3$ is given later in Figure \[figure:areas\]. We first present a preparatory result that is used later to estimate of the spectrum of the Riemannian Hessian. The following estimate holds for all $\mathbf{z}\in {\mathbb{S}^{n-1}}$: $$\label{eq:lem-05} z_{0}^2 \leq \min_{{\bf v} \in \mathcal T_{\bf z}\mathcal M \cap {\mathbb{S}^{n-1}}} \sum_{k \in [n]} v_k^2z_k^2 \leq \frac{n}{n-1}(1-z_{0}^2)z_{0}^2,$$ where $ z_{0}^2 := \min_{k \in [n]} z_k^2$. \[lem:large-0\] If there exists $k \in [n]$ with $z_k = 0$, then the optimal value of the latter optimization problem is $0$ and it is attained for ${\bf v} = e_k$. Next, let us assume $z_0^2 > 0$ and let $i \in [n]$ be given with $z_i^2 = z_0^2$. We define $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ via $$\begin{aligned} v_{i}=-\frac{(n-1)c}{z_{1}},\quad v_{k}=\frac{c}{z_{k}}, \quad c=\left[\frac{(n-1)^{2}}{z_{i}^{2}}+{\sum}_{k\neq i}\frac{1}{z_{k}^{2}}\right]^{-1/2}, \quad k \neq i.\end{aligned}$$ Using the reverse Hölder inequality, we have $\sum_{k\neq i} z_k^{-2} \geq (n-1)^2 (\sum_{k \neq i} z_k^2)^{-1}$ and thus, we obtain $$\begin{aligned} \sum_{k\in[n]}z_{k}^{2}v_{k}^{2}=n(n-1)c^{2} \leq \frac{n(n-1)}{\frac{(n-1)^{2}}{z_{0}^{2}}+\frac{(n-1)^2}{{1-z_{0}^{2}}}} \leq \frac{n}{n-1}(1-z_{0}^{2})z_{0}^{2}.\end{aligned}$$ This establishes the upper bound in . The lower bound follows from $\|{\bf v}\| = 1$ and $z_k^2 \geq z_0^2$ for all $k \in [n]$. In the following lemma, we verify that the objective function has the strong convexity property on the region $\mathcal{R}_{1}$. Let $\gamma > 0$ be given and suppose that $\beta \geq 2(1+\gamma)\rho n$. Then, for all ${\bf z} \in \mathcal R_1$ and all $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ it holds that $H_{f}(\mathbf{z})[\mathbf{v}] \geq \gamma \rho$. \[lem:large-1\] Let us define $w_{k} := z_{k}^{2}-\frac{1}{n}$. Hence, due to $\sum_{k\in[n]} w_{k}=0$, we obtain: $$\begin{aligned} \sum_{k\in[n]}z_{k}^{4}=\sum_{k\in[n]}\left[w_{k}+\frac{1}{n}\right]^{2}=\frac{1}{n}+\sum_{k\in[n]} w_{k}^{2} \leq \frac{5}{4n}.\end{aligned}$$ Next, by Lemma \[lem:large-0\] and ${\bf z} \in \mathcal R_1$, we have $$\begin{aligned} \min_{\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}}\sum_{k\in[n]}z_{k}^{2}v_{k}^{2} \geq \min_{k \in [n]} z_k^2 \geq \frac{1}{2n}.\end{aligned}$$ Combining the last results and using $\mathbf{v}^{T}A\mathbf{v}-\mathbf{z}^{T}A\mathbf{z}\geq-\rho$, it follows $$\begin{aligned} \min_{\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}}H_{f}(\mathbf{z})[\mathbf{v}]&\geq-\rho+\frac{2\beta}{n} \left [ \frac32 - \frac54 \right] \geq \gamma \rho $$ for all $\mathbf{z}\in\mathcal{R}_{1}$. The next lemma shows that the norm of the Riemannian gradient is strictly larger than zero – uniformly – on the region $\mathcal{R}_{2}$. Let $\beta \geq 8(1+\frac{1}{n-1}) (1+\gamma) \rho n^{3/2}$ be given for some $\gamma > 0$. Then, for all $\mathbf{z}\in\mathcal{R}_{2}$, it holds that $\lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert \geq\frac{\gamma}{\sqrt{2}}\rho$. \[lem:large-2\] First, we split the norm of the Riemannian gradient ${\mathrm{grad\!\;}}{f(\mathbf{z})}$ as follows $$\begin{aligned} \lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert & \geq2\beta\lVert\operatorname*{diag}(|{\bf z}|^2){\bf z}- \|{\bf z}\|^4_4 \cdot \mathbf{z} \rVert - \lVert A{\bf z}-(\mathbf{z}^{T}A\mathbf{z})\cdot \mathbf{z}\rVert \\ & = 2\beta \sqrt{\|{\bf z}\|_6^6 - \|{\bf z}\|_4^8} - \lVert A{\bf z}-(\mathbf{z}^TA\mathbf{z})\cdot \mathbf{z}\rVert .\end{aligned}$$ We now estimate the minuend and subtrahend in the latter expression separately. Let us set $\mathbf{z}=\sum_{k} \alpha_{k}\mathbf{p}_{k}$, where $\{\mathbf{p}_{k}\}_k$ is the set of orthogonal eigenvectors of the matrix $A$ with corresponding eigenvalues $\lambda_{1},...,\lambda_{n}$. Then, it holds that $\sum_{k}\alpha_{k}^{2}=1$ and $$\begin{aligned} \lVert A{\bf z}-(\mathbf{z}^{T}A\mathbf{z})\cdot {\bf z}\rVert^{2}&=\mathbf{z}^{T}A^{2}\mathbf{z}-(\mathbf{z}^{T}A\mathbf{z})^{2}=\sum_{k\in[n]}\alpha_{k}^{2}\lambda_{k}^{2}-\left[{\sum}_{k\in[n]}\alpha_{k}^{2}\lambda_{k}\right]^{2} \\ &=\sum_{k\in[n]}\alpha_{k}^{2}\lambda_{k}^{2} -\sum_{k,\ell \in[n]}\alpha_{k}^{2} \alpha_\ell^2 \lambda_{k} \lambda_\ell \\ & =\frac{1}{2}\sum_{k, \ell \in[n]}\alpha_{k}^{2} \alpha_{\ell}^{2}(\lambda_{k}-\lambda_{\ell})^{2} \leq \frac{1}{2} \rho^2 \sum_{k, \ell \in [n]} \alpha_{k}^{2} \alpha_{\ell}^{2} = \frac{1}{2}\rho^{2}.\end{aligned}$$ We continue with bounding the term $\lVert\mathbf{z}\rVert_{6}^{6}-\lVert\mathbf{z}\rVert_{4}^{8}$. As before using the spherical constraint $\sum_k z_k^2 = 1$, we obtain: $$\begin{aligned} \lVert\mathbf{z}\rVert_{6}^{6}-\lVert\mathbf{z}\rVert_{4}^{8} = \sum_{k\in[n]} z_{k}^{6}-\left[ {\sum}_{k\in[n]}z_{k}^{4}\right]^{2} = \frac12 \sum_{k,\ell\in[n]} z_{k}^{2}z_{\ell}^{2}(z_{k}^{2}-z_{\ell}^{2})^{2}.\end{aligned}$$ In the following and without loss of generality, we assume that the components of ${\bf z}$ are ordered and satisfy $z_1^2 \leq z_2^2 \leq ... \leq z_n^2$. Notice that $\mathbf{z}\in\mathcal{R}_{2}$ implies $z_1^2 \neq z_n^2$ and let us choose $t \in [n-1]$ such that $z_t^2 \leq \frac12(z_1^2 + z_n^2) \leq z_{t+1}^2$. Then, we have $$\begin{aligned} 1=\sum_{k=1}^{t}z_{k}^{2}+\sum_{k=t+1}^{n}z_{k}^{2}\geq tz_{1}^{2}+\frac{n-t-1}{2}(z_{1}^{2}+z_{n}^{2}) + z_n^2,\end{aligned}$$ which yields $$\begin{aligned} t\geq\frac{(n-1)(z_{n}^{2}+z_{1}^{2}) + 2z_n^2 -2}{z_{n}^{2}-z_{1}^{2}} \geq \frac{(n-1)z_1^2 + z_n^2 - 1}{z_n^2 - z_1^2}.\end{aligned}$$ Consequently, we get $$\begin{aligned} \sum_{k,\ell\in[n]}z_{k}^{2}z_{\ell}^{2}(z_{k}^{2}-z_{\ell}^{2})^{2} & \geq z_1^2 \sum_{k \in [n]} z_k^2 (z_k^2 - z_1^k)^2 + z_n^2 \sum_{k \in [n]} z_k^2 (z_k^2 - z_n^2)^2 \\ &\geq \frac{1}{4}(z_{n}^{2}-z_{1}^{2})^{2} \left[ z_{1}^{2} \sum_{k=t+1}^{n} z_{k}^{2}+z_{n}^{2}\sum_{k=1}^{t}z_{k}^{2} \right] \\ &=\frac{1}{4}(z_{n}^{2}-z_{1}^{2})^{2}\left [z_{1}^{2}+(z_{n}^{2}-z_{1}^{2})\sum_{k=1}^{t}z_{k}^{2}\right]\\ &\geq\frac{1}{4}(z_{n}^{2}-z_{1}^{2})^{2}[z_{1}^{2}+(z_{n}^{2}-z_{1}^{2})t z_{1}^{2}] \geq \frac{n}{4} (z_n^2-z_1^2)^2 z_1^4. $$ Using $z_{1}^{2}\geq\frac{n-1}{4n^{2}}$ and $\max_{k}\lvert z_{k}^{2}-\frac{1}{n}\rvert = \max\{|z_1^2 - \frac1n|,|z_n^2 - \frac1n|\}\geq\frac{1}{2n}$, it follows $$\begin{aligned} 2[\|{\bf z}\|_6^6 - \|{\bf z}\|_4^8] \geq \frac{n}{4}(z_{n}^{2}-z_{1}^{2})^{2}z_{1}^{4}\geq\frac{n}{4}\left(\frac{1}{2n}\right)^{2}\left(\frac{n-1}{4n^{2}}\right)^{2}=\left(\frac{n-1}{16n^{5/2}}\right)^{2}\end{aligned}$$ and finally, combining the last results, we obtain $$\lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert \geq\frac{1}{\sqrt{2}}\left[ 2\beta\cdot\frac{n-1}{16n^{5/2}}-\rho\right ] \geq \frac{\gamma}{\sqrt{2}} \rho,$$ as desired. Finally, we show that we can always find a negative curvature direction if ${\bf z}$ belongs to the set $\mathcal{R}_{3}$. Let $\gamma > 0$ be arbitrary and let $\beta \geq 2(1+\gamma)\rho n$ be given. Then, for all $\mathbf{z}\in\mathcal{R}_{3}$, there exists $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ such that $H_{f}(\mathbf{z})[\mathbf{v}]\leq-\gamma\rho$. \[lem:large-3\] Using the bound $\mathbf{v}^{T}A\mathbf{v}-\mathbf{z}^{T}A\mathbf{z}\leq\rho$ for $\mathbf{z},\mathbf{v}\in\mathbb{S}^{n-1}$ and Lemma \[lem:large-0\], it follows $$\begin{aligned} \min_{\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}}H_{f}(\mathbf{z})[\mathbf{v}]&\leq\rho+2\beta\left[\frac{3n}{n-1}(1-z_0^2) z_{0}^{2}- \|{\bf z}\|_4^4\right], $$ where $z_0^2 := \min_{k \in [n]} z_k^2$. Moreover, due to $\lVert\mathbf{z}\rVert_{4}^{4}\geq\frac{1}{n}\lVert\mathbf{z}\rVert^{4}=\frac{1}{n}$ and $\mathbf{z}\in\mathcal{R}_{3}$, we can infer $$\min_{\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}}H_{f}(\mathbf{z})[\mathbf{v}] \leq \rho + 2\beta \left[ \frac{3n}{n-1}z_0^2 - \frac{1}{n} \right] \leq \rho - \frac{\beta}{2n} \leq - \gamma\rho,$$ which finishes the proof of Lemma \[lem:large-3\]. Combining the last lemmata, we can derive the following strict-saddle property. Suppose that the coefficient $\beta$ satisfies $\beta \geq \frac{8n}{n-1}(1+\gamma)\rho n^{3/2}$ for some given $\gamma > 0$. Then, the function $f$ has the $(C_\gamma \rho,\frac{\gamma}{\sqrt{2}}\rho,C_\gamma \rho)$-strict-saddle property with $C_\gamma := \frac{4}{n-1}(1+\gamma)n^{3/2} - 1 $. \[theo:large\] As a consequence of the strict-saddle property, we can prove that each component of $\mathcal{R}_{1}$ contains exactly one local minimizer when $\beta$ is sufficiently large. In the next lemma, we first discuss uniqueness of local minimizer if the Riemannian Hessian is positive definite on a certain subset of the sphere. Let $\nu \in (0,1]$ and ${\bf z}_0 \in {\mathbb{S}^{n-1}}$ be given and let us define $\mathcal R_\nu := \{\mathbf{z}\in\mathbb{S}^{n-1}:\lVert\mathbf{z}-\mathbf{z}_{0}\rVert^{2}\leq \nu\}$. If the Riemannian Hessian ${\mathrm{Hess\!\;}}{f}$ of $f$ is positive definite on $\mathcal R_\nu$, i.e., if we have $${\bf v}^T {\mathrm{Hess\!\;}}{f({\bf z})} {\bf v} = H_f({\bf z})[{\bf v}] > 0, \quad \forall~{\bf v} \in \mathcal T_{{\bf z}}\mathcal M \backslash \{0\}, \quad \forall~{\bf z} \in \mathcal R_\nu,$$ then the problem $\min_{{\bf z} \in \mathcal R_\nu} f({\bf z})$ has at most one local minimizer. \[lem:convex\] Suppose that there exist two different local minima $\mathbf{z}_{1},\mathbf{z}_{2}$ of $f$ in the set $\mathcal{R}_\nu$. Let us consider the geodesic curve $\ell : {\mathbb{R}}\to {\mathbb{S}^{n-1}}$ on the sphere connecting $\mathbf{z}_{1}$ and $\mathbf{z}_{2}$. As shown in [@Absil2009Optimization Example 5.4.1], the curve $\ell$ can be represented as follows $$\ell(t) := {\bf z}_1 \cos(t)+{\bf v}\sin(t), \quad {\bf v}:=\frac{{\bf z}_{2}-({\bf z}_{2}^{T}{\bf z}_{1}){\bf z}_{1}}{\|{\bf z}_{2}-({\bf z}_{2}^{T}{\bf z}_{1}){\bf z}_{1}\|} \in \mathcal T_{{\bf z}_1}\mathcal M \cap {\mathbb{S}^{n-1}},$$ where $T\in(0,2\pi)$ is chosen such that $\ell(T) = {\bf z}_2$. Multiplying $\ell(T)={\bf z}_{2}$ with ${\bf v}^{T}$ from the left yields $\sin(T)={\bf v}^{T}{\bf z}_{2}\geq0$ and hence, we have $T\in(0,\pi]$. Then, for all $t \in (0,T)$, it follows $\ell(t)^T \ell^\prime(t) = 0$, $\|\ell^\prime(t)\| = 1$, and $$\begin{aligned} \lVert\ell(t)-\mathbf{z}_{0}\rVert^{2}=2-2\left(\mathbf{z}_{0}^{T}{\bf z}_1\cdot\cos(t)+\mathbf{z}_{0}^{T}{\bf v}\cdot\sin(t)\right)=2-2M\cos(t+\theta),\end{aligned}$$ where $M=\sqrt{(\mathbf{z}_{0}^{T}{\bf z}_{1})^{2}+(\mathbf{z}_{0}^{T}{\bf v})^{2}}$ and $\theta=\arccos(\mathbf{z}_{0}^{T}{\bf z}_{1})$. Since $\ell(0)$ and $\ell(T)$ are both elements of $\mathcal{R}_\nu$, we know that $$\begin{aligned} \lVert\mathbf{z}_{0}-\ell(0)\rVert^{2}\leq \nu\leq1,\quad\lVert\mathbf{z}_{0}-\ell(T)\rVert^{2}\leq \nu\leq1,\end{aligned}$$ which lead to $\cos(\theta),\cos(\theta+T)>0$. Due to $T\in(0,\pi]$, we have $\theta,\theta+T\in\left(-\frac{\pi}{2}+2k\pi,\frac{\pi}{2}+2k\pi\right)$ for some $k\in\mathbb{Z}$. The landscape of $\cos(t)$ in this range shows that the maximal value of $\lVert\mathbf{z}_{0}-\ell(t)\rVert^{2}$ is achieved by the endpoints $t=0$ or $t=T$. Hence, it holds that $\lVert\mathbf{z}_{0}-\ell(t)\rVert^{2}\leq\nu$ and $\ell(t)\in\mathcal{R}_\nu$ for all $t\in[0,T]$. The special form of the curve $\ell$ yields $\ell^{\prime\prime}(t)=-\ell(t)$ for all $t \in {\mathbb{R}}$ and thus, the second-order derivative of the continuous function $g(t):=f(\ell(t))$ is given by $$\begin{aligned} g^{\prime\prime}(t)&=\ell^{\prime\prime}(t)^{T}\nabla f(\ell(t))+\ell^{\prime}(t)^{T}\nabla^{2}f(\ell(t))\ell^\prime(t)\\ &=\ell^\prime(t)^{T}\left[\nabla^{2}f(\ell(t))-\ell(t)^{T}\nabla f(\ell(t))\cdot I\right]\ell^\prime(t)\\ &=\ell^\prime(t)^{T}{\mathrm{Hess\!\;}}{f(\ell(t))}\ell^\prime(t)>0\end{aligned}$$ for all $t\in(0,T)$, which implies that $g$ is strictly convex on $[0,T]$. Per assumption the points $\mathbf{z}_{1}$ and $\mathbf{z}_{2}$ are local minima of the problem $\min_{\mathbf{z}\in\mathcal{R}_\nu} f(\mathbf{z})$ and thus, also of the problem $\min_{t\in[0,T]}g(t)$. However, this contradicts the strict convexity of $g$. If $\beta > \rho n^{2}$, the problem has at least $2^{n}$ local minima. Furthermore, if $\beta>\frac{18n^{3}}{n-1}\rho$, then the problem has exactly $2^{n}$ local minima. \[cor:large\] Without loss of generality, we can assume $\lambda_n(A) = \lambda_n = 0$. We first prove that there exists at least one local minimizer in each component of the region $\mathcal{R}_{1}$ if $\beta\geq\rho n^{2}$. Notice that we have ${\bf z} \in \mathcal R_1$ if and only if $z_k^2 \in [\frac{1}{2n},\frac{3}{2n}]$ for all $k \in [n]$. Let $\sigma \in \{-1,+1\}^n$ be a given binary vector and let us define the sets $\mathcal B := \{\mathbf{z}\in\mathbb{S}^{n-1}: \| |{\bf z}|^2 - \frac{1}{n} {\mathds 1}\| \leq \frac1n \}$ and $\mathcal P_\sigma := \prod_{k \in [n]} \sigma_k {\mathbb{R}}_+$. We now show that for each possible choice of $\sigma$ the set $\mathcal B \cap \mathcal P_\sigma$ contains a local minimizer of $f$. Let us note that, for all ${\bf z} \in {\mathbb{S}^{n-1}}$, the condition $\||{\bf z}|^2 - \frac{1}{n}{\mathds 1}\| \leq \frac{1}{n}$ is equivalent to $\|{\bf z}\|_4^4 \leq \frac1n + \frac{1}{n^2}$. This observation can be used to establish $z_k \neq 0$ for all ${\bf z} \in \mathcal B \cap \mathcal P_\sigma$ and consequently, the sets $\mathcal B \cap \mathcal P_\sigma$ and $\mathcal B \cap \mathcal P_\nu$ do not intersect for different binary vectors $\sigma \neq \nu$. Now, for all ${\bf z} \in \mathcal D := \{{\bf z} \in {\mathbb{S}^{n-1}}: \|{\bf z}\|_4^4 = \frac1n + \frac{1}{n^2}\}$, we have $f({\bf z}) \geq \frac{\beta}{2}(n^{-1} + n^{-2})$ and setting $\mathbf{z}_{\sigma}=\sigma/\sqrt{n} \in \mathcal B \cap \mathcal P_\sigma$, we obtain $$\begin{aligned} \label{eq:loc-bound-sig} f(\mathbf{z}_{\sigma})=\frac{1}{2}\mathbf{z}_{\sigma}^{T}A\mathbf{z}_{\sigma}+\frac{\beta}{2n}\leq\frac{\rho}{2}+\frac{\beta}{2n} < \frac{\beta}{2}\left[\frac{1}{n}+\frac{1}{n^{2}}\right].\end{aligned}$$ Hence, the global minimizer ${\bf y}$ of the problem $\min_{{\bf z} \in \mathcal B \cap \mathcal P_\sigma} f({\bf z})$ satisfies $\|{\bf y}\|_4^4 < \frac1n + \frac{1}{n^2}$. This implies that the Lagrange multiplier associated with the constraint $\|{\bf z}\|_4^4 \leq \frac1n + \frac{1}{n^2}$ is zero, and the KKT conditions for ${\bf y}$ reduce to ${\mathrm{grad\!\;}}{f({\bf y})} = 0$. Due to $\mathcal B \cap \mathcal P_\sigma \subset \mathcal R_1 \cap \mathcal P_\sigma$, the Riemannian Hessian ${\mathrm{Hess\!\;}}{f({\bf y})}$ is positive definite on the tangent space $\mathcal T_{{\bf y}}\mathcal M \backslash \{0\}$ and thus, by Lemma \[lemma:son-sos\], the points [**y**]{} is one of at least $2^n$ isolated local minimum of problem . Next, we consider the case when $\beta>\frac{18n^{3}}{n-1}\rho$. We introduce the following refined versions of the $\mathcal{R}_{1}$ and $\mathcal{R}_{2}$: $$\begin{aligned} &\bar{\mathcal{R}}_{1}=\bigcup_{\sigma \in \{\pm 1\}^n} \left\{\mathbf{z}\in\mathbb{S}^{n-1}: \| {\bf z} - {\textstyle \frac{1}{\sqrt{n}}}\sigma\| \leq {\textstyle \frac{2}{9\sqrt{n}}} \right \} ,\\ &\bar{\mathcal{R}}_{2}=\bigcap_{\sigma \in \{\pm 1\}^n} \left \{\mathbf{z}\in\mathbb{S}^{n-1}: \| {\bf z} - {\textstyle \frac{1}{\sqrt{n}}}\sigma\| \geq {\textstyle \frac{2}{9\sqrt{n}}},{\min}_{k\in[n]}z_{k}^{2}\geq {\textstyle\frac{n-1}{4n^{2}}} \right\}.\end{aligned}$$ It can be easily seen that the set $\bar{\mathcal{R}}_{1}$ consists of $2^{n}$ non-intersecting components and that the three regions $\bar{\mathcal{R}}_{1}$, $\bar{\mathcal{R}}_{2}$, and $\mathcal R_3$ cover the sphere $\mathbb{S}^{n-1}$. Now, let $\mathbf{z}\in \bar{\mathcal{R}}_{1}$ be arbitrary. Then, there exists $\sigma \in \{\pm 1\}^n$ such that $$\begin{aligned} \left\lvert z_{k}-\frac{\sigma_{k}}{\sqrt{n}}\right\rvert\leq\frac{2}{9\sqrt{n}} \quad \text{and} \quad \left| z_k + \frac{\sigma_k}{\sqrt{n}}\right| \leq \frac{20}{9\sqrt{n}}, \quad\forall~k\in[n].\end{aligned}$$ Hence, it follows $\||{\bf z}|^2 - \frac{1}{n} {\mathds 1}\|_\infty \leq \frac{40}{81n} < \frac{1}{2n}$, which implies $\bar{\mathcal{R}}_{1} \subset \mathcal{R}_{1}$. Thus, the strong convexity property also holds on $\bar{\mathcal{R}}_{1}$. We now prove the Riemannian gradient is lower bounded on $\bar{\mathcal{R}}_{2}$. For every $\mathbf{z}\in\bar{\mathcal{R}_{2}}$, there exists $\sigma \in \{\pm1\}^n$ such that $\sigma_k z_k = |z_k|$ for all $k \in [n]$. Consequently, we obtain $$\begin{aligned} \||{\bf z}|^2 - {n}^{-1}{\mathds 1}\|_\infty = \max_{k \in [n]} \left| z_k - \frac{\sigma_k}{\sqrt{n}} \right| \left| |z_k| + \frac{1}{\sqrt{n}} \right| \geq \frac{\|{\bf z} - \frac{\sigma}{\sqrt{n}} \|_\infty}{\sqrt{n}} \geq \frac{2}{9n\sqrt{n}} \end{aligned}$$ and by mimicking the steps and estimates in the proof Lemma \[lem:large-2\], we get $$\begin{aligned} 2[\|{\bf z}\|_6^6 - \|{\bf z}\|_4^8] \geq\frac{n}{4}\left(\frac{2}{9n\sqrt{n}}\right)^{2}\left(\frac{n-1}{4n^{2}}\right)^{2}=\left(\frac{n-1}{36n^{3}}\right)^{2}. $$ Thus, the norm of the Riemannian gradient is lower bounded by $$\begin{aligned} \lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert \geq\frac{1}{\sqrt{2}}\left[2\beta\sqrt{2[\|{\bf z}\|_6^6 - \|{\bf z}\|_4^8]}-\rho\right] \geq\frac{1}{\sqrt{2}}\left(\beta\cdot\frac{n-1}{18n^{3}}-\rho\right)>0\end{aligned}$$ and all local minima should need to be located in the set $\bar{\mathcal{R}}_{1}$. Applying Lemma \[lem:convex\], each connected component of $\bar{\mathcal{R}}_{1}$ contains at most one local minimizer and hence, problem has exactly $2^{n}$ local minima. Although the local minima of problem are not unique, all the local minima have similar objective function values. Suppose that $\beta > \frac{18n^3}{n-1}\rho$. Then, it follows $$\label{eq:loc-glob-diff} f(\mathbf{y})- \min_{{\bf z} \in {\mathbb{S}^{n-1}}} f(\mathbf{z}) \leq \frac{1}{18{n}} \cdot \min_{{\bf z} \in {\mathbb{S}^{n-1}}}f(\mathbf{z}),$$ for all local minimizer ${\bf y} \in {\mathbb{S}^{n-1}}$ of problem . Without loss of generality, we can assume that the smallest eigenvalue of the matrix $A$ is zero. Then, for all ${\bf z} \in {\mathbb{S}^{n-1}}$, we have $f({\bf z}) \geq \frac{\beta}{2} \|{\bf z}\|_4^4 \geq \frac{\beta}{2n}$. According to the analysis in Corollary \[cor:large\], each component ${\mathcal B} \cap \mathcal P_\sigma$, $\sigma \in \{\pm1\}^n$, contains exactly one local minimizer ${\bf y}$ of problem which is also the unique global minimizer of the restricted problem $\min_{{\bf z} \in \mathcal B \cap \mathcal P_\sigma} f({\bf z})$. Together with , this yields $$ f({\bf y}) \leq f({\bf z}_\sigma) \leq \frac{\rho}{2} + \frac{\beta}{2n} \leq \left[ 1 + \frac{n\rho}{\beta} \right] f({\bf z}), \quad \forall~{\bf z} \in {\mathbb{S}^{n-1}}.$$ Finally, the estimate can be established via minimizing the latter expression with respect to ${\bf z}$ and using the bound on $\beta$. In the remainder of this section, we present an example demonstrating that the bound $\beta \approx C \rho n^{3/2}$ and the dependence on $n^{3/2}$ can not be further improved and that the strict-saddle property is violated whenever a smaller coefficient is chosen. Let $C>0$ and $\epsilon > 0$ be given constants and suppose $\beta = C n^{3/2-\epsilon}$. In the following, we construct a specific matrix $A \in {\mathbb{R}}^{n \times n}$ and a point ${\bf z} \in \mathbb{S}^{n-1}$ such that the three conditions of the strict-saddle property do not hold at ${\bf z}$. We set $$z_1 = \left[\frac{1}{3 n- 2}\right]^\frac12, \quad z_k = \left[\frac{3}{3 n - 2}\right]^\frac12, \quad k \geq 2,$$ ${\bf u} = 2\beta \|{\bf z}\|_4^4 \cdot {\bf z} - 2\beta \operatorname*{diag}(|{\bf z}|^2) {\bf z}$, and $A = \alpha {\bf w}{\bf w}^T + {\bf z}{\bf u}^T + {\bf u}{\bf z}^T$, where $$w_1 = - \left[ \frac{3n-3}{3n-2} \right]^\frac12, \quad w_k = \left[ \frac{1}{(3n-2)(n-1)} \right]^\frac12, \, k \geq 2, \quad \alpha = - \frac{16\beta}{(3n-2)^2}.$$ Then, we have $\|{\bf z}\| = \|{\bf w}\| = 1$, $\|{\bf z}\|_4^4 = (9n-8) / (3n-2)^2$, and $$\begin{aligned}  \|{\bf u}\|^2 & = 4\beta^2 [\|{\bf z}\|^6_6 - \|{\bf z}\|_4^8] = 4\beta^2 \left[ \frac{1+27(n-1)}{(3 n-2)^{3}}-\|{\bf z}\|_4^8 \right] = \frac{48(n-1)}{(3 n-2)^{4}} \cdot \beta^2. \end{aligned}$$ Furthermore, it holds that ${\bf u}^T {\bf z} = {\bf w}^T {\bf z}= 0$ which implies $A{\bf z} = {\bf u}$ and ${\mathrm{grad\!\;}}{f({\bf z})} = 0$. The eigenvalues of the matrices $\alpha {\bf w}{\bf w}^T$ and ${\bf z}{\bf u}^T + {\bf u}{\bf z}^T$ are $0$, ..., $0$, $\alpha$ and $\|{\bf u}\|$, $0$, ..., $0$, $- \|{\bf u}\|$, respectively. Thus, by Weyl’s inequality, it follows $$\rho = \lambda_1(A) - \lambda_n(A) \leq 2 \|{\bf u}\| - \alpha = 8\beta \cdot \frac{2 + \sqrt{3n-3}}{(3n-2)^2} \leq \bar C n^{-\epsilon} = O(n^{-\epsilon}),$$ for some constant $\bar C$. Let us now consider an arbitrary vector ${\bf v} \in \mathcal T_{\bf z}\mathcal M \cap {\mathbb{S}^{n-1}}$. We have $\sum_{k=2}^n v_k^2 = 1 - v_1^2$, $v_1 + \sqrt{3} \sum_{k=2}^n v_k = 0$, and $${\bf v}^T {\bf w} = v_1 w_1 - \frac{1}{\sqrt{3}} \left[ \frac{1}{(3n-2)(n-1)} \right]^\frac12 v_1 = -\left[ \frac{3n-2}{3n-3} \right]^{\frac12} \cdot v_1.$$ Hence, we obtain $$\begin{aligned} H_f({\bf z})[{\bf v}] & = \alpha ({\bf v}^T{\bf w})^2 + 6\beta {\bf v}^T \operatorname*{diag}(|{\bf z}|^2) {\bf v} - 2\beta \|{\bf z}\|_4^4 \\ & = \frac{3n-2}{3n-3} \alpha v_1^2 - \frac{12\beta}{3n-2} v_1^2 + \frac{18\beta}{3n-2} - 2\beta \|{\bf z}\|_4^4. \end{aligned}$$ Since $H_f({\bf z})[{\bf v}]$ is linear in $v_1^2$, its maximum and minimum value are attained at the boundary of the range interval $[\underline{\nu},\overline{\nu}]$ of $v_1^2$. Notice that we have $\underline{\nu} = 0$ and $\overline{\nu}$ can be found by discussing the optimization problem $$\min_{\bf v}~-v_1 \quad {\mathrm{s.\,t.}}\quad v_1 + \sqrt{3} \sum_{k=2}^n v_k = 0, \quad \|{\bf v}\| = 1$$ and its associated KKT conditions. In particular, it can be shown that $\overline{\nu} = (3n-3)/(3n-2)$. In the case $v_1^2 = 0$, we obtain $H_f({\bf z})[{\bf v}] > 0$ and for $v_1^2 = \overline{\nu}$, we have $$H_f({\bf z})[{\bf v}] = \alpha - \frac{36\beta (n-1)}{(3n-2)^2} + \frac{18\beta}{3n-2} - \frac{2\beta(9n-8)}{(3n-2)^2} = \alpha + \frac{16\beta}{(3n-2)^2} = 0.$$ Consequently, we can infer $\min_{{\bf v} \in \mathcal T_{\bf z}\mathcal M \cap {\mathbb{S}^{n-1}}} H_f({\bf z})[{\bf v}] = 0$ and $H_f({\bf z})[{\bf v}] \geq 0$ for all ${\bf v} \in \mathcal T_{\bf z}\mathcal M \cap {\mathbb{S}^{n-1}}$, which shows that none of the conditions of the strict-saddle property hold at ${\bf z}$. Thus, the order $n^{3/2}$ can not be improved in the deterministic case. -- -- -- -- Small interaction coefficient {#sec:small} ----------------------------- In the following, we discuss the geometric landscape of problem (\[eqn:obj\]) for small interaction coefficients. We additionally assume that there is a positive spectral gap $\delta=\lambda_{n-1}-\lambda_{n}>0$ between the two smallest eigenvalues of the matrix $A$. As shown by Marvcenko [@marvcenko1967distribution], this condition holds with probability $1$ when $A$ is a Gaussian random matrix. Let us recall that the eigenvalue decomposition of $A$ is given by $A = P \Lambda P^T$, where $\Lambda = \operatorname*{diag}(\lambda_1,\lambda_2,...,\lambda_n)$ and $P = (\mathbf{p}_{1},\mathbf{p}_{2},...,\mathbf{p}_{n})$ is an orthogonal matrix. Similar to the method used in section \[sec:large\], we now divide $\mathbb{S}^{n-1}$ into three sub-regions: - (Strong convexity) $\mathcal{R}_{1}: =\{\mathbf{z} \in \mathbb{S}^{n-1}: {\bf z} =\sum_{k=1}^n \alpha_{k}\mathbf{p}_{k}, \, \alpha_{n}^{2}\geq\frac{(2+\gamma)\beta+\rho}{\delta+\rho}\}$. - (Large gradient) $\mathcal{R}_{2} :=\{\mathbf{z} \in\mathbb{S}^{n-1}: \|A{\bf z}\|^2 - ({\bf z}^T A{\bf z})^2 \geq (\frac23+\gamma)^2\beta^{2}\}$. - (Negative curvature) $\mathcal{R}_{3} := \{\mathbf{z} \in\mathbb{S}^{n-1}: {\bf z}=\sum_{k = 1}^n \alpha_{k}\mathbf{p}_{k}, \, \alpha_{n}^{2}\leq\frac{\delta-(4+\gamma)\beta}{\delta+\rho}\}$. An exemplary illustration of the sets $\mathcal R_1$–$\mathcal R_3$ is given in Figure \[figure:areas\]. We first show that the Riemannian Hessian is uniformly positive definite on $\mathcal{R}_{1}$. Suppose that the gap between the two smallest eigenvalues of the matrix $A$ satisfies $\delta:=\lambda_{n-1}-\lambda_{n}>0$ and let $\beta, \gamma > 0$ be given. Then, for all $\mathbf{z}\in\mathcal{R}_{1}$ and all $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$, it follows $H_{f}(\mathbf{z})[\mathbf{v}]\geq\gamma\beta$. \[lem:small-1\] Let ${\bf v} \in \mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ be arbitrary with $\mathbf{v}=\sum_{k\in[n]} \nu_{k}\mathbf{p}_{k}$. Then, we have $\sum_{k\in[n]}\alpha_{k} \nu_{k}=0$ and $\sum_{k\in[n]}\nu_{k}^{2}=1$ and the Cauchy inequality implies $$\begin{aligned} \vert \nu_{n}\alpha_{n}\rvert&=\left\lvert{\sum}_{k=1}^{n-1}\nu_{k}\alpha_{k}\right\rvert\leq \left[{\sum}_{k=1}^{n-1}\nu_{k}^{2}\right]^\frac12\left[{\sum}_{k=1}^{n-1}\alpha_{k}^{2}\right]^\frac12 = \sqrt{(1-\nu_{n}^{2})(1-\alpha_{n}^{2})}.\end{aligned}$$ Thus, it follows $\nu_{n}^{2}\alpha_{n}^{2}\leq(1-\nu_{n}^{2})(1-\alpha_{n}^{2})$ and $\alpha_{n}^{2}\leq1-\nu_{n}^{2}$ and, due to $\|{\bf z}\|_4^4 \leq 1$, we obtain $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}]&=\mathbf{v}^{T}A\mathbf{v}-\mathbf{z}^{T}A\mathbf{z}+6\beta{\sum}_{k=1}^n z_{k}^{2}v_{k}^{2} - 2\beta \|{\bf z}\|^4_4 \\ &\geq {\sum}_{k=1}^n \lambda_{k}(\nu_{k}^{2}-\alpha_{k}^{2})-2\beta \\ &\geq{\sum}_{k=1}^{n-1}(\lambda_{n}+\delta)\nu_{k}^{2}+\lambda_{n}\nu_{n}^{2}-{\sum}_{k=1}^{n-1}(\lambda_{n}+\rho)\alpha_{k}^{2}-\lambda_{n}\alpha_{n}^{2}-2\beta\\ &=(1-\nu_{n}^{2})\delta-(1-\alpha_{n}^{2})\rho-2\beta \\ & \geq \alpha_{n}^{2}\delta-(1-\alpha_{n}^{2})\rho-2\beta = \alpha_{n}^{2}(\delta+\rho)-\rho-2\beta \geq \gamma\beta,\end{aligned}$$ as desired. Next, we prove that the norm of the Riemannian gradient is strictly larger than zero on the set $\mathcal{R}_{2}$. For all $\mathbf{z}\in\mathcal{R}_{2}$, it holds that $\lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert_{2}\geq \gamma\beta$ . \[lem:small-2\] As in the proof of Lemma \[lem:large-2\], we have $$\begin{aligned} \left\lVert\left(\operatorname*{diag}(|{\bf z}|^2)- \|{\bf z}\|_4^4 \cdot I_{n}\right)\mathbf{z}\right\rVert ^{2} = \|{\bf z}\|_6^6 - \|{\bf z}\|_4^8 \leq \|{\bf z}\|_4^6 (1 - \|{\bf z}\|_4^2) \leq \frac14 \left[ \frac34\right]^3,\end{aligned}$$ where the last estimate follows from the fact that the mapping $x \mapsto x^6-x^8$ attains its global maximum at $x = \pm \frac{\sqrt{3}}{2}$. Hence, we obtain $$\begin{aligned} \lVert{\mathrm{grad\!\;}}{f(\mathbf{z})}\rVert_{2}&=\left\lVert(A-\mathbf{z}^{T}A\mathbf{z}\cdot I_{n})\mathbf{z}+2\beta\left(\operatorname*{diag}(|{\bf z}|^2)- \|{\bf z}\|_4^4 \cdot I_{n}\right)\cdot\mathbf{z}\right\rVert\\ &\geq\lVert(A-\mathbf{z}^{T}A\mathbf{z}\cdot I_{n})\mathbf{z}\rVert-2\beta\left\lVert\left(\operatorname*{diag}(|{\bf z}|^2)- \|{\bf z}\|_4^4 \cdot I_{n}\right)\mathbf{z}\right\rVert \\ &\geq\lVert(A-\mathbf{z}^{T}A\mathbf{z}\cdot I_{n})\mathbf{z}\rVert -\frac{3\sqrt{3}}{8} \beta= \left[ \frac23 - \frac{3\sqrt{3}}{8} \right ] \beta + \gamma\beta \geq \gamma\beta. $$ Finally, for points in the region $\mathcal{R}_{3}$, we construct directions along which the curvature of the objective function is strictly negative. Suppose that the gap between the two smallest eigenvalues of the matrix $A$ satisfies $\delta:=\lambda_{n-1}-\lambda_{n}>0$ and let $\gamma > 0$ be given. If $\beta \leq (4+\gamma)^{-1}\delta$, then for all $\mathbf{z}\in\mathcal{R}_{3}$ there exists $\mathbf{v}\in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$ such that $H_{f}(\mathbf{z})[\mathbf{v}]\leq-\gamma\beta$. \[lem:small-3\] By the Cauchy’s inequality and using the estimate $|3x-x^{2}| \leq 2$, $x \in [0,1]$, we have $$\begin{aligned} \sum_{k\in[n]} 3z_{k}^{2}v_{k}^{2}-z_{k}^{4} \leq 3\lVert\mathbf{z}\rVert_{4}^{2}\lVert\mathbf{v}\rVert_{4}^{2}-\lVert\mathbf{z}\rVert_{4}^{4}\leq3\lVert\mathbf{z}\rVert_{4}^{2}-\lVert\mathbf{z}\rVert_{4}^{4}\leq2. $$ Next, we choose a specific direction $\mathbf{v}=\sum_{k\in[n]} \nu_{k}\mathbf{p}_{k}$ that satisfies $H_{f}(\mathbf{z})[\mathbf{v}]\leq- \gamma\beta$, ${\bf z}^T {\bf v} = \sum_{k\in[n]}\alpha_{k}\nu_{k}=0$ and $\|{\bf v}\|^2 = \sum_{k \in [n]} \nu_k^2 = 1$. *Case 1. $\alpha_{n}=0$. Let us set $\nu_{n}=1$ and $\nu_{k}=0$ for all $ k\in [n-1]$.* Then, we obtain $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}]&=\mathbf{v}^{T}A\mathbf{v}-\mathbf{z}^{T}A\mathbf{z}+2\beta{\sum}_{k=1}^n(3z_{k}^{2}v_{k}^{2}-z_{k}^{4}) \\ & \leq {\sum}_{k = 1}^n \lambda_{k}(\nu_{k}^{2}-\alpha_{k}^{2})+4\beta=\lambda_{n}-{\sum}_{k=1}^{n-1}\lambda_{k}\alpha_{k}^{2}+4\beta\\ &\leq\lambda_{n}- (\lambda_n + \delta ){\sum}_{k=1}^{n-1} \alpha_{k}^{2}+4\beta=-\delta+4\beta\leq- \gamma \beta, $$ where the last inequality immediately follows from $\beta \leq \frac{\delta}{4+\gamma}$. *Case 2. $0< \alpha_{n}^{2}\leq\frac{\delta-(4+\gamma)\beta}{\delta+\rho}$.* In this situation, we set $$\begin{aligned} \nu_{k} = \eta \alpha_{k},\quad \forall~k \in [n-1],\quad \nu_{n}=- \sqrt{1-\alpha_{n}^{2}}, \quad \eta = \frac{\alpha_{n}}{\sqrt{1-\alpha_{n}^{2}}}. $$ With this choice, we have ${\bf z}^T{\bf v} = \eta(1-\alpha_n^2) - \alpha_n \sqrt{1-\alpha_n^2} = 0$ and $\|{\bf v}\|^2 = (\eta^2 + 1)(1-\alpha_n^2) = 1$, i.e., it holds that $\mathbf{v} \in\mathcal{T}_{\mathbf{z}}\mathcal{M}\cap\mathbb{S}^{n-1}$. Similar to the calculations in the proof of Lemma \[lem:small-1\], we now get $$\begin{aligned} H_{f}(\mathbf{z})[\mathbf{v}]&=\mathbf{v}^{T}A\mathbf{v}-\mathbf{z}^{T}A\mathbf{z}+2\beta{\sum}_{k=1}^{n}(3z_{k}^{2}v_{k}^{2}-z_{k}^{4})\\ &\leq {\sum}_{k=1}^{n}\lambda_{k}(\nu_{k}^{2}-\alpha_{k}^{2})+4\beta\\ &\leq {\sum}_{k=1}^{n-1}((\lambda_{n}+\rho)\nu_{k}^{2}-(\lambda_{n}+\delta)\alpha_{k}^{2})+\lambda_{n}(\nu_{n}^{2}-\alpha_{n}^{2})+4\beta\\ &=\rho \cdot {\sum}_{k=1}^{n-1} \nu_{k}^{2}-\delta \cdot {\sum}_{k=1}^{n-1} \alpha_{k}^{2}+4\beta=\rho(1-\nu_{n}^{2})-\delta(1-\alpha_{n}^{2})+4\beta\\ &=\rho\eta^{2}(1-\alpha_{n}^{2})+\delta \alpha_{n}^{2}-\delta+4\beta=(\delta+\rho)\alpha_{n}^{2}-\delta+4\beta\leq-\gamma\beta. $$ Combining the latter two cases, we can conclude the proof of Lemma \[lem:small-3\]. We now verify that $f$ has the strict-saddle property whenever $\beta$ is chosen sufficiently small. Suppose that the gap between the two smallest eigenvalues of the matrix $A$ satisfies $\delta:=\lambda_{n-1}-\lambda_{n}>0$ and let $\gamma > 0$ be given. If $\beta \leq [2(\frac73+\gamma) + (\frac23 + \gamma)\frac{\rho}{\delta}]^{-1}\delta =: b_\gamma$, then $f$ has the $(\gamma\beta,\gamma\beta,\gamma\beta)$-strict-saddle property. \[theo:small\] By Lemmata \[lem:small-1\]-\[lem:small-3\], we know that the function $f$ satisfies the strong convexity, large gradient, and negative curvature property on the three different set $\mathcal R_1$, $\mathcal R_2$, and $\mathcal R_3$, respectively. To finish the proof, we need to show that those regions actually cover the whole sphere $\mathbb{S}^{n-1}$. Combining these observations, we can then conclude that $f$ has the $(\gamma\beta,\gamma\beta,\gamma\beta)$-strict-saddle property. In order to prove $\mathcal{R}_{1}\cup\mathcal{R}_{2}\cup\mathcal{R}_{3}=\mathbb{S}^{n-1}$, we only need to verify that for all $\mathbf{z}=\sum_{k\in[n]} \alpha_{k}\mathbf{p}_{k}\in\mathbb{S}^{n-1}$ with $$\label{eq:b-alpha} \frac{\delta-(4+\gamma)\beta}{\delta +\rho}\leq \alpha_{n}^{2}\leq\frac{(2+\gamma)\beta+\rho}{\delta+\rho},$$ we have $\|A{\bf z}\|^2 - ({\bf z}^T A {\bf z})^2 \geq (\frac{2}{3} + \gamma)^2\beta^{2}$. Using the bounds , it follows $$\begin{aligned} \|A{\bf z}\|^2 - ({\bf z}^T A {\bf z})^2 & = \frac{1}{2} \sum_{k,j\in[n]} \alpha_{k}^{2} \alpha_{j}^{2}(\lambda_{k}-\lambda_{j})^{2} \\ & \geq \sum_{k\in[n-1]} \alpha_{k}^{2}\alpha_{n}^{2}(\lambda_{k}-\lambda_{n})^{2} \geq\sum_{k\in[n-1]}\alpha_{k}^{2}\alpha_{n}^{2}\delta^{2} \\ & =(1-\alpha_{n}^{2})\alpha_{n}^{2}\delta^{2} \geq\frac{(\delta-(4+\gamma)\beta)(\delta-(2+\gamma)\beta)}{(\delta+\rho)^{2}}\delta^{2}.\end{aligned}$$ where the first identity was established in the proof of Lemma \[lem:large-2\]. Rearranging the terms in the latter estimate, we see that our claim is satisfied if $$\ell(\beta) := \left[ ({\textstyle\frac23} + \gamma)^2(\delta+\rho)^2 - (4+\gamma)(2+\gamma) \right] \beta^2 + 2(3+\gamma) \delta^3 \beta - \delta^4 \leq 0$$ for all $\beta \leq b_\gamma$. Since the unique nonnegative zero of the quadratic polynomial $\ell$ is given by $$\bar \beta = \left[{3+\gamma + \sqrt{ 1+ ({\textstyle\frac23} + \gamma)^2(1+{\textstyle\frac{\rho}{\delta}})^2}} \right]^{-1} \delta,$$ we can finish the proof by noticing $\bar \beta \geq b_\gamma$. Finally, as a counterpart of Corollary \[cor:large\], we can establish the uniqueness of local minima as a consequence of the strict-saddle property. Under the conditions of Theorem \[theo:small\], the problem has exactly two local minima which are also global minima. Note that all local minima locate in $\mathcal{R}_{1}$ and that $\mathcal{R}_{1}$ consists of two symmetrical non-intersecting subsets. We now consider one of the subsets $\bar{\mathcal{R}}_{1}:=\{\mathbf{z}\in{\mathbb{S}^{n-1}}:\mathbf{z}=\sum_{k\in[n]}\alpha_{k}\mathbf{p}_{k},\alpha_{n}\geq\sqrt{\nu}\}$ where $ \nu := ({\delta+\rho})^{-1}[(2+\gamma)\beta+\rho]$. Using $\|{\bf z}-\mathbf{p}_{n}\|^{2}=2-2\alpha_{n}$, for ${\bf z} \in \bar {\mathcal R}_1$, an equivalent definition of this subset is given by $$\begin{aligned} \bar{\mathcal{R}}_{1}=\left\{\mathbf{z}\in{\mathbb{S}^{n-1}}:\|\mathbf{z}-\mathbf{p}_{n}\|^{2}\leq2-2\sqrt{\nu}\right\}.\end{aligned}$$ In order to apply Lemma \[lem:convex\], we need to verify $2-2\sqrt{\nu}\leq1$ or, equivalently, $\nu \geq \frac{1}{4}$. However, due to $\gamma,\beta>0$ and $\rho\geq\delta>0$, we have $\nu \geq\frac{\rho}{\delta+\rho}\geq\frac{1}{2}>\frac{1}{4}$. Similary, the second subset can be represented by $\bar{\mathcal R}_2 := \{{\bf z} \in {\mathbb{S}^{n-1}}: \|{\bf z} + {\bf p}_n\|^2 \leq 2 - 2\sqrt{\nu}\}$. Then, by Lemma \[lem:convex\], there exist exactly two equivalent local minima which are also global minima of problem . Estimation of the Kurdyka-Łojasiewicz Exponent {#sec:KLExp} ============================================== In this section, we estimate the Kurdyka-Łojasiewicz (KL) exponent of problem . Specifically, we want to find the largest $\theta\in(0,\frac{1}{2}]$ such that for all stationary points $\mathbf{z}$ of problem (\[eqn:obj\]), the Łojasiewicz inequality, $$\label{eq:kl} \lvert f(\mathbf{y})-f(\mathbf{z})\rvert^{1-\theta}\leq\eta_{\mathbf{z}}\lVert{\mathrm{grad\!\;}}{f(\mathbf{y})}\rVert,\quad\forall~\mathbf{y}\in B(\mathbf{z},\delta_{\mathbf{z}})\cap\mathbb{CS}^{n-1},$$ holds with some constants $\delta_{\mathbf{z}},\eta_{\mathbf{z}}>0$. The largest possible $\theta$ is called the KL exponent of problem . As already mentioned, the [Ł]{}ojasiewicz inequality plays a fundamental role in nonconvex optimization and is frequently utilized to analyze the local convergence properties of nonconvex optimization methods, [@attouch2009convergence; @AttBolRedSou10; @AttBolSva13; @bolte2014proximal; @OchCheBroPoc14; @BonLorPorPraReb17; @liu2017quadratic; @li2018calculus]. In [@attouch2009convergence], Attouch and Bolte derived an abstract KL-framework based on the Łojasiewicz inequality that allows to establish global convergence and local convergence rates for general optimization approaches satisfying certain function reduction and asymptotic step size safe-guard conditions. In particular, if $\theta\geq\frac{1}{2}$, then the corresponding iterates can be shown to converge linearly. Otherwise, if $\theta\in(0,\frac{1}{2})$, the iterates converge at a sublinear rate $O(t^{-\frac{\theta}{1-2\theta}})$. By introducing the auxiliary problem $$\begin{aligned} \min_{{\bf z} \in {\mathbb{C}^n}}~\hat f({\bf z}), \quad \hat{f}(\mathbf{z}) := \begin{cases} f(\mathbf{z}) & \text{if } \mathbf{z}\in{\mathbb{CS}^{n-1}},\\ +\infty & \text{otherwise},\\ \end{cases} \label{eq:nonsmooth}\end{aligned}$$ the original problem can be treated as the minimization of an extended real-valued, proper, and lower semicontinuous function which allows to apply existing results and the rich KL theory for nonsmooth problems, see, e.g., [@lojasiewicz1963propriete; @Kur98; @BolDanLew-MS-06; @BolDanLew06; @BolDanLewShi07]. In the nonsmooth setting, the Riemannian gradient, appearing in , is typically substituted by the nonsmooth slope of $\hat f$ which is based on the Fréchet and limiting subdifferential of $\hat f$. In our case, if problem is restricted to the real space ${\mathbb{R}^n}$, the limiting subdifferential and nonsmooth slope of $\hat{f}$ at $\mathbf{z}\in{\mathbb{S}^{n-1}}$ can be expressed as $\partial\hat{f}(\mathbf{z})= \{\nabla f(\mathbf{z})+\mu\mathbf{z}:\mu\in\mathbb{R}\}$ and $$\operatorname*{arg\,min}_{\mathbf{g}\in\partial\hat{f}(\mathbf{z})}\|\mathbf{g}\|= (I - {\bf z}{\bf z}^T)\nabla f({\bf z}) = {\mathrm{grad\!\;}}{f(\mathbf{z})}.$$ Hence, the Riemannian-type [Ł]{}ojasiewicz inequality coincides with the standard notion and KL framework used in nonsmooth optimization. Our goal is now to show that the KL exponent of is at least $\frac{1}{4}$ under suitable conditions. Throughout this section, we assume that $\mathbf{z}\in\mathbb{CS}^{n-1}$ is a stationary point of problem . Furthermore, $\mathbf{y}\in\mathbb{CS}^{n-1}$ denotes a neighboring point of $\mathbf{z}$ and we set $\Delta=\mathbf{y}-\mathbf{z}$. We now collect and present some preparatory notations and computational results that will be used in the following derivations. Since $\mathbf{z}$ is a stationary point of problem (\[eqn:obj\]), we have $$\begin{aligned} H\mathbf{z} = A\mathbf{z}+2 \beta\operatorname*{diag}(\lvert\mathbf{z}\rvert^{2})\mathbf{z}-2\lambda\mathbf{z} = 0, \quad 2\lambda = {\bf z}^*A{\bf z} + 2\beta\|{\bf z}\|_4^4 \label{eqn:6-1}\end{aligned}$$ and as proved in , it holds that $$\begin{aligned} f(\mathbf{y})-f(\mathbf{z})=\frac{1}{2}\mathbf{y}^{*}H\mathbf{y}+\frac{\beta}{2} \|\tau\|^2, \quad \tau_k := \lvert y_{k}\rvert^{2}-\lvert z_{k}\rvert^{2}, \quad k \in [n]. \end{aligned} \label{eqn:6-2}$$ The norm of the Riemannian gradient can be expressed as follows $$\begin{aligned} \nonumber \|{\mathrm{grad\!\;}}{f({\bf y})}\|^2 & = {\frac{1}{2}}\| P^\bot_{\bf y}[H + 2\beta \operatorname*{diag}(\tau)]{\bf y}\|^2 \\\nonumber & = {\frac{1}{2}}\|[H + 2\beta \operatorname*{diag}(\tau)] {\bf y}\|^2 - {\frac{1}{2}}({\bf y}^*[H + 2\beta \operatorname*{diag}(\tau)]{\bf y})^2 \\ \nonumber &=\frac{1}{2}\mathbf{y}^{*}H^{2}\mathbf{y}-\frac{1}{2}(\mathbf{y}^{*}H\mathbf{y})^{2}+ 2\beta^{2} \mathbf{y}^{*}\operatorname*{diag}(|\tau|^2)\mathbf{y}-2 \beta^{2}(\mathbf{y}^{*}\operatorname*{diag}(\tau)\mathbf{y})^{2}\\ &\hspace{4ex}+2\beta\mathbf{y}^{*}H\operatorname*{diag}(\tau)\mathbf{y}-2\beta(\mathbf{y}^{*}H\mathbf{y})(\mathbf{y}^{*}\operatorname*{diag}(\tau)\mathbf{y}), \label{eqn:6-4}\end{aligned}$$ where $P_{\mathbf{y}}^{\perp}=I_{n}-\mathbf{yy}^{*}$ is the orthogonal projection onto the space $[\mathrm{span}\,\{\mathbf{y}\}]^\bot$ Finally, let us define the index sets $$\begin{aligned} \mathcal{A}=\{k:z_{k}=0\},\quad\mathcal{I}=\{k:z_{k}\neq0\}\end{aligned}$$ and $r_{+}=\min_{k\in\mathcal{I}}r_{k}>0$. Notice that we have $\tau_{k}=t_{k}^{2}\geq0$ for all $k\in\mathcal{A}$. We first show that the Łojasiewicz inequality holds with $\theta=\frac{1}{4}$ at those stationary points where $H=0$ (we also refer to the remark after this lemma). \[lemma:kl-slim\] Suppose $\mathbf{z}$ is an arbitrary point on $\mathbb{CS}^{n-1}$. Then, the inequality $$\begin{aligned} \lVert\tau\rVert^{\frac{3}{2}}\leq\eta_{\mathbf{z}}\lVert P_{\mathbf{y}}^{\perp}\operatorname*{diag}(\tau)\mathbf{y}\rVert,\quad\forall~\mathbf{y}\in B(\mathbf{z},\delta_{\mathbf{z}})\cap\mathbb{CS}^{n-1} \label{eqn:6-5}\end{aligned}$$ holds for some constants $\eta_{\mathbf{z}},\delta_{\mathbf{z}}>0$. \[lem:6-2\] In the case $\tau=0$, we have $\lVert\tau\rVert = \lVert P_{\mathbf{y}}^{\perp}\operatorname*{diag}(\tau)\mathbf{y}\rVert = 0$ and consequently, the inequality holds trivially with $\theta=\frac{1}{2}$. Next, let us assume $\mathcal{A}\neq\emptyset$ and let us introduce the polar coordinates $z_k = r_k e^{i \theta_k}$, $y_k = t_k e^{i \phi_k}$ for $r_k, t_k \geq 0$, $\theta_k, \phi_k \in [0,2\pi]$, and all $k \in [n]$. A straightforward calculation yields $|\Delta_k|^2 = t_k^2 - 2 r_k t_k \cos(\theta_k - \phi_k) + r_k^2 \geq (t_k -r_k)^2$ and hence, it follows $$|\tau_k| = (t_k + r_k) |t_k - r_k| \leq (2r_k + |\Delta_k|)|\Delta_k|, \quad \forall~k \in {\mathcal{I}},$$ and $\tau_k = t_k^2 = |\Delta_k|^2$ for all $k \in {\mathcal{A}}$ . Using $\sum_{k \in \mathcal I} \tau_k = - \sum_{k \in \mathcal A} \tau_k = - \|\tau_{\mathcal{A}}\|_1$, the estimates $\|\tau_{\mathcal{A}}\|_1^2 \leq |{\mathcal{A}}|\|\tau_{\mathcal{A}}\|^2$ and $$\begin{aligned} \nonumber \frac12 \sum_{k,j \in \mathcal I}(\tau_k- \tau_j)^2 & = |\mathcal I| \sum_{k \in \mathcal I} \tau_k^2 - \left[ {\sum}_{k \in \mathcal I} \tau_k \right]^2 \\ \label{eq:diff-norm}& = |\mathcal I| \|\tau_{\mathcal{I}}\|^2 - \left[ {\sum}_{k \in \mathcal A} \tau_k \right]^2 = |\mathcal I| \|\tau_{\mathcal{I}}\|^2 - \|\tau_{\mathcal{A}}\|_1^2, \end{aligned}$$ and setting $m := |\mathcal I | \geq 1$, we obtain $$\|\tau\|^2 \leq \frac{1}{2m} \sum_{k,j \in \mathcal I}(\tau_k - \tau_j)^2 + \frac{n}{m} \|\tau_{\mathcal{A}}\|^2 \leq \frac{n}{m} \left [ {\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2 + \|\tau_{\mathcal{A}}\|^2 \right].$$ Furthermore, it holds that $$\begin{aligned} 2 \| P_{\bf y}^\bot\operatorname*{diag}(\tau){\bf y}\|^2 & = 2 {\bf y}^* \operatorname*{diag}(|\tau|^2){\bf y} - 2 ({\bf y}^*\operatorname*{diag}(\tau){\bf y})^2 \\ &= 2 \left[ \sum t_k^2 \tau_k^2 - \left( \sum t_k^2 \tau_i \right)^2 \right] = \sum_{k,j = 1}^n t_k^2 t_j^2 (\tau_k - \tau_j)^2 \\ & = \sum_{k,j \in \mathcal A} \tau_k \tau_j (\tau_k - \tau_j)^2 + \sum_{k,j \in \mathcal I} t_k^2 t_j^2 (\tau_k - \tau_j)^2 \\ & \hspace{4ex} + 2 \sum_{k \in \mathcal A} \sum_{j \in \mathcal I} \tau_k t_j^2 (\tau_k - \tau_j)^2. \end{aligned}$$ Next, defining $ \delta_{\bf z} := \min \left\{1,\min\{r_+^2,1\}/6 \right\}$, it follows $\|\tau_{\mathcal{I}}\|_\infty \leq \min\{r_+^2,1\} / 2$ for all ${\bf y} \in B_{\delta_{\bf z}}({\bf z})$. Moreover, the latter condition implies $|y_k|^2 = t_k^2 > {r^2_+}/{2}$ for all $k \in \mathcal I$ and thus, it holds that $$\begin{aligned} 2 \sum_{k \in {\mathcal{A}}} \sum_{j \in {\mathcal{I}}} \tau_k t_j^2 (\tau_k - \tau_j)^2 & \geq r_+^2 \sum_{k \in {\mathcal{A}}} \sum_{j \in {\mathcal{I}}} [\tau_k^3 - 2 \tau_k^2 \tau_j + \tau_k \tau_j^2] \\ & = r_+^2 \sum_{k \in {\mathcal{A}}} \left [ m \tau_k^3 - 2 \tau_k^2 \left( {\sum}_{j \in {\mathcal{I}}} \tau_j \right) + \tau_k \|\tau_{\mathcal{I}}\|^2 \right] \\ & = r_+^2 m \|\tau_{\mathcal{A}}\|_3^3 + r_+^2 \|\tau_{\mathcal{A}}\|_1 [ \|\tau\|^2 + \|\tau_{\mathcal{A}}\|^2] \geq r_+^2 m \|\tau_{\mathcal{A}}\|_3^3. \end{aligned}$$ Similarly, we can show $\sum_{k,j \in {\mathcal{A}}} \tau_k\tau_j(\tau_k-\tau_j)^2 = 2 \|\tau_{\mathcal{A}}\|_1 \|\tau_{\mathcal{A}}\|_3^3 - 2 \|\tau_{\mathcal{A}}\|^4 \geq 0$. Due to $\|\tau_{\mathcal{I}}\|_\infty \leq 1/2$, we have $|\tau_k - \tau_j| \leq 1$ for all $k,j \in {\mathcal{I}}$ and consequently, it follows $\sum_{k,j \in {\mathcal{I}}} (\tau_k - \tau_j)^2 \geq \sum_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^3$. Together and using the estimate $\|{\bf w}\|^3 \leq \sqrt{p} \|{\bf w}\|_3^3$, ${\bf w} \in {\mathbb{C}}^p$, we finally obtain $$\begin{aligned} \| P_{\bf y}^\bot\operatorname*{diag}(\tau){\bf y}\|^2 & \geq \frac{r_+^2}{2} \min \left\{\frac{r_+^2}{4},m\right\} \left[ {\sum}_{k,j \in \mathcal I} |\tau_k - \tau_j|^3 + \|\tau_{\mathcal{A}}\|_3^3 \right] \\ & \geq \underbracket{\begin{minipage}[t][5.5ex][t]{20ex}\centering$\displaystyle\frac{r^2_+\min \left\{{r_+^2}/{4},m\right\}}{2\sqrt{m^2 + n - m}} $ \end{minipage}}_{=: c} \left[ {\sum}_{k,j \in \mathcal I} |\tau_k - \tau_j|^2 + \|\tau_{\mathcal{A}}\|^2 \right]^{\frac32} . \end{aligned}$$ Thus, the [Ł]{}ojasiewicz-type inequality is satisfied with $\eta_{\bf z} := (\frac{n}{m})^{1.5}c^{-1}$. If $\mathbf{z}\in\mathbb{CS}^{n-1}$ is a stationary point of problem with $H=0$, then we have $f({\bf y}) - f({\bf z}) = (\beta/2) \|\tau\|^2$ and $\|{\mathrm{grad\!\;}}{f({\bf z})}\| = \sqrt{2}\beta \|P_{{\bf y}}^\bot \operatorname*{diag}(\tau){\bf y}\|$ and thus, Lemma \[lemma:kl-slim\] implies that the Łojasiewicz inequality holds with $\theta=\frac{1}{4}$. Moreover, this result can be used to show that the KL exponent can not be larger than $\frac14$ for general stationary points. The proof of Lemma \[lemma:kl-slim\] implies that the exponent $\frac32$ can be improved to $1$ if the index set $\mathcal A$ is empty. More generally, it can be shown that the Łojasiewicz-type inequality holds with exponent $1$ along directions ${\bf y} \in B({\bf z},\delta_{\bf z}) \cap {\mathbb{CS}^{n-1}}$ with $y_{\mathcal{A}}= 0$. Next, we prove that the KL exponent is at least $\frac{1}{4}$ when the matrix $A$ is diagonal. Let $A = \operatorname*{diag}({\bf a}) \in {\mathbb{C}}^{n \times n}$, ${\bf a} \in {\mathbb{R}^n}$, be a diagonal matrix. Then, the KL exponent of problem is $\frac{1}{4}$. In the case $\tau=0$, we have $f({\bf y}) - f({\bf z}) = {\frac{1}{2}}{\bf y}^* H {\bf y}$ and $2 \|{\mathrm{grad\!\;}}f({\bf y})\|^2 = {\bf y}^*H^2 {\bf y} - ({\bf y}^*H{\bf y})^2$. We now decompose ${\bf y}$ as follows $${\bf y} = {\bf u} + {\bf v}, \quad H{\bf u} = 0, \quad \|H{\bf v}\| \geq \sigma_{-}(H) \|{\bf v}\|, \quad \text{and} \quad {\bf u}^*{\bf v} = 0,$$ where $\sigma_{-}(H)$ denotes smallest positive singular value of $H$. Furthermore, let $\sigma_{+}(H) \geq 0$ be the maximum singular value of $H$. It holds that $$f({\bf y}) - f({\bf z}) \leq \frac{\sigma_{+}(H) }{2} \|{\bf v}\|^2, \quad \|{\mathrm{grad\!\;}}f({\bf y})\|^2 \geq \frac{\sigma_-(H)^2}{2} \|{\bf v}\|^2 - \frac{\sigma_{+}(H)^2}{2} \|{\bf v}\|^4.$$ Thus, due to $\|{\bf v}\|^2 \leq \|\Delta\|^2$, the inequality is satisfied with exponent $\theta = \frac12$. Next, we consider the general case $\tau \neq 0$. In this situation, we have $A_{[{\mathcal{A}}{\mathcal{I}}]} = 0$, $H_{[{\mathcal{I}}{\mathcal{A}}]} = A_{[{\mathcal{I}}{\mathcal{A}}]} = 0$ and the KKT conditions imply $$a_k + 2\beta |z_k|^2 - 2\lambda = 0 \quad \forall~k \in {\mathcal{I}}\quad \implies \quad H_{[{\mathcal{I}}{\mathcal{I}}]} = 0$$ and hence, due to $\tau_k = |\Delta_k|^2 = |y_k|^2 = t_k^2$ for all $k \in {\mathcal{A}}$, it follows $$\begin{aligned} {\bf y}^*\operatorname*{diag}(\tau)H{\bf y} & = y_{\mathcal{A}}^* \operatorname*{diag}(\tau_{\mathcal{A}}) H_{[{\mathcal{A}}\cdot]}\Delta + y_{\mathcal{I}}^* \operatorname*{diag}(\tau_{\mathcal{I}})H_{[{\mathcal{I}}\cdot]}\Delta \\ & = \Delta_{\mathcal{A}}^* \operatorname*{diag}(\tau_{\mathcal{A}}) A_{[{\mathcal{A}}{\mathcal{A}}]} \Delta_{\mathcal{A}}- 2\lambda \|\tau_{\mathcal{A}}\|^2 = \sum_{k \in {\mathcal{A}}} (a_k - 2\lambda) \tau_k^2. \end{aligned}$$ Using Young’s inequality, $t_k^2 \leq 1$, and $\tau_k = t_k^2$ for all $k \in {\mathcal{A}}$, it follows $${\bf y}^*H{\bf y} \cdot {\bf y}^*\operatorname*{diag}(\tau){\bf y} = {\bf y}^*H{\bf y} \left [\|\tau_{\mathcal{A}}\|^2 + \sum_{k \in {\mathcal{I}}} t_k^2 \tau_k\right] \leq \frac{2|{\bf y}^* H {\bf y}|^q }{q} + \frac{\|\tau_{\mathcal{A}}\|^{2p} + \|\tau_{\mathcal{I}}\|_1^p}{p}$$ for $p > 1$ and $q = 1 + \frac{1}{p-1}$. Let us now introduce the index set $\mathcal B := \{k \in \mathcal A: a_k - 2\lambda \neq 0\}$ and let us define $h_- := \min_{k \in \mathcal B} |a_k - 2\lambda|$ and $h_+ := \max_{k \in \mathcal B} |a_k - 2\lambda|$. Then, due to $\|\tau_{\mathcal{B}}\|^2 = \|y_{\mathcal{B}}\|^4_4$ and , and applying the estimates derived in the proof of Lemma \[lemma:kl-slim\], we obtain $$\begin{aligned} \|{\mathrm{grad\!\;}}f({\bf y})\|^2 & \\ & \hspace{-8ex} \geq \frac{h_-^2}{2} \|y_{\mathcal{B}}\|^2 - \frac{h_+^2}{2} \|y_{\mathcal{B}}\|^4 - \frac{4\beta}{q} h_+^q \|y_{\mathcal{B}}\|^{2q} - \frac{2\beta}{p} [\|\tau_{\mathcal{A}}\|^{2p} + \|\tau_{\mathcal{I}}\|_1^p] \\ & \hspace{-5ex} - 2\beta h_+ \|\tau_{\mathcal{B}}\|^2 + 2\beta^2 {\bf y}^* \operatorname*{diag}(|\tau|^2){\bf y} - 2\beta^2 ({\bf y}^*\operatorname*{diag}(\tau){\bf y})^2 \\ & \hspace{-8ex} \geq \frac{h_-^2}{2} \|y_{\mathcal{B}}\|^2 - o(\|y_{\mathcal{B}}\|^2) + \frac{\beta^2r_+^2}{4} {\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2 + \beta^2 r_+ m \|\tau_{\mathcal{A}}\|_3^3 \\ & \hspace{-5ex} - 2\beta p^{-1} [\|\tau_{\mathcal{A}}\|^{2p} + \|\tau_{\mathcal{I}}\|_1^p] \end{aligned}$$ for $\|y_{\mathcal{B}}\| \to 0$, $m := |{\mathcal{I}}|$, and ${\bf y}$ sufficiently close to ${\bf z}$. By , we have $$\begin{aligned} \|\tau_{\mathcal{I}}\|_1^4 \leq m^2 \|\tau_{\mathcal{I}}\|^4 & \leq \left[ {\frac{1}{2}}{\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2 + |{\mathcal{A}}| \|\tau_{\mathcal{A}}\|^2 \right]^2 \\ & \leq \left[{\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2\right ]^2 + 2(n-m)^2 \|\tau_{\mathcal{A}}\|^4 \end{aligned}$$ and consequently, setting $p = 4$, we can choose $\|\Delta\|$ (and thus $\|\tau\|$ and $\|y_{\mathcal{B}}\|$) sufficiently small, such that $$\begin{aligned} \|{\mathrm{grad\!\;}}f({\bf y})\|^2 & \geq \eta_1 \left [ \|y_{\mathcal{B}}\|^2 + {\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2 + \|\tau_{\mathcal{A}}\|_3^3 \right] \\ & \geq \eta_2 \left [ \|y_{\mathcal{B}}\|^2 + {\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2 + \|\tau_{\mathcal{A}}\|^2 \right]^{\frac32} \end{aligned}$$ and $f({\bf y}) - f({\bf z}) \leq \eta_3 [ \|y_{\mathcal{B}}\|^2 + {\sum}_{k,j \in {\mathcal{I}}} |\tau_k - \tau_j|^2 + \|\tau_{\mathcal{A}}\|^2 ]$ for suitable $\eta_1, \eta_2, \eta_3 > 0$. This shows that the [Ł]{}ojasiewicz inequality is satisfied with $\theta = \frac14$. Finally, we derive the KL exponent in the real case for global minimizers characterized by the positive semidefiniteness condition in Theorem \[theorem:glob-suff\]. Suppose $A \in {\mathbb{R}}^{n\times n}$ is a symmetric matrix and $\mathbf{z}$ is a stationary point of problem satisfying $$\begin{aligned} H \succeq0,\end{aligned}$$ where $H$ is defined in . Then, the KL exponent of problem at $\mathbf{z}$ is at least $\frac{1}{4}$. Without loss of generality we assume $\beta=1$. Let ${\bf y} \in {\mathbb{S}^{n-1}}$ be arbitrary and let us set $\Delta = {\bf y} - {\bf z}$, and $$\begin{aligned} &\gamma_{1} = \sum_{k\in {\mathcal{I}}}z_{k}^{3}\Delta_{k},\quad \gamma_{2} = \sum_{k\in{\mathcal{I}}} z_{k}^2\Delta_{k}^2, \quad \gamma_{3}=\sum_{k\in {\mathcal{I}}} z_{k}\Delta_{k}^{3},\quad \gamma_{4}= \|\Delta\|_4^4.\end{aligned}$$ Based on the representation ${\mathrm{grad\!\;}}{f({\bf y})} = P_{\bf y}^\bot [H + 2 \operatorname*{diag}(\tau)]{\bf y}$, we now introduce the following decompositions $$\begin{aligned} & 2\operatorname*{diag}(\tau){\bf y} = {\bf w} + c_{1} {\bf y}, \quad {\bf w} = 2 P_{\bf y}^\bot \operatorname*{diag}(\tau){\bf y} , \quad c_{1} = 2 {\bf y}^T \operatorname*{diag}(\tau){\bf y} \\ & {\Delta} = {\bf u} + {\bf v}, \quad H{\bf u} = 0, \quad {\bf u}^T {\bf v} = 0, \quad \|H{\bf v}\| \geq \sigma_-(H)\|{\bf v}\|,\end{aligned}$$ where $\sigma_-(H)$ denotes the smallest positive singular value of $H$. Using the latter decomposition and $H{\bf y} = H\Delta$, we can express the norm of the Riemannian gradient as follows $$\label{eq:grad-calc-kl} \|{\mathrm{grad\!\;}}{f({\bf y})}\|^2 = \|H{\Delta} + {\bf w}\|^2 - ({\Delta}^T H{\Delta})^2 = \|H{\bf v} + {\bf w}\|^2 - ({\bf v}^T H{\bf v})^2$$ Let $\lambda_+(H)$ be the largest eigenvalue of $H$. Then, by definition of ${\bf v}$, we obtain $$\begin{aligned} |{\bf v}^{T}H{\bf v}| & \leq \lambda_+(H) \|{\bf v}\|^2 \leq \lambda_+(H){\sigma_{-}(H)}^{-2} \|H{\bf v}\|^2 =: \bar \sigma^{-1} \cdot {\bf v}^{T} H^{2} {\bf v}. \label{eqn:6-10}\end{aligned}$$ Moreover, Lemma \[lem:6-2\] yields $$\begin{aligned} \|\tau\|^\frac32 \leq \eta_1 \|{\bf w}\| \label{eqn:6-11}\end{aligned}$$ for some constant $\eta_{1}>0$ and for all ${\bf y} \in {\mathbb{S}^{n-1}}$ sufficiently close to ${\bf z}$. Let $\epsilon > 0$ be an arbitrary small positive constant. We now discuss three different cases. *Case 1.* $\lVert\mathbf{w}\rVert \geq (1+\epsilon) \lVert H{\bf v}\rVert$ or $\lVert\mathbf{w}\rVert \leq (1-\epsilon)\lVert H{\bf v}\rVert$. In this case, we have $$\begin{aligned} \lVert H{\bf v} + {\bf w}\rVert^{2} \geq \frac{\epsilon^{2}}{2} \left(\lVert{\bf w}\rVert^{2}+\lVert H{\bf v}\rVert^{2} \right).\end{aligned}$$ Let us set $\delta_{\bf z} > 0$ sufficiently small such that $\|{\bf v}\|^2 \leq \epsilon^2 \bar \sigma / (4\lambda_+(H))$ and $|{\bf v}^TH{\bf v}| \leq 1$. Using , , and the estimates $({\bf v}^{T}H{\bf v})^{2}\leq\lambda_{+}^{2}(H)\lVert{\bf v}\rVert^{4}$ and $[\frac 12 |a + b|]^{3/2} \leq [|a|^{3/2} + |b|^{3/2}] / \sqrt{2}$, $a,b \in {\mathbb{R}}$, it follows $$\begin{aligned} \lVert{\mathrm{grad\!\;}}{f(\mathbf{y})}\rVert^{2}&\geq\frac{\epsilon^{2}}{2}(\lVert\mathbf{w}\rVert^{2}+\lVert H{\bf v}\rVert^{2})-\lambda_{+}^{2}(H)\lVert{\bf v}\rVert^{4}\geq\frac{\epsilon^{2}}{4}\left(2\lVert\mathbf{w}\rVert^{2}+\lVert H{\bf v}\rVert^{2}\right)\\ &\geq \frac{\epsilon^2}{2} \min\left \{\frac{1}{\eta_{1}},\frac{\bar\sigma}{2} \right \} \left[ \|\tau\|^3 + |{\bf v}^T H {\bf v}|^\frac32 \right ]\\ &\geq \frac{\epsilon^2\sqrt{2}}{2} \min\left \{\frac{1}{\eta_{1}},\frac{\bar\sigma}{2} \right \} \cdot | f({\bf y}) - f({\bf z}) |^\frac32. $$ Thus, we can infer that the KL exponent of problem at ${\bf z}$ is $\frac{1}{4}$. *Case 2.* $(2-{\epsilon})r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert\geq\lVert\Delta\rVert^2$ or $\gamma_1 \leq (2-\epsilon)r_+^2 \|\Delta_{\mathcal{I}}\|^2 \|\Delta\|^{-2}$. First, due to $\Delta^T {\bf y} = {\frac{1}{2}}\|\Delta\|^2$, we have $P_{\mathbf{y}}^{\perp}\Delta=\Delta-\frac{1}{2}\lVert\Delta\rVert_{2}^{2}\mathbf{y}$. Defining $t_\Delta := \Delta^T P_{\mathbf{y}}^{\perp}\Delta = \|\Delta\|^2 - \frac14 \|\Delta\|^4$, we will work with the following decompositions $$\begin{aligned} & H\Delta = \Delta^{T}H\Delta \cdot \mathbf{y}+c_{2} P_{\mathbf{y}}^{\perp}\Delta+\mathbf{w}_{1}, \quad c_{2}=\frac{1-\frac{1}{2}\lVert\Delta\rVert^{2}}{t_\Delta}\Delta^{T}H\Delta,\\ & \mathbf{w}=c_{3} P_{\mathbf{y}}^{\perp}\Delta+\mathbf{w}_{2}, \quad c_{3}= \frac{1}{t_\Delta}[2\Delta^T\operatorname*{diag}(\tau){\bf y} - \|\Delta\|^2 {\bf y}^T\operatorname*{diag}(\tau){\bf y}] $$ where $\mathbf{w}_{1}$ and $\mathbf{w}_{2}$ are vectors orthogonal to $\mathbf{y}$ and $\Delta$. Hence, by , it holds that $$\lVert{\mathrm{grad\!\;}}{f(\mathbf{y})}\rVert^{2} = (c_{2}+c_{3})^{2} \cdot t_\Delta +\lVert\mathbf{w}_{1}+\mathbf{w}_{2}\rVert^{2} \geq (c_{2}+c_{3})^{2} \cdot t_\Delta.$$ Using the definitions introduced in the first case, we can express $t_\Delta \cdot c_{3}$ via $$t_\Delta \cdot c_{3} = (2-\|\Delta\|^2) \gamma_4 + (6 - 4 \|\Delta\|^2) \gamma_3 + (4-5\|\Delta\|^2) \gamma_2 - 2 \|\Delta\|^2 \gamma_1.$$ Now, if $(2-{\epsilon})r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert \geq\lVert\Delta\rVert^{2}$, we readily obtain $(4-5\|\Delta\|^2) \gamma_2 - 2\|\Delta\|^2 \gamma_1 \geq (2\epsilon - 5\|\Delta\|^2) r_+^2 \|\Delta_{\mathcal{I}}\|^2$ and thus, it follows $$\label{eq:low-tdelta} t_\Delta \cdot c_{3} \geq \gamma_4 + \epsilon r_+^2 \|\Delta_{\mathcal{I}}\|^2 + o(\|\Delta_{\mathcal{I}}\|^2) \geq \|\Delta\|_4^4 + \frac{\epsilon r_+^2}{2} \|\Delta_{\mathcal{I}}\|^2$$ for $\Delta$ sufficiently small. Otherwise, if $\gamma_1 \leq (2-\epsilon)r_+^2 \|\Delta_{\mathcal{I}}\|^2 \|\Delta\|^{-2}$, then we also have $(4-5\|\Delta\|^2) \gamma_2 - 2 \|\Delta\|^2 \gamma_1 \geq (2\epsilon - 5\|\Delta\|^2) r_+^2 \|\Delta_{\mathcal{I}}\|^2$ and thus, holds in both sub-cases. Consequently, due to the positive semidefiniteness of $H$ and $\|\Delta\|^2 = - 2 \Delta^T {\bf z} = -2\Delta_{\mathcal{I}}^T z_{\mathcal{I}}\leq 2 \|\Delta_{\mathcal{I}}\|$, we can infer $$\begin{aligned} \lVert{\mathrm{grad\!\;}}{f(\mathbf{y})}\rVert & \geq|c_{2}+c_{3}| \sqrt{t_\Delta} = \left[\left(1-{\frac{1}{2}}\|\Delta\|^2\right) \Delta^{T}H\Delta + t_\Delta c_{3}\right] t_\Delta^{-1/2} \\ & \geq [\Delta^{T}H\Delta + 2 \|\Delta\|_4^4 +\epsilon r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert^{2}] \cdot (2\|\Delta\|)^{-1} \\ & \geq [\Delta^{T}H\Delta + 2 \|\Delta\|_4^4 +\epsilon r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert^{2}]^{\frac34} \cdot \frac{\epsilon^{\frac14}\sqrt{r_{+}\lVert\Delta_{\mathcal{I}}\rVert}}{2\|\Delta\|} \\ & \geq \eta_2 [\Delta^{T}H\Delta + 2 \|\Delta\|_4^4 +\epsilon r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert^{2}]^{\frac34}, $$ where $\eta_2 := \epsilon^{\frac14}\sqrt{r_{+}} / 2\sqrt{2}$. Next, utilizing and $|\tau_k| \leq 2 |\Delta_k|$ for all $k \in {\mathcal{I}}$, we finally obtain $$\begin{aligned} \lvert f(\mathbf{y})-f(\mathbf{z})\rvert&=\frac{1}{2}[ \Delta^{T}H\Delta+ \|\tau\|^2] \\ & \leq \frac{1}{2}\left[\Delta^{T}H\Delta+ \|\Delta_{\mathcal{A}}\|_4^4 + 4\lVert\Delta_{\mathcal{I}}\rVert^{2} \right] \leq \eta_{3} \lVert{\mathrm{grad\!\;}}{f(\mathbf{y})}\rVert^{\frac43}\end{aligned}$$ for some constant $\eta_{3}>0$ and for all ${\bf y}$ sufficiently close to ${\bf z}$. Hence, the KL exponent is $\frac{1}{4}$ in this case. *Case 3.* $(1-\epsilon)\lVert H{\bf v}\rVert\leq\lVert\mathbf{w}\rVert\leq(1+\epsilon)\lVert H{\bf v}\rVert$, $\gamma_{1}\geq(2-\epsilon)r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert^{2}\|\Delta\|^{-2}$ and $\left(2-\frac{\epsilon}{2}\right)r_{+}^{2}\lVert\Delta_{\mathcal{I}}\rVert\leq\lVert\Delta\rVert^{2}\leq2\lVert\Delta_{\mathcal{I}}\rVert$. In this case, we have $$\begin{aligned} \gamma_{1} = \Theta(\|\Delta\|^2) ,\quad \gamma_{2} = \Theta( \|\Delta\|^4), \quad \gamma_{3}=O(\|\Delta\|^{6}),\quad \gamma_{4}=\Theta(\|\Delta\|^{4}). \label{eq:order-gamma}\end{aligned}$$ Let us set $m := |{\mathcal{I}}|$ and define $\sigma_{k}= \Delta_{k} z_{k}+\frac{1}{2m}\|\Delta\|^{2}$ for all $k\in {\mathcal{I}}$ and $\sigma_k = 0$ for all $k \in {\mathcal{A}}$. Then, it follows $$\begin{aligned} \sum_{k \in [n]} \sigma_k = 0, \quad \|\sigma\|_1 \geq \sum_{k\in {\mathcal{I}}} z_{k}^{2}\sigma_{k} = \gamma_1 + \frac12 \|\Delta\|^2,\end{aligned}$$ and $\|\sigma\| = \Theta(\|\Delta\|^2)$. We now express $\|\mathbf{w}\|^{2}$ in terms of $\gamma_{1}$, $\gamma_2$, $\gamma_3$, and $\gamma_{4}$. Specifically, by utilizing , we obtain $$\begin{aligned} \frac14 \|\mathbf{w}\|^{2} & = {\bf y}^T \operatorname*{diag}(|\tau|^2){\bf y} - ({\bf y}\operatorname*{diag}(\tau){\bf y})^2 \\ & = \sum_{k \in [n]} \Delta_k^2 [\Delta_k^4 + 6\Delta_k^3z_k + 13\Delta_k^2z_k^2 + 12 \Delta_kz_k^3 + 4z_k^4] - 4\gamma_1^2 - 20\gamma_1\gamma_2 \\ & \hspace{2ex} - 25\gamma_2^2 - 8\gamma_3(5\gamma_2+2\gamma_1) - 16 \gamma_3^2 - 2 \|\Delta\|_4^4(4\gamma_3+5\gamma_2+2\gamma_1) - \|\Delta\|_4^8 \\ & = 4 \sum_{k \in [n]} \Delta_k^2 z_k^4 - 4\gamma_1^2 + O(\|\Delta\|^6) = 2 \sum_{k, j \in {\mathcal{I}}} z_k^2 z_j^2 (\sigma_k - \sigma_j)^2 + O(\|\Delta\|^6) \\ & = \Theta(\|\sigma\|^2) + O(\|\Delta\|^6), $$ which implies $\|\mathbf{w}\|=\Theta(\|\Delta\|^{2})$ and $\| H{\bf v}\|=\Theta(\|\Delta\|^{2})$. As a consequence, we get ${\bf v}^{T}H{\bf v}=\Theta(\|\Delta\|^{4})$ and $$\begin{aligned} \label{eqn:case3-diff} 2 \lvert f(\mathbf{y})-f(\mathbf{z})\rvert=\lvert{\bf v}^{T}H{\bf v} + \|\Delta\|_4^4 +4\gamma_{3}+4\gamma_{2}\rvert = \Theta(\|\Delta\|^{4}).\end{aligned}$$ For some index sets $\mathcal K, \mathcal J \subset [n]$, let $H_{\mathcal K\mathcal J} \in {\mathbb{R}}^{|\mathcal K| \times |\mathcal J|}$ denote the submatrix $H_{\mathcal K\mathcal J} = (H_{kj})_{k \in \mathcal K, j \in \mathcal J}$. Due to the positive semidefiniteness of $H$, we have $H_{\mathcal{I}\mathcal{I}}\succeq0$ and $H_{\mathcal{A}\mathcal{A}}\succeq0$. Furthermore, due to and , we obtain $$\begin{aligned} \nonumber \|{\mathrm{grad\!\;}}{f({\bf y})}\|^2 & = \|H\Delta + {\bf w} \|^2 - \Theta(\|\Delta\|^8) \\ \nonumber & = \|H_{{\mathcal{I}}{\mathcal{A}}}^T\Delta_{\mathcal{I}}+ H_{{\mathcal{A}}{\mathcal{A}}} \Delta_{\mathcal{A}}+ 2\operatorname*{diag}(|\Delta_{\mathcal{A}}|^2)\Delta_{\mathcal{A}}- c_1 \Delta_{\mathcal{A}}\|^2 \\ \label{eq:grad-split}& \hspace{4ex} + \| H_{{\mathcal{I}}{\mathcal{I}}} \Delta_{\mathcal{I}}+ H_{{\mathcal{I}}{\mathcal{A}}} \Delta_{\mathcal{A}}+ {w}_{\mathcal{I}}\|^2 - \Theta(\|\Delta\|^8).\end{aligned}$$ We define $$\begin{aligned} \nonumber {\bf g}_1 & := H_{{\mathcal{I}}{\mathcal{I}}} \Delta_{\mathcal{I}}+ H_{{\mathcal{I}}{\mathcal{A}}} \Delta_{\mathcal{A}}+ {w}_{\mathcal{I}}, \\ {\bf g}_2 &:= H_{{\mathcal{I}}{\mathcal{A}}}^T\Delta_{\mathcal{I}}+ H_{{\mathcal{A}}{\mathcal{A}}} \Delta_{\mathcal{A}}+ 2\operatorname*{diag}(|\Delta_{\mathcal{A}}|^2)\Delta_{\mathcal{A}}- c_1 \Delta_{\mathcal{A}}\end{aligned}$$ Next, let $\eta_{4},\eta_{5}>0$ and $\mu \in (0,\frac{1}{2})$ be given constants and suppose $\Delta_{\mathcal{I}}^{T}(H_{\mathcal{I}\mathcal{I}}\Delta_{\mathcal{I}}+H_{\mathcal{I}\mathcal{A}}\Delta_{\mathcal{A}})\geq-\eta_{4}\|\Delta\|^{4+\mu}$. Then, using ${\frac{1}{2}}c_1 = \|\Delta\|_4^4 + 4\gamma_3 + 5\gamma_2 + 2\gamma_1$, $\Delta_{\mathcal{I}}^T y_{\mathcal{I}}= - {\frac{1}{2}}\|\Delta\|^2$, and $\Delta_{\mathcal{I}}^T w_{\mathcal{I}}= 2\|\Delta\|_4^4 + 6 \gamma_3 + 4\gamma_2 - c_1 \Delta_{\mathcal{I}}^T y_{\mathcal{I}}$, it follows $$\begin{aligned} \Delta_{\mathcal{I}}^{T}{\bf g}_{1}=\Delta_{\mathcal{I}}^{T}(H_{\mathcal{I}\mathcal{I}}\Delta_{\mathcal{I}}+H_{\mathcal{I}\mathcal{A}}\Delta_{\mathcal{A}})+ \Delta_{\mathcal{I}}^T w_{\mathcal{I}}\geq \Theta(\|\Delta\|^{4}).\end{aligned}$$ Similarly, in the case $\Delta_{\mathcal{I}}^{T}(H_{\mathcal{I}\mathcal{I}}\Delta_{\mathcal{I}}+H_{\mathcal{I}\mathcal{A}}\Delta_{\mathcal{A}})\leq-\eta_{5}\|\Delta\|^{4-\mu}$ and if $\|\Delta\|$ is sufficiently small, we get $$\Delta_{\mathcal{I}}^{T}{\bf g}_{1} \leq -\frac{\eta_{5}}{2}\|\Delta\|^{4-\mu},$$ Combining both cases, we can infer $\|\Delta_{\mathcal{I}}\|\|{\bf g}_{1}\|\geq|\Delta_{\mathcal{I}}^{T}{\bf g}_{1}| \geq \eta_6 \|\Delta\|^4$ for some $\eta_6 > 0$ and for all ${\bf y}$ sufficiently close to ${\bf z}$. This implies $\|{\bf g}_{1}\| \geq [(2-\frac{\epsilon}{2})r_+^2]^{-1}\eta_6 \|\Delta\|^2$ and hence, by , we obtain $\|{\mathrm{grad\!\;}}f({\bf y})\| \geq \Theta(\|\Delta\|^2)$. Considering equation , the KL exponent has to be $\theta = \frac{1}{2}$ in these two cases. Now, let us suppose $-\eta_{5}\|\Delta\|^{4-\mu}\leq\Delta_{\mathcal{I}}^{T}(H_{\mathcal{I}\mathcal{I}}\Delta_{\mathcal{I}}+H_{\mathcal{I}\mathcal{A}}\Delta_{\mathcal{A}})\leq-\eta_{4}\|\Delta\|^{4+\mu}$. Due to $\Delta_{\mathcal{I}}^{T}H_{\mathcal{I}\mathcal{I}}\Delta_{\mathcal{I}}\geq0$, this directly yields $-\eta_{5}\|\Delta\|^{4-\mu}\leq\Delta_{\mathcal{I}}^{T}H_{\mathcal{I}\mathcal{A}}\Delta_{\mathcal{A}}<0$ and $$\begin{aligned} \Delta_{\mathcal{A}}^{T}H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}}\geq\Delta_{\mathcal{A}}^{T}(H_{\mathcal{I}\mathcal{A}}^{T}\Delta_{\mathcal{I}}+H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}})\geq\Delta^{T}H\Delta=\Theta(\|\Delta\|^{4}).\end{aligned}$$ If we have $\Delta_{\mathcal{A}}^{T}H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}}\geq\eta_{7}\|\Delta\|^{4-2\mu}$ for some constant $\eta_{7}>0$, then it holds that $$\begin{aligned} \Delta_{\mathcal{A}}^{T}{\bf g}_{2}&=\Delta_{\mathcal{A}}^{T}(H_{\mathcal{I}\mathcal{A}}^{T}\Delta_{\mathcal{I}}+H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}})+2\|\Delta_{\mathcal{A}}\|_{4}^{4}-c_1\|\Delta_{\mathcal{A}}\|^{2}\\ &\geq\eta_{7}\|\Delta\|^{4-2\mu}-\eta_{5}\|\Delta\|^{4-\mu}+O(\|\Delta\|^{4})\geq\frac{\eta_{7}}{2}\|\Delta\|^{4-2\mu}\end{aligned}$$ if $\|\Delta\|$ is sufficiently close to zero. As before, this estimate can be utilized to show $\|{\mathrm{grad\!\;}}{f({\bf y})}\| \geq \Theta(\|\Delta\|^{3-2\mu})$ and consequently, the KL exponent is $\frac{1+2\mu}{4}$. Finally, we consider $\eta_{8}\|\Delta\|^{4}\leq\Delta_{\mathcal{A}}^{T}H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}}\leq\eta_{7}\|\Delta\|^{4-2\mu}$, where $\eta_{8} > 0$ is chosen such that $\Delta^{T}H\Delta\geq\eta_{8}\|\Delta\|^{4}$. Let us define the decompositions $$\begin{aligned} \Delta_{\mathcal{A}}=\psi_{1}+\xi_{1},\quad\psi_{1}\in\mathrm{null}~H_{\mathcal{A}\mathcal{A}},\quad\xi_{1} \in [\mathrm{null}~H_{\mathcal{A}\mathcal{A}}]^\bot ,\\ \operatorname*{diag}(|\Delta_{\mathcal{A}}|^2)\Delta_{\mathcal{A}}=\psi_{2}+\xi_{2},\quad\psi_{2}\in\mathrm{null}~H_{\mathcal{A}\mathcal{A}}, \quad \xi_{2} \in [\mathrm{null}~H_{\mathcal{A}\mathcal{A}}]^\bot,\end{aligned}$$ where $\mathrm{null}~M$ is the null space of matrix $M$. We then have $\|\xi_{1}\|=\Theta(\|H_{\mathcal{A}\mathcal{A}}\xi_{1}\|)=\Theta(\sqrt{\xi_{1}^{T}H_{\mathcal{A}\mathcal{A}}\xi_{1}})=O(\|\Delta\|^{2-\mu})$ and $\lVert\psi_{1}\rVert=O(\|\Delta\|)$. Notice that such decompositions exist due to the symmetry of $H_{{\mathcal{A}}{\mathcal{A}}}$. Since $H$ is positive semidefinite, we can show that $\mathrm{null}\,H_{\mathcal{A}\mathcal{A}}\subset\mathrm{null}\,H_{\mathcal{I}\mathcal{A}}$. If $H_{\mathcal{I}\mathcal{A}}=0$, then this claim is certainly true. Otherwise, if we assume that the statement is false, the set $S = \mathrm{null}\,H_{\mathcal{A}\mathcal{A}}\cap [\mathrm{null}\,H_{\mathcal{I}\mathcal{A}}]^{\bot}$ is nonempty and there exists $\psi \in S$ and $\xi \in [\mathrm{null}\,H_{\mathcal{I}\mathcal{A}}^{T}]^{\bot}$. Then it holds $$\begin{aligned} \begin{bmatrix}\xi \\ \nu\psi \end{bmatrix} \begin{bmatrix} H_{{\mathcal{I}}{\mathcal{I}}} & H_{{\mathcal{I}}{\mathcal{A}}} \\ H_{{\mathcal{I}}{\mathcal{A}}}^T & H_{{\mathcal{A}}{\mathcal{A}}} \end{bmatrix} \begin{bmatrix}\xi &\nu\psi\end{bmatrix}=\xi^{T}H_{\mathcal{I}\mathcal{I}}\xi +2\nu \psi^{T}H_{\mathcal{I}\mathcal{A}}^{T}\xi \geq0,\quad\forall~\nu\in\mathbb{R}.\end{aligned}$$ But since $\psi^{T}H_{\mathcal{I}\mathcal{A}}^{T}\xi \neq0$, we can choose $\nu$ such that $\xi^{T}H_{\mathcal{I}\mathcal{I}}\xi+2\nu\psi^{T}H_{\mathcal{I}\mathcal{A}}^{T}\xi<0$, which is a contradiction. Hence, due to $H_{\mathcal{I}\mathcal{A}}^{T}\Delta_{\mathcal{I}} \in \mathrm{ran}\,H_{{\mathcal{I}}{\mathcal{A}}}^T = [\mathrm{null}\,H_{\mathcal{I}\mathcal{A}}]^\bot$, we can infer $H_{\mathcal{I}\mathcal{A}}^{T}\Delta_{\mathcal{I}} + H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}}\in [\mathrm{null}\,H_{\mathcal{A}\mathcal{A}}]^{\perp}$. Consequently, ${\bf g}_2$ can be written as ${\bf g}_2 = {\bf g}_3 + {\bf d}$ where ${\bf g}_3 \in [\mathrm{null}\,H_{\mathcal{A}\mathcal{A}}]^{\perp}$ and ${\bf d}=2\psi_{2}-c_{1}\psi_{1} \in \mathrm{null}\,H_{\mathcal{A}\mathcal{A}}$. If $\|{\bf d}\|\geq\frac{\eta_{8}}{2}\|\Delta\|^{3}$, then we obtain $\|{\mathrm{grad\!\;}}{f({\bf y})}\|\geq\|{\bf g}_{2}\|\geq\|{\bf d}\|\geq\frac{\eta_{8}}{2}\|\Delta\|^{3}$ and, by , the KL exponent is $\frac{1}{4}$. Otherwise, if $\|{\bf d}\|\leq\frac{\eta_{8}}{2}\|\Delta\|^{3}$, it follows $$\begin{aligned} 2\lVert\Delta_{\mathcal{A}}\rVert_{4}^{4}-c_{1}\lVert\Delta_{\mathcal{A}}\rVert^{2} &= 2\psi_{1}^{T}\psi_{2}-c_{1}\lVert\psi_{1}\rVert^{2}+2\xi_{1}^{T}\xi_{2}-c_{1}\lVert\xi_{1}\rVert^{2}\\ &=\psi_{1}^{T}\mathbf{d}+2\xi_{1}^{T}\xi_{2}-c_{1}\lVert\xi_{1}\rVert^{2} \geq-\frac{\eta_{8}}{2}\|\Delta\|^{4}+O(\lVert\Delta\rVert^{5-\mu})\end{aligned}$$ and $$\begin{aligned} \Delta_{\mathcal{A}}^{T}{\bf g}_{2}&=\Delta_{\mathcal{A}}^{T}(H_{\mathcal{I}\mathcal{A}}^{T}\Delta_{\mathcal{I}}+H_{\mathcal{A}\mathcal{A}}\Delta_{\mathcal{A}})+2\lVert\Delta_{\mathcal{A}}\rVert_{4}^{4}-c_{1}\lVert\Delta_{\mathcal{A}}\rVert^{2}\\ &\geq\Delta^{T}H\Delta-\frac{\eta_{8}}{2}\|\Delta\|^{4}+O(\lVert\Delta\rVert^{5-\mu}) \\ &\geq\frac{\eta_{8}}{2}\|\Delta\|^{4}+\Theta(\lVert\Delta\rVert^{5-\mu})\geq\frac{\eta_{8}}{4}\|\Delta\|^{4}.\end{aligned}$$ Thus, we have $\|{\mathrm{grad\!\;}}{f({\bf y})}\|\geq \Theta(\|\Delta\|^{3})$ and the KL exponent is $\frac{1}{4}$. Conclusions {#sec:conclusion} =========== In this paper, we analyze the geometric properties of a class of quartic-quadratic optimization problems under a single spherical constraint. When the matrix $A$ in the quadratic form is diagonal, the stationary points and local minima can be fully characterized and we show that the minimization problem does not possess any spurious local minima. Furthermore, a closed-form expression for global minimizer is available which is based on the projection onto the $n$-simplex. If $A$ is a rank-one matrix, a similar analysis can be performed and we derive characteristic properties of associated local minima and uniqueness of global minima up to a certain phase shift. We verify that the problem satisfies a Riemannian-type strict-saddle property in the real space when the interaction coefficient is at least of order $O(n^{3/2})$ or sufficiently small which corresponds to the case where either the quartic or the quadratic part is the leading term of the objective function. Finally, we estimate the Kurdyka-Łojasiewicz exponent $\theta$ of problem and show that $\theta$ is $\frac{1}{4}$ for all stationary points ${\bf z}$ if $A$ is diagonal or if the problem is restricted to the real space and ${\bf z}$ fulfills a certain global optimality condition. [^1]: H$.$ Zhang is partly supported by the elite undergraduate training program of School of Mathematical Sciences in Peking University. Z$.$ Wen is supported in part by the NSFC grants 11831002 and 11421101. A$.$ Milzarek is partly supported by the Fundamental Research Fund – Shenzhen Research Institute for Big Data (SRIBD) Startup Fund JCYJ-AM20190601.

ArXiv

--- abstract: 'In a landscape of compactifications with different numbers of macroscopic dimensions, it is possible that our universe has nucleated from a vacuum where some of our four large dimensions were compact while other, now compact, directions were macroscopic. From our perspective, this shapeshifting can be perceived as an anisotropic background spacetime. As an example, we present a model where our universe emerged from a tunneling event which involves the decompactification of two dimensions compactified on the two-sphere. In this case, our universe is of the Kantowski-Sachs type and therefore homogeneous and anisotropic. We study the deviations from statistical isotropy of the Cosmic Microwave Background induced by the anisotropic curvature, with particular attention to the anomalies. The model predicts a quadrupolar power asymmetry with the same sign and acoustic oscillations as found by WMAP. The amplitude of the effect is however too small given the current estimated bound on anisotropic curvature derived from the quadrupole.' author: - Julian Adamek - David Campo - 'Jens C. Niemeyer' bibliography: - 'decompact.bib' title: 'Anisotropic Kantowski-Sachs Universe from Gravitational Tunneling and its Observational Signatures' --- Introduction ============ If inflation lasted sufficiently long, all curvature scales imprinted on our universe by pre-inflationary physics are pushed to undetectable distances beyond our current horizon. If, on the other hand, inflation ended soon after the required amount of accelerated expansion to allow a later epoch of rich structure formation in our local universe, this cosmic amnesia may have been only partial. The largest cosmological scales observable today would then potentially show traces of the initial conditions for our inflating universe, including curvature , anisotropies [@GCP; @PPU1; @PPU2], or nontrivial topology [@LachiezeRey:1995kj]. This is the setting in which the scenario we propose can have observational consequences. While it has often been stated that fine-tuning the amount of inflation to this extent is unnatural, the landscape paradigm combined with the difficulty to find long-lasting inflationary solutions in string theory provide sufficiently strong counterarguments to take this possibility seriously [@DeSimone:2009dq]. Like the collision of bubbles in scenarios of false vacuum inflation (see [@Aguirre:2009ug] for a status report), it offers a remote chance to probe the landscape of string theory by observations. What are plausible initial conditions for our local inflationary patch in the landscape? Compactification in string theory is often treated kinematically, as part of the construction of the effective four dimensional theory. With four macroscopic plus a number of microscopic compact dimensions fixed once and for all, the transition between metastable vacua can be described by the spontaneous nucleation of bubbles with open homogeneous and isotropic spatial sections [@CDL; @Aguirre:2005nt]. The observable consequences of false vacuum bubble nucleation followed by a brief period of slow-roll inflation have been studied extensively in the context of open inflation and are well understood by now. From the dynamical perspective, however, compactification becomes a problem of string cosmology, for the compactified dimensions can spontaneously open up and become large [@Giddings]. This substantially widens the parameter space for initial conditions of our local universe, since there are now alternative channels to populate the landscape which can be viewed as transitions between vacua with differing numbers of macroscopic dimensions. This was first explicitly spelled out by Carroll, Johnson, and Randall [@CJR] in the context of dynamical compactification from a higher dimensional spacetime to our effectively four-dimensional one. The opposite process, dynamical decompactification, was studied in [@BlancoPillado:2009di], and in [@GiddingsMyers] in a different context. Although these previous studies were all concerned with the (de-)compactification of higher dimensions, there is no reason to exclude that the three macroscopic dimensions of our present universe may themselves be the result of such a process [@BrandyVafa]. This is the starting point of our work[^1]. It is intuitively clear that decompactification allows the existence of a preferred direction in the sky if only one or two directions are compact, and therefore gives rise to anisotropic cosmologies. We specifically consider the case where two of our macroscopic dimensions are compact. Before inflation, these dimensions were microscopic, leaving one macroscopic direction which may still play a preferred role for cosmological observations today if inflation was short. As a concrete example, we present a four dimensional model with two dimensions compactified on a sphere by the flux of an Abelian gauge field and a cosmological constant. The solutions of the Euclidean Einstein-Maxwell equations with spherical symmetry are well known. They describe pair creation of charged black holes in de Sitter space [@BoussoHawking1; @BoussoHawking2; @MannRoss]. We note that the causal patch beyond the cosmological horizon in the Lorentzian spacetime is an anisotropic inflating Kantowski-Sachs (KS) universe. This initial state can be placed in the broader context of the landscape in different ways. It can be viewed as an intermediate phase in a progressive opening up of initially compact dimensions as described in [@GiddingsMyers], or as the temporarily final step in a sequence of transitions starting from a higher dimensional geometry that triggered the decompactification of two previously compact dimensions. We comment briefly on these scenarios in Sec. \[ref:motivations\] but we will leave the discussion of further implications for future work. The remainder of this article is structured as follows. In Sec.\[sec:model\], the model is introduced and its connection to black hole pair production is discussed. Sec.\[sec:perturbations\] presents a preliminary exploration of Cosmic Microwave Background (CMB) signatures in our model, which are due to the anisotropy of KS spacetime. Owing to the technical complexity of a full analysis, we only consider the perturbations of a test scalar field and outline the qualitative modifications of CMB temperature anisotropies. With these results we try to assess some of the observed CMB anomalies which seem to indicate a violation of statistical isotropy in our universe. We conclude and discuss some directions for further research along these lines in Sec. \[sec:conclusions\]. The shapeshifting universe {#ref:motivations} ========================== Rearrangement of macroscopic dimensions --------------------------------------- The hypothesis of a string landscape [@Susskind; @BP] leads to the following scenario. The effective potential of the moduli (scalar fields describing the size and shapes of compact dimensions) possesses local minima, created from the competing effects of a positive cosmological constant, the curvature of the compact spaces, fluxes, wrapped branes, etc., but also presents flat directions. The compactified configurations, which are only metastable as a consequence of gravitational tunneling, may thus open up [@Giddings; @GiddingsMyers]. Consequently, in the general case the four-dimensional macroscopic spacetime has one, two, or three compact spatial dimensions, and is anisotropic (one exceptional, and interesting case, is the compactification on a flat torus as in [@OVV], see also [@McInnes]). More precisely, we consider the mechanism where $q$ dimensions are compactified on a $q$-sphere whose radius is stabilized using the flux of an appropriate $q$-form field [@Grana:2005jc; @Krishnan:2005su]. Many such gauge fields are available in string theories, each one coming with its own gauge coupling. One therefore obtains an entire landscape of possible flux compactifications [@BP]. The vacua in this landscape (the minima of the effective potential) differ in the number of large dimensions and in the effective value of their vacuum energy, which is determined by the quantum numbers of the flux fields. These vacua are rendered metastable by the possibility to tunnel through the barriers in the effective potential of the moduli. In such a process, a bubble is formed containing a new vacuum configuration, which differs from the parent vacuum in the values of the fluxes and, possibly, in the number of compactified dimensions. Elementary transitions where compact dimensions destabilize and start to open up (decompactification) and the inverse process (compactification) have been investigated by [@GiddingsMyers] and [@CJR], respectively. It has been recognized that both processes can be described with a generic type of instanton, and that the two different interpretations follow by exchanging the labels for “parent” and “daughter” vacuum. More generally, however, a vacuum transition may re-arrange the configuration of compact directions in an arbitrary way, such that it is possible that some dimensions compactify while others decompactify at the same time[^2]. We are thus confronted with the interesting possibility of a shapeshifting universe: some of our macroscopic dimensions may have been destabilized in the final vacuum transition which spawned our present universe and have been opening up ever since. The outcome of such re-arrangements is that the macroscopic space dimensions have a nontrivial topology, for instance $\mathbb{R}\times \mathrm{S}_2$ corresponding to a KS spacetime, or $\mathrm{H}_2 \times \mathrm{S}_1$ corresponding to Bianchi III. Schematically, this history can be described as $$\mathrm{dS}_D \times \mathrm{S}_2 \times \mathcal{M}_d \longrightarrow \mathrm{KS}_{\left(4\right)} \times \mathcal{M}'_{d + D - 2}~,$$ where $\mathrm{dS}_D$ is our effective $D$-dimensional parent vacuum, $\mathrm{KS}_{\left(4\right)}$ is the Kantowski-Sachs spacetime describing our macroscopic universe, and $\mathcal{M}_d$, $\mathcal{M}'_{d + D - 2}$ are additional compactification manifolds. The corresponding Penrose-Carter diagram is shown in Fig. \[fig:shape-shift\]. Owing to the reconfiguration of the microscopic dimensions, the effective vacuum energy is changed with respect to the parent vacuum and may lie in the anthropic window if the usual conditions for the smallness of incremental changes apply [@BP; @CJR]. ![\[fig:shape-shift\] Spacetime diagram of a generic tunneling process giving rise to a four-dimensional Kantowski-Sachs universe. The effective $D$-dimensional parent vacuum is marked as $\mathrm{dS}_D$ (left and right edge of the diagram should be identified). In this region, the two compact directions of our macroscopic universe approach a constant microscopic radius, as they are stabilized on $\mathrm{S}_2$ by a magnetic flux. In the Kantowski-Sachs spacetime, marked as $\mathrm{KS}_{\left(4\right)}$, these directions have destabilized and become large. Vice versa, there are $D-2$ large dimensions in the parent vacuum which approach a compact microscopic configuration in our universe. The “static” regions interpolate between the parent vacuum and our bubble. The interpolation smoothly re-arranges all the moduli fields to their new vacuum configuration. The case where the parent vacuum shares two large dimensions with our universe has recently been studied in [@BPS]. In this case, the $\mathrm{KS}_{\left(4\right)}$ region with topology $\mathbb{R}\times \mathrm{S}_2$ is replaced by a Bianchi III spacetime, which has spatial topology $\mathrm{H}_2 \times \mathrm{S}_1$. The familiar result of an open FRW universe in place of $\mathrm{KS}_{\left(4\right)}$ is obtained if all of our macroscopic dimensions are shared by our parent vacuum. ](generic-PCD.ps){width="85mm"} Description of the model ------------------------ The dynamics of tunneling processes involving several interacting fields (the moduli) can be very complicated (e.g., [@Johnson:2008kc]). Here, however, we are mainly interested in the geometrical properties of the corresponding instanton. Furthermore, we will treat the microscopic dimensions in our present universe as mere spectators, and therefore we can effectively work in a four-dimensional description. The tunneling event describing the decompactification of two dimensions wrapped onto a two-sphere is then completely equivalent to the pair creation of charged black holes. It was already pointed out in [@CJR] that the mechanism of dynamical compactification, using magnetic fluxes, is in some sense a generalization of the idea of black hole pair creation. The connection between these ways of thinking was further elucidated in [@BSV], where different types of compactifications from a six-dimensional de Sitter vacuum have been studied. Ignoring the additional dimensions in our description comes at the cost of not being able to represent the parent vacuum correctly, nor making quantitative statements about the tunneling rates. However, our simplified picture should accurately capture the essential geometrical properties of the daughter vacuum. In particular, within our scenario where two directions are wrapped onto a sphere, the tunneling process gives rise to a Kantowski-Sachs universe of topology $\mathbb{R} \times \mathrm{S}_2$. If the scales comparable to the size of the compact directions have recently entered our horizon, we may gather information about the topology of our universe by examining, e.g., the temperature correlations in the CMB. Inflation is still a necessary ingredient in this class of models in order to dilute the spatial curvature, produce a phenomenologically viable perturbation spectrum, and heat up the universe. However, any detection of the CMB signatures that we discuss in Sec. IV implies that inflation ended before they were redshifted too far beyond the horizon. Whether or not anthropic pressure makes this parameter range likely is an interesting debate (e.g., [@DeSimone:2009dq]) to which we have nothing new to add. For the purposes of this work, we simply accept it as a plausible possibility. As the authors of [@CJR] have shown, it is possible to accommodate a period of slow-roll inflation following the transdimensional tunneling. In their example, this is achieved by coupling an inflaton field to the curvature and flux. It was demonstrated that the tunneling process can indeed trigger slow-roll inflation by setting free the inflaton, which was previously trapped in a minimum by the configuration of the parent vacuum. In this case, the initial conditions for inflation are governed by the tunneling process. In order to simplify our analysis, we do not consider a realistic model for slow-roll inflation and absorb the inflaton potential energy into the effective cosmological constant. Thus, in our model $\Lambda$ is the sum of the inflaton energy density and of the observed value of the cosmological constant. Kantowski-Sachs universe from gravitational tunneling in $4$D Einstein-Maxwell theory {#sec:model} ===================================================================================== We now introduce our cosmological model and outline the links with the pair creation of charged black holes during inflation. Most of the content of this section is well-known and was originally presented in the context of black hole pair creation (see, e.g., [@BoussoHawking1; @BoussoHawking2; @MannRoss] and references therein). In our notation and terminology, we closely follow [@MannRoss]. We present our model using a unified picture where pair creation of charged black holes can equivalently be understood as Coleman-De Luccia tunneling between two different vacua of a lower dimensional effective theory. We will therefore use both interpretations interchangeably, depending on which of them is more useful in a particular situation. As a starting point we take the action of 4D Einstein-Maxwell theory, $$\mathcal{S} = \frac{1}{16 \pi} \int\!\dd^4 x \sqrt{-g} \Bigl[\mathcal{R}[g_{\mu\nu}] - 2 \Lambda - F_{\mu\nu} F^{\mu\nu}\Bigr]~,$$ where a cosmological constant $\Lambda > 0$ is included in order to model inflation in a simple way. $\mathcal{R}[g_{\mu\nu}]$ is the Ricci scalar of the metric tensor $g_{\mu\nu}$, $g$ is the metric determinant, and $F_{\mu\nu}$ is the field strength tensor of an Abelian gauge field. We look for solutions with two of the three spatial dimensions compactified on a sphere of radius $R$. The line element therefore takes the form $$\dd s^2 = \gamma_{ab}(\mathbf{x}) \dd x^a \dd x^b + R^2(\mathbf{x}) \left[ \dd \theta^2 + \sin^2\theta \dd \phi^2\right]~,$$ where $\gamma_{ab}$ is the metric of a 1+1-dimensional Lorentzian manifold with coordinates $\mathbf{x}$, i.e., the Latin indices take values $0$ and $1$ only. The gauge field should also respect the symmetries of our metric ansatz. The magnetic solutions of Maxwell’s equations are therefore proportional to the volume form of $\mathrm{S}_2$, $$\mathbf{F} = Q \sin\theta \dd \theta \wedge \dd \phi~,$$ where $Q$ is the magnetic charge. Our next step will be the dimensional reduction of the theory by integration over the coordinates $\theta$, $\phi$ of $\mathrm{S}_2$. Decomposing the full Ricci scalar $\mathcal{R}[g_{\mu\nu}]$ into contributions from the $\mathrm{S}_2$-curvature and from the Ricci scalar of $\gamma_{ab}$, hereafter denoted as $\mathcal{R}[\gamma_{ab}]$, the action can be rewritten as $$\begin{gathered} \mathcal{S} = \frac{1}{4} \int \dd^2 x \sqrt{-\gamma} \Bigl[ R^2 \mathcal{R}[\gamma_{ab}] + 2 \gamma^{ab} \nabla_a R \nabla_b R \Bigr. \\ \Bigl. + 2 - 2 \Lambda R^2 - \frac{2 Q^2}{R^2}\Bigr]~,\end{gathered}$$ after an integration by parts has removed second derivatives on $R$. Variation with respect to $R$ and $\gamma_{ab}$ yields two coupled second order equations, $$\label{EOM} \frac{1}{2} \mathcal{R}[\gamma_{ab}] R - \square R - \Lambda R + \frac{Q^2}{R^3} = 0~,$$ and $$\begin{gathered} \label{EEQ} 2 R \nabla_a \nabla_b R - 2 \gamma_{ab} R \square R - \gamma_{ab} \gamma^{cd} \nabla_c R \nabla_d R \\+ \gamma_{ab} \Bigl( 1 - \Lambda R^2 - \frac{Q^2}{R^2} \Bigr) = 0~.\end{gathered}$$ Following the usual procedure, we now solve the corresponding Euclidean equations with the boundary conditions appropriate to describe a tunneling Riemannian geometry. In section \[sec:cosmology\], we consider the nucleated $4$D Lorentzian geometries which contain a KS spacetime. These solutions describe a universe with two compact spatial dimensions which are destabilized and start to grow. Euclidean geometries {#sec:instantons} -------------------- A formal analytic continuation to imaginary time takes the action to its Euclidean counterpart. The Euclidean equations remain formally identical to the Lorentzian ones (\[EOM\]) and (\[EEQ\]), with the important difference that the metric $\gamma_{ab}$ now has Euclidean signature $(++)$. We consider all the solutions of the $2$D Euclidean equations which have $O(2)$-symmetry, meaning that all quantities only depend on the Euclidean distance $\chi$ from the symmetry point: $$\begin{array}{r c l} \gamma_{ab} \dd x^a \dd x^b &=& \dd \chi^2 + \rho^2(\chi) \dd \varphi^2~, \\ R(\chi, \varphi) &=& R(\chi)~. \end{array} \quad \text{(Euclidean)}$$ Here, the coordinate $\varphi$ is an angular coordinate with period $2 \pi$. We will relate these solutions to the instantons describing pair creation of charged black holes in de Sitter space. At the symmetry point $\chi = 0$ the Euclidean scale factor $\rho$ is zero. The zeros of $\rho$ are sometimes called “poles” since they may represent the origin of a polar coordinate system. In this paper, we will also call the symmetry point $\chi = 0$ the “south pole” of the geometry. If $\rho$ has a second zero at finite $\chi = \chi_{\mathrm{max}}$, we will call this the “north pole”. Regularity of the geometry at the poles requires $\rho'(0) = -\rho'(\chi_{\mathrm{max}}) = 1$, otherwise there will be a conical singularity. Using the $O(2)$-symmetric ansatz, eqs. (\[EOM\]) and (\[EEQ\]) read, respectively, $$\label{EucEOM} \frac{\rho''}{\rho} R + R'' + \frac{\rho'}{\rho} R' + \Lambda R - \frac{Q^2}{R^3} = 0~,$$ and $$\label{EucEEQ} R'^2 + 2 \frac{\rho'}{\rho} R' R - 1 + \Lambda R^2 + \frac{Q^2}{R^2} = 0~.$$ These two equations can be combined to show that $\dd \ln \rho = \dd \ln R'$, which implies $\rho \propto R'$. This result can be used to eliminate $\rho$ in eq. (\[EucEEQ\]). One finds $$2 R R'' + R'^2 - 1 + \Lambda R^2 + \frac{Q^2}{R^2} = 0~,$$ and its first integral $$\label{ECL} R'^2 - \frac{Q^2}{R^2} + \frac{2 M}{R} - 1 + \frac{\Lambda}{3} R^2 = 0~.$$ These solutions are indeed the Euclidean analogs of Reissner-Nordström-de Sitter (RNdS) black holes, where the constant of integration $M$ is the Misner-Sharp mass. The positive roots of the potential $$V(R) \equiv - \frac{Q^2}{R^2} + \frac{2 M}{R} - 1 + \frac{\Lambda}{3} R^2$$ correspond to the horizons of the black hole spacetime. These are, in ascending order, the inner Cauchy horizon $R_i$, the black hole event horizon $R_b$, and the cosmological horizon $R_c$. There is also one root at negative $R$ which has no physical significance. ![\[fig:potential\] Qualitative shape of the potential $V(R)$ for three different mass parameters $M$ at fixed $\Lambda$ and $Q$, with $Q < 3 / \sqrt{48 \Lambda}$. The roots of $V(R)$ correspond to the three different horizons in a Reissner-Nordström-de Sitter spacetime. The lower dashed line (smallest value of $M$) shows the special case when the black hole event horizon $R_b$ and the inner Cauchy horizon $R_i$ coincide (extremal Reissner-Nordström black hole). This potential may characterize a “cold” instanton as well as the $\mathrm{H}_2 \times \mathrm{S}_2$ instanton, which has no finite action. The upper dashed line (largest value of $M$) shows the special case when $R_b$ coincides with the cosmological horizon $R_c$. This would be characteristic for a “charged Nariai” instanton. The line in between shows the case when $V'(R_b) = -V'(R_c)$, corresponding to a “lukewarm” instanton.](RNdS-potential.ps){width="85mm"} The number of positive roots can of course be less than three. For a given $\Lambda$ and a sufficiently small $Q$, there is a minimal and a maximal value for $M$ at which two roots exactly coincide, such that $V(R)$ has a double root. These two situations are realized when $R_i \equiv R_b$ and $R_b \equiv R_c$, respectively. For the intermediate range of $M$, all three roots are distinct. This range for $M$ shrinks as $Q$ increases, and at $Q = 1 / \sqrt{4 \Lambda}$, all three positive roots coincide for $M = 2 / \sqrt{18 \Lambda}$. For $Q$ larger than that, or for $M$ outside the intermediate range, only one positive root exists, and no instanton of finite action can be obtained. Let us first examine the cases where $M$ is chosen such that $V(R)$ has a double root. If $R_b$ and $R_c$ coincide, they designate a minimum of the potential, see Fig. \[fig:potential\]. Hence, there is a solution $R = \mathrm{const.} = R_b (\equiv R_c)$. In this case $\rho = H^{-1} \sin H \chi$, with $H^2 = V''(R_b) / 2$. \[The easiest way to see this is to write $R = R_b + \delta R$ and expand eq. (\[ECL\]) to second order in the perturbation $\delta R$. The functional form of $\rho$ then follows from $\rho \propto \delta R'$, and since the amplitude of $\rho$ is entirely fixed by the boundary conditions $\rho(0) = 0$, $\rho'(0) = 1$, it is independent of the amplitude of the perturbation.\] The Euclidean solution is a sphere with radius $H^{-1}$. This particular solution is called the charged Nariai instanton, since the analytic continuation to Lorentzian signature turns the $\mathrm{S}_2 \times \mathrm{S}_2$, obtained in the full 4D picture, into a $\mathrm{dS}_2 \times \mathrm{S_2}$ (charged) Nariai spacetime. If, on the other hand, $R_b$ and $R_i$ coincide, they designate a maximum of the potential. There is again a solution $R = \mathrm{const.} = R_b (\equiv R_i)$, but this time one finds that $\rho = \omega^{-1} \sinh \omega \chi$, with $\omega^2 = - V''(R_b) / 2$. The geometry is now a hyperbolic plane $\mathrm{H}_2$ instead of a sphere. Since the hyperbolic plane has infinite volume, this instanton does not have a finite action. This is the reason why it is usually not discussed in the context of black hole pair creation. The spacetime one would get from analytic continuation of this instanton is $\mathrm{AdS}_2 \times \mathrm{S_2}$, again in the 4D picture. We note that in the 2D picture the compactification vacuum has negative effective vacuum energy, and hence the universe is Anti-de Sitter. This seems to be a generic feature of two-dimensional vacua obtained by flux compactification. However, this does not exclude the possibility that other compactification mechanisms may give rise to effective two-dimensional solutions with positive vacuum energy. Then there would also be two-dimensional de Sitter vacua which could play the role of our parent vacuum. It is evident from Fig. \[fig:potential\] that in the case $R_b \equiv R_i$, a second instanton should exist. It is obtained by starting from $R = R_c$ at the south pole of the geometry. As $\chi$ increases, $R$ rolls through the potential well and approaches $R = R_b$. However, since $R_b$ is a double root, $R$ reaches this value only at $\chi = \infty$. The resulting geometry is a two-dimensional $O(2)$-symmetric space which interpolates between a sphere with Gaussian curvature $V''(R_c) / 2$ at the south pole and a pseudosphere with (negative) Gaussian curvature $V''(R_b) / 2$ as $\chi \rightarrow \infty$. It is called the “cold” instanton, since it describes pair creation of extremal (and therefore cold) black holes. The Euclidean action of this instanton is finite and was computed, e.g., in [@MannRoss]. Another special class of solutions exists when all three horizons coincide, $R_i \equiv R_b \equiv R_c$. Solutions of this type are called “ultracold”. Because they are only obtained on a single point in parameter space, namely $Q = 1/\sqrt{4\Lambda}$, we do not discuss them in more detail. Let us finally turn to the case where there are three distinct horizons. Starting again with $R = R_c$ at the south pole, one can see from Fig. \[fig:potential\] that now $R$ reaches the value $R_b$ at a finite Euclidean distance $\chi = \chi_\mathrm{max}$. Imposing the regularity condition $\rho'(0) = 1$ at the south pole, we see from eq. (\[ECL\]) and $\rho \propto R'$ that a conical singularity at the north pole can only be avoided if $V'(R_c) \equiv -V'(R_b)$, which is only possible for $Q < 3 / \sqrt{48 \Lambda}$ and fixes the mass to $M = Q$. The instanton obtained this way is regular and has topology $\mathrm{S}_2$, and its action was also computed in [@MannRoss]. It describes pair creation of non-extremal charged black holes and is therefore called the “lukewarm” solution. In the terminology of [@CJR], it corresponds to the “interpolating solution”. This completes our catalog of $O(2)$-symmetric solutions of the 2D Euclidean equations. The lukewarm solutions are the only ones that we shall consider in the following, though the precise values of the parameters $Q$ and $M$ are not important for our discussion. We want to point out here that additional compact dimensions would not change the geometrical properties of the instanton from the 2D point of view. The additional dimensions would add new moduli fields which complicate the equations, but the $O(2)$-symmetry and the 2D topology of the instanton would remain unchanged. Cosmology in the spacetime of pair-created charged black holes {#sec:cosmology} -------------------------------------------------------------- In the language of gravitational tunneling, the Euclidean geometries describe the solution in the classically forbidden region of configuration space. A turning-point configuration, i.e., a spacelike hypersurface which contains the initial data to solve the Cauchy problem in the classically allowed region, is given by a maximal section of the instanton, running perpendicular to the circles of constant $\chi$ from the south pole to the north pole and back again. On this surface, which has topology $\mathrm{S}_1 \times \mathrm{S}_2$ in the 4D picture, the Euclidean solution connects to a classical Lorentzian spacetime. The $O(2)$-symmetry of the Euclidean $\gamma_{ab}$ carries over to an $O(1, 1)$-symmetry of the Lorentzian counterpart. Therefore, the metric solution is simply obtained by formally taking $\varphi \rightarrow i t$. The line element then reads $$\dd s^2 = -\rho^2(\chi) \dd t^2 + \dd \chi^2 + R^2(\chi) \left[\dd \theta^2 + \sin^2 \theta \dd \phi^2\right]~.$$ Using $R'^2 = -V(R) \equiv f(R)$, $\rho \propto R'$, and absorbing the constant of proportionality into the definition of $t$, this can be written in the familiar form of Schwarzschild-type coordinates, $$\label{RNdSmetric} \dd s^2 = - f(R) \dd t^2\\ + f^{-1}(R) \dd R^2 + R^2 \left[\dd \theta^2 + \sin^2 \theta \dd \phi^2\right]~.$$ A causal diagram of the full classical spacetime is shown in Fig. \[fig:pcd\]. Strictly speaking, the analytic continuation only covers the patch $R_b < R < R_c$. However, the metric (\[RNdSmetric\]) can simply be extended to allow $R > R_c$ or $R < R_b$, respectively IV and II on Fig. \[fig:pcd\], in order to cover the entire physical manifold. In these patches, $R$ is a timelike coordinate (since $f(R)$ is negative there), and $t$ is a spacelike coordinate, so that the metric in manifestly homogeneous (while it is static in the domain between the horizons). Furthermore, using $R$ as a time coordinate one obtains a foliation $\mathbb{R}_{\mathrm{time}} \times (\mathbb{R} \times \mathrm{S}_2)_{\mathrm{space}}$, which means that the two patches may be considered as two separate KS universes. ![\[fig:pcd\] Penrose-Carter diagram of Reissner-Nordström-de Sitter spacetime – left and right side (regions II and VII) should be identified. The line marked $\Sigma_0$ shows the location of the turning point three-geometry in the diagram. It also contains the “north pole” (NP) and the “south pole” (SP) of the instanton. Region I and V are the static regions between the cosmological horizons and the event horizons of the two black holes. Region II is the interior of the black holes, which make up a wormhole. It is a Kantowski-Sachs universe evolving from a Big Bang at $R = R_b$ to a Big Crunch at $R = R_i$. Region IV lies beyond the cosmological horizons of the black holes. It is another Kantowski-Sachs universe which has its Big Bang at $R = R_c$ and starts to inflate indefinitely due to the presence of a cosmological constant. Regions III and VI lie beyond the inner Cauchy horizons $R = R_i$ and exist only in the mathematical solution, as well as region VII, which would lie in another universe. In regions I-IV, curves of constant $R$ are sketched as dashed lines. We have omitted the time-reversed copy of the diagram below $\Sigma_0$, which would be present in a full classical solution due to time reflection symmetry. After the gravitational tunneling event has taken place, classical evolution “begins” on $\Sigma_0$.](RNdS-PCD.ps){width="85mm"} The cosmology of our own universe is described by the KS spacetime of region IV. For a general tunneling process in the landscape of compactifications, region II will be replaced by a patch containing the parent vacuum, which may have a number of additional large dimensions, cf. Fig. \[fig:shape-shift\]. Regions I and V then interpolate between the parent vacuum and our own universe. We assume that this does not drastically alter the cosmology in our universe, though the early evolution may change somewhat as the moduli describing our microscopic dimensions settle to their new vacuum configuration. The KS spacetime of region IV begins with a Big Bang which is a “pancake” singularity where only the scale factor of the non-compact direction is zero (since $R = R_c$ at this time). After a short period of curvature domination, the universe starts to inflate. In our simple model, inflation is driven by a cosmological constant and therefore it never ends. This can be overcome by replacing $\Lambda$ by an inflaton field with appropriate properties. We will assume that the initial conditions for inflaton, which are determined by the tunneling process, allow for a period of slow-roll inflation in the newly nucleated universe. Explicit examples for such a scenario have been constructed in [@CJR]. In this case, the inflaton will eventually decay and the universe will reheat. We have presented a scenario of gravitational tunneling in which our universe would have homogeneous but anisotropic spacelike hypersurfaces with topology $\mathbb{R} \times \mathrm{S}_2$. In addition to this anisotropic curvature, the expansion is also anisotropic. The anisotropy of the background is reflected in observables such as the correlation of multipole coefficients in the CMB. We now present a simplified analysis of these observable signatures. Observable signatures of the anisotropy in a Kantowski-Sachs universe {#sec:perturbations} ===================================================================== The non-trivial topology and geometry of the background introduces several changes with respect to Friedman-Robertson-Walker cosmologies. First, because of the compact dimensions, the power spectrum on large scales is modified. In the event that the period of inflation was short, these scales would enter our horizon today. Second, the theory of cosmological perturbations is more intricate because scalar, vector, and tensor perturbations are coupled via the shear (see, e.g., [@GCP; @PPU2; @Watanabe:2010fh] for related work in Bianchi I models). Third, the free streaming of the photons is also anisotropic, i.e., the redshift factor between the last scattering and today depends on the direction of observation (an “anisotropic Sachs-Wolfe effect”). These three sources of primary anisotropies cannot be treated independently in a self-consistent way because they have a common origin. We will nevertheless examine them separately by means of the following two simplifications. First, we will not present the full theory of cosmological perturbations but only examine the power spectrum of a test scalar field. Second, we disentangle the “early” from the “late” time effects, that is those resulting from the modifications of the inflationary power spectrum and the anisotropic Sachs-Wolfe effect. This is achieved by assuming that the expansion rate is isotropic after the end of inflation. In sec. \[sec:power phi\], we solve the mode equation of a massless scalar field in KS spacetime, motivate the choice of the state, and give an analytic expression of the corresponding two-point correlation function at equal time. In sec. \[sec:power alm\] , we project this two-point correlation function onto a sphere and calculate the corresponding multipole correlators $\langle a_{lm} a^\ast_{l'm'} \rangle$ for $l,l' = 2,3$ and $4$. In sec. \[sec:rafinements\], we discuss the additional corrections introduced by the remaining sources of anisotropies. In sec. \[sec:anomalies\] we examine whether some of the reported CMB anomalies can be accounted for by the model. Power spectrum of a massless scalar field in a Kantowski-Sachs spacetime {#sec:power phi} ------------------------------------------------------------------------ We begin by introducing a new set of coordinates in the inflationary KS universe of RNdS. It is important for the quantization and the definition of the ground state to work with a time coordinate which can be continued analytically beyond the cosmological horizons (into I and V of Fig. \[fig:pcd\]). The following choice satisfies this requirement: $$\label{metricmode} \dd s^2 = H^{-2} \left[-\dd \tau^2 + V(R) \dd z^2\right] + R^2 \left[\dd \theta^2 + \sin^2\theta \dd \phi^2\right]~,$$ where we have introduced the inflationary Hubble parameter, $H \equiv \sqrt{\Lambda / 3}$. In this equation, $R$ should be read as a function of $\tau$. Note that $V(R)$ is positive and plays the role of a scale factor for the non-compact direction $z$. As can be seen by comparing with eq. (\[RNdSmetric\]), the spacelike coordinate is defined by $\dd z = H\, \dd t$, and the (dimensionless) time coordinate by $\dd \tau = H\, \dd R/\sqrt{V(R)}$. $R$ is thus a growing function of $\tau$. Let us consider a minimally coupled massless real scalar field $\Phi$, $$\mathcal{S}_{\Phi} = -\frac{1}{2} \int\!\dd^4 x \sqrt{-g} g^{\mu\nu} \partial_\mu \Phi \partial_\nu \Phi~.$$ Given a Cauchy surface of the entire spacetime such as $\Sigma_0$ on Fig. \[fig:pcd\], $\Phi$ and its conjugate momentum are required to satisfy the canonical commutation relations. We note the Klein-Gordon product $$\left( f,g \right) = i \int_\Sigma\!\!d\sigma^\mu \, \sqrt{h_\Sigma} \left( f^* \partial_\mu g - g \partial_\mu f^* \right)$$ and decompose the field into creation and annihilation operators $\mathrm{a}_n^\dagger, \mathrm{a}_n$ defined by $${\mathrm{a}}_n = \left( f_n, \Phi \right)~,$$ where $\left\{f_n\right\}$ is a complete family of positive norm solutions of the Klein-Gordon equation to be specified below. Following [@Sasaki], we carry out this program in the following two steps. First we find the normalized solutions of the field equation in the KS wedge. Then we analytically continue the solutions beyond the cosmological horizon (into regions I and V of Fig. \[fig:pcd\]) and demand that they be regular there. This way, we avoid normalizing the solutions of the Klein-Gordon equation in regions I and V, which would require the calculation of a complicated integral. Exploiting the $\mathbb{R} \times \mathrm{S}_2$ symmetry of the spacelike hypersurfaces of KS, we introduce the decomposition $$\label{modeexpansion} \Phi(\tau, z, \theta, \phi) = \sum_{\ell, m} \int\!\frac{\dd k}{\sqrt{2 \pi}} \, \Phi_{k\ell m}(\tau) e^{i k z} Y_{\ell m}(\theta, \phi)~.$$ We introduce the rescaled field $\Psi = \vert g \vert^{1/4} \Phi= R V^{1/4} \Phi$, in terms of which the action reads $$\begin{gathered} \mathcal{S}_{\Psi} = \frac{1}{2} \sum_{\ell, m} \int\!\dd k \dd \tau \biggl\{\dot{\Psi}_{k\ell m} \dot{\Psi}^\ast_{k\ell m} - \biggl[\frac{k^2}{V(R)} + \frac{\ell \left(\ell + 1\right)}{H^2 R^2}\biggr.\biggr.\\ \biggl. \biggl. - \frac{1}{H^2} \left(\frac{V'(R)}{R} - \frac{V'(R)^2}{16 V(R)} + \frac{V''(R)}{4}\right)\biggr] \Psi_{k\ell m} \Psi^\ast_{k\ell m}\biggr\}~,\end{gathered}$$ where we have used $\ddot{R} = V'(R) / 2 H^2$ and dropped the boundary term of an integration by parts. We further decompose the solutions into $$\label{ladderops} {\Psi}_{k\ell m}(\tau) = {\mathrm{a}}_{k\ell m} \, u_{k\ell}(\tau) + (-1)^{m}\, {\mathrm{a}}^\dagger_{-k\ell -m}\, u^\ast_{k\ell}(\tau)~.$$ where the functions $u_{k\ell}$ satisfy the equation $$\begin{gathered} \label{modeeq} \ddot{u}_{k\ell} + \biggl[\frac{k^2}{V(R)} + \frac{\ell \left(\ell + 1\right)}{H^2 R^2}\biggr.\\ \biggl. - \frac{1}{H^2} \left(\frac{V'(R)}{R} - \frac{V'(R)^2}{16 V(R)} + \frac{V''(R)}{4}\right)\biggr] u_{k\ell} = 0\end{gathered}$$ The mode functions defined by (\[modeexpansion\])–(\[modeeq\]) form a complete family of solutions of the Klein-Gordon equation and are normalized provided $$\label{Wronskian} \mathrm{Im}\left(u_{k\ell} \dot{u}^\ast_{k\ell}\right) = 1~.$$ The particular solution of (\[modeeq\]) is fixed by demanding that the analytic continuation of the functions $u_{k \ell}$ to the entire spacetime is regular on a Cauchy surface, say, $\Sigma_0$ of Fig. \[fig:pcd\] for definiteness. In order to carry out this calculation explicitly, we make the approximation $V(R) \sim H^2 R^2 - 1$. One may notice that this corresponds to work with a $\mathbb{R} \times \mathrm{S}_2$ foliation of 4D de Sitter space (see the appendix) and one expects that it is good only if there is a separation of scales $R_c \gg R_b, R_i$. With this approximation, we have $R(\tau) = H^{-1} \cosh \tau$. We can now eliminate $R$ from (\[modeeq\]), $$\label{pseudoradial} \ddot{u}_{k\ell}(\tau) + \left[\frac{k^2 + 1/4}{\sinh^2 \tau} + \frac{\ell \left(\ell + 1\right)}{\cosh^2 \tau} - \frac{9}{4}\right] u_{k\ell}(\tau) = 0~.$$ This equation can be solved exactly by substituting $s \equiv \sinh^2 \tau$. The resulting equation is recognized as Riemann’s differential equation whose general solution can be written in terms of the hypergeometric function ${}_2F_1$: $$\begin{gathered} \label{modesol} u_{k\ell} = s^{1/4} s^{i k / 2} \sqrt{\left(s + 1\right)^{\ell + 1}}\\ \times \left[ {}_2F_1\!\left(\frac{\ell + i k}{2}, \frac{\ell + 3 + i k}{2}; 1 + i k; -s\right) A_{k\ell}\right.\\ \left.+ \left(-s\right)^{-i k} {}_2F_1\!\left(\frac{\ell - i k}{2}, \frac{\ell + 3 - i k}{2}; 1 - i k; -s\right) B_{k\ell}\right]\end{gathered}$$ for $\ell \geq 1$. Here and in the following, we place the branch cuts such that all functions are analytic in the entire upper half of the complex plane of $s$, including the real line but allowing for isolated singular points. The case $\ell = 0$ has to be treated separately, the general solution in this case is $$\begin{gathered} \label{modesol0} u_{k0} = s^{1/4} s^{i k / 2} \left[ {}_2F_1\!\left(\frac{i k - 1}{2}, 1 + \frac{i k}{2}; 1 + i k; -s\right) A_{k0}\right.\\ \left.+ \left(-s\right)^{-i k} {}_2F_1\!\left(\frac{-i k - 1}{2}, 1 - \frac{i k}{2}; 1 - i k; -s\right) B_{k0}\right]~.\end{gathered}$$ The constants of integration, $A_{k\ell}$ and $B_{k\ell}$, are fixed by the conditions of regularity and of normalization[^3]. To implement the former, we note that we can cover the static regions of RNdS (I and V in Fig. \[fig:pcd\]) by integrating $\dd \tau = H\, \dd R/\sqrt{V(R)}$ to values $R < R_c$. The imaginary part of $\tau$ then plays the role of a radial coordinate. In a RNdS spacetime, we would impose regularity at $R = R_b$, the “north pole” of the geometry. Since we are using de Sitter space as an approximation, we instead use this condition at $R = 0$, i.e. at the poles of Euclidean de Sitter space $\tau \to\pm i\pi/2$. As seen in (\[pseudoradial\]), for $\ell \geq 1$ the solutions near the poles asymptote to $u_{k\ell} \sim \rho^{\ell+1}$ or $u_{k\ell} \sim \rho^{-\ell}$, where $\tau = \pm i(\pi/2 - \rho)$. The second solution is not normalizable and is therefore excluded from the physical spectrum. Expanding the solution (\[modesol\]) around $s=-1$, one sees that it remains regular if the constants of integration fulfill the relation $$B_{k\ell} = -A_{k\ell} \frac{\Gamma\left(1 + i k\right) \Gamma\left(\frac{\ell - i k}{2}\right) \Gamma\left(\frac{\ell + 3 - i k}{2}\right)}{\Gamma\left(1 - i k\right) \Gamma\left(\frac{\ell + i k}{2}\right) \Gamma\left(\frac{\ell + 3 + i k}{2}\right)}~.$$ For $\ell = 0$, we impose $$\label{choice l=0} \frac{\dd u_{k0}}{\dd R}\Big|_{R = 0} = 0$$ such that the derivative is continuous at the poles. The constants of integration in (\[modesol0\]) are therefore related by $$B_{k0} = -A_{k0} \frac{\Gamma\left(1 + i k\right) \Gamma\left(\frac{-1 - i k}{2}\right) \Gamma\left(1 - \frac{i k}{2}\right)}{\Gamma\left(1 - i k\right) \Gamma\left(\frac{-1 + i k}{2}\right) \Gamma\left(1 + \frac{i k}{2}\right)}~.$$ Using these relations, one can show that the normalization of the Wronskian, eq. (\[Wronskian\]), implies $$\left|A_{k\ell}\right|^2 = \frac{1}{k \left(e^{2 \pi k} - 1\right)}~,$$ including the case $\ell = 0$. This fixes the vacuum mode functions up to an irrelevant overall phase. The two-point correlation function (at equal time) is finally given by $$\begin{gathered} \label{2ptfunction} \left<\Phi(\tau, 0, 0, 0) \Phi(\tau, z, \theta, 0)\right> = \\ \sum_\ell \frac{\left(2 \ell + 1\right)}{16 \pi^2} \int\!\dd k \, \frac{\left|u_{k\ell}\right|^2}{R^2 \sqrt{V(R)}} P_\ell\left(\cos \theta\right) e^{-i k z}~,\end{gathered}$$ where we used homogeneity of space to shift one of the points to the origin of the coordinate system $(z = 0, \theta = 0)$, and the rotational symmetry of $\mathrm{S}_2$ to set the azimuthal coordinate $\phi$ of the second point to zero. We call $\left|u_{k\ell}\right|^2 / R^2 \sqrt{V(R)}$ the “power spectrum”. It becomes time-independent at late time $\tau \rightarrow \infty$, $$\label{power1} \frac{\left| u_{k\ell} \right|^2}{R^2 \sqrt{V(R)}} \rightarrow \frac{H^2 \left|\Gamma\left(\frac{\ell + i k}{2}\right)\right|^2}{2 \left(k^2 + \left(\ell + 1\right)^2\right) \left|\Gamma\left(\frac{\ell + 1 + i k}{2}\right)\right|^2} \equiv \mathcal{P}_{k\ell}~,$$ for $\ell > 0$, and $$\label{power2} \frac{\left| u_{k0} \right|^2}{R^2 \sqrt{V(R)}} \rightarrow \frac{H^2 \tanh\frac{\pi k}{2}}{k + k^3} \equiv \mathcal{P}_{k0}~,$$ for $\ell = 0$. These results are most conveniently obtained by taking the limit $s \rightarrow \infty$ of eqs. (\[modesol\]) and (\[modesol0\]), and noting that $R^2 \sqrt{V(R)} \rightarrow H^{-2} s^{3/2}$. Multipole coefficients {#sec:power alm} ---------------------- The CMB anisotropies are described by the two-point correlation function on the two-dimensional intersection of the past light-cone of the observer with a constant time hypersurface. Anisotropic expansion during the matter dominated era causes an angular dependent redshift but for simplicity we shall neglect this additional corrections to the standard angular power spectrum in this section. Namely, we evaluate the two-point correlation function given in the previous section on a surface at fixed comoving distance and calculate the correlations between the multipole coefficients $\left< a_{lm} a^\ast_{l'm'}\right>$. Our approximation amounts to assume that the post-inflationary phase of expansion is isotropic, $$\label{latemetric} \dd s^2 \sim - \dd t^2 + R^2(t) \left[\dd z^2 + \dd \theta^2 + \sin^2 \theta \dd \phi^2\right]~.$$ We can choose the coordinate system so that the observer is at the origin $(z = 0, \, \theta = 0)$. The direction of observation $\mathbf{n}$ defines a point $(z = r \, \mathbf{z}\cdot\mathbf{n}, \, \theta = r\, \sqrt{1 - (\mathbf{z}\cdot\mathbf{n})^2})$ on the last scattering surface, where $\mathbf{z}$ is the unit vector pointing along the $z$-direction, and $r$ is the comoving radius of the last scattering surface, $$\label{important} r = \int_{t_{\rm rec}}^{t_0}\!\!\frac{dt}{R(t)} \simeq \frac{3.5}{\dot{R}(t_0)} = 3.5\, \sqrt{\Omega_{\rm curv}}~,$$ see the appendix for the definitions. We also record the relation between the ratio $r$ and the number of $e$-folds of inflation. Since the shear is small, the radius of the last scattering surface is in the first approximation equal to its value in isotropic cosmologies $l_{\rm LSS} \simeq 0.5 H_0^{-1}$. Let us define the origin of the number of e-folds as the instant where the radius of the compact directions $R$ is of the order of the Hubble radius today $H_0^{-1} \simeq R$. If inflation lasts $N_{\rm extra} > 0$ additional $e$-folds, this radius is $e^{N_{\rm extra}}$ larger. Hence $$\label{r} r = \frac{l_{\rm LSS}}{R} \simeq 0.5 \, e^{-N_{\rm extra}}~.$$ A long period of inflation therefore corresponds to $r \ll 1$. The correlations between the multipole coefficients are given by $$\label{correlations a_lm} \left< a_{lm} a^\ast_{l'm'} \right> = \int \!\dd \Omega \dd \Omega' Y_{lm}(\mathbf{n}) Y^\ast_{l'm'}(\mathbf{n}') \left< \Phi(\mathbf{n}, r) \Phi(\mathbf{n'}, r) \right>~.$$ Before we start to calculate, we note that some of them vanish by invariance under point-reflection, which is a symmetry of the KS model. The two-point function in the above expression is indeed even under parity and the correlations between multipole coefficients of opposite parity, i.e. such that $l+l'$ is odd, vanish identically. Parity is also a symmetry of the Bianchi III models considered in [@BPS] and [@GHR], which explains why they also find no correlations for the multipole coefficients where $l+l'$ is odd. We now introduce intrinsic spherical coordinates $(\vartheta, \varphi)$ on the last scattering surface. For simplicity, we choose the polar axis aligned with the $z$-direction such that $\mathbf{z}\cdot\mathbf{n} = \cos\vartheta$. As we have already seen in the previous section, homogeneity of space guarantees that the two-point function in the above equation only depends on two parameters: the separation along the flat direction $z$, given by $r (\cos \vartheta - \cos \vartheta')$, and the angle $\theta$ subtended by an arc on the compact dimensions, that is $\cos \theta = \cos (r \sin \vartheta) \cos (r \sin \vartheta') + \sin (r \sin \vartheta) \sin (r \sin \vartheta') \cos (\varphi - \varphi')$, as some simple geometric considerations show. In order to simplify our expressions, we will use the identity $$P_\ell\left(\cos \theta\right) = \frac{4 \pi}{2 \ell + 1} \sum_{n = -\ell}^{\ell} Y^\ast_{\ell n}(r \sin \vartheta, \varphi) Y_{\ell n}(r \sin \vartheta', \varphi')~.$$ Using eq. (\[2ptfunction\]), the multipole correlations in terms of these intrinsic coordinates become $$\begin{gathered} \left< a_{lm} a^\ast_{l'm'} \right> = \int \!\dd \Omega \dd \Omega' \sum_\ell \int\!\dd k Y_{lm}(\vartheta, \varphi) Y^\ast_{l'm'}(\vartheta', \varphi') \\ \times \frac{\mathcal{P}_{k\ell}}{4 \pi} \sum_{n = -\ell}^{\ell} Y^\ast_{\ell n}(r \sin \vartheta, \varphi) Y_{\ell n}(r \sin \vartheta', \varphi')\\ \times e^{-i k r \left(\cos \vartheta - \cos \vartheta'\right)}~.\end{gathered}$$ Since KS spacetime is axially symmetric around the $z$-axis, and since our choice of intrinsic coordinates respects that symmetry, the coefficients $a_{lm}$ with different $m$ are uncorrelated. One can show this explicitly by integrating out the angles $\varphi$ and $\varphi'$, $$\begin{gathered} \label{final-alm} \left< a_{lm} a^\ast_{l'm'} \right> = \delta_{mm'} \sqrt{2 l + 1} \sqrt{2 l' + 1} \sqrt{\frac{\left(l - m\right)!}{\left(l + m\right)!} \frac{\left(l' - m\right)!}{\left(l' + m\right)!}} \int_0^\pi\!\sin\vartheta \dd \vartheta \int_0^\pi\!\sin\vartheta' \dd \vartheta' \int\!\dd k P_{lm}(\cos\vartheta) P_{l'm}(\cos\vartheta')\\ \times \sum_\ell \mathcal{P}_{k\ell} \frac{2 \ell + 1}{16 \pi} \frac{\left(\ell - m\right)!}{\left(\ell + m\right)!} P_{\ell m}\Bigl(\cos (r \sin \vartheta)\Bigr) P_{\ell m}\Bigl(\cos (r \sin \vartheta')\Bigr) e^{-i k r \left(\cos \vartheta - \cos \vartheta'\right)}~.\end{gathered}$$ We calculate the remaining coefficients $\left< a_{lm} a^\ast_{l'm'} \right>$ numerically. We compare these results with the statistically isotropic case of a flat Friedman-Robertson-Walker universe with a Harrison-Zel’dovich spectrum $$\label{FRWresult} \left< a_{lm} a^\ast_{l'm'} \right>_\mathrm{iso} = \frac{H^2}{2 \pi} \, \frac{1}{l \left(l + 1\right)}\, \delta_{ll'} \delta_{mm'} ~.$$ This expression is obtained by computing the vacuum expectation value in the Bunch-Davies vacuum using the flat slicing of de Sitter space. We present our results in the following form $$\begin{gathered} \label{almvariance} \left< a_{lm} a^\ast_{l'm'} \right> = \frac{H^2}{2\pi} \delta_{mm'} \left(\delta_{ll'} + \delta C_{ll'mm}\right) \\ \times \max\left\{\frac{1}{l \left(l + 1\right)}, \frac{1}{l' \left(l' + 1\right)}\right\}~,\end{gathered}$$ and give the coefficients $\delta C_{ll'mm}$ in table \[tab:numerics\] for the case $r=0.5$, which corresponds to $\Omega_{\rm curv} \simeq 0.02$, see eq. (\[important\]). The scaling of the coefficients $\delta C_{ll'mm}$ with $r$ is illustrated in Fig. \[fig:scaling\]. Our findings are the following. First, the amplitude of the corrections is suppressed by $r^2 \propto \Omega_{\rm curv}$. In the limit of a long period of inflation, $r \rightarrow 0$, i.e., when the observer only has access to scales much smaller than the radius of $\mathrm{S}_2$, the anisotropies become unobservable. Second, as already pointed out the correlations vanish by parity for $l+l'$ odd. This result does not depend on our choice of coordinate system. Third, the amplitude of the corrections is largest for $l=l'$, and generally decreases with growing $\left|l - l'\right|$. It is also noteworthy that the amplitude, at fixed $\left|l - l'\right|$, is almost independent of $l$, which means that the correlations extend uniformly up to arbitrarily high multipoles. Fourth, for $l=l'$, the $m$-dependence of the corrections obeys $$\label{parabola} \delta C_{llmm} = \delta C_{ll00} \left(1 - \frac{3 m^2}{l \left(l + 1\right)}\right)$$ to a good approximation. As a result, the corrections cancel in the averaged multipole $$\label{C_l} C_l \equiv \sum_m \frac{\left<a_{lm} a^\ast_{lm}\right>}{2 l + 1} = \frac{H^2}{2\pi } \frac{1}{l(l+1)}~.$$ Fifth, from eq. (\[important\]) we deduce $$\label{amplitude} \delta C_{ll00} \simeq \Omega_{\rm curv}\, .$$ A mild dependence on $l$ of the prefactor is possible. We emphasize two remarkable properties of these spectra. The first one is the scale invariance of the power spectrum, and in particular the absence of an infrared cutoff which one could have naively expected from the asymptotic form of the power spectrum (\[power1\]) and (\[power2\]) at $k\to 0$ or by analogy with closed FRW models. The second is the fact that, as we already pointed out, at a given value of $l$, the anisotropy manifests itself by a redistribution of the power amongst the various $a_{lm}$ according to the “sum rule” $$\label{sum rule} \sum_m \delta C_{llmm} = 0\, .$$ In sec. \[sec:anomalies\], we show how eqs. (\[parabola\]) and (\[sum rule\]) translate into observables (the so-called bipolar coefficients). Equations (\[parabola\])-(\[sum rule\]) and table \[tab:numerics\] are the central results of this section.[^4] ![\[fig:scaling\] Amplitude of the anisotropic correction to the multipole correlators for different values of $r$ (three examples). We found a scaling consistent with $\delta C_{ll'mm'} \propto r^2$, as indicated by the dashed lines.](dCscaling.ps){width="85mm"} $\delta C_{ll'mm}$ $m = 0$ $m = 1$ $m = 2$ $m = 3$ -------------------- ----------------- ----------------- ----------------- ----------------- $l = 2$, $l' = 2$ [$+0.02(40)$]{} [$+0.01(16)$]{} [$-0.02(48)$]{} — $l = 2$, $l' = 3$ [$0$]{} [$0$]{} [$0$]{} — $l = 2$, $l' = 4$ [$-0.005(4)$]{} [$-0.005(0)$]{} [$-0.003(3)$]{} — $l = 3$, $l' = 3$ [$+0.02(29)$]{} [$+0.01(62)$]{} [$-0.00(04)$]{} [$-0.02(69)$]{} $l = 3$, $l' = 4$ [$0$]{} [$0$]{} [$0$]{} [$0$]{} : \[tab:numerics\] Relative deviations from a flat Friedman-Robertson-Walker cosmology of the multipole correlations. The numerical results are for $r = 0.5$. Solid convergence was achieved for the leading digit only. We found that for the nonzero entries $\delta C_{ll'mm'}$ is roughly proportional to $r^2$, cf. Fig. \[fig:scaling\]. In [@BPS], a table similar to our table \[tab:numerics\] can be found for a Bianchi III model. The deviations due to anisotropy are controlled in that case by the comoving radius of the last scattering surface, $\rho_\star$ in their notation, and scale like $\rho_\star^2 \propto \Omega_{\rm curv}$. Their table and ours are remarkably similar[^5], the only difference appears to be the overall sign of the corrections. It is tempting to conjecture that this is related to the opposite sign of the curvature of Bianchi III models as compared to KS models. Note as well that eq. (\[parabola\]) then should hold for both cases, as seems to be implied also by Fig. 3 of [@BPS]. The resulting cancellation of the corrections when taking the sum over $m$ was, however, not reported in [@BPS]. Corrections to the Sachs-Wolfe effect {#sec:rafinements} ------------------------------------- Our analysis of anisotropic signatures was simplified in two respects. First, we only consider a test field. Without a detailed analysis, it is hard to determine whether the theory of cosmological perturbations in KS spacetime would give primordial spectra qualitatively different from the ones of perturbed FRW spacetimes, besides the ones already found for test scalar fields. We leave this question for future work. Our second simplification is that we neglected the anisotropic redshift after last scattering. In a KS spacetime, this effect is controlled by the ratio of the shear, $\delta$, to the isotropic expansion rate $H$ (see the appendix for the definitions). This ratio decreases during inflation but grows again during radiation and matter domination, so that it becomes significant only at small redshifts $z = \mathcal{O}(1)$. These late time effects are thus determined by the value of the ratio $$\sigma = \frac{\delta_0}{H_0}\,\,.$$ To get a handle on the additional corrections brought by the anisotropic redshift of photons since last scattering, note that this effect has two origins in our model. One is caused by the anisotropic expansion and the second by the curvature of the homogeneous surfaces. To see the effect of the anisotropic expansion alone, consider a Bianchi I model $ds^2_{\rm B \, I} = -dt^2 + a^2(t)dx^2 + b^2(t)(dy^2 + dz^2)$. The null geodesic equation is readily solved and gives the temperature distribution in the direction $(\vartheta,\varphi)$ $$\label{ellipse} T_{\rm B \, I} (\vartheta,\varphi,t_O) = T_{\rm em} \, \left( \frac{x^2}{a_{\rm em}^2} + \frac{y^2 + z^2}{b_{\rm em}^2} \right)^{-1/2}$$ where $T_{\rm em}$ is the temperature of the Planck spectrum at the time of emission, and the cartesian coordinates $(x,y,z)$ are related to the intrinsic coordinates on the unit sphere by $x=\sin \vartheta \cos \varphi$, $y=\sin \vartheta \sin \varphi$, and $z=\cos \vartheta$. The quadrupole character of this distribution is obvious (recall that when written in cartesian coordinates, the spherical harmonics of weight $l$ are homogeneous polynomials of degree $l$). We conclude that the anisotropic Sachs-Wolfe effect takes the form $$\label{quadmodulation} \left( \frac{\delta T}{T}\right)_{\rm B \, I} \sim \frac{\Psi}{3} \left( 1 + \sum_{m=-2}^{m=2} \mathcal{O}(\sigma) Y_{2m}(\vartheta,\varphi) \right) + \ldots$$ where $\Psi$ is the gravitational potential in the Newtonian gauge and the ellipsis stands for the corrections from the velocity potential of matter and the integrated Sachs-Wolfe term, both of which receive similar quadrupolar [*modulations*]{}. Inserted into the expression (\[correlations a\_lm\]) of the correlation functions, this quadrupole generates additional correlations between the $\left< a_{lm} a^\ast_{l'm'} \right>$ with $l' = l, l\pm 2$. The non Euclidean character of the surface of homogeneity in Kantowski-Sachs spacetimes is responsible for additional corrections to the geodesic equation. As a result, the temperature distribution cannot be written in closed form as in the simple expression (\[ellipse\]). Nevertheless, inspection of the geodesic equations is sufficient to show that these corrections scale with $\sigma$ as one could expect. Indeed, the null component of the momentum is given by $$\frac{dp_0}{d\lambda} = \delta \left\{ - 2 e^{4\Delta} \bar p_z^2 + e^{-2\Delta} \left( \bar p_\theta^2 + \frac{\bar p_\phi^2}{\sin^2 \theta} \right)\right\}$$ where $\Delta = \int^t\!\!\delta$ and $\bar p_{z,\theta,\varphi}$ are integration constants. A detailed calculation of the bolometric flux, which includes all of these corrections, was done in [@GHR] for Bianchi III spacetimes. It can easily be adapted to the Kantowski-Sachs spacetime and confirms the qualitative results we have just derived. Moreover, it is noteworthy that the diagonal corrections $\delta C_{llmm}$ retain the same parametric form (\[parabola\]) and therefore verify the sum rule (\[sum rule\]), as can be seen from eq. (47) of [@GHR], with the use of their equations (B1) and (B3). Since we expect that both early and late modifications of the power spectra add at leading order in $\Omega_{\rm curv}$, the total correction should therefore also verify eqs. (\[parabola\]) and (\[sum rule\]). CMB anomalies {#sec:anomalies} ------------- We now ask whether the correlations of the multipole coefficients we have computed in the previous sections can account for some of the anomalies found in the CMB (see [@WMAP7anomalies; @Copi:2010na] for recent reviews and references). We can, in fact, discuss both Kantowski-Sachs (KS) and Bianchi III (BIII) models because they have important features in common and therefore make very similar predictions. Moreover, although we have not calculated explicitly the corrections from the recent history of the universe, we saw that the parametric form eq. (\[parabola\]) holds for both, and therefore for the total correction of the CMB spectra. Here we make no statement regarding either the cold spots, which would require an analysis far beyond the scope of this paper, or the alignment of the dipole and quadrupole with the ecliptic, which is unlikely to have a cosmological explanation. Also, from the scale invariance of the primordial power spectrum, and in particular the absence of a cutoff (see eq. (\[C\_l\])) we conclude that the KS and BIII models seem unable to explain the lack of power at large scales (of the quadrupole and of the correlation function) and the disparity between odd and even values of $l$. However, the quadrupole receives a non-stochastic contribution in anisotropic models which, in a complete calculation, may have a number of different sources (see sec. \[sec:bound\] for further details.) The KS and BIII models can, however, in principle account for the alignment of the quadrupole and octopole and the quadrupolar power asymmetry. A statistical analysis reveals that the effect of anisotropy on the alignment is not significant. On the other hand, we find good indications that the quadrupolar asymmetry can be explained by the KS model, but not by BIII. However, as we now explain the amount of anisotropic curvature needed to produce a signal with the observed amplitude, $\Omega_{\rm curv} \simeq 10^{-2}$, is probably inconsistent with the non-detection of a CMB quadrupole at the same level. ### The quadrupole moment and the bound on curvature {#sec:bound} In anisotropic models, the late-time expansion couples multipoles $l$ and $l+2$ [@GHR]. The quadrupole thus receives a [*non-stochastic*]{} contribution from the monopole proportional to $T_0 \Omega_{\rm curv}$. Given a model with a scale invariant power spectrum of amplitude $\delta T / T_0 \sim 10^{-5}$, and in the absence of other sources of quadrupolar anisotropy, the value of anisotropic curvature is thus expected to be [@boundOmega] $$\label{bound} \Omega_{\rm curv} \lesssim 10^{-4}~.$$ For the purpose of model building we note that, through the relationships (\[important\]) and (\[r\]), it implies a lower bound on the duration of inflation, $N_{\rm extra} \gtrsim 2.7$. Moreover, we will see that this bound, if correct, is too stringent for the model to account for the quadrupolar asymmetry in the CMB. It is, however, important to stress that this bound is a rough estimate obtained under the assumption of adiabatic perturbations [@boundOmega]. A detailed calculation might show this estimate to be too naive, since compensations might occur between the Doppler, proper Sachs-Wolfe and integrated Sachs-Wolfe terms in a KS and BIII universe, as is the case in, e.g., open inflation scenarios [@GZopen1; @GZopen2]. Moreover, a mixture of adiabatic and isocurvature perturbations presumably also modifies the bound (\[bound\]). Thus one must bear in mind that eq. (\[bound\]) depends on the details of the model and might be too stringent. ### Alignment of the quadrupole and octopole We study the problem of the alignment with the help of the so-called multipole vectors [@multipolevectors]. The two quadrupole vectors define a plane whose normal $\mathbf{v}$ is obtained by taking the cross-product of the two vectors. Similarly, all possible pairings of the three octopole vectors define three different planes, with normals $\mathbf{w}^{\left(1\right)}$, $\mathbf{w}^{\left(2\right)}$, $\mathbf{w}^{\left(3\right)}$. One says that the octopole is planar when the three $\mathbf{w}^{\left(i\right)}$ are almost collinear, and one says that the quadrupole and octopole are aligned when they are also almost collinear with $\mathbf{v}$. The degree of alignment is measured by the statistics $$S \equiv \frac{1}{3} \sum_i |\mathbf{v} \cdot \mathbf{w}^{\left(i\right)}|$$ A large value of $S$ indicates an alignment. The observed value (taken from the Doppler-corrected ILC maps, see [@Sarkar:2010yj]) is $S_{\rm ILC7} = 0.736$. [^6] For a Gaussian and statistically isotropic distribution of $a_{lm}$ we found that only about $1\%$ of the realizations have a value of $S$ larger than $S_{\rm ILC7}$. According to [@WMAP7anomalies], this “remarkable degree of alignment” lacks a compelling theoretical explanation. Unfortunately, we will now show that the class of KS and BIII models do not provide one. For each model and each set of parameters, we generated $10^{6}$ independent realizations of a quadrupole and octopole. To this end, the coefficients $a_{lm}$ are treated as independent (remember that the correlations between quadrupole and octopole vanish) Gaussian random variables with zero mean[^7] and variance $\left< a_{lm} a^\ast_{lm} \right>$ given in eq. (\[almvariance\]). For each realization, we calculate the corresponding multipole vectors from the coefficients $a_{lm}$ using the algorithm of [@Weeks:2004cz], and compute from them the value of $S$. We set $\delta C_{3300} \simeq \delta C_{2200}$, as indicated from our table \[tab:numerics\], and with the use of (\[parabola\]) we then study how the distribution of $S$ varies as a function of the single parameter $\delta C_{2200} \simeq\Omega_{\rm curv}$. For both KS and BIII spacetimes (that is for either sign of $\delta C_{2200}$), the mean of the distribution shifts towards larger values, indicating a tendency of the quadrupole and octopole to align. However, the distribution of $S$ turns out to be fairly robust. In particular, the mean value and variance of $S$ change very little. For instance, in KS the mean value of $S$ increases by only $\sim 1\%$ for $\delta C_{2200} \simeq 0.4$. The effect is about twice as large for BIII. The skewness depends more strongly on $\delta C_{2200}$, and in particular on its sign. It increases for $\delta C_{2200} < 0$ (BIII) and decreases for $\delta C_{2200} > 0$ (Kantowski-Sachs). For $\delta C_{2200} \simeq \pm 0.2$ the skewness changes by $\sim 20 \%$ in both cases. As a result of the increase of both the mean and skewness, the number of realizations which exceed the threshold $S_{\rm ILC7}$ in BIII is increased by $\sim (5 \%, 18 \%, 62 \%)$ for $\delta C_{2200} \simeq (-0.1, -0.2, -0.4)$, while it hardly changes for KS. The number of realizations nevertheless remains of the same order of magnitude as in the isotropic case. In conclusion, we find that the modified multipole correlations in KS and BIII are unable to account for the alignment of the quadrupole and octopole at a statistically significant level. ### Quadrupolar power asymmetry A practical way to check whether the two-point correlation function ${\cal C}(\mathbf{n},\, \mathbf{n}')$ is invariant under rotation is to consider its representation in the basis of the total angular momentum operator of eigenvalues $L$ and $M$ (sometimes referred to as bipolar spherical harmonics [@Hajian:2003qq]). Using the ‘bra-ket’ notation, we have $$\label{totangmoment} \langle \mathbf{n}, \mathbf{n}' \,\vert \, {\cal C} \rangle = \sum_{l,l',L,M} N^L_{ll'} A^{LM}_{ll'} \, \langle \mathbf{n}, \mathbf{n}'\, \vert l l'; L M \rangle~,$$ where we follow the normalization convention chosen by the WMAP team [@WMAP7anomalies] in order to facilitate comparison: $$N^L_{ll'} \equiv \sqrt{\frac{\left(2 l + 1\right) \left(2 l' + 1\right)}{2 L + 1}} \, \langle l, 0, l', 0 \vert l l';L 0\rangle~.$$ The $A^{LM}_{ll'}$ are the sum over $m$ and $m'$ of the $\left< a_{lm} a^\ast_{l'm'} \right>$ weighted by the appropriate Clebsch-Gordan coefficients. Since the eigenstates $\vert l l'; L M \rangle$ generate a $(2 L+1)$-dimensional representation of the rotation group, the condition that ${\cal C}$ is statistically isotropic reads $$\begin{gathered} {\cal C}({\cal R}\mathbf{n},\, {\cal R}\mathbf{n}') = {\cal C}(\mathbf{n},\, \mathbf{n}') \quad \forall \, \mathcal{R} \in SO(3) \\ \quad \Leftrightarrow \quad A^{LM}_{ll'} = 0 \quad \forall \, L>0 ~, \end{gathered}$$ where ${\cal R}$ is an arbitrary rotation. If some of the $A^{LM}_{ll'}$ are non-zero for $L=1$, the CMB temperature presents a dipole modulation. If some coefficients are non-zero for $L=2$, the temperature is modulated by a quadrupole as, for example, we found for the late time effects eq. (\[quadmodulation\]) (see the appendix of [@WMAP7anomalies] for other illustrations). Note that the $A^{LM}_{ll'}$ are in general not independent of the choice of a coordinate system. The WMAP team uses a preferred coordinate system in which all coefficients with $M \neq 0$ are consistent with zero. In KS spacetime this coincides with the coordinate system we use in our calculations. With the use of eq. (\[parabola\]) we find that $A^{00}_{ll} = C_l$ and $$\label{modulation} A^{20}_{ll} = \frac{C_l}{\sqrt{5}} \left(4 - \frac{3}{l \left(l + 1\right)}\right) \delta C_{ll00}~.$$ There is no dipole modulation because of the parity symmetry of the background which entails vanishing correlations for $l' = l \pm 1$. The important remark concerning (\[modulation\]) is that these coefficients are proportional to $C_l$. Therefore they should exhibit acoustic oscillations as well. The WMAP team found that $A^{20}_{ll} \approx -2 A^{20}_{l-2,l}$. As illustrated in the appendix of [@WMAP7anomalies], this relation occurs when the anisotropy does not induce any correction to the average power spectrum but only redistributes the power amongst the $a_{lm}$. So we expect to find it in our model. Having calculated the correlations only up to $l=4$, we can only check the relation for that value. From our table \[tab:numerics\] we obtain $A^{20}_{24} \simeq -0.001\times H^2 / (2 \pi)$. With $\delta C_{4400} \simeq \delta C_{3300}$ we find indeed that $A^{20}_{44} \approx -2 A^{20}_{24}$. In conclusion, the anisotropy in the KS and BIII models shows all of the characteristic features found in the WMAP data. In particular, the appearance of the acoustic peak in the $A_{ll}^{20}$ naturally follows from such models. The predictions of the models however differ in the [*sign*]{} of the bipolar coefficients $A^{20}_{ll}$ which is determined by the sign of the curvature. We find that KS predicts the correct sign. Unfortunately, if the bound on anisotropic curvature (\[bound\]) is correct, the predicted amplitude is two orders of magnitude too small. Indeed, from Fig. $16$ of [@WMAP7anomalies], we get $A^{20}_{ll}/C_l \simeq 4 \cdot 10^{-2}$, which according to expression (\[modulation\]) corresponds to $\delta C_{ll00}$, and therefore $\Omega_{\rm curv}$, of the order of a few percents. However, as we recalled in section \[sec:bound\] the anisotropic curvature has another effect: it adds a non-stochastic component to the quadrupole of the order of $T_0 \Omega_{\rm curv}$ [@boundOmega; @GHR], which leads to the bound $\Omega_{\rm curv} \lesssim 10^{-4}$. With a value of the anisotropic curvature in that range, the quadrupolar anisotropy cannot be accounted for by the model. Summary and conclusions {#sec:conclusions} ======================= In the landscape of string theory, vacua with differing numbers of macroscopic dimensions and topologies are connected by tunneling events. In this paradigm, our universe has been produced as the last step in a potentially long chain of events, by tunneling from a parent vacuum with a greater, equal, or smaller number of macroscopic dimensions. Like [@GHR; @BPS] which appeared when our work was near completion, we consider the latter case. Our approach differs in the specific model under examination and in its embedding into a cosmological scenario, which we dubbed the “shapeshifting universe”. Specifically, we take the example of a spacetime where two of our current macroscopic dimensions are compactified on a two-sphere by a magnetic flux. The existence of a long-lived parent vacuum with two macroscopic dimensions, while all others are stably compactified, is not a necessary requirement for this scenario. All we need is a precursor with two of our four large dimensions compactified while other, currently compact, directions may well have been macroscopic. This allows a direct connection of anisotropic cosmologies produced by decompactification with the vast possibilities of transdimensional vacuum transitions described in [@CJR; @BSV] and, furthermore, admits histories with greatly varying effective values of $\Lambda$. In one particular possible scenario, our parent vacuum was a $\mathrm{dS}_D \times \mathrm{S}_2$ spacetime with higher $\Lambda_{\rm eff}$ and different compact directions than our universe, connected to ours by a combined spontaneous compactification of $D-2$ dimensions which are now small and decompactification of the $\mathrm{S}_2$. We have not explicitly constructed the instanton describing this shapeshifting event but we know of no reasons why it shouldn’t exist. It is also plausible that models can be constructed where inflation is triggered by the transition, as described in [@CJR]. This line of investigation raises a number of new and interesting questions. In the context of the cosmological measure problem, the contribution of regions with different numbers of macroscopic dimensions and transitions in between is still largely unexplored (see [@CJR; @BSV; @BPS] for some ideas). We hope to return to this question in future work. From the phenomenological point of view, on the other hand, decompactification offers a concrete framework for parameter studies of anisotropic cosmological models. Our ability to detect statistical anisotropy hinges on the assumption of a sufficiently short period of inflation, which again may find some justification in the string theory landscape. Given the observational constraints, a small but nonzero window for anisotropic curvature remains (see, e.g., [@GHR]). Precisely how strong the constraints already are, and how much can be gained by combining future large-scale galaxy surveys and CMB data, can only be answered by a more detailed investigation of the cosmological signatures of models like the one proposed here. It is also interesting to note that having embedded the anisotropic universe model in a more complete scenario of quantum tunneling, where the KS spacetime corresponds to only a wedge of a large spacetime, solves the ambiguity in the ground state (compare with the discussions of [@GCP; @PPU2]). Our preliminary analysis of the deviations from statistical isotropy of the CMB shows the following trends. Some of the off-diagonal correlations $\langle a_{lm} a_{l'm'}^* \rangle$, which vanish in the isotropic case, are now non-zero. Because of the invariance under point reflection of the background, the correlations with $l$ and $l'$ of different parities vanish. The nonzero correlations receive contributions from both the inflationary phase and the recent expansion of the universe. Both scale as $\Omega_{\rm curv}$ in the observations. The central results are the table \[tab:numerics\] and the equations (\[parabola\]) and (\[sum rule\]). We showed that as a direct consequence of this, and in contrast to Bianchi III models, the Kantowski-Sachs models predict a quadrupolar modulation in the CMB with all the features observed in the data, namely the sign of the bipolar coefficients $A_{ll'}^{LM}$, their acoustic oscillations, and a particular relation between them. To obtain the observed amplitude requires, however, an anisotropic curvature at a level of $\Omega_{\rm curv} \simeq 10^{-2}$, which is incompatible with the bound derived from the value of the quadrupole, namely $\Omega_{\rm curv} \lesssim 10^{-4}$. We briefly argued that this bound could be mitigated. The price to pay is unfortunately an increased complexity of the model, which we leave for future work. We conclude that the model in its present, most simple, form is unable to account for any of the anomalies. As an additional line for future investigation we mention the generation of primordial magnetic fields. Since KS spacetimes are not conformally flat, we indeed expect that long wavelength magnetic fields are produced from vacuum fluctuations during inflation, without recourse to noncanonical couplings of masses. This generation could perhaps be sufficient to explain the primordial magnetic fields necessary to seed galactic dynamos. We thank Michael Salem and Matthew Johnson for valuable comments on the first preprint of this article. JCN would also like to thank Raphael Bousso, Ben Freivogel, Roni Harnik, Matthew Johnson, Antony Lewis, and I-Sheng Yang for helpful discussions, and acknowledges the hospitality of the Berkeley Center for Theoretical Physics and the Aspen Center for Physics where many of them took place. We further wish to thank the anonymous referee for useful comments and Aleksandar Rakić for some insights on the CMB anomalies. The work of JA was supported by the German Research Foundation (DFG) through the Research Training Group 1147 “Theoretical Astrophysics and Particle Physics”. Kantowski-Sachs spacetimes {#app:KS} ========================== Geometry -------- We collect some useful properties of Kantowski-Sachs (KS) spacetimes [@KS]. They are (locally) homogeneous, that is they are invariant under an isometry group $G$ which acts transitively on the spacelike hypersurfaces (i.e. for two points $x$ and $y$, there exists an isometry $g \in G$ such that $g \circ x = y$). The group has four parameters and admits a three-parameter subgroup isomorphic to $SO(3)$. The Killing vectors therefore verify the Lie algebra $\left[ X_i, \, X_j \right] = \epsilon_{ijk} X_k$ with $i=1,2,3$ and $\left[ X_4,\, X_i \right] = 0$. Moreover there exists locally a coordinate system $(t,z,\theta,\phi)$ which diagonalizes the metric, $$\label{metricKS} ds^2 = - dt^2 + a^2(t) dz^2 + b^2(t) \left( d\theta^2 + \sin^2\theta\, d\phi^2 \right) ~,$$ where $z$ is the coordinate of the flat direction, and $(\theta,\phi)$ are the intrinsic coordinates on a two-sphere. Then $X_4 = \partial/\partial z$ and $X_i$ are the familiar generators of the rotation group. Because of the spherical symmetry, and since it is not conformaly flat, KS is of Petrov type D. The principal null directions are in the plane $(t,z)$. It is good to make a pause here and compare with the perhaps better known Bianchi classes. Bianchi class I, V, VII and IX have a universal covering with constant spatial curvature. Bianchi class I is probably the most studied anisotropic space to date because it has locally Euclidean spatial sections [@DulaneyGresham; @PPU1; @PPU2; @GCP]. Classes II, IV and VIII do not have an obvious cosmological application. Finally class III resembles closest KS spacetimes. The spatial metric is $$ds_{\rm B \, III}^2 = -dt^2 + a^2(t) dz^2 + b^2(t) \left( d\chi^2 + \sinh^2\chi \, d\phi^2 \right) ~.$$ Like KS spacetime, the orbits are two dimensional. The only difference seems to be the sign of the curvature of the latters, the topology of the spacelike hypersurfaces being $\mathbb{R} \times \mathrm{H}_2$. But the metric is deceptive for one can show that the isometry group of Bianchi III admits a second three parameter subgroup whose orbits are three dimensional, which is not the case if the two dimensional orbits have a positive curvature (see in particular [@KSalgebra] for details). In that sense, Bianchi III and KS spacetimes form two distinct classes. Note finally that Bianchi III spaces can be compactified by identifying points on the spacelike surfaces with respect to a discrete subgroup of $G$, yielding a multiconnected space, for instance a torus [@Fagundes]. Evolution --------- Returning to KS spacetimes, the spatial sections are described by the Ricci scalar ${}^3 \mathcal{R} = 1/b^2$ and the intrinsic curvature tensor is $K^{i}_{\,\,j} = \frac{1}{2} h^{ik} \partial_t h_{kj} = {\rm diag}\left( \alpha, \beta, \beta \sin^2\theta \right)$, with $\alpha = \dot a/a$ and $\beta = \dot b / b$. Its trace, the local expansion rate in the frame (\[metricKS\]), is $K = \gamma^{ij} K_{ij} = \alpha + 2 \beta$. The shear, $\sigma^{i}_{\,\,j} = K^{i}_{\,\,j}-\delta^{i}_{\,\, j} K/3 = \frac{1}{3}{\rm diag}\left( - 2\delta, \delta, \delta \right)$ with $\delta = \beta - \alpha$, characterizes the anisotropy of the expansion. Einstein’s equations with a perfect fluid are $$\begin{aligned} &&2\alpha \beta + \beta^2 + \frac{1}{b^2} = T_{t t} = \rho \\ &&2 \frac{\ddot b}{b} + \beta^2 + \frac{1}{b^2} = - T_{z z} = - p_z \\ &&\frac{\ddot a}{a} + \frac{\ddot b}{b} + \alpha \beta = - T_{\theta \theta} = - T_{\phi \phi} = -p_\theta\end{aligned}$$ and the Bianchi identities are $$\begin{aligned} \label{Bianchy id} \dot \rho + K \rho + \alpha p_z + 2 \beta p_\theta = 0 \, .\end{aligned}$$ We can combine the first three equations to obtain equations for the local expansion rate and shear $$\begin{aligned} \label{eq1} &&\frac{1}{3} \left( K^2 - \delta^2 \right) + \frac{1}{b^2} = \rho \\ \label{eq2} &&\dot K + \frac{K^2}{3} + \frac{2}{3} \delta^2 = - \frac{1}{2} (\rho + P) \\ \label{eq3} &&\dot \delta + K \delta + \frac{1}{b^2} = 0 \end{aligned}$$ where we note $P = p_z + 2 p_\theta$. The autonomous system formed by these equations has been analyzed in [@KSsing] and the possible singularities were described. Assuming the equation of state $p_i \simeq - \rho$, the case relevant for our inflationary model is a pancake singularity, i.e.  $a \to 0$ and $b \to {\rm cte}$ for $t \to - \infty$, where only the non compact direction grows in the first stage dominated by curvature. The physical variables behave as ${}^3\mathcal{R} \sim K^2/3 \sim \delta^2/3 \sim \rho$. Once inflation starts, the curvature term $1/b^2$ in equation (\[eq1\]) becomes negligible and the universe becomes effectively isotropic and asymptotes to de Sitter space, $a \sim b \sim e^{Ht}$. To see this, use (\[eq1\]) to eliminate $\delta^2$ from (\[eq2\]) and write $K = 3\dot\Omega/\Omega$ to get $\ddot \Omega / \Omega = 2 \rho /3 \simeq {\rm cte}$. Hence, $K \simeq {\rm cte} = 3H$ and $\Omega \simeq \Omega_0 \cosh \left( H (t-t_0)\right)$. Finally substitution of this solution in (\[eq3\]) or (\[eq1\]) gives $\delta \propto e^{-2Ht}$. In view of this solution, it is reasonable to expect that inflation is an attractor solution as in isotropic cosmologies. The asymptotic behavior translates as follows in the coordinate system (\[RNdSmetric\]). For large enough $R$, $f(R) \simeq 1 - \Lambda R^2 / 3$ becomes a good approximation and (\[RNdSmetric\]) coincides with the static de Sitter metric $$\begin{gathered} \label{dSmetric} \dd s^2 = -\left(1 - H^2 R^2\right) \dd t_{\rm s}^2 + \left(1 - H^2 R^2\right)^{-1} \dd R^2\\ + R^2 \left[\dd \theta^2 + \sin^2 \theta \dd \phi^2\right]~,\end{gathered}$$ de Sitter space indeed admits a foliation $\mathbb{R} \times \mathrm{S}_2$ obtained simply by extending (\[dSmetric\]) to the region where $R$ is timelike, $R > R_c \equiv H^{-1}$, followed by an obvious change of coordinates which puts then the metric in the form $$\dd s^2 = - \dd t^2 + \frac{\sinh^2 Ht}{H^2} \dd z^2 + \frac{\cosh^2 Ht}{H^2}\left(\dd \theta^2 + \sin^2 \theta \dd \phi^2\right) \, .$$ Cosmological models based on de Sitter space are usually studied in foliations of a Friedman-Robertson-Walker type, i.e. which are both homogeneous and isotropic, either closed, open, or flat. This choice is not available in the spacetime of black hole pairs because symmetry dictates that homogeneous fields are functions of $R$, and therefore homogeneous slices are slices of constant $R$. From the previous results, we can assume that reheating is an isotropic process, and therefore the CMB is such that $p_z = p_\theta = \rho/3$. The equation of conservation of the energy (\[Bianchy id\]) is then solved trivially $\rho \propto \Omega^{-4}$. From Stefan’s law we get that a surface of constant temperature is $$T = {\rm cte} \quad \Leftrightarrow \quad \Omega = {\rm cte} \quad \Leftrightarrow \quad t= {\rm cte}$$ Since $\delta \ll 1$, eqs. (\[eq1\]) and (\[eq2\]) with the terms $\delta^2$ neglected are the familiar Friedman equations. For an equation of state $p=(\gamma - 1)\rho$ we thus get in the first approximation $a \sim b \sim t^{2/3\gamma}$. The general solution of equation (\[eq3\]) is $$\begin{aligned} \label{sol delta} \delta &=& \delta_0 \left( \frac{\Omega_0}{\Omega}\right)^3 + \frac{1}{ab^2}\int_{t_0}^{t}\!\!dt'a(t') \nonumber \\ &\sim& \frac{3\gamma}{3\gamma+2}t^{1- 4/3\gamma}~.\end{aligned}$$ The second line is obtained after substituting the solutions of $a$ and $b$ previously found and noticing that the homogeneous solution is subleading. We could then use (\[sol delta\]) to iterate an asymptotic expansion for $a$, $b$ and $\delta$. Thus during matter domination, the ratio of the shear to the isotropic expansion rate grows like $$\label{delta/K} \frac{\delta}{H} \sim t^{2/3} \sim \Omega~.$$ This can be reexpressed in terms of the curvature parameter defined by $$\Omega_{\rm curv} = \frac{1}{b^2 H^2} \sim t^{2/3}~.$$ The relation with the parameter $r$ of eq. (\[important\]) follows from their respective definitions: $r= l_{\rm LSS}/R$ is the radius of the last scattering surface in the units of the radius of ${\mathrm{S}_2}$, and $\Omega_{\rm curv} = (l_{H_0}/R)^2$ is the square of the Hubble length in those units. The relation (\[important\]) follows from the proportionality of $l_{\rm LSS}$ and $l_{H_0}$. We note finally that these results are readily adapted to the Bianchi type III since only the sign of the curvature term $1/b^2$ changes. By contrast in Bianchi I, where the spatial hypersurfaces are Euclidean, only the first term on the r.h.s. of (\[sol delta\]) is present, and (\[delta/K\]) is replaced by $\delta/K \sim \Omega^{-3/2}$. [^1]: Shortly before completion of our work, two articles ([@GHR] and [@BPS]) were posted which have a significant amount of overlap with ours. We will comment on the similarities and differences at various places in the main text. [^2]: This has already been speculated in [@BSV]. [^3]: In a first preprint of this paper we used a state corresponding to the solutions $\sim e^{-ik\eta}$, with $d\eta = d\tau/\sqrt{V[R(\tau)]}$ a conformal time coordinate. We see from the limiting form of eq. (\[pseudoradial\]) near the horizon, i.e. $\tau \to 0$, that they correspond to $A_{k\ell}=0$. Some of the integrals in (\[final-alm\]) are, however, infrared divergent for that spectrum. This infrared contribution dominated the (truncated) integrals and was responsible for a different scaling of the corrections reported there ($\propto r$ instead of $\propto r^2$). We wish to thank Mike Salem for helpful correspondence concerning the appropriate definition of the vacuum state. [^4]: We also note that because of the sum rule (\[sum rule\]), the corrections to the (cosmic) variance of the estimator $\hat{C}_l \equiv \sum_m \left|a^{\rm obs}_{lm}\right|^2 / (2 l + 1)$ are only quadratic in $\Omega_{\rm curv}$. Indeed with use of eq. (\[parabola\]) one obtains $$\mathrm{Var}\left[\hat{C}_l\right] = \frac{2 C_l^2}{2 l + 1} \left[1 + \frac{1}{5} \left(4 - \frac{3}{l \left(l + 1\right)}\right) \delta C^2_{ll00}\right]~, \nonumber$$ where it is still assumed that the field is Gaussian. [^5]: It is worth pointing out that their calculations differ significantly from ours. While we calculated the integrals (\[final-alm\]) numerically, the authors of [@BPS] made a series of approximations to simplify the corresponding expression (their eq. ($4.33$)) before resorting to numerics. The agreement between our results and theirs, where they overlap, gives a good indication that these approximations, rather hard to justify rigorously, are correct. [^6]: In order to be able to compare our results with others, we followed the general practice where the vectors $\mathbf{v}$ and $\mathbf{w}^{\left(i\right)}$ are [*not*]{} normalized. The use of normalized vectors instead affects the significance level by an order of magnitude. It was argued that the effective weighting introduced by non normalized vectors accounts for “how well” the plane is defined by the two vectors of the cross product, a badly defined plane being one where the angle between these vectors is small (since the norm of the vector is proportional to the sine of this angle) [@Copi:2010na]. This position is untenable, because a plane is geometrically defined by two non-collinear vectors. Hopefully, the bias introduced by the weighting is less important for the relative statistical significance, i.e. when comparing two models, than it is for the absolute statistical significance, i.e. for a given model. [^7]: Since the late-time expansion in anisotropic models couples multipoles $l$ and $l+2$ [@GHR], the quadrupole receives a non-stochastic contribution from the monopole proportional to $T_0 \Omega_{\rm curv}$. We limit however our analysis to whether the correlations of the primordial spectrum can themselves favor an alignment.

ArXiv

--- abstract: 'We prove statistical stability and robustness of the variance in the Central Limit Theorem (CLT) for a family of Lorenz attractors with a $C^{1+\alpha}$ stable foliation.' address: 'Department of Mathematical Sciences, Loughborough University, Loughborough, Leicestershire, LE11 3TU, UK' author: - Wael Bahsoun - Marks Ruziboev title: 'On the stability of statistical properties for Lorenz attractors with a $C^{1+\alpha}$ stable foliation' --- [^1] Introduction ============ In his seminal work [@Lo] Lorenz introduced the following system of equations $$\label{LS} \begin{cases} \dot x = -10x+10y \\ \dot y = 28x-y-xz \\ \dot z = -\frac{8}{3}z+xy \end{cases}$$ as a simplified model for atmospheric convection. Numerical analysis performed by Lorenz showed that the above system exhibits sensitive dependence on initial conditions and has a non-periodic “strange" attractor. A rigorous mathematical framework of similar flows was initiated with the introduction of the so called geometric Lorenz flow in [@ABS; @GW]. Nowadays it is well known that the geometric Lorenz attractor, whose vector field will be denoted by $X_0$, is robust in the $C^1$ topology [@AP]. This means that vector fields $X_\eps$ that are sufficiently close in the $C^1$ topology to $X_0$ admit invariant contracting foliations $\mathcal F_{\eps}$ on the Poincaré section $\Sigma$ and they admit strange attractors. More precisely, there exists an open neighbourhood $U$ in $\mathbb R^3$ containing $\Lambda$, the attractor of $X_0$, and an open neighbourhood $\mathcal U$ of $X_0$ in the $C^1$ topology such that for all vector fields $X_\eps\in \mathcal U$, the maximal invariant set $\Lambda_{X_\eps}=\cap_{t\ge 0}X_\eps^t(U)$ is a transitive set which is invariant under the flow of $X_\eps$ [@MPP]. The papers [@Tu1; @Tu] provided a computer-assisted proof that the classical Lorenz flow; i.e., the flow defined in has a robustly transitive invariant set containing an equilibrium point. In [@LMP] it was proved that the Lorenz flow is mixing. Recently, Araújo and Melboune proved in [@AM1] that the stable foliation of the Lorenz flow is $C^{1+\alpha}$. Moreover, they showed that this property is robust; i.e., $C^1$ perturbations $ X_{\eps}$ of the Lorenz flow admit a $C^{1+\alpha}$ stable foliation, and such foliations are close to the stable foliation of the original Lorenz flow in the $C^1$-topology. Further, in another paper Araújo and Melbourne [@AM2] showed that the Lorenz system is exponentially mixing and that this property is robust in the $C^1$-topology. In this paper we study a family of perturbations $X_\eps$ which are consistent with the results of [@AM1; @AM2; @AP; @Tu] and prove statistical stability and variance continuity of Lorenz attractors with a $C^{1+\alpha}$ stable foliation. For precise statements see Theorem \[main\] and Theorem \[main3\] in section \[setup\]. Previous results on the statistical stability of Lorenz attractors was announced in [@AS] but only for flows with $C^2$ stable foliations. In [@GL] among other things, statistical stability of Poincaré maps for ‘$BV$-like’ Lorenz attractors is studied. Both [@AS] and [@GL] obtain their results under the assumption that the corresponding $1$-d maps are strongly statistically stable. In our work, we only assume $C^{1+\alpha}$ stable foliation for the family of flows and we prove that the corresponding $1$-d maps are strongly statistically stable. We then obtain statistical stability for the family of flows and we obtain robustness of the variance in the CLT of the corresponding 1-d maps. Subsequently, with a martingale coboundary decomposition assumption on observables of the flows, we show that the variance in the CLT for the family of flows is also robust. Our proofs allow the discontinuities in the base of the corresponding Poincaré maps to change with the perturbation (See Figure \[1dfig\] for an illustration). In fact, this is the main issue with perturbations of Lorenz systems since the derivative blows up at the discontinuity point. The rest of the paper is organised as follows. In section \[setup\] we introduce the family of flows that we study in this paper. The statement of our main results (Theorem \[main\] and Theorem \[main3\]) are also in this section. In section \[pf\] we provide proofs of our results in a series of lemmas and propositions. The proof of Theorem \[main2\] is given in the Appendix. Setup and statement of main result {#setup} ================================== A geometric Lorenz flow with a $C^{1+\alpha}$ stable foliation -------------------------------------------------------------- Let $X_0:\mathbb R^3\to\mathbb R^3$ be a vector field associated with a flow that has an equilibrium point at $0$. We assume that $X_0$ satisfies the following assumptions:\ $\bullet$ The differential $DX_0(0)$ has three real eigenvalues $0<-\lambda_3<\lambda_1<-\lambda_2$.\ $\bullet$ Let $$\Sigma :=\left\{(x, y, 1)|-\frac 1 2\le x, y \le\frac 1 2 \right\}$$ and let $\Gamma=\{(0, y, 1)| -\frac 1 2 \le y \le \frac 1 2\}$ be the intersection of $\Sigma$ with the local stable manifold of the fixed point $0$. The segment $\Gamma$ divides $\Sigma$ into two parts $$\Sigma^+=\{(x, y, 1)\in \Sigma:\, x>0\} \,\, \text{and} \,\, \Sigma^-=\{(x, y, 1)\in \Sigma:\, x<0\}.$$ There exists a well defined Poincaré map $F:\Sigma\to\Sigma$ such that the images of $\Sigma^{\pm}$ by this map are curvilinear triangles $S^\pm$ without the vertexes $(\pm 1, 0, 0)$ and every line segment in $\mathcal F=\{(x, y, 1) \in \Sigma \mid x=\text{const}\}$ except $\Gamma$ is mapped into a segment $\{(x, y, 1) \in S^\pm \mid x=\text{const}\}$. The return time $\tau : \Sigma\setminus \Gamma\to \mathbb R$ to $S^\pm$ is given by $\tau(x, y, 1)=-\frac 1 {\lambda_1} \log|x|$.\ $\bullet$ The flow maps $\Sigma$ into $S^\pm$ in a smooth way so that the Poincaré map $F(x, y)= X_0^{\tau(x, y, 1)}((x, y, 1), t)$ has the form $$F(x, y)=(T(x), g(x, y)),$$ - $T:I\to I$, where $I:=[-\frac{1}{2},\frac12]$, has discontinuity at $x=0$ with side limits $T(0^+)=-\frac 1 2$ and $T(0^-)=\frac 1 2$; - $T$ is piecewise monotone increasing $C^{1+\alpha}$, $0<\alpha\le 1$, on $I\setminus\{0\}$; i.e. $T$ is $C^1$ on $I\setminus\{0\}$, while $\frac{1}{T'}$ is $\alpha$-Hölder on $I_i$, $i=1,2;$ $I_1:=[-1, 0]$ and $I_2:=[0, 1]$; - $\lim_{x\to 0^\pm}T'(x)=+\infty$; - There are $C>0$ and $\theta>1$ such that $(T^n)'(x)\ge C\theta^{n}$ for any $n\in \mathbb N$; - $T$ is transitive; - $g$ preserves $\mathcal F$ and it is uniformly contracting; i.e., there exists $K>0$ and $0<\rho<1$ such that for any given leaf $\gamma$ of the foliation and $\xi_1, \xi_2\in\gamma$ and $n\ge 1,$ $$\text{dist}(F^n(\xi_1), F^n(\xi_2))\le K\rho^n\text{dist}(\xi_1, \xi_2).$$ Universally bounded $p$-variation --------------------------------- We now define a space that captures the regularity of $\frac{1}{T'}$. For $p\ge 1$, we say $f:I \to \mathbb R$ is a function of universally bounded $p$-variation if $$V_p(f):= \sup_{-1/2\le x_0<...<x_n\le 1/2} \left( \sum_{i=1}^{n}\left| f(x_i)-f(x_{i-1})\right|^p \right)^{1/p} <+\infty.$$ The space of universally bounded $p$-variation functions is denoted by $UBV_p(I)$ and it will play a key role in studying perturbations of $X_0$. Perturbations: a family $\mathcal X$ of flows with $C^{1+\alpha}$ stable foliations ----------------------------------------------------------------------------------- We now consider perturbations of $X_0$ which are consistent with the results of [@AM1; @AP]. Let $\mathcal X$ be the family of $C^1$ perturbations of $X_0$; i.e., there exists an open neighbourhood $U$ in $\mathbb R^3$ containing $\Lambda$, the attractor of $X_0$, and an open neighbourhood $\mathcal U$ of $X_0$ containing $\mathcal X$ such that 1. for each $X_\eps\in \mathcal X,$ the maximal forward invariant set $\Lambda_{X_\eps}$ is contained in $U$ and is an attractor containing a hyperbolic singularity; 2. for each $X_\eps\in\mathcal X,$ $\Sigma$ is a cross-section for the flow with a return time $\tau_\eps$ and a Poincaré map $F_\eps;$ 3. for each $X_\eps\in\mathcal{X},$ the map $F_\eps$ admits a $C^{1+\alpha}$ uniformly contracting invariant foliation $\mathcal F_\eps$ on $\Sigma$; 4. $F_\eps:\Sigma\to\Sigma$ is given by $$F_\eps(x,y)=(T_\eps(x), g_{\eps}(x,y));$$ 5. \[expansion\] the map $T_\eps:I\to I$ is transitive piecewise $C^{1+\alpha}$ expanding with two branches and a discontinuity point $O_\eps$ such that $T(O_{\eps}^+)=-\frac12$, $T(O_{\eps}^-)=\frac12$ and $\lim_{x\to O_\eps^\pm} T'_\eps(x)=+\infty$ (see Figure \[1dfig\] for an illustration); 6. for any $\eta>0$ there exists $\eps_0$ and an interval $H(\eps_0)\subset I$ such that for all $0\le \eps<\eps_0$, $\{O_\eps, 0\}\in H, \,\, |H|=2|O_{\eps_0}| <\eta;$ and $$d(T, T_\eps):=\displaystyle\sup_{x\in H^c} \{|T_\eps(x)-T(x)|+|T'_\eps(x)-T'(x)| \} \le C\eta,$$ where $H^c:=I\setminus H$; 7. there are uniform (in $\eps$ and $x$) constants $C>0$ and $\theta>1$ such that $(T_\eps^n)'(x)\ge C\theta^{n}$ except at the discontinuity point $O_\eps$; 8. \[VI\] there is a uniform (in $\eps$) constant $W>0$ such that $${\underset{i\in{1,2}}{\max}}V_{\frac{1}{\alpha}|_{ I_{i,\eps}}}(\frac{1}{T'_{\eps}})\le W,$$ where $I_{i,\eps}$ is a monotonicity interval of $T_\eps$, and $V_{\frac{1}{\alpha}}(\cdot)$ is the $\frac{1}{\alpha}$-variation; 9. For any $n>0$ let $\mathcal P_\eps^{(n)}:=\vee_{j=0}^{\ell-1}T^{-j}_\eps(\mathcal P_\eps)$, where $\mathcal P_\eps=\{I_{1,\eps}, I_{2,\eps}\}$. There exists $\delta_n>0$ independent of $\eps$ such that $\min_{J\in \mathcal P_\eps^{(n)}}|J|\ge \delta_n$; 10. there is a constant $C>0$ such that for each $X_\eps\in\mathcal X$ $$\tau_\eps(\xi)\le -C\log|\pi_\eps(\xi)-O_\eps|,$$ where $\pi_\eps$ is the projection along the leaves of $\mathcal F_\eps$ onto $I$. ![A graph of $T_{\eps}$ with discontinuity $O_\eps$ versus the graph of $T$ with discontinuity at $0$.[]{data-label="1dfig"}](Lorenz1d){width="2.5in"} Before stating our main results, we define an appropriate Banach space, which was first introduced by Keller [@K], that will play a key role in our analysis. From now on we are going to set $p:=\frac{1}{\alpha}$. A Banach space -------------- Let $S_\rho(x):=\{y\in I: |x-y|<\rho\}$ and $f:I\to \mathbb R$ be any function defined on $I.$ Let $$\operatorname{osc}(f, \rho, x):= \operatorname{ess sup}\{|f(y_1)-f(y_2)|: y_1, y_2\in S_\rho(x)\},$$ and $$\operatorname{osc}_1(f, \rho)=\|\operatorname{osc}(f, \rho, x)\|_1,$$ where the essential supremum is taken with respect to the two dimensional Lebesgue measure on $I\times I$ and $\|\cdot\|_1$ is the $L^1$- norm with respect to Lebesgue measure on $I.$ Fix $\rho_0>0$ and let $BV_{1, 1/p}\subset L^1$ be the Banach space equipped with the norm $$\|f\|_{1, 1/p}=V_{1, 1/p}(f)+\|f\|_1,$$ where $$V_{1, 1/p}(f)=\sup_{0<\rho\le \rho_0}\frac{\operatorname{osc}_1(f, \rho)}{\rho^{1/p}}.$$ Notice that $V_{1, 1/p}(\cdot)$ depends on $\rho_0$. The fact that $BV_{1, 1/p}$ is a Banach space is proved in [@K]. Moreover, it is proved in [@K] that the unit ball of $BV_{1, 1/p}$ is compact in $L^1$. We now list several inequalities, involving functions in $BV_{1, 1/p}$, that were proved in [@K]. For any $p\ge 1$ and $f\in UBV_p(I)$ we have $$\label{fact2} V_{1, 1/p}(f)\le 2^{1/p}V_p(f).$$ Moreover, if $f\in BV_{1, 1/p}$ and $Y\subset I$ is an interval with $|Y|\ge 4\rho_0$, then for each $0<\rho\le \rho_0$ we have $$\label{fact1} \operatorname{osc}_1(f\cdot \mathbf 1_{Y}, \rho)\le \left(2+\frac{8\rho_0}{|Y|-2\rho_0}\right)\int_Y\operatorname{osc}(f|_{Y}, \rho, x)dx+\frac{4\rho}{|Y|}\int_Y|f(x)|dx.$$ Further, if $Y, Z\subset I$ are intervals such that $T_\eps:Y\to Z$ is differentiable then $$\begin{aligned} \label{fact3} \begin{split} &\int_Z\operatorname{osc}(\frac{f}{T_\eps'}\circ T_\eps^{-1}|_{Z}, \rho, z)dx\le \int_Y\operatorname{osc}(f|_Y, (C\theta)^{-1}\rho, y)dy \\ &\hskip 0.6cm+ 5\int_Z\operatorname{osc}(\frac{1}{T_\eps'}\circ T_\eps^{-1}|_{Z}, \rho, z)dx\left(\frac{1}{|Y|}\int_Y|f|dx+\frac{1}{\rho_0}\int_Y\operatorname{osc}(f|_Y, \rho_0, y)dy\right). \end{split}\end{aligned}$$ Statements of the main results ------------------------------ \[main\] Let $X_\eps \in \mathcal X$. Then 1. $X_{\eps}$ admits a unique invariant probability SRB measure $\mu_\eps$. 2. For any continuous $\varphi:\mathbb R^3\to\mathbb R$ we have $$\lim_{\eps\to 0}\int\varphi d\mu_{\eps}=\int\varphi d\mu;$$ where $\mu$ is the SRB measure associated with $X_0$; i.e. the family $\mathcal X$ is statistically stable. 1\) in Theorem \[main\] is well known. See for instance [@APPV]. We prove 2) in the following section. \[main2\] Let $T_\eps$ be the family of the $1$-d maps corresponding to the family $\mathcal X$. Then: $T_{\eps}$ satisfies a CLT with variance $\sigma^2_{T_\eps}$; i.e., for ${\bar{\psi_{\eps}}}:I\to\mathbb R$, ${\bar{\psi_{\eps}}}\in BV_{1,1/p}$ with $\int{\bar{\psi_{\eps}}}d\bar\mu_\eps=0$, where $\bar\mu_\eps$ is the $T_\eps$ $(\eps\ge 0)$ absolutely continuous invariant measure $$\frac{1}{\sqrt{n}}\sum_{i=0}^{n-1}{\bar{\psi_{\eps}}}(T_\eps^ix){\overset{\text{law}}{\longrightarrow}}\mathcal N(0, \sigma_{T_\eps}^2)$$ and $\sigma_{T_\eps}^2>0$ if and only if ${\bar{\psi_{\eps}}}\not= c+\nu\circ T-\nu$, $\nu\in BV_{1,1/p}$, $c\in \mathbb R$.\ Let ${\bar{\psi_{\eps}}}, {\bar{\psi}}\in BV_{1,1/p}$ such that ${\|{\bar{\psi_{\eps}}}\|}_{\infty}$ is uniformly bounded in $\eps$ and $\int |{\bar{\psi_{\eps}}}-{\bar{\psi}}|dx \to 0$ as $\eps \to 0$ [^2]. Then $$\lim_{\eps\to 0}|\sigma_{T_\eps}^2({\bar{\psi_{\eps}}})-\sigma_{T}^2({\bar{\psi}})|=0.$$ 1\) in Theorem \[main2\] is well known and follows from the fact that the systems enjoys exponential decay of correlations. We prove 2) in the Appendix. Denote by $P_{\eps,\bar\mu_\eps}$ the transfer operator associated with $T_\eps$, defined with respect to $\bar\mu_\eps$, i.e. for $f\in L^1(I, \bar\mu_\eps)$ and $g\in L^\infty (I, \bar\mu_\eps)$ $$\int P_{\eps,\bar\mu_\eps}f\cdot gd \bar\mu_\eps = \int f \cdot g\circ T_\eps d\bar\mu_\eps.$$ Let $\mu_{F_\eps}$ denote the SRB measure of Poincaré map $F_\eps :\Sigma\to \Sigma$. Given $\psi_\eps:\mathbb R^3\to \mathbb R$, with $\int_{\mathbb R^3}\psi_\eps d\mu_{\eps}=0$, define $$\Psi_\eps(\xi):=\int_0^{\tau_\eps(\xi)}\psi_{\eps}(X_\eps(t))dt.$$ \[main3\] Assume that $\tau_\eps\in L^a(\Sigma, \mu_{F_\eps})$ and $\psi_\eps\in L^b(U, \mu_\eps)$ for some $(1-1/a)(1-1/b)\ge 1/2$ . Further, assume that $\Psi_\eps$ satisfies the following martingale co-boundary decomposition: $$\Psi_\eps={\bar{\psi_{\eps}}}\circ\pi_\eps+\chi_\eps\circ F_\eps-\chi_\eps,$$ where $\chi_\eps\in L^{a}(\Sigma,\mu_{F_\eps})$, ${\bar{\psi_{\eps}}}\in BV_{1,1/p}$ with ${\bar{\psi_{\eps}}}\in \text{Ker}P_{\eps,\bar\mu_\eps}$. Then - $X_\eps(t)$ satisfies a CLT; i.e., $$\frac{1}{\sqrt t}\int_{0}^{t}\psi_\eps\circ X_{\eps}(s) ds {\overset{\text{law}}{\longrightarrow}} \mathcal {N}(0, \sigma^2_{X_\eps})$$ with $$\sigma^2_{X_\eps}(\psi_\eps)=\frac{\sigma^2_{F_\eps}(\Psi_\eps)}{\int\tau_\eps d\mu_{F_\eps}}.$$ - Moreover, if $\lim_{\eps\to 0}\int |{\bar{\psi_{\eps}}}- {\bar{\psi}}|dx =0$, then $\lim_{\eps\to 0}|\sigma^2_{X_\eps}(\psi_\eps)-\sigma^2_{X}(\psi)|=0.$ The integrability assumption on $\tau_\eps$ and $\psi_\eps$ is needed to use the machinery of Melbourne-Török [@MelbTor] to obtain a CLT for the suspension flow from the Poincaré map. The martingale coboundary decomposition is natural and it is used to obtain CLT for the Poincaré map from the corresponding 1-d map. The assumption $\lim_{\eps\to 0}\int |{\bar{\psi_{\eps}}}- {\bar{\psi}}|dx =0$ is also natural and it is in some sense an assumption on the regularity in $\eps$, of the stable foliations. Proofs {#pf} ====== Statistical stability of the family $\mathcal X$ ------------------------------------------------ Let $\bar\mu_\eps$ and $\bar\mu_0$ denote the unique absolutely continuous invariant measures of the one dimensional maps $T_\eps,T_0$ respectively. Let $h_\eps, h_0$ denote the densities corresponding to $\bar\mu_\eps, \bar\mu_0$ respectively. For $\eps\ge 0$, let $$P_{\eps}:L^1(I)\to L^1(I)$$ denote the transfer operator(Perron-Frobenius) associated with $T_\eps$ [@Ba; @BG]; i.e., for any $f\in L^1(I)$ $$P_\eps f(x)=\sum_{i=1}^{2}(f\cdot\frac{1}{T'_\eps})(T_\eps|I_{i,\eps})^{-1}(x)\cdot \mathbf{1}_{T_\eps(I_{i,\eps})}(x).$$ Our first goal is to prove that $\lim_{\eps\to 0}||h_\eps-h||_1=0.$ This will be achieved by showing that $P_\eps$, when acting on $BV_{1, 1/p}$, satisfies a uniform (in $\eps)$ Lasota-Yorke inequality and that $\eps\mapsto P_\eps$ is continuous at $\eps=0$ in an appropriate topology. We will be then in a setting where we can apply the spectral stability result of [@KL], and hence achieve our first goal. ### A uniform Lasota-Yorke inequality In this subsection, we show that $P_\eps$ admits a uniform (in $\eps$) Lasota-Yorke inequality when acting on $BV_{1,1/p}$. We first start with two lemmas to control, uniformly in $\eps$, the $p$-variation of $\frac{1}{(T^n_\eps)'}$. \[Tell\] For any $\ell\in\mathbb N$ $$\max_{J\in\mathcal P^{(\ell+1)}_\eps}V_{p| J}\left(\frac{1}{|(T_\eps^{\ell+1})'|}\right)<\frac{W(\ell)}{C\theta^{\ell}},$$ where $W(\ell):= \frac{C^{\ell}+C-2}{C^{\ell-1}(C-1)}W.$ The proof is by induction on $\ell$. Conclusion holds for $\ell=0$ by h). Suppose it holds for $\ell$. Since $J\in \mathcal P_{\eps}^{(\ell+1)}$ is a subset of some $I_{i, \eps} \in\mathcal P_\eps$ and $T_{\eps}(J)\in \mathcal P_\eps^\ell$, using the standard properties of variation we get: $$\begin{split} V_p\left(\frac{1}{|(T^{\ell+1}_\eps)'|}\right) &\le \sup_{x\in I} \frac{1}{|T_\eps'|}V_p\left(\frac{1}{|(T_\eps^\ell)'\circ T_\eps|}\right)+ \sup_{x\in I}\left(\frac{1}{|(T_\eps^\ell)'\circ T_\eps|}\right)V_p\left(\frac{1}{|T_\eps'|}\right)\\ &\le \frac{1}{C\theta}V_p\left(\frac{1}{|(T_\eps^\ell)'\circ T_\eps|}\right)+ \frac{1}{C\theta^\ell}W\\ &\le \frac{1}{C\theta}\cdot\frac{1}{C\theta^{\ell-1}}\cdot\frac{C^{\ell-1}+C-2}{C^{\ell-2}(C-1)}W+\frac{1}{C\theta^{\ell}}W\\ &\le\frac{1}{C\theta^\ell}\frac{C^\ell+C-2}{C^{\ell-1}(C-1)}W. \end{split}$$ \[pvariation\] Fix $\ell$ such that $\bar\theta:=C\theta^{\ell-1}>1$. Then for any $\eps$ and $n$ the following holds $$\max_{J\in\mathcal P^{(n\ell)}} V_{p | J}\left(\frac{1}{(T^{n\ell}_\eps)'}\right)\le \frac{n}{\bar\theta^{n}}W(\ell-1),$$ where $W(\ell)$ is as in Lemma \[Tell\]. The proof is again by induction on $n.$ For $n=1$, by Lemma \[Tell\], we have $$\max_{J\in\mathcal P^{(\ell)}_\eps}V_{p | J}\left(\frac{1}{|(T_\eps^{\ell})'|}\right)\le W(\ell-1)/(C\theta^{\ell-1})=W(\ell-1)/\bar{\theta}.$$ Suppose that $$\max_{J\in \mathcal P^{((n-1)\ell)}_\eps}V_{p | J}\left(\frac{1}{|(T_\eps^{(n-1)\ell})'|}\right)\le \frac{n-1}{\bar\theta^{n-1}}W(\ell-1).$$ Let $J$ be any interval in $\mathcal P^{(n\ell)}_\eps$ then $T_\eps^\ell J\in\mathcal P^{(n-1)\ell}_\eps$. Hence, we get: $$\begin{aligned} V_{p, J}\left(\frac{1}{|(T_\eps^{n\ell})'|}\right) &\le \sup_{x\in J} \frac{1}{|(T^{\ell}_\eps)'|}V_{p, J}\left(\frac{1}{|(T_\eps^{(n-1)\ell)})'\circ T_\eps^\ell|}\right) \\ & +\sup_{x\in J}\left(\frac{1}{|(T_\eps^{(n-1)\ell})'\circ T_\eps|}\right)V_{p, J}\left(\frac{1}{|(T^\ell_\eps)'|}\right)\\ &\le \frac{1}{C\theta^\ell}\frac{n-1}{\bar\theta^{n-1}}W(\ell-1)+ \frac{1}{(C\theta^\ell)^{n-1}}\frac{1}{\bar\theta}W(\ell-1)\\ &\le\frac{n-1}{\bar\theta^{n}}W(\ell-1)+\frac{1}{\bar\theta^{n}}W(\ell-1) =\frac{n}{\bar\theta^{n}}W(\ell-1). \end{aligned}$$ \[uniform\] There exits[^3] $\rho_0$, $0<A_1, A_2<\infty$, $0<\kappa<1$, such that for any $0<\rho<\rho_0$, $n\in\mathbb N$ and for any $f\in BV_{1, 1/p}$ the following holds $$||P^n_\eps f||_{1,1/p}\le A_1\kappa^n||f||_{1, 1/p}+A_2||f||_1.$$ We first obtain an inequality for $n=1$. Let $$\begin{aligned} \underline K:=\inf_{0\le\eps< \eps_0}\{|T_{\eps}I_{1, \eps}|, |T_{\eps}I_{2, \eps}|\}.\end{aligned}$$ Fix $\rho_0$ such that $$\label{ch.ofepsilon} \rho_0< \underline K/{10}.$$ Since $\operatorname{ess sup}(f+g)\le \operatorname{ess sup}(f)+\operatorname{ess sup}(g)$ we have $$\label{oscestimate} \begin{aligned} &\operatorname{osc}_1(P_\eps f, \rho) \le \sum_{i=1}^{2}\operatorname{osc}_1\left(f \cdot \frac{1}{T_\eps'}\circ(T_\eps|I_{i, \eps})^{-1}\cdot \mathbf{1}_{T_\eps (I_{\eps, i}),}\,\, \rho \right)\\ &\le \sum_{i=1}^{2} \left( \left(2+\frac{8\rho_0}{|T(I_{i, \eps})|-2\rho_0} \right) \right. \int_{T_\eps(I_{i, \eps})}\operatorname{osc}(\frac{f}{T_\eps'}\circ(T_\eps|I_{i, \eps})^{-1}|_{T_\eps(I_{i, \eps})}, \rho, x)dx \\ & \left. \hskip 5.5cm+\frac{4\rho}{|T_\eps(I_{i, \eps})|}\int_{T_\eps(I_{i, \eps})}|f\circ (T|I_{i, \eps})^{-1}|dx\right). \end{aligned}$$ The last inequality follows from inequality . Now, using the notation $T_\eps|I_{i, \eps}=T_{i, \eps}$ and change of variable formula we have $$\label{changeofvar} \int_{T(I_{i, \eps})}\left(|f|\circ T_{i, \eps}^{-1}\right) \left(\frac{1}{T_{i, \eps}'}\circ T_{i, \eps}^{-1}\right)dx=\int_{I_{i, \eps}}|f|dx.$$ Inequality implies that $$\label{deltave0} 2+\frac{8\rho_0}{|T_\eps(I_{i, \eps})|-2\rho_0} < 3.$$ On the other hand from it follows that $$\label{oscsemifinal} \begin{aligned} \int_{T_\eps (I_{i, \eps})}\operatorname{osc}\left(\frac{f}{T_\eps'}\circ T_{i, \eps}^{-1}|_{T_\eps(I_{i, \eps})}, \rho, x\right)dx \le \int_{I_{i, \eps}}\operatorname{osc}(f|_{I_{i, \eps}}, (C\theta)^{-1}\rho, x)dx \\ +5\int_{T_\eps (I_{i, \eps})}\operatorname{osc}\left(\frac{1}{T_{i, \eps}'}\circ T^{-1}_{i, \eps}|_{T_\eps (I_{i, \eps})}, \rho, x\right) dx \\ \cdot \left(\frac{1}{|I_{i, \eps}|}\int_{I_{i, \eps}}|f|dx+\frac{1}{\rho_0}\int_{I_{i, \eps}}\operatorname{osc}(f|_{I_{i, \eps}}, \rho_0, y)dy \right). \end{aligned}$$ Using the relation between $L^p$ norms[^4], the definition of $V_{1, 1/p}(\cdot)$ and lead to $$\begin{aligned} &\int_{T_\eps (I_{i, \eps})}\operatorname{osc}\left(\frac{1}{T_{i, \eps}'}\circ T^{-1}_{i, \eps}|_{T_\eps (I_{i, \eps})}, \rho, x\right) dx\\ &\le \left(\int_{T_\eps (I_{i, \eps})}\operatorname{osc}\left(\frac{1}{T_{i, \eps}'}\circ T^{-1}_{i, \eps}|_{T_\eps I_{i, \eps}}, \rho, x\right) dx \right)^{1/p} |T_\eps (I_{i, \eps})|^{1-1/p} \\ &\le \rho^{1/p}V_{1, 1/p}\left(\frac{1}{T_{i, \eps}'}\circ T^{-1}_{i, \eps}|_{T(I_{i, \eps})}\right)\le (2\rho)^{1/p}V_p\left(\frac{1}{T_{i, \eps}'}\circ T^{-1}_{i, \eps}|_{T(I_{i, \eps})}\right)\\ & \le (2\rho)^{1/p}V_p\left((T_{i, \eps}')^{-1}|_{I_{i, \eps}}\right). \end{aligned}$$ Therefore, $$\label{oscfinal} \operatorname{osc}_1\left(\frac{1}{T_{i, \eps}'}\circ T_{i, \eps}^{-1}|_{T_\eps (I_{i,\eps})}, \rho\right)\le 2\rho^{1/p}W.$$ Substituting equation first into and then substituting , and into and using property h) gives $$\begin{aligned} \operatorname{osc}_1(Pf, \rho)& \le 3\sum_{i=1}^{2}\int_{I_i}\operatorname{osc}(f|_{I_{i, \eps}}, (C\theta)^{-1}\rho, y)dy\\ &+30\rho^{1/p}W\sum_{i=1}^2\left(\frac{1}{|I_{i, \eps}|}\int_{I_{i, \eps}}|f|dm+\frac{1}{\rho_0}\int_{I_{i, \eps}}\operatorname{osc}(f|_{I_{i, \eps}}, \rho_0, y)dy \right)\\ &+\sum_{i=1}^{2} \frac{4\rho}{|T_\eps(I_{i, \eps})|}\int_{I_{i, \eps}}|f|dx\\ &\le 3\int_I \operatorname{osc}(f, \frac{\rho}{C\theta}, x)dx \\ &+30\rho^{1/p} \left( \frac{1}{\delta_1}\int_{I}|f|dx+\frac{1}{\rho_0}\int_I\operatorname{osc}(f, \rho_0, x) dx \right)+ \frac{4\rho}{ \underline K }\int_I|f|dx.\end{aligned}$$ Therefore, $$\begin{aligned} \frac{\operatorname{osc}_1(P_\eps, \rho)}{\rho^{1/p}}\le \left(\frac{3}{C\theta} +30 W\rho^{1/p-1}\right)V_{1, 1/p}(f)\\ +\left(30 W\frac{1}{\delta_1}+\frac{4\rho}{ \underline K }\right)\|f\|_1. \end{aligned}$$ Consequently, we have $$\label{step1} \begin{split} V_{1, 1/p}(P_\eps f) &\le \left(\frac{3}{C\theta} +30 W\rho^{1/p-1}\right)V_{1, 1/p}(f)\\ &\hskip 4cm+\left(30 W\frac{1}{\delta_1}+\frac{4\rho}{ \underline K }\right)\|f\|_1. \end{split}$$ We now prove an inequality for all $n$ as stated in the lemma. Fix $\ell\in\mathbb N$ such that $\bar\theta:=C\theta^{\ell-1}>1$. By Lemma \[pvariation\] and applied to $(P_\eps^\ell)^k$ we get $$\label{step2} \begin{aligned} V_{1, 1/p}((P_\eps^\ell)^k f) \le \left(\frac{3}{\bar\theta^k} +\frac{30}{\bar\theta^{k}} k W(\ell-1)\rho^{1/p-1}\right)V_{1, 1/p}(f)+\\ +\left(30 W(\ell-1)\frac{1}{\bar\theta^{k}\delta_{k\ell}}+\frac{4\rho}{ \underline K_{k\ell} }\right)\|f\|_1, \end{aligned}$$ where $\underline K_{k\ell}:=\sup_{0<\eps< \eps_0}\min_{J\in \mathcal P_\eps^{(k\ell)}}\{|T_{\eps}^{k\ell} J|\}$ and $\delta_{k\ell}>0$ by assumption i). Since $\ell$ is fixed we can choose $k$ large enough so that $$\beta:=\left(\frac{3}{\bar\theta^k} +\frac{30}{\bar\theta^{k}} k W(\ell-1)\rho^{1/p-1}\right)<1$$ and let $$K:=\left(30 W(\ell-1)\frac{1}{\bar\theta^{k}\delta_{k\ell}}+\frac{4\rho}{ \underline K_{k\ell} }\right).$$ Thus, we have $$\label{ULYforTell} V_{1, 1/p}((P_\eps^\ell)^k f)\le \beta V_{1, 1/p}(f)+K\| f\|_1.$$ Similar to , by using and Lemma \[pvariation\], for any $j\in \mathbb N$ we have $$\label{step3} \begin{aligned} V_{1, 1/p}(P_\eps^j f)\le \left(\frac{3}{(C\theta)^j} +30 \frac{W(j-1)}{C\theta^{j-1}}\rho^{1/p-1}\right)V_{1, 1/p}(f)\\ +\left(30\frac{W(j-1)}{C\theta^{j-1}}\frac{1}{\delta_j}+\frac{4\rho}{ \underline K_j }\right)\|f\|_1. \end{aligned}$$ Set $k_0=\ell k$, where $k$ and $\ell$ are chosen so that $P_\eps^{k_0}$ satisfies . Then for any $n\in\mathbb N$ we can write $n=k_0m+j$ for some $j=1, ..., k_0-1.$ Applying consecutively implies $$\begin{aligned} V_{1, 1/p}(P^n_\eps f) &=V_{1, 1/p}((P_\eps^{k_0})^m\circ P_\eps^jf)\le \beta V_{1, 1/p}((P_\eps^{k_0})^{m-1}\circ P_\eps^j f)+K\|f\|_1 \\ &\le ...\le \beta^mV_{1, 1/p}(P_\eps^j f)+K\frac{\beta}{1-\beta}\|f\|_1.\end{aligned}$$ Using and setting $$A_1:=\max_{1\le j\le k_0}\left\{\frac{3}{(C\theta)^j} +30 \frac{W(j-1)}{C\theta^{j-1}}\rho^{1/p-1}\right\}\cdot\beta^{-j/k_0},$$ $$A_2:=\max_{1\le j\le k_0}\left\{30\frac{W(j-1)}{C\theta^{j-1}}\frac{1}{\delta_j}+ \frac{4\rho}{ \underline K_j }\right\}+K\frac{\beta}{1-\beta} +1, \text{ and } \kappa:=\beta^{1/k_0},$$ we obtain $$||P^n_\eps f||_{1,1/p} \le A_1\kappa^n ||f ||_{1, 1/p}+A_2\|f\|_1.$$ ### Estimating the difference of the transfer operators in the mixed norm Define the following operator ‘mixed’ norm: $$|||P_\eps|||:= \sup_{ \|f\|_{1, 1/p} \le 1}||P_\eps f||_1.$$ To apply the spectral stability result of [@KL], we still need to prove that $\lim_{\eps\to 0}|||P_\eps-P|||=0.$ Firstly, we start with a simple lemma that is similar[^5] to Lemma 11 in [@Kel82]. \[skeller\] For any $f \in BV_{1, 1/p}(I)$ and $u\in L^1(I)$ $$\left| \int_{I} f\cdot u dx \right| \le (1+\rho_0^{1/p}) \|f\|_{1, 1/p} \sup_{z\in I} \int_{x\le z} u(x)dx.$$ We prove the lemma when $u$ is a simple function. Then general case follows, since $BV_{1, 1/p}\subset L^1$ and any $L^1$ function can be approximated by a sequence of simple functions. Let $-\frac{1}{2}=a_0 < a_2 < ... < a_n=\frac{1}{2}$ be a partition of $I$ and suppose that $u$ is constant on each $J_i=(a_{i-1}, a_{i}),$ $i=1, 2, ..., n.$ Denote by $G(x):=\int_{-\frac{1}{2}}^{x} udx$. Then $$\begin{aligned} &\left| \int f\cdot udx \right| =\left| \sum_{i=1}^n \int_{J_i} f\cdot udx \right| = \left| \sum_{i=1}^n u|J_i \int_{J_i} fdx \right| \\ &=\left| \sum_{i=1}^n u|J_i y_i|{J_i}|\right| \quad \text{where} \quad y_i\in \overline{\text{conv}(f(J_i))} \\ &=\left| \sum_{i=1}^n y_i \int_{J_i} udx \right|=\left| \sum_{i=1}^n y_i(G(a_i)-G(a_{i-1}))\right| \\ &\le \sum_{i=2}^n |y_i-y_{i+1}||G(a_i)| +|G(-1/2)y_1|+|G(1/2)y_n| \\ &\le \sup_{z\in I}|G(z)|\sum_{i=1}^n |y_i-y_{i+1}| +\|f\|_{\infty}\|u\|_1. \end{aligned}$$ In the last inequality we used the facts $G(-1/2)=0$ and $G(1/2)\le \|u\|_1.$ By definition of $\operatorname{osc}(f, \rho, x)$ we have $$|y_i-y_{i-1}|\le \operatorname{osc}(f|{J_i}, 1, z)$$ which implies $$\int \sum_{i=1}^n |y_i-y_{i+1}|\le \operatorname{osc}(f, 1) \le \rho_0^{1/p}V_{1, 1/p}(f).$$ Substituting the latter into above equation finishes the proof. \[triplemorm0\] $$\lim_{\eps\to 0}|\|P_\eps-P_0\||=0.$$ For any $f \in {BV}_{1, 1/p}$ we have $$\label{triplenorm}\begin{aligned} &\|P_\eps f -Pf \|_1 \le \|P_\eps f -P_{\eps}(\textbf{1}_{H^c}f) \|_1 +\|P_{\eps}(\textbf{1}_{H^c} f) -P(\textbf{1}_{H^c}f) \|_1\\ &\hskip 6cm+\|P(\textbf{1}_{H^c} f) -Pf \|_1. \end{aligned}$$ We first estimate the first and the last term in . By linearity of $P$ we have $$\begin{aligned}\label{PTh-PT} \|P(\textbf{1}_{H^c} f )-Pf \|_1&= \| P(\textbf{1}_{H^c}f- f)\|_1\le\| \textbf{1}_{H}f\|_1\\ &\le \|f\|_{1, 1/p}|H|\le 2|O_{\eps_0}| \|f\|_{1, 1/p} \le \eta\|f\|_{1, 1/p}. \end{aligned}$$ Similarly, for the first term we have $$\label{PTh-PTeps} \|P_\eps f -P_{\eps}(\textbf{1}_{H^c}f )\|_1\le \eta\|f\|_{1, 1/p}.$$ It remains to prove that the second term in equation goes to zero as $\eps\to 0.$ Let $u=\text{sgn}(P_T({\bf 1}_{H^c}f)-P_{T_\eps}({\bf 1}_{H^c}f))$. Using the dual operators of $P$, $P_\eps$ and Lemma \[skeller\], we have: $$\begin{aligned}\label{PT-PTeps} \|P(\textbf{1}_{H^c} f) -P_{\eps}(\textbf{1}_{H^c}f) \|_1 = \left |\int u(P(\textbf{1}_{H^c} f) -P_{\eps}(\textbf{1}_{H^c}f))dx\right | \\ =\left |\int \textbf{1}_{H^c} u\circ T f dx- \int \textbf{1}_{H^c}u\circ T_\eps f dx\right | = \left |\int \textbf{1}_{H^c} f (u\circ T-u\circ T_\eps)dx\right | \\ \le (1+\rho_0^{1/p})\|f\|_{1, 1/p}\sup_{z}\int_{-1/2}^z{\bf 1}_{H^c}(u\circ T-u\circ T_\eps)dx \\ = (1+\rho_0^{1/p})\|f\|_{1, 1/p}\sup_{z}\int_{[-1/2, z]\cap H^c}(u\circ T-u\circ T_\eps)dx \\ =(1+\rho_0^{1/p})\|f\|_{1, 1/p}\sup_{z}\sum_{i=1}^2\left( \int _{T_i(H_z^c)}\frac{u(y)}{T'(T^{-1}_i(y))}dy- \int_{T_{i, \eps}(H_z^c)}\frac{u(y)}{T_\eps'(T^{-1}_{i, \eps}(y))} dy\right) \end{aligned}$$ where $H_z^c=[-1/2, z]\cap H^c$ and we used change of variables $y=T(x)$ and $y=T_\eps(x)$ for the first and second summands respectively. For $i=1, 2$ we have $$\begin{aligned}\label{intt-intteps} &\int _{T_i(H_z^c)}\frac{u(y)}{T'(T^{-1}_i(y))}dy- \int_{T_{i, \eps}(H_z^c)}\frac{u(y)}{T_\eps'(T^{-1}_{i, \eps}(y))} dy \\ &= \int _{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)} \left(\frac{u(y)}{T'(T^{-1}_i(y))} - \frac{u(y)}{T_\eps'(T^{-1}_{i, \eps}(y))} \right)dy \\ &+\int _{T_i(H_z^c)\setminus T_{i, \eps}(H_z^c)}\frac{u(y)}{T'(T^{-1}_i(y))}dy- \int _{T_{i, \eps}(H_z^c)\setminus T_{i}(H_z^c)} \frac{u(y)}{T_\eps'(T^{-1}_{i, \eps}(y))} dy\\ &:= E_1+E_2+E_3 \end{aligned}$$ Now we estimate each of the terms in the right hand side separately. Using $T'_\eps\ge C\theta$ and the fact that $||u||_{\infty}\le 1$ we have $$\begin{aligned} E_1 &:= \left| \int _{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)} \left(\frac{u(y)}{T'(T^{-1}_i(y))} - \frac{u(y)}{T_\eps'(T^{-1}_{i, \eps}(y))} \right)dy \right|\\ &\le \int_{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)}\left|\frac{1}{T'(T^{-1}_i(y))} - \frac{1}{T'(T^{-1}_{i, \eps}(y))} \right|dy \\ &+ \int_{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)}\left|\frac{1}{T'(T^{-1}_{i, \eps}(y))} - \frac{1}{T_\eps'(T^{-1}_{i, \eps}(y))} \right|dy\\ &\le\|\frac{1}{T'}\|_\alpha \int _{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)}|T^{-1}_{i, \eps}(y)- T^{-1}_i(y)|^\alpha dy \\ &+\frac{1}{(C\theta)^2}\int _{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)}|T_\eps'(T^{-1}_{i, \eps}(y))- T'(T^{-1}_{i, \eps}(y))|dy \\ &\le \|\frac1{T'}\|_{\alpha}\int_{T_i(H_z^c)\cap T_{i, \eps}(H_z^c)}|T^{-1}_{i, \eps}(y)- T^{-1}_i(y)|^\alpha dy\\ &+ \frac{1}{(C\theta)^2}|T_i(H_z^c)\cap T_{i, \eps}(H_z^c)|d(T, T_\eps). \end{aligned}$$ Notice that for $y\in T(H^c_z)\cap T_\eps(H_z^c)$ there exists $x\in H_z^c$ such that $T_\eps(x)=y$ we have $$\label{t-teps} |T^{-1}_{i, \eps}(y)- T^{-1}_i(y)|= |x-T_i^{-1}(T_\eps(x))|\le\operatorname{Lip}(T^{-1}_i)d(T, T_\eps) \le \frac{1}{ C\theta}d(T, T_\eps).$$ Taking into account the fact that $|T_i(H_z^c)\cap T_{i, \eps}(H_z^c)|\le 1$ the relation implies that $$E_1 \le C_2\eta^\alpha.$$ Now note that by assumption f) for any $x\in H^c$ we have $|T_\eps(x)-T(x)|< C\eta$. This implies $|T_\eps(H^c) \bigtriangleup T(H^c)| \le 2C\eta$. Hence, for all sufficiently small $\eps$ we have $$E_2:=\left| \int _{T_i(H_z^c)\setminus T_{i, \eps}(H_z^c)}\frac{u(y)}{T'(T^{-1}_i(y))}dy \right| \le \frac{1}{C\theta} |T_i(H_z^c)\setminus T_{i, \eps}(H_z^c)| =C_3\eta.$$ Similarly, $$E_3:= \left| \int _{T_{i, \eps}(H_z^c)\setminus T_{i}(H_z^c)} \frac{u(y)}{T_\eps'(T^{-1}_{i, \eps}(y))} dy \right| \le C_3\eta.$$ Substituting estimates for $E_1$, $E_2$ and $E_3$ first into equation and then substituting the result into gives $$\|P(\textbf{1}_{H^c} f) -P_{\eps}(\textbf{1}_{H^c}f) \|_1 \le (1+\rho_0^{1/p})(2C_2+4C_3)\|f\|_{1, 1/p}\eta^{\alpha}.$$ Substituting this and equations and into implies $$\|P_\eps f -Pf \|_1\le C\eta^\alpha\|f\|_{1, 1/p}$$ which finishes the proof. We are now ready to prove that the $1$-d family $T_\eps$ is strongly statistically stable. Firstly, we set some notation. Consider the set $$V_{\delta,r}(P)=\{z\in\mathbb C: |z|\le r\text{ or dist}(z,\sigma(P))\le\delta\},$$ where $\sigma(P)$ is the spectrum of $P$ when acting on $BV_{1,1/p}$. \[1dstability\] $$\lim_{\eps\to 0}||h_{\eps}-h||_1=0.$$ By Lemma \[uniform\] and Lemma \[triplemorm0\], for any $z\in V_{\delta,r}(P)$ the Keller-Liverani [@KL] stability result implies $$\lim_{\eps \to 0}|||(z I-P_{\eps})^{-1}-(z I-P)^{-1}|||=0.$$ Consequently, $$\lim_{\eps \to 0}|||\Pi_{1,\eps}- \Pi_{1}|||=0,$$ where $\Pi_{1,\eps}$ and $\Pi_{1}$ are the spectral projections of $P_\eps$ and $P$ associated with the eigenvalue $1$. This completes the proof since both $\Pi_{1,\eps}$ and $\Pi_{1}$ have rank $1$. ### Statistical stability: from the 1-d family, to the Poincaré maps, to the family of flows {#ssfrom1d} We now discuss how to obtain continuity of the SRB measures (3) of Theorem \[main\]) from Proposition \[1dstability\]. We first show how the absolutely continuous invariant measures of the family of $1$-d maps are related to the SRB measures of the family of the flows via the Poincaré maps. This construction is well known (see for instance [@APPV]). Let ${\tilde\psi}: \Sigma\to \mathbb R$ be any bounded function. Notice that $\Sigma$ is foliated by stable manifolds, and any $x\in I$ defines unique stable manifold $\pi^{-1}(x)$. Therefore $\psi^{+}_\eps:I\to \mathbb R$ and $\psi^{-}_\eps:I\to \mathbb R$ are well defined by $$\psi^{+}_{\eps}(x):=\sup_{\xi\in\pi_{\eps}^{-1}(x)}{\tilde\psi}(\xi) \quad\text{and} \quad \psi_{\eps}^{-}(x):=\inf_{\xi\in\pi_{\eps}^{-1}(x)}{\tilde\psi}(\xi).$$ There exists a unique $F_\eps$-invariant probability measure $\mu_{F_\eps}$ on $\Sigma$ such that for every continuous function ${\tilde\psi}:\Sigma \to \mathbb R$ $$\int {\tilde\psi}d\mu_{F_\eps}=\lim_{n\to \infty}\int_{\mathcal F_\eps}({\tilde\psi}\circ F_\eps^n)^{+}_\eps d\bar\mu_\eps=\lim_{n\to \infty}\int_{\mathcal F_\eps}({\tilde\psi}\circ F_\eps^n)^{-}_\eps d\bar\mu_\eps,$$ where $\bar\mu_\eps$ is the $T_\eps$-invariant absolutely continuous measure. To pass from the Poincaré map to the flow we use standard procedure: first consider suspension flow from the Poincaré map and then embed the suspension flow into the original flow. To apply the standard construction we first need to prove the following For every $X_\eps\in\mathcal X$ let $F_\eps:\Sigma\to \Sigma$ be its Poincaré map and define $\mu_{F_\eps}$ as above. Then $\tau_\eps$ is $\mu_{F_\eps}$- integrable. Let $\tau_{N, \eps}=\min\{N, \tau_\eps\}$. Then $\tau_{N, \eps}$ is monotone increasing in $N$ and it converges to $\tau_\eps$ almost everywhere. Since $\tau_{N, \eps}$ is continuous and $d\bar\mu_\eps/dm$ is uniformly bounded, in $\eps$, we have $$\begin{aligned} \int \tau_{N, \eps} d\mu_{F_\eps}=\lim_{n\to \infty}\int (\tau_{N, \eps}\circ F_\eps^n)_\eps^{+}d\bar\mu_\eps =\lim_{n\to \infty}\int (\tau_{N, \eps}\circ F^n)_\eps^{+}(d\bar\mu_\eps/ dm)dm\\ \le C\|d\bar\mu_\eps/dm\|_\infty|\int_{I} \log |x- O_\eps|dx<+\infty, \end{aligned}$$ which implies that $\lim_{N\to \infty}\int \tau_{N, \eps} d\mu_F$ exists and finite. Hence by monotone converges theorem $\int\tau_{\eps}d\mu_{F_\eps}<\infty.$ Let $$\Sigma_{\tau_\eps}=\Sigma\times [0, +\infty)/\sim,$$ where $\sim$ is the equivalence relation on $\Sigma\times [0, +\infty)$ generated by $(\xi, \tau_\eps(\xi))\sim (F_\eps(\xi), 0).$ Then there is a natural projection $\pi_{\tau_\eps} : \Sigma\times [0, \infty)\to \Sigma _{\tau_\eps}$ which induces a topology and a Borel $\sigma$-algebra on $\Sigma_{\tau_\eps}.$ The suspension flow of $F_\eps$ with return time $\tau_\eps$ is the semi-flow $(X_\eps^t)_{t\ge 0}$ defined on $\Sigma_{\tau_\eps}$ as $$X_\eps^t(\pi_{\tau_\eps}(\xi, s))=\pi_{\tau_\eps}(\xi, s+t) \quad \text{for any} \quad (\xi, s)\in \Sigma\times [0, +\infty).$$ ([@APPV], Lemma 6.7.) $X_\eps^t$ has unique invariant probability measure $\bar\mu_{X_\eps}$ that is defined by $$\int\varphi d\bar \mu_{X_\eps}=\frac {1}{\mu_{F_\eps}(\tau_\eps)}\int\int_0^{\tau_\eps(\xi)}\varphi(\pi_{\tau_\eps}(\xi, t))dtd\mu_{F_\eps}(\xi)$$ for every bounded measurable $\varphi: \Sigma_{\tau_\eps}\to \mathbb R,$ where $\mu_{F_\eps}(\tau)=\int\tau_\eps d\mu_{F_\eps}.$ Now, we can define the unique SRB measure of the original flow $X_\eps(\xi, t)$. Define $$\Phi_\eps :\Sigma\times [0, +\infty)\to U \quad \text{by letting} \quad \Phi_\eps(\xi, t)=X_\eps(\xi, t).$$ Since $\Phi_\eps(\xi, \tau_\eps(\xi))=(F_\eps(\xi), 0)\in \Sigma\times\{0\}$, map $\Phi_\eps$ induces a map $$\phi_\eps:\Sigma_{\tau\eps}\to U, \quad \text{such that} \quad \phi_\eps\circ X_\eps^t=X_\eps(\cdot, t)\circ \phi_\eps, \quad \text{for} \quad t\ge 0.$$ via the identification $\sim$. Now the invariant measure of $X_\eps^t$ is naturally transferred to an invariant measure for $X_\eps(\cdot, t)$ via pushing it forward $\mu_{\eps}={\phi_{\eps}}_\ast\bar\mu_{X_\eps}$ (see [@APPV], Section 7). Hence, we can define the SRB measure of the flow as follows: \[measureforflow\] The flow of each $X_\eps\in\mathcal X$ has a unique SRB measure $\mu_{\eps}$. In particular, for any continuous map $\varphi:U\to \mathbb R$ $$\int\varphi d\mu_{\eps}=\frac {1}{\mu_{F_\eps}(\tau_\eps)}\int\int_0^{\tau_\eps(\xi)}\varphi \circ \phi_\eps \circ \pi_\eps(\xi, t)dtd\mu_{F_\eps}(\xi)$$ where $\mu_{F_\eps}(\tau_\eps)=\int\tau_\eps d\mu_{F_\eps}.$ To finish the proof of 2) of Theorem \[main\], we use the above construction, Proposition \[1dstability\], and section[^6] 3 of [@AS]. Variance continuity for the family of $1$-d maps $T_{\eps}$ ----------------------------------------------------------- Let ${\bar{\psi_{\eps}}}:I\to\mathbb R$ be in[^7] $BV_{1,1/p}$ with $\int{\bar{\psi_{\eps}}}d\bar\mu_\eps=0$. Since $(T_\eps, \bar\mu_{\eps})$ is mixing, by Lemma \[uniform\], the system enjoys exponential decay of correlations and hence it satisfies a Central Limit Theorem: $$\frac{1}{\sqrt{n}}\sum_{i=0}^{n-1}{\bar{\psi_{\eps}}}(T_\eps^ix){\overset{\text{law}}{\longrightarrow}}\mathcal N(0, \sigma_{T_\eps}^2)$$ and $\sigma_{T_\eps}^2>0$ if and only if ${\bar{\psi_{\eps}}}\not= c+\nu\circ T-\nu$, $\nu\in BV_{1,,1/p}$, $c\in \mathbb R$.\ Moreover, $$\label{eq_me} \sigma_{T_\eps}^2:=\lim_{n\to\infty}\frac1n\int_{I}\left(\sum_{i=0}^{n-1}{\bar{\psi_{\eps}}}(T_\eps^ix)\right)^2d\bar\mu_\eps.$$ In fact, since the system enjoys exponential decay of correlations, the limit in exists and $\sigma_{T_\eps}^2$ admits a Green-Kubo formula[^8]: $$\sigma_{T_\eps}^2=\int{\bar{\psi_{\eps}}}^2d\bar\mu_\eps+2\sum_{n=1}^{\infty}\int{\bar{\psi_{\eps}}}\cdot ({\bar{\psi_{\eps}}}\circ T^n_{\eps}) d\bar\mu_\eps.$$ Using the fact that $d\bar\mu_\eps:=h_{\eps}dx$ and the duality property of $P_\eps$, one can rewrite $\sigma_{T_\eps}^2$ as $$\label{eq2} \begin{split} \sigma_{T_\eps}^2&=\int_I{\bar{\psi_{\eps}}}^2 h_\eps dx+2\sum_{i=1}^{\infty}\int_IP_\eps^i({\bar{\psi_{\eps}}}h){\bar{\psi_{\eps}}}dx\\ &=-\int_I{\bar{\psi_{\eps}}}^2 h_\eps dx+2\int_I{\bar{\psi_{\eps}}}(I- P_\eps)^{-1}({\bar{\psi_{\eps}}}h_\eps) dx. \end{split}$$ ### Proof of Theorem \[main2\] part 2) See Appendix. [^9] Variance continuity for the family $\mathcal X$ ----------------------------------------------- Limit theorems (ASIP and hence CLT) for Lorenz attractors with return function $\tau= -\frac{1}{\lambda_1}\log|x|$ were obtained in [@HM]. Then rapid mixing for the Lorenz attractor and statistical limit laws for their time-1 maps was obtained in [@AMV].\ We now present a classical argument on obtaining the CLT of the family of flows from the CLT of the corresponding family of $1$-d maps. We then show that the variance in the CLT of the family of flows is also stable. Let $\psi_\eps:\mathbb R^3\to \mathbb R$, with $\int_{\mathbb R^3}\psi_\eps d\mu_{\eps}=0$, and define $$\Psi_\eps(\xi):=\int_0^{\tau_\eps(\xi)}\psi_{\eps}(X_\eps(t))dt.$$ Since $\Psi_\eps$ satisfies martingale co-boundary decomposition: $$\Psi_\eps={\bar{\psi_{\eps}}}\circ\pi_\eps+\chi_\eps\circ F_\eps-\chi_\eps,$$ with $\chi_\eps\in L^{a}(\Sigma,\mu_{F_\eps})$, ${\bar{\psi_{\eps}}}\in BV_{1,1/p}$, ${\bar{\psi_{\eps}}}\in \text{Ker}P_{\eps,\bar\mu_\eps}$ and $(T_\eps, {\bar{\psi_{\eps}}})$ satsfies a CLT: $$\frac{1}{\sqrt{n}}\sum_{i=0}^{n-1}{\bar{\psi_{\eps}}}(T_\eps^ix) {\overset{\text{law}}{\longrightarrow}} \mathcal N(0, \sigma_{T_\eps}^2),$$ this implies that $$\frac{1}{\sqrt{n}}\sum_{i=0}^{n-1}\Psi_\eps(F_\eps^i\xi) {\overset{\text{law}}{\longrightarrow}}\mathcal N(0, \sigma_{F_\eps}^2)$$ with $\sigma_{F_\eps}^2(\Psi_\eps)=\sigma_{T_\eps}^2({\bar{\psi_{\eps}}})$ (see for instance [@Sin]). Further, since we have $\tau_\eps\in L^a(\Sigma, \mu_{F_\eps})$ and $\psi_\eps\in L^b(U, \mu_\eps)$ for some $(1-1/a)(1-1/b)\ge 1/2$ by [@MelbTor] $(X_\eps(t), \psi_\eps)$ satisfies a CLT i.e. $$\frac{1}{\sqrt t}\int_{0}^{t}\psi_\eps\circ X_{\eps}(s) ds {\overset{\text{law}}{\longrightarrow}} \mathcal {N}(0, \sigma^2_{X_\eps})$$ with $$\sigma^2_{X_\eps}(\psi_\eps)=\frac{\sigma^2_{F_\eps}(\Psi_\eps)}{\int\tau_\eps d\mu_{F_\eps}};$$ We now show how to obtain continuity of the variance $\sigma_{X_\eps}$ at $\eps=0$. The following is a preparatory lemma: \[tauntau\] $$\lim_{\eps \to 0}\int \tau_\eps d\mu_{F_\eps}=\int \tau d\mu_F.$$ We start by defining the following cut off’s of the return times. $$\tau_N(\xi)=\min\{\tau(\xi), N\} \quad \text{and} \quad \tau_{\eps, N}(\xi)=\min\{\tau_\eps(\xi), N\}.$$ For each fixed $N\ge 1$ the functions $\tau_N$ and $\tau_{\eps, N}$ are bounded and continuous and $\tau_{\eps, N}$ converges uniformly to $\tau_N$ as $\eps\to 0$. To prove this we use the following standard estimate $$\label{variation} |X_\eps(\xi, t)-X(\xi, t)| \le t{\|X_\eps - X\|}_{\infty} \exp({\|X\|}_{C^1}t),$$ where $X_\eps(\xi, t)$ and $X_\eps(\xi, t)$ denotes the flow of a vector field $X_\eps\in \mathcal X$ for $\eps>0$ and $\eps=0$ respectively[^10]. Inequality implies that for each fixed $N$ we can choose $\eps$ small enough so that the distance between from $X(\xi, t)$ and $X_\eps(\xi, t)$ is small. Hence if one of the trajectories crosses $\Sigma$ at a certain moment another will also cross $\Sigma$ within controlled time interval by rectification theorem. The rest of the proof is the same as the proof of Lemma 4.1 in [@AS]. \[Proof of Theorem \[main3\], part 2)\] Using the relation between the variance of the flow and the variance of the corresponding $1$-d map, we have: $$\begin{aligned}\label{27} &|\sigma^2_{X_\eps}(\psi_\eps)-\sigma^2_{X_0}(\psi)|=\left| \frac{\sigma_{T_\eps}^2( {\bar{\psi_{\eps}}})}{\int \tau_\eps d\mu_{F_\eps}}- \frac{\sigma_{T}^2( \bar\psi)}{\int \tau d\mu_{F}}\right|\\ &\le \sigma_{T}^2( \bar\psi) \left| \frac{1}{\int \tau_\eps d\mu_{F_\eps}}- \frac{1}{\int \tau d\mu_{F}}\right| +\frac{1}{\int \tau_\eps d\mu_{F_\eps}} \left| \sigma_{T_\eps}^2( {\bar{\psi_{\eps}}})- \sigma_{T}^2( \bar\psi) \right|. \end{aligned}$$ The first term in converges to zero by Lemma \[tauntau\]. By the same lemma $\frac{1}{\int \tau_\eps d\mu_{F_\eps}}$ is uniformly bounded in $\eps$, since $\frac{1}{\int \tau_\eps d\mu_{F_\eps}}\to \frac{1}{\int \tau d\mu_{F}}.$ The second term converges to $0$ by Theorem \[main2\]. Appendix {#app} ======== We first prove two lemmas that will be used to prove Theorem \[main2\]. \[le1\] For ${\bar{\psi_{\eps}}}\in BV_{1,1/p}$, $\eps\ge 0$, we have $$\lim_{\eps\to 0}|\int_I({\bar\psi}^2 h-{{\bar\psi}_{\eps}}^2 h_\eps)dx|\to 0.$$ We have $$|\int_I({\bar\psi}^2_{\eps}-{\bar\psi}^2)hdx|=|\int_I({\bar\psi}_{\eps}-{\bar\psi})({\bar\psi}_{\eps}+{\bar\psi})h dx| \le ||h({\bar{\psi_{\eps}}}+{\bar\psi})||_{\infty}\cdot||{\bar{\psi_{\eps}}}-{\bar\psi}||_1.\\ $$ Consequently, $$|\int_I({\bar\psi}^2 h-{{\bar\psi}_{\eps}}^2 h_\eps)dx|\le ||h({\bar{\psi_{\eps}}}+{\bar\psi})||_{\infty}\cdot||{\bar{\psi_{\eps}}}-{\bar\psi}||_1+||{{\bar\psi}_{\eps}}^2||_\infty|| h- h_\eps||_1\to 0\\$$ as $\eps\to 0$. \[le2\] For any $l\ge1$ we have $$\begin{split} & |\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({{\bar\psi}_{\eps}}h_{\eps}){{\bar\psi}_{\eps}}-P^i({\bar\psi}h){\bar\psi}\right)dx|\le C \sum_{i=1}^{l-1}||P^i({\bar{\psi_{\eps}}}h_\eps)||_{1,1/p}\cdot ||{{\bar\psi}_{\eps}}-{\bar\psi}||_1\\ &+(l-1)||{\bar\psi}||_\infty\cdot ||h_\eps||_\infty|| {\bar{\psi_{\eps}}}-{\bar\psi}||_1+ (l-1)||{\bar\psi}||^2_\infty ||h_\eps-h ||_1\\ &+||{\bar\psi}||_{\infty}\cdot |||P_{\eps}-P|||\cdot \sum_{i=1}^{l-1}\sum_{j=0}^{i-1}||P^{j}({\bar\psi}h)||_{1,1/p}. \end{split}$$ $$\begin{split} &|\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({{\bar\psi}_{\eps}}h_{\eps}){{\bar\psi}_{\eps}}-P^i({\bar\psi}h){\bar\psi}\right)dx|\\ &\le |\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({{\bar\psi}_{\eps}}h_{\eps}){{\bar\psi}_{\eps}}-P^i_\eps({\bar\psi}h){\bar\psi}\right)dx|+|\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({\bar\psi}h){\bar\psi}-P^i({\bar\psi}h){\bar\psi}\right)dx|\\ &\le |\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({{\bar\psi}_{\eps}}h_{\eps}){{\bar\psi}_{\eps}}-P^i_\eps({\bar{\psi_{\eps}}}h_\eps){\bar\psi}\right)dx|+ |\sum_{i=1}^{l-1}\int_I\left( P^i_\eps({\bar{\psi_{\eps}}}h_\eps){\bar\psi}-P^i_\eps({\bar\psi}h){\bar\psi}\right)dx|\\ &+|\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({\bar\psi}h){\bar\psi}-P^i({\bar\psi}h){\bar\psi}\right)dx|\\ &:=(I)+(II) +(III). \end{split}$$ We have $$\label{es1} (I)\le\sum_{i=1}^{l-1}||P^i({\bar{\psi_{\eps}}}h_\eps)||_{\infty}\cdot ||{{\bar\psi}_{\eps}}-{\bar\psi}||_1\le C \sum_{i=1}^{l-1}||P^i({\bar{\psi_{\eps}}}h_\eps)||_{1,1/p}\cdot ||{{\bar\psi}_{\eps}}-{\bar\psi}||_1. $$ We now estimate $(II)$: $$\label{es2} \begin{split} (II)&\le |\sum_{i=1}^{l-1}\int_I\left( P^i_\eps({\bar{\psi_{\eps}}}h_\eps){\bar\psi}-P^i_\eps({\bar\psi}h){\bar\psi}\right)dx|\le ||{\bar\psi}||_\infty\sum_{i=1}^{l-1}|| P^i_\eps({\bar{\psi_{\eps}}}h_\eps)-P^i_\eps({\bar\psi}h)||_1\\ &\le ||{\bar\psi}||_\infty\cdot ||h_\eps||_\infty\sum_{i=1}^{l-1}|| {\bar{\psi_{\eps}}}-{\bar\psi}||_1+ ||{\bar\psi}||^2_\infty\sum_{i=1}^{l-1} ||h_\eps-h ||_1. \end{split}$$ Finally we estimate $(III)$ $$\label{es3} \begin{split} (III)&\le||{\bar\psi}||_{\infty}\sum_{i=1}^{l-1}\sum_{j=0}^{i-1}||P_{\eps}^{i-1-j}(P_\eps-P)P^j({\bar\psi}h)||_1\\ &\le ||{\bar\psi}||_{\infty}\cdot |||P_{\eps}-P|||\cdot \sum_{i=1}^{l-1}\sum_{j=0}^{i-1}||P^{j}({\bar\psi}h)||_{1,1/p}. \end{split}$$ Combining estimates , and completes the proof of the lemma. (Proof of Theorem \[main2\]) We will show that for any $\eta> 0$ there exists an $\eps_\ast>0$ such that for all $\eps<\eps_\ast$ the inequality $|\sigma^2_{T_\eps}({\bar{\psi_{\eps}}})-\sigma^2_{T}({\bar\psi})|< \eta$ holds. $$\begin{split} &|\sigma^2_{T_\eps}({\bar{\psi_{\eps}}})-\sigma^2_{T}({\bar\psi})|\le |\int_I({\bar\psi}^2 h-{{\bar\psi}_{\eps}}^2 h_\eps)dx|\\ &+ 2|\sum_{i=1}^{l-1}\int_I\left(P^i_{\eps}({{\bar\psi}_{\eps}}h_{\eps}){{\bar\psi}_{\eps}}-P^i({\bar\psi}h){\bar\psi}\right)dx|\\ &+2||{\bar\psi}||_{\infty}\sum_{i=l}^{\infty}||P^i({\bar\psi}h)||_{1}+2||{{\bar\psi}_{\eps}}||_{\infty}\sum_{i=l}^{\infty}||P_\eps^i({\bar{\psi_{\eps}}}h_\eps)||_{1}\\ &:=(I)+(II)+(III)+(IV). \end{split}$$ We start with $(III)$. Since $\int_I{\bar\psi}hdx=0$, there are constants $C>0$ and $\kappa<\beta <1$ such that $$||P^i({\bar\psi}h)||_{1}\le||P^i({\bar\psi}h)||_{1,1/p}\le ||{\bar\psi}h||_{1,1/p}C \beta^{i} $$ Thus, choosing $ l_*$ large enough $$\label{esl} 2||{\bar\psi}||_{\infty}\sum_{i=l_*}^{\infty}||P^i({\bar\psi}h)||_{1}\le\frac{\eta}{4}.$$ Similarly we get $(IV)\le\frac{\eta}{4}$. Fix $l_*$ as in . Notice that we can choose $l_\ast$ for all $\eps<\eps_\ast$ uniformly, since $P_\eps$ satisfy the uniform Lasota-York inequalities. Now using Lemmas \[le1\] and \[le2\], we can find $\eps_*$ such that for all $\eps <\eps_\ast$ $$|\int_I({\bar\psi}^2 h-{\bar\psi}_{\eps}^2 h_{\eps})dx|+ 2|\sum_{i=1}^{l_*-1}\int_I\left(P^i_{\eps}({{\bar\psi}_{\eps}}h_{\eps}){{\bar\psi}_{\eps}}-P^i({\bar\psi}h){\bar\psi}\right)dx|\le\frac{\eta}{2}.$$ This completes the proof of the theorem. [99]{} V.S. Afraimovic, V.V. Bykov, L.P. Sili’nikov. The origin and structure of the Lorenz attractor. Dokl. Akad. Nauk SSSR 234 (1977), no. 2, 336 - 339. J. F. Alves, M. Soufi. Statistical stability of geometric Lorenz attractors. *Fund. Math.* 224 (2014), no. 3, 219–231. V. Araújo, I. Melbourne. Existence and smoothness of the stable foliation for sectional hyperbolic attractors. Available at https://arxiv.org/abs/1604.06924 V. Araújo, I. Melbourne. Exponential decay of correlations for nonuniformly hyperbolic flows with a $C^{1+\alpha}$ stable foliation, Including the classical Lorenz attractor. *Ann. Henri Poincaré*. 17 (2016), no. 11, 2975–3004. V. Araújo, I. Melbourne, P. Varandas. Rapid mixing for the Lorenz attractor and statistical limit laws for their time-1 maps.*Comm. Math. Phys.* 340 (2015), no. 3, 901–938. V. Araújo, M. J. Pacífico. Three-dimensional flows. In: Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics \[Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics\], vol. 53. Springer, Berlin, Heidelberg (2010). V. Araújo, M. J. Pacífico, E. R. Pujals, M. Viana. Singular-hyperbolic attractors are chaotic. *Trans. Amer. Math. Soc.* 361 (5) (2009), 2431–2485. W. Bahsoun, S. Galatolo, I. Nisoli, X. Niu. Rigorous approximation of diffusion coefficients for expanding maps. *J. Stat. Phys*. 163 (2016), no. 6, 1486–1503. V. Baladi. *Positive transfer operators and decay of correlations*. Advanced Series in Nonlinear Dynamics, 16. World Sci. Publ., NJ, 2000. A. Boyarsky, P. Góra. *Laws of Chaos, Invariant measures and Dynamical Systems in one dimension*. Birkhäuser, (1997). S. Galatolo, R. Lucena. Spectral Gap and quantitative statistical stability for systems with contracting fibres and Lorenz like maps. Available at https://arxiv.org/abs/1507.08191 J. Gukenheimer, R.F. Williams. Structural stability of Lorenz attractors. Inst. Hautes Études Sci. Publ. Math., 50:59-72, 1979. F. Hofbauer, G. Keller. Ergodic properties of invariant measures for piecewise monotonic transformations. *Math. Z.* 180 (1982), no. 1, 119–40. M. Holland, I. Melbourne. Central limit theorems and invariance principles for Lorenz attractors. *J. Lond. Math. Soc.* (2) 76 (2007), no. 2, 345–364. G. Keller. Gerhard Stochastic stability in some chaotic dynamical systems. *Monatsh. Math.* 94 (1982), no. 4, 313–333. G. Keller. Generalized bounded variation and applications to piecewise monotonic transformations. *Z. Wahrsch. Verw. Gebiete* 69 (1985), no. 3, 461–478. G. Keller, C. Liverani. Stability of the spectrum for transfer operators. *Ann. Scuola Norm. Sup. Pisa Cl. Sci.* (4) 28 (1999), no. 1, 141–152. C. Liverani. Central limit theorem for deterministic systems. International Conference on Dynamical Systems (Montevideo, 1995), 56-75, Pitman Res. Notes Math. Ser., 362, Longman, Harlow, 1996. E. D. Lorenz. Deterministic nonperiodic flow. *J. Atmosph. Sci.* 20 (1963) 130–141. S. Luzzatto, I. Melbourne, F. Paccaut. The Lorenz attractor is mixing. *Commun. Math. Phys.* 260, (2005), 393–401. I. Melbourne, A. Török. Statistical limit theorems for suspension flows, *Israel Math. J.* 144 (2004), 191-209. C. A. Morales, M. J. Pacífico, E. R. Pujals. Robust transitive singular sets for 3-flows are partially hyperbolic attractors or repellers. *Ann. of Math. (2)*, vol. 160, (2004), no. 2, 375–432. Y. G. Sinai. Gibbs measures in ergodic theory, *Russ. Math. Surv.* 27 (1972), 21-70. W. Tucker. The Lorenz attractor exists. C. R. Acad. Sci. Paris S´er. I Math. 328, (1999) 1197–1202. W. Tucker. A rigorous ODE solver and Smale’s 14th problem. *Found. Comput. Math.* 2 (2002) 53–117. [^1]: WB and MR would like to thank The Leverhulme Trust for supporting their research through the research grant RPG-2015-346. [^2]: For instance this assumption is automatically satisfied when ${\bar{\psi_{\eps}}}:=\hat\psi-\int\hat\psi d\bar\mu_\eps$ and ${\bar{\psi}}:=\hat\psi-\int\hat\psi d\bar\mu$ for some $\hat\psi\in BV_{1,1/p}$. [^3]: All the constants in this lemma are independent of $\eps$. [^4]: $\| f \|_1 \le \mu(X)^{1-1/p}\|f\|_p$, where $X$ is the space and $\mu$ is a measure on it. [^5]: Lemma 11 in [@Kel82] was proved for $f\in BV$, the space of one dimensional functions of bounded variation. Here we deal with functions in $BV_{1, 1/p}$. To keep the paper self contained, we include a proof. [^6]: Note that in [@AS] it is assumed that the stable foliation is $C^2$, and the $1$-d expanding maps are uniformly expanding. However, these assumptions are not used in section 3 of [@AS]. They were only used by [@AS] to use existing literature on strong statistical stability for piecewise $C^2$ and uniformly expanding $1$-d maps. [^7]: For the 1-d maps one can work with any observable in $BV_{1,1/p}$ which is not necessarily Hölder continuous. [^8]: The fact that exponential decay of correlations implies CLT, and that the variance admits a Green-Kubo formula follows from general theory. See for instance [@HK; @Li]. [^9]: The proof is similar to that of [@BGNX]. However in [@BGNX] a special structure on the observables was considered and the space under the consideration was $BV$ we include the proof in the appendix for completeness. [^10]: To prove the inequality we note that the corresponding solution started at $\xi$ satisfies the integral equality $$\label{gronwal} X_\eps(\xi, t)=\xi+\int_{0}^tX_\eps(X_\eps(\xi, s))ds, \quad\text{for} \quad \eps \ge 0.$$ Therefore, $$\begin{aligned} |X_\eps(\xi, t)-X(\xi, t)| &\le \int_{0}^t|X_\eps(X_\eps(\xi, s))-X(X_\eps(\xi, s))|ds\\ &+\int _0^t|X(X_\eps(\xi, s))-X(X(\xi, s))|ds \\ & \le \operatorname{dist}(X_\eps, X)t+{\|X\|}_{C^1}\int_{0}^{t}|X_\eps(\xi, s)-X(\xi, s)| ds.\end{aligned}$$ Applying Gronwal-Bellman inequality implies the desired estimate.

ArXiv

--- abstract: 'We prove that a divergence-free and $C^1$-robustly transitive vector field has no singularities. Moreover, if the vector field is $C^4$ then the linear Poincaré flow associated to it admits a dominated splitting over $M$.' author: - 'Mário Bessa [^1] and Jorge Rocha [^2]' title: '**On $C^1$-robust transitivity of volume-preserving flows**' --- *MSC 2000:* primary 37D30, 37D25; secondary 37A99.\ *keywords:* Volume-preserving flows; Robust transitivity; Dominated splitting; Ergodicity.\ [Introduction and statement of the\ results]{} It is well known that, in the $C^1$-topology, robust transitivity of a dynamical system defined on a compact manifold always implies some form of (weak) hyperbolicity. In fact in the early 1980s Mañé ([@M4]) proved that a $C^1$-robustly transitive two-dimensional diffeomorphism is uniformly hyperbolic. Mañé’s Theorem was generalized first by Díaz, Pujals and Ures ([@DPU]) showing that $C^1$-robustly transitive three-dimensional diffeomorphism are partially hyperbolic, and then by Bonatti, Díaz and Pujals ([@BDP]) obtaining that a $C^1$-robustly transitive diffeomorphism has dominated splitting. In the symplectomorphism case Horita and Tahzibi ([@HT]) showed that $C^1$-robust transitivity implies partial hyperbolicity in any dimension. Concerning the vector field context Doering ([@D]) transposed Mañé’s result to three-dimensional flows. Then, generalizing this result, Vivier ([@V]) showed that, in any dimension, $C^1$-robustly transitive vector fields do not have singularities, and Bonatti, Gourmelon and Vivier ([@BGV]) proved that the linear Poincaré flow of a $C^1$-robustly transitive vector field admits a dominated splitting. In the three-dimensional and volume-preserving case, Arbieto and Matheus ([@AM]) showed that a $C^1$-robustly transitive vector field is Anosov. Finally, Vivier ([@V2]) proved that any Hamiltonian vector field defined on a four-dimensional sympletic manifold and admitting a robustly transitive regular energy surface is hyperbolic on this energy surface. In this paper we consider the conservative flows setting (or, equivalently, the divergence-free vector fields scenario) and obtain the same kind of results of Vivier and of Bonatti, Gourmelon and Vivier mentioned above. Concerning the ergodic theoretical point of view we mention that, using the Mañé, Bochi and Viana strategies ([@M2] and [@BV]), in  [@B1] is proved that generically conservative linear differential systems have, for almost every point, zero Lyapunov exponents or else a dominated splitting. Before stating precisely our results let us introduce some definitions. Let $M$ be a compact, connected and boundaryless smooth Riemannian manifold of dimension $n \geq 4$. We denote by $\mu$ the Lebesgue measure induced by the Riemannian volume form on $M$. We say that a vector field $X$ is *divergence-free* if its divergence is equal to zero or equivalently if the measure $\mu$ is invariant for the associated flow, $X^t$, $t \in \mathbb{R}$. In this case we say that the flow is *conservative* or *volume-preserving*. We denote by $\mathfrak{X}_\mu^r(M)$ ($r\geq 1$) the space of $C^r$ divergence-free vector fields of $M$ and endow this set with the usual $C^1$-topology. A vector field $X$ is said to be *transitive* if its flow has a dense orbit in $M$. Moreover, $X$ is $C^1$-*robustly transitive* if there exists a $C^1$-neighbourhood of $X$ in $\mathfrak{X}_\mu^1(M)$ such that all its elements are transitive. Let us now state our first result. \[T2\] Let $X \in \mathfrak{X}_\mu^1(M)$ be a $C^1$-robustly transitive vector field. Then $X$ has no singularities. We denote by $Sing(X)$ the set of *singularities* of $X$ and by $R:=M\setminus Sing(X)$ the set of *regular* points. Given $x\in R$ we consider its normal bundle $N_{x}=X(x)^{\perp}\subset T_{x}M$ and define the *linear Poincaré flow* by $P_{X}^{t}(x):=\Pi_{X^{t}(x)}\circ DX^{t}_{x}$ where $\Pi_{X^{t}(x)}:T_{X^{t}(x)}M\rightarrow N_{X^{t}(x)}$ is the projection along the direction of $X(X^{t}(x))$. Let $\Lambda \subset R$ be an $X^t$-invariant set and $N=N^{1}\oplus N^{2}\oplus ... \oplus N^{k}$ be a $P_{X}^{t}$-invariant splitting over $\Lambda$ such that all the subbundles have constant dimension. We say that this splitting is an $\ell$-*dominated splitting* for the linear Poincaré flow if there exists an $\ell \in {\mathbb N}$ such that, for all $0 \leq i < j \leq k$ and for all $x\in \Lambda$ we have: $$\|P_{X}^{\ell}(x)|_{N^{j}_{x}}\| . \|(P_{X}^{\ell}(x)|_{N^{i}_{x}})^{-1}\| \leq \frac{1}{2}.$$ Previous result guaranties that a $C^1$-robustly transitive vector field has no singularities. Moreover, next theorem shows that if these vector fields are of class $C^4$ then they exhibit some type of weak hyperbolicity. \[T1\] Let $X \in \mathfrak{X}_\mu^4(M)$ be a $C^1$-robustly transitive vector field. Then $X$ has no singularities and the linear Poincaré flow of $X$ admits a dominated splitting over $M$. We point out that this theorem requires that the $C^1$-robustly transitive vector field $X$ is of class $C^4$ and not just of class $C^1$ as it would be expected. This hypothesis is a technical assumption needed to make $C^1$-conservative perturbations of the initial vector field $X$. Actually, the proofs of Theorems \[T2\] and \[T1\] are based on the ones made by Vivier ([@V]) and by Bonatti, Gourmelon and Vivier ([@BGV]) but, as the perturbations are made in the conservative class, we need to develop some appropriate $C^1$ perturbation lemmas, namely a kind of conservative Franks’ lemma, and for that we need to begin with $C^4$ regularity. We also refer that one of the main tools to get these perturbation lemmas is the Arbieto and Matheus Pasting Lemma ([@AM]). Let us now state some corollaries of the two theorems above. As, for $r \geq 2$, $C^r$ divergence-free vector fields are $C^1$-dense in $\mathfrak{X}_\mu^1(M)$ (Zuppa, [@Z]) we obtain the following corollary. Let $\mathcal{T}$ be the open set of $C^1$-robustly transitive vector fields $X \in \mathfrak{X}_\mu^1(M)$. There exists a $C^1$-dense subset $\mathcal{U}$ of $\mathcal{T}$ such that if $X \in \mathcal{U}$ then $\text{Sing}(X)=\emptyset$ and the linear Poincaré flow of $X$ admits a dominated splitting over $M$. Given a probability measure $\nu$ invariant for the flow $X^t$ we say that $\nu$ is an *ergodic* measure for $X$ if any measurable set that is invariant by the flow has zero or full measure. Equivalently, $\nu$ is ergodic if, for every observable continuous function $\varphi\colon M \rightarrow \mathbb{R}$ and for $\nu$-a.e. point $x \in M$, one has $$\int_M \varphi(z) d\nu(z)= \underset{t\rightarrow \infty}{\lim} \frac{1}{t}\int_0^t \varphi(X^s(x))ds.$$ We say that a $C^r$ vector field $X$ is $C^1$-*stably ergodic* with respect to a probability measure $\nu$, $r \geq 1$, if there exists a $C^1$-neighbourhood of $X$, $\mathcal{U}$, such that $\nu$ is an ergodic measure for $Y$, for all $Y \in \mathcal{U}$. In this paper we only consider ergodicity and stable ergodicity for the Lebesgue measure. It is well known that, for conservative systems, ergodicity implies transitivity. Therefore, as an immediate consequence of Theorem \[T1\], we get the following corollary. Let $X \in \mathfrak{X}_\mu^4(M)$ be a $C^1$-stably ergodic vector field. Then $\text{Sing}(X)=\emptyset$ and the linear Poincaré flow of $X$ admits a dominated splitting over $M$. Let $X \in \mathfrak{X}_\mu^1(M)$ be a vector field without singularities and let $E^1 \oplus E^2 ... \oplus E^k=N$ be a dominated splitting for the linear Poincaré flow associated to $X$. We consider the integrated Jacobian of $X$ restricted to each $E^i$ defined by $$\Sigma^i(X)=\int_M \log |\det P_X^1(x)|_{E^i} |d\mu(x),\,\,i \in \{1,2,...,k\}.$$ In [@BR] we prove that a $C^1$-stably ergodic vector field $X$, such that $M \setminus Sing(X)$ is partially hyperbolic and all the singularities are *linear hyperbolic*, can be $C^1$-approximated by a $C^2$-divergence-free vector field $Y$ such that $\Sigma^c(Y)\not=0$, where $\Sigma^c(Y)$ denotes the integrated Jacobian of $Y$ restricted to the central subbundle $E^c$. Actually, Theorem \[T2\] implies that a $C^1$-stably ergodic vector field does not have singularities. We also remark that the proof given on [@BR] only requires the existence of a dominated splitting. This fact was already observed in [@BB] in the diffeomorphism context. Hence, Theorem 1 of  [@BR] can be reformulated as follows. \[nbr\] Let $X \in \mathfrak{X}^{1}_{\mu}(M)$ be a stably ergodic flow and $E^1 \oplus E^2 ... \oplus E^k=N$ be a dominated splitting for the linear Poincaré flow of $X$ over $M$. Then $X$ may be approximated, in the $C^{1}$-topology, by $Y \in \mathfrak{X}^{\infty}_{\mu}(M)$ for which $\Sigma^i(Y) \not=0$, for all $i \in \{1,2,...,k\}$. Note that if $X \in \mathfrak{X}^{4}_{\mu}(M)$ then the previous result applies directly to the dominated splitting given by Theorem \[T1\]. We recall that a conservative vector field $X$ is said to be *nonuniformly hyperbolic* if all the Lyapunov exponents are a.e. different from zero. In particular if $M$ is four dimensional, $X \in \mathfrak{X}_\mu^4(M)$ is $C^1$-stably ergodic and admits a dominated splitting with three nontrivial subbundles then the previous corollary assures that $X$ can be $C^1$-approximated by a nonuniformly hyperbolic vector field. If $M$ is a four dimensional manifold and $X \in \mathfrak{X}_\mu^4(M)$ is stably ergodic then $M$ admits a dominated splitting, $E \oplus F$, for the linear Poincaré flow associated to $X$. Since the vector field is divergence-free and has no singularities it is straightforward to see that the one-dimensional invariant subbundle is hyperbolic and the other subbundle is hyperbolic in volume, that is the splitting is a partially hyperbolic one. Therefore, using again Zuppa’s Theorem ([@Z]), we obtain the following result. A $C^1$-stably ergodic vector field $X \in \mathfrak{X}_\mu^1(M^4)$ can be $C^1$-approximated by a partially hyperbolic vector field. This paper is organized as follows. In section \[tour\] we state three results (Proposition \[V1\], Proposition \[V2\] and Proposition \[VF\]) and deduce the theorems from the first two. Proposition \[V1\] is an easy adaptation of Proposition 4.1 of Vivier and we show that Proposition \[VF\] implies Proposition \[V2\]. In section  \[pl\] we obtain the perturbation lemmas needed to prove Proposition \[V1\] and Proposition \[VF\] and, finally, in section \[MG\] we prove Proposition \[VF\]. [A *tour* on the proofs of the theorems]{}\[tour\] In this section we prove Theorems \[T2\] and \[T1\] following the strategy used by Vivier in  [@V] and by Bonatti, Gourmelon and Vivier in  [@BGV] and adapting some of their results to the conservative setting. For that we begin by stating, in the divergence-free vector fields scenario, two main results of the references above. \[V1\] Let $X \in \mathfrak{X}_\mu^1(M)$ be a robustly transitive vector field such that $Sing(X)\not= \emptyset$. Then there exists an arbitrarily $C^1$-close vector field $Y \in \mathfrak{X}_\mu^4(M)$ such that the linear Poincaré flow of $Y$ does not admit any dominated splitting over $M\setminus Sing(Y)$. \[V2\] Let $X \in \mathfrak{X}_\mu^4(M)$ be a robustly transitive vector field. Then there exists a dominated splitting for the linear Poincaré flow of $X$ over $M\setminus Sing(X)$. Let us now explain how we derive Theorems \[T2\] and \[T1\] from the previous propositions. Fix a robustly transitive vector field $X \in \mathfrak{X}_\mu^1(M)$; let $\mathcal{U}$ be a $C^1$ open neigbourhood of $X$ such that all $Y \in \mathcal{U}$ are transitive, hence robustly transitive. Let us assume that $Sing(X)\not=\emptyset$. Applying Lemma \[linear\], we obtain $Y \in \mathfrak{X}_\mu^{\infty}(M) \cap \mathcal{U}$ such that $Sing(Y)\not= \emptyset$ and $Y$ has at least �ne hyperbolic singularity. Therefore, by Proposition \[V1\], we obtain a vector field $Z \in \mathfrak{X}_\mu^4(M) \cap \mathcal{U}$ such that $Sing(Z)\not= \emptyset$ and the linear Poincaré flow of $Z$ does not admit any dominated splitting over $M\setminus Sing(Z)$, which is in contradiction with Proposition \[V2\] applied to $Z$. Therefore $Sing(X)=\emptyset$, which proves Theorem \[T2\]. Now let $X \in \mathfrak{X}_\mu^4(M)$ be a $C^1$-robustly transitive vector field; the previous argument shows that $Sing(X)=\emptyset$ and then Proposition \[V2\] guarantees that $M$ admits a dominated splitting for the linear Poincaré flow of $X$, thus proving Theorem \[T1\]. Proposition \[V1\], up to a minor detail, is a consequence of Proposition 4.1 of [@V]. To see this let us first recall that a singularity $p$ of a given vector field $X$ is said to be a *linear hyperbolic singularity* if it is a hyperbolic singularity and there exist smooth local coordinates that conjugate $X$ and $DX_p$ in a neighbourhood of $p$. In Lemma \[linear\] we prove that any $X \in \mathfrak{X}_\mu^1(M)$ having a singularity $p$ can be $C^1$-approximated by a vector field $Y \in \mathfrak{X}_\mu^{\infty}(M)$ such that $p$ is a linear hyperbolic singularity of $Y$. Since $Y$ is a divergence-free vector field it follows that $p$ is of the saddle-type. Now Proposition \[V1\] is a direct consequence of the following result. (Proposition 4.1 of [@V]) If $Y \in \mathfrak{X}^1(M)$ admits a linear hyperbolic singularity of saddle-type, then the linear Poincaré flow of $Y$ does not admit any dominated splitting over $M\setminus Sing(Y)$. Proposition \[V2\] is a consequence of the following result, which is an adaptation to the conservative setting of Corollary 2.22 of [@BGV], whose proof is postponed to section \[MG\]. \[VF\] Let $X \in \mathfrak{X}_\mu^4(M)$ be a $C^1$-robustly transitive vector field and let $\mathcal{U}$ be a small $C^1$-neighbourhood of $X$. There exist $\ell, \varrho \in \mathbb{R}_0^+$ such that for any $Y \in \mathcal{U}$ and any periodic orbit $x$ of $Y^t$, of period $\pi(x) \geq \varrho$, the linear Poincaré flow of $Y$ admits an $\ell$-dominated splitting over the $Y^t$-orbit of $x$. Let us explain how Proposition \[V2\] is deduced from the previous result. Fix a robustly transitive vector field $X \in \mathfrak{X}_\mu^4(M)$ and let $\mathcal{U}$ be a $C^1$-neighbourhood of $X$ as in the previous proposition and such that any $C^1$ vector field $Y \in \mathcal{U}$ is also robustly transitive. Consider $\ell$ and $\varrho$ given by Proposition \[VF\]. Let $x \in M$ be a point with dense $X^t$-orbit. Using Pugh and Robinson’s volume-preserving closing lemma ([@PR]) we get a sequence of vector fields $X_n \in \mathfrak{X}_\mu^4(M)$, converging to $X$ in the $C^1$-topology, and, for each $n \in \mathbb{N}$, $X_n$ has a periodic orbit $\Gamma_n=\Gamma_n(t)$ of period $\pi_n$ such that $\lim_{n \rightarrow +\infty} \Gamma_n(0)=x$. In particular $\lim_{n \rightarrow +\infty} \pi_n=+\infty$. Therefore, for large $n$, we can apply Proposition \[VF\] to conclude that there is an $\ell$-dominated splitting for the linear Poincaré flow of $X_n$ over the orbit $\Gamma_n$; taking a subsequence if necessary, we can assume that the dimensions of the invariant bundles do not depend on $n$. Finally, as $\bigcap_N(\overline{\cup_N^{+\infty}\Gamma_n})=M$, by well known results on dominated splittings (see for example [@BDV]) it follows that there exists an $\ell$-dominated splitting for the linear Poincaré flow of $X$ over $M\setminus Sing(X)$. [Perturbation Lemmas]{}\[pl\] In this section we state and prove three perturbation lemmas needed to obtain the main results of this article. In particular the Main Perturbation Lemma (Lemma \[mpl\]) is a kind of Franks’ Lemma ([@F]) for conservative flows. As we mention before, for technical reasons, we required that the vector fields involved are of class $C^4$. One of the main tools to obtain this result is the Arbieto and Matheus Pasting Lemmas ([@AM]). We refer that their result and our Main Perturbation Lemma make use of a key result of Dacorogna and Moser ([@DM]). We fix a vector field $X \in \mathfrak{X}_\mu^4(M)$, $\tau >0$, and a point $p \in M$ such that $X^t(p)\not=p$, for all $t \in [0, \tau]$. Define $\Gamma(p, \tau)=\{X^t(p);\,\, t \in [0, \tau]\}$. In the sequel up to a smooth conservative change of coordinates defined on a neighbourhood of $\Gamma(p,\tau)$ (see [@Mo]) we can assume that we are working on $\mathbb{R}^n$ and that $\frac{1}{\|X(p)\|}X(p)=\frac{\partial}{\partial x_1}=v$. Let $W\subset \mathbb{R}^n$ be the $(n-1)$-dimensional vectorial subspace orthogonal to the unitary vector $v$. Given $r>0$ let $B_r(p)$ denote the ball of radius $r$, centered at $p$ and contained in $N_p=X(p)^{\perp}=W$. For $r>0$ and $\delta >0$ define $$\mathcal{T}=\mathcal{T}(p,\tau,r, \delta)=\bigcup_{t \in ]-\delta, \tau+\delta[} X^t(B_r(p)).$$ If $r>0$ and $\delta >0$ are small enough the set $\mathcal{T}$ is an open neighbourhood of $\Gamma(p,\tau)$; by definition this neighbourhood is foliated by orbits of the flow so we call it a *flowbox*. We fix a linear isometry $\iota_p \colon N_p \rightarrow W$ and choose a family $\{\iota_t\}_{t \in ]-\delta, \tau+\delta[}$, such that, for each $t \in ]-\delta, \tau+\delta[$, $\iota_t$ is a linear isometry from $N_{X^t(p)}$ onto $W$, $\iota_0=\iota_p$, and this family is $C^1$ on the parameter $t$. Given $q \in \mathcal{T}$, we write $q=\lambda_q v+w_q$, where $w_q \in W$ and $\lambda_q \in \mathbb{R}$. Define $\ell(t)=\int_0^t \|X(X^s(p))\|ds$; there exists $t_q \in ]-\delta, \tau + \delta[$ such that $\ell(t_q)=\lambda_q$. We note that $t_p=0$. Let us now define the *Poincaré flow* $\hat{X}^t$ associated to $X$ on $\mathcal{T}$. For $t$ such that $t_q+t \in ]-\delta, \tau + \delta[$ define $$\hat{X}^t(q)= \ell(t_q+t)v+ \iota_{(t_q+t)}\circ P_X^{t}(X^s(p))\circ \iota_{t_q}^{-1}(w_q).$$ It is straightforward to see that $\hat{X}^0 \equiv Id$ and that $\hat{X}^{t+t^\prime}(q)=\hat{X}^t(\hat{X}^{t^\prime}(q))$, when defined. Let $\hat{X}$ be the vector field associated to the flow $\hat{X}^t$. $\hat{X}$ is of class $C^2$ and it is divergence-free. To see this we first recall the Liouville formula $$exp \left(\int_0^t div(\hat{X}(\hat{X}^s(q)))ds\right) =\det D\hat{X}^t(q).$$ Now a direct computation gives that the matrix of $D\hat{X}^t(q)$ relatively to the decomposition $\mathbb{R}^n=W\oplus \langle v \rangle $ is $$\begin{pmatrix} \frac{\|X(X^{t_q+t}(p))\|}{\|X(X^{t_q}(p))\|} & 0 \\ \ast & \iota_{t_q+t}\circ P_X^{t}(X^{t_q}(p))\circ \iota_{t_q}^{-1} \end{pmatrix}.$$ As $X^t$ is volume-preserving and the maps $\iota_s$ are linear isometries we get $$\det D\hat{X}^t(q)=\frac{\|X(X^{t_q+t}(p))\|}{\|X(X^{t_q}(p))\|} \times \det P_X^{t}(X^{t_q}(p))=1,\,\,\forall t.$$ Thus, according to Liouville’s formula, it follows that $div(\hat{X})=0$. We also observe that $P_{\hat{X}}^t(q)=\iota_{t_q+t}\circ P_X^{t}(X^{t_q}(p))\circ \iota_{t_q}^{-1}$; in particular $P_{\hat{X}}^t(0)=\iota_{t}\circ P_X^{t}(p) \circ \iota_0^{-1}$. \[hat\] Let $X \in \mathfrak{X}_\mu^4(M)$, $\tau >0$, and $p \in M$ such that $X^t(p)\not=p$, $\forall t \in [0, \tau]$. There exists a $C^2$-conservative change of coordinates $\Phi$, defined on a neighbourhood of $\Gamma(p,\tau)$, such that $$\hat{X}=\Phi_{*}X \text{ and } \Phi(X^t(p))=\hat{X}^t(0),\,\,\forall t \in [0, \tau].$$ We recall that $\Phi_{*}X(y)=D\Phi_{\Phi^{-1}(y)}X(\Phi^{-1}(y))$. By Lemma 2.1 of [@BR] we know that there exists a conservative $C^{2}$ diffeomorphism $\Psi$ defined on a flowbox containing $\Gamma(p,\tau)$ such that $T=\Psi_{*}X$, where $T= \frac{\partial}{\partial x_1}$. Exactly in the same way there exists a conservative $C^{2}$ diffeomorphism $\hat{\Psi}$ defined on a flowbox containing $\Gamma(0,\tau)$ such that $T=\hat{\Psi}_{*}\hat{X}$. Up to translations defined on the hyperplane $\langle \frac{\partial}{\partial x_1} \rangle^{\perp}$ and shrinking the neigbourhoods of the definition of these maps, we can assume that $\Psi(p)=0$ and that $\hat{\Psi}(0)=0$. Finally, we define $\Phi=\hat{\Psi}^{-1}\circ \Psi$. Now let $V, V^{\prime} \subset N_p$, $dim(V)=j$, $2 \leq j \leq n-1$, and $N_p=V \oplus V^{\prime}$. A *one-parameter linear family* $\{A_t\}_{t\in \mathbb{R}}$ associated to $\Gamma(p, \tau)$ and $V$ is defined as follows: - $A_t\colon N_p \rightarrow N_p$ is a linear map, for all $t\in \mathbb{R}$, - $A_t=Id$, for all $t\leq 0$, and $A_t=A_{\tau}$, for all $t\geq \tau$, - $A_t|_V \in SL(j, \mathbb{R})$, and $A_t|_{V^{\prime}}\equiv Id$, $\forall t \in [0, \tau]$, in particular we have $\det(A_t)=1$, for all $t\in \mathbb{R}$, and - the family $A_t$ is $C^\infty$ on the parameter $t$. (Main perturbation lemma)\[mpl\] Given $\epsilon>0$ and a vector field $X \in \mathfrak{X}_\mu^4(M)$ there exists $\xi_0=\xi_0(\epsilon,X)$ such that $\forall \tau \in [1,2]$, for any periodic point $p$ of period greater than $2$, for any sufficient small flowbox $\mathcal{T}$ of $\Gamma(p, \tau)$ and for any one-parameter linear family $\{A_t\}_{t \in [0, \tau]}$ such that $\| A_t^{\prime} A_t^{-1}\|<\xi_0$, $\forall t \in [0, \tau]$, there exists $Y \in \mathfrak{X}_\mu^1(M)$ satisfying the following properties 1. $Y$ is $\epsilon$-$C^1$-close to $X$; 2. $Y^t(p)=X^t(p)$, for all $t \in \mathbb{R}$; 3. $P_Y^\tau(p)=P_X^\tau(p) \circ A_{\tau}$, and 4. $Y|_{\mathcal{T}^c}\equiv X|_{\mathcal{T}^c}$. Using Lemma \[hat\] we get a $C^2$ change of coordinates $\Phi$, defined in a flowbox $\mathcal{T}=\mathcal{T}(p, \tau, r, \delta)$ and such that $\Phi_{*}X=\hat{X}$ and $\Phi(X^t(p))=\hat{X}^t(0)$, $\forall t \in [0, \tau]$, where $\hat{X}^t$ is the Poincaré flow associated to $X$ on $\mathcal{T}$ defined above. To obtain the vector field $Y$ we first construct a $C^2$ divergence-free vector field $\hat{Y}$ defined on $\Phi(\mathcal{T})$ and such that \(a) $\hat{Y}$ is $\hat{\epsilon}$-$C^1$-close to $\hat{X}$; \(b) $\hat{Y}^t(0)=\hat{X}^t(0)$, when defined; \(c) $P_{\hat{Y}}^t(0)=P_{\hat{X}}^t(0) \circ B_{t}$, where $B_{t}=\iota_p \circ A_t$, $t\in [0, \tau]$, and \(d) $\hat{Y}|_{\hat{\mathcal{T}}^c}\equiv \hat{X}|_{\hat{\mathcal{T}}^c}$, where ${\hat{\mathcal{T}_2}}=\Phi(\mathcal{T}(p, \tau, r_2, \delta_2))$, for some $0<r_2<r$ and $0<\delta_2<\delta$ to be fixed. The positive real number $\hat{\epsilon}$ depends only on $\Phi$ and $\mathcal{T}$ and assures that if $Z$ is $\hat{\epsilon}$-$C^1$-close to $\hat{X}$ on $\Phi(\mathcal{T})$ then ${\Phi^{-1}}_{\ast}(Z)$ is $\epsilon$-$C^1$-close to $X$ on $\mathcal{T}$. Once we get $\hat{Y}$ we define $\tilde{Y}={\Phi^{-1}}_{\ast}(\hat{Y})$ and, as $\tilde{Y}\equiv X$ on $\mathcal{T} \setminus \mathcal{T}(p, \tau, \frac{r}{2}, \frac {\delta} {2})$, we consider $Y=\tilde{Y}$ on $\mathcal{T}$ and $Y=X$ on $\mathcal{T}^c$. We observe that we can only guaranty that $Y$ is of class $C^1$. From this construction it follows immediately that items *1.*, *2.* and *4.* of the lemma are a direct consequence of conditions *(a)*, *(b)* and *(d)* on $\hat{Y}$, respectively. To get item *3.* we observe that our construction of $\hat{Y}$ will imply that $P_{\hat{Y}}^t(p)=\iota_t \circ P_{\tilde{Y}}^t(p) \circ \iota_p^{-1}$ where, recall, $\tilde{Y}={\Phi^{-1}}_{\ast}(\hat{Y})$. Therefore, as $P_{\hat{X}}^t(p)=\iota_t \circ P_{X}^t(p) \circ \iota_p^{-1}$ and $B_{t}=\iota_p \circ A_t$, from condition *(c)* we obtain *3.*. Let us now explain how to construct the vector field $\hat{Y}$ defined on $\Phi(\mathcal{T})$. The linear variational equation associated to the linear Poincaré flow of $\hat{X}$ is $$[P_{\hat{X}}^t(0)]^\prime=(\Pi \circ D\hat{X}_{\hat{X}^t(0)})(P_{\hat{X}}^t(0)),$$ where $^\prime$ denotes the time derivative, $D$ is the spacial derivative and $\Pi$ is the orthogonal projection onto $N_{\hat{X}^t(0)}$. To get $\hat{Y}$ we begin by considering an analogous linear variational equation associated to $P_{\hat{Y}}^t(0)$ in order to obtain $D\hat{Y}$ along the orbit of the point $0$ and then define in a linear way the flow $\hat{Y}^t$. Since we require that $P_{\hat{Y}}^t(0)=P_{\hat{X}}^t(0) \circ B_{t}$ we have that $$\begin{aligned} [P_{\hat{Y}}^t(0)]^\prime&=&[P_{\hat{X}}^t(0) \circ B_{t}]^\prime=[P_{\hat{X}}^t(0)]^\prime \circ B_{t} + P_{\hat{X}}^t(0) \circ {B_{t}}^{\prime}\\ &=& (\Pi \circ D\hat{X}_{\hat{X}^t(0)})(P_{\hat{X}}^t(0)) \circ B_t + P_{\hat{X}}^t(0) \circ {B_{t}}^{\prime}\\ &=& \left[\Pi \circ D\hat{X}_{\hat{X}^t(0)}+ (P_{\hat{X}}^t(0) \circ {B_{t}}^{\prime}) \circ (B_t^{-1} \circ P_{\hat{X}}^{-t}(\hat{X}^t(0)))\right]\circ P_{\hat{Y}}^t(0).\end{aligned}$$ This equation allows us to define, along the $\hat{X}^t$ orbit of $0$, the infinitesimal generator $D\hat{X}+\mathcal{H}$ of the desired perturbation, where $\mathcal{H}_{\hat{X}^t(0)}(v)=\vec{0}$ and $$\label{C} \Pi \circ \mathcal{H}_{\hat{X}^t(0)}=(P_{\hat{X}}^t(0) \circ {B_{t}}^{\prime}) \circ (B_t^{-1} \circ P_{\hat{X}}^{-t}(\hat{X}^t(0)))=C_t.$$ With this definition the previous differential equation can be written as: $$\label{D} u^{\prime}(t)=\Pi \circ (D\hat{X}+ \mathcal{H})_{\hat{X}^t(0)}(u(t)).$$ Let $\mathcal{P}(\lambda v, w)=(0, C_t(w))$, where $t$ is given by $\int_0^t\|\hat{X}(\hat{X}^s(0))\|ds= \lambda$; observe that $$\label{W} D_w\mathcal{P}(\hat{X}^t(0))=\Pi \circ \mathcal{H}_{\hat{X}^t(0)},$$ and that $$\label{U} D_w\mathcal{P}(\hat{X}^t(0))(0,u)= \Pi \circ D\mathcal{P}_{\hat{X}^t(0)}(0,u).$$ Now we define the $C^2$ vector field $\hat{Y}(q)=(\hat{X} + \mathcal{P})(q)$, for $q \in \Phi(\mathcal{T}(p,\tau, r_1, \delta_1))$, where $0<r_1<r$ and $0<\delta_1<\delta$ will be fixed. In order to extend this vector field let us first prove that it is divergence-free. As $\hat{Y}=\hat{X} + \mathcal{P}$ and $\hat{X}$ is divergence-free, using (\[W\]), (\[C\]) and the definition of the maps $B_t$, it follows that $$div(D\hat{Y})=div(D\mathcal{P})=tr(C_t)=tr(B_t^{\prime} \circ B_t^{-1})=tr(A_t^{\prime} \circ A_t^{-1}).$$ Now, as $\det(A_t)=1, \forall t \in \mathbb{R}$, the result follows observing that $$0=(\det(A_t))^{\prime}=tr(A_t^{\prime}A_t^{-1})\det(A_t)=tr(A_t^{\prime}A_t^{-1}).$$ Now, to extend $\hat{Y}$ to a conservative vector field, we apply the Arbieto and Matheus *$C^{1+\alpha}$ -Pasting Lemma* (Theorem 3.1, [@AM]) which guaranties that there are $0<r_1<r_2<r$ and $0<\delta_1<\delta_2<\delta$ such that $\hat{Y}$ has a divergence-free $C^2$ extension to $\Phi(\mathcal{T}(p,\tau, r, \delta))$, that we also denote by $\hat{Y}$, with $\hat{Y}(q)=\hat{X}(q)$, for all $q \in \Phi(\mathcal{T}(p,\tau, r, \delta)) \setminus \Phi(\mathcal{T}(p,\tau, r_2, \delta_2))$. Let us now prove that this vector field $\hat{Y}$ satisfies properties *(a)*-*(d)*. Condition *(d)* is a direct consequence of the way we made the extension of $\hat{Y}$. To get *(b)* just observe that $$\begin{aligned} \hat{Y}(\hat{X}^t(0))&=&\hat{X}(\hat{X}^t(0))+\mathcal{P}\left(\int_0^t \|\hat{X}( \hat{X}^s(0))\|ds,\vec{0}\right)\\ &=&\hat{X}(\hat{X}^t(0))+(0,C_t(\vec{0}))=\hat{X}(\hat{X}^t(0)).\end{aligned}$$ To get *(c)* let us first remark that the linear Poincaré flow of $\hat{Y}$ at $0$, $P_{\hat{Y}}^t(0)$, is the solution of the differential equation $u^{\prime}(t)=\Pi \circ D\hat{Y}_{\hat{Y}^t(0)}(u(t))$. By equations (\[U\]) and (\[W\]) we have that $$\Pi \circ D\hat{Y}_{\hat{Y}^t(0)}(u(t))=\Pi \circ D(\hat{X}+\mathcal{P})_{\hat{X}^t(0)}=\Pi \circ D\hat{X}_{\hat{X}^t(0)}+ \Pi \circ \mathcal{H}_{\hat{X}^t(0)}.$$ Hence, by (\[C\]) and (\[D\]), we get that $P_{\hat{Y}}^t(0)=P_{\hat{X}}^t(0) \circ B_{t}$. To prove condition *(a)* we begin by observing that the Pasting Lemma guaranties that given $\hat{\epsilon}$ there exists $\sigma>0$ such that if the vector field $\hat{Y}$ we constructed is $\sigma$-$C^1$ close to $\hat{X}$ on $\Phi(\mathcal{T}(p,\tau, r_1, \delta_1))$ then its extension to $\Phi(\mathcal{T}(p,\tau, r, \delta))$ is $\hat{\epsilon}$-$C^1$ close to $\hat{X}$, and to take $r_1$ and $\delta_1$ smaller does not change this $C^1$ closeness. So let us prove that $\hat{Y}$ is $\sigma$-$C^1$ close to $\hat{X}$ on $\mathcal{T}_1=\Phi(\mathcal{T}(p,\tau, r_1, \delta_1))$ for sufficiently small $r_1$, $\delta_1$ and $\xi_0$. Recalling that $\hat{Y}=\hat{X}+ \mathcal{P}$, that $\mathcal{P}(\hat{X}^t(0), \vec{0})=0$ and that $\mathcal{P}$ is continuous, to choose $r_1$ and $\delta_1$ small is enough to assure $\|\hat{Y}-\hat{X}\|_0< \sigma$ on $\mathcal{T}_1$. We observe that the matrix of $D\mathcal{P}(\lambda,w)$ depends only on map $C_t$ and on real numbers $$\left|\frac{\partial [C_t(w)]^i}{\partial t} \frac{\partial t}{\partial \lambda}w_i \right|, \,\,\,i \in \{1,...,n-1\},$$ where $w=(w_1,...,w_{n-1})$ and $\|w\|<r_1$. By the definition of $C_t$ (see (\[C\])), up to constants that depend only on $X$, its norm is given by $$\|B^{\prime}_t\circ B_t^{-1}\|=\|\iota_p \circ A^{\prime}_t\circ A_t^{-1} \circ \iota_p^{-1}\|= \|A^{\prime}_t\circ A_t^{-1}\|,$$ because the map $\iota_p$ is an isometry. Hence, by the hypothesis, it is enough to take $\xi_0$ sufficiently small to get that $\|D\mathcal{P}\| < \sigma$. This ends the proof of the lemma. \[linear\] Let $p$ be a singularity of $X \in \mathfrak{X}_\mu^1(M)$. For any $\epsilon >0$ there exists $Y \in \mathfrak{X}_\mu^{\infty}(M)$, such that $Y$ is $\epsilon$-$C^1$-close to $X$ and $p$ is a linear hyperbolic singularity of $Y$. Let $(U,\phi)$ be a conservative chart given by Moser’s Theorem ([@Mo]) such that $p \in U$ and $\phi(p)=0$. Let $A=DX_p$ and, for arbitrarily small $\tilde{\delta}>0$, choose a linear and hyperbolic isomorphism $H=H_{\delta}$ such that $\|A-H\|<\delta$. We fix small $r>0$ such that $B(0;r)\subset \phi(U)$ and consider the pull-back of $H$, $Z=Z_{\tilde{\delta},r}={(\phi^{-1})}_\ast H$, defined on $\phi^{-1}(B(0;r))$. For any small $\delta>0$ there are $\tilde{\delta}>0$ and $r>0$ such that $X$ and $Z$ are $\delta$-$C^1$-close. Therefore a straightforward application of the $C^1$-pasting lemma (Theorem 3.2, [@AM]) to $X$ and $\epsilon$ (which give a $\delta$) guarantees that there exists a vector field $Y \in \mathfrak{X}_\mu^{\infty}(M)$ such that $Y$ is $\epsilon$-$C^1$-close to $X$ and $Y|_W=Z_W$, where $W=\phi^{-1}(B(0; \frac{r}{2}))$. From the construction it follows that $p$ is a linear hyperbolic singularity of $Y$. \[linear2\] Let $X \in \mathfrak{X}_\mu^1(M)$ and assume that $p \in M$ is a periodic elliptic point of period $\pi(p)> 1$. Then, for any $\epsilon >0$ there exist $Z \in \mathfrak{X}_\mu^{\infty}(M)$ and a tubular neighbourhood $\mathcal{U}$ of the $X^t$-orbit of $p$, such that $Z$ is $\epsilon$-$C^1$-close to $X$ and $\mathcal{U}$ is $Z^t$-invariant. Let $p \in M$ is a periodic elliptic point of period $\pi(p)> 1$ of $X$ and denote its orbit by $\gamma=\Gamma(p, \pi(p))$. As $p$ is elliptic all the eigenvalues of the linear Poincaré map $P_X^{\pi(p)}(p):N_p \rightarrow N_p$ have modulus one. Let us first assume that the map $P_X^{\pi(p)}(p)$ admits a basis formed by eigenvectors. We consider the inner product defined on $N_p$ and associated to this basis, that is the one that orthonormalizes the fixed basis. For $r>0$ let $D(0;r) \subset N_p$ denote the $(n-1)$-disk centered at $0 \in N_p$ and of radius $r$ for the distance associated to this inner product. Note that $D(0;r)$ is $P_X^{\pi(p)}(p)$-invariant. Let $Y$ be the divergence-free linear vector field associated to the flow obtained by suspending $P_X^{\pi(p)}(p)$ along $\gamma$; this vector field is defined in a tubular neighbourhood of $\gamma$, $\mathcal{U}(r)$, which is homeomorphic to $\gamma \times D(0;r)$. Given $\delta>0$ we can choose a small $r$ such that $X$ and $Y$ are $\delta$-$C^1$-close on $\mathcal{U}(r)$. Now, for fixed $\epsilon$ and an appropriate $\delta$ we apply the $C^1$-pasting lemma (Theorem 3.2, [@AM]) to get a vector field $Z \in \mathfrak{X}_\mu^{\infty}(M)$ such that $X$ and $Z$ are $\epsilon$-$C^1$-close on $M$ and $Z|_{\mathcal{U}}=Y|_{\mathcal{U}}$, where $\mathcal{U}=\mathcal{U}(\frac{r}{2})$. It follows from this construction that $\mathcal{U}$ is $Z^t$-invariant. Assume now that $P_X^{\pi(p)}(p)$ does not admit a basis formed by eigenvectors, that is there exists at least one eigenvalue whose multiplicity is bigger than the dimension of the associated eigenspace. Let us first explain how to deal with the simplest case, that is when $dim(M)=3$, $P_X^{\pi(p)}(p)$ has only one eigenvalue, say equal to $1$, and the associated eigenspace is one-dimensional. In this case we will perturb $X$ in order to get complex eigenvalues. For that and exactly as before we begin by $\frac{\epsilon}{2}$-$C^1$-approximate $X$ by a $C^{\infty}$ vector field $Y$ linear in a neighbourhood of $\gamma$ and such that $P_X^{\pi(p)}(p)=P_Y^{\pi(p)}(p)$. Now, there exists a basis of $N_p$ such that relatively to this basis $P_Y^{\pi(p)}(p)$ has matrix $$\begin{pmatrix} 1&0\\1&1 \end{pmatrix}.$$ Consider the one-parameter linear family $$A_t= \begin{pmatrix} 1&-\delta \alpha(t) \\0&1 \end{pmatrix},$$ where $\alpha(t)$ is a $C^{\infty}$-bump function, $\alpha(t)=1$, for all $t \geq 1$, $\alpha(t)=0$, for all $t \leq 0$, and $\delta>0$ is arbitrarily small. As $\|A_t^{\prime} \circ A_t^{-1}\| \leq \delta |\alpha^{\prime}(t)|$, choosing $\delta$ small enough we can apply Lemma \[mpl\] to the arc $\Gamma(p,1)$ in order to obtain a new vector field $Z \in \mathfrak{X}_\mu^{\infty}(M)$, $\frac{\epsilon}{2}$-$C^1$-close to $Y$, such that $p$ is a periodic orbit of $Z$ of period $\pi(p)$ and $P_Z^1(p)=P_Y^1(p)\circ A_1=P_X^1(p)\circ A_1$. As $$P_Z^{\pi(p)}(p)= P_Z^{\pi(p)-1}(Z^1(p)) \circ P_Z^1(p)=P_X^{\pi(p)-1}(X^1(p)) \circ P_X^1(p) \circ A_1,$$ it follows that the matrix of $P_Z^{\pi(p)}(p)$ with respect to the basis we fixed above is $$\begin{pmatrix} 1&-\delta\\1&1-\delta \end{pmatrix}.$$ Therefore $p$ is an elliptic point of $Z$ and $P_Z^{\pi(p)}(p)$ has two complex eigenvalues. Moreover, it is clear that $Z$ and $X$ are $\epsilon$-$C^1$-close. If the eigenvalue is equal to $-1$ we proceed in the same way considering the matrix $$A_t= \begin{pmatrix} 1&\delta \alpha(t) \\0&1 \end{pmatrix}.$$ Finally, to deal with the general case (several eigenvalues, real or complex, whose multiplicity is greater than the dimension of the corresponding eigenspace) we just have to apply the previous argument a finite number of times in order to get a new vector field $Z$ such that $p$ is a periodic elliptic orbit, of period $\pi(p)$, and such that $P_Z^{\pi(p)}(p)$ admits a basis of eigenvectors. [Proof of Proposition \[VF\]]{}\[MG\] [Conservative Linear Differential Systems]{} We begin this section by recalling some definitions introduced in [@BGV]. Let $X \in \mathfrak{X}_\mu^1(M)$ and consider a set $\Sigma \subset M$ which is a countable union of periodic orbits of $X^t$. A *Linear Differential System* (LDS) is a four-tuple $\mathcal{A}=(\Sigma, X^t, N_{\Sigma}, A^{\star})$, where $N_{\Sigma}$ is the restriction to $\Sigma$ of the normal bundle of $X$ over $M\setminus Sing(X)$ and $A^{\star} \colon \Sigma \rightarrow GL(n-1,\mathbb{R})$ is a continuous map. In fact, for $x \in \Sigma$, $A^{\star}_x$ is a linear map of $N_x$ and we identify this space with $\mathbb{R}^{n-1}$. The natural LDS associated to the dynamics of the vector field is obtained by taking $A^{\star}_x=\Pi \circ DX_x$. Given an LDS $\mathcal{A}=(\Sigma, X^t, N_{\Sigma}, A^{\star})$ the *linear variational equation* associated to it is $$\label{lve} \dot{u}(t,x)=A^{\star}(X^t(x)) \cdot u(t,x).$$ The solution of (\[lve\]) with initial condition $u(0,x)=Id$ is, for each $t$ and $x$, a linear map $\Phi_{A^{\star}}^t(x) \colon N_x \rightarrow N_{X^t(x)}$. The map $A^{\star}$ is called the *infinitesimal generator* of $\Phi_ {A^{\star}}$; it is easy to see that $\Phi_{A^{\star}}^t(x)=P_X^t(x)$ when the infinitesimal generator is $\Pi \circ DX$. The LDS $\mathcal{A}=(\Sigma, X^t, N_{\Sigma}, A^{\star})$ is *bounded* if there exists $K>0$ such that $\|A^{\star}_x\| \leq K$, for all $x \in \Sigma$. The LDS $\mathcal{A}$ is said to be a *large period system* if the number of orbits of $\Sigma$ with period less or equal to $\tau$ is finite, for any $\tau >0$. We say that the LDS $\mathcal{A}$ is *conservative* if $$| \det\Phi_{A^{\star}}^t(x) | = \frac{\|X(x)\|}{\|X(X^t(x))\|},\,\, \forall x \in \Sigma.$$ We observe that from Liouville’s formula it follows that $$\label{LF} \det\Phi_{A^{\star}}^t(x) = \exp\left(\int_0^t tr(A^{\star}(X^s(x)))ds\right).$$ A LDS $\mathcal{B}=(\Sigma, X^t, N_{\Sigma}, B^{\star})$ is a *conservative perturbation* of a bounded $\mathcal{A}$ if, for every $\epsilon >0$, $\|A^{\star}_x - B^{\star}_x\| < \epsilon$, up to points $x$ belonging to a finite number of orbits, and $\mathcal{B}$ is conservative. In view of (\[LF\]) it follows that $\mathcal{B}$ is conservative if and only if $tr(B^{\star})=tr(A^{\star})$. A direct application of the Gronwall inequality gives that $$\|\Phi_{A^{\star}}^t(x)-\Phi_{B^{\star}}^t(x)\| \leq \exp(K|t|)\|A^{\star}_x - B^{\star}_x\|.$$ In particular $\Phi_{B^{\star}}^1$ is a perturbation of $\Phi_{A^{\star}}^1$ in the sense introduced in [@BGV] for the discrete case. A bounded LDS $\mathcal{A}$ is *strictly without dominated decomposition* if the only invariant subsets of $\Sigma$ that admit a dominated splitting for $\Phi_{A^{\star}}^t$ are finite sets. Let us now present a key result about linear differential systems which is the conservative version of Theorem 2.2 of [@BGV]. \[2.2\] Let $\mathcal{A}$ be a conservative, large period and bounded LDS. If $\mathcal{A}$ is strictly without dominated decomposition then there exist a conservative perturbation $\mathcal{B}$ of $\mathcal{A}$ and an infinite set $\Sigma^{\prime} \subset \Sigma$ which is $X^t$-invariant such that for every $x \in \Sigma^{\prime}$ the linear map $\Phi_{B^{\star}}^{\pi(x)}(x)$ as all eigenvalues real and with the same modulus (thus equal to $1$ or to $-1$). The perturbations used in the proof of Theorem 2.2 of [@BGV] are rotations and directional homotheties (diagonal linear maps for a fixed basis). They are made in the linear cocycle setting and (discrete/continuous-time) Franks’ Lemma allows to realize them as perturbations of a fixed diffeomorphism or vector field. Once we have a dictionary to pass from linear cocycles (discrete case) to conservative linear differential systems (*conservative* continuous-time case) and we obtained the Main Perturbation Lemma (Lemma \[mpl\]) which allows to realize these kind of conservative perturbations of linear differential systems as conservative perturbations of vector fields, the proof given by Bonatti, Gourmelon and Vivier can be carried on to our setting without additional obstructions. Therefore, to illustrate how this can be done, we show how to perturb along a periodic orbit of a conservative vector field in order to get real eigenvalues for the linear Poincaré map in the period. This is obtained by first making a conservative perturbation of the LDS associated to the orbit and then, using the Main Perturbation Lemma, realize it as a conservative perturbation of the vector field. Let $X \in \mathfrak{X}_\mu^4(M)$ and fix small $\epsilon_0 >0$. There exists $\pi_0$ such that for any periodic orbit $x$ with period $\pi(x)>\pi_0$ there is $Y \in \mathfrak{X}_\mu^1(M)$ satisfying - $Y$ is $\epsilon_0$-$C^1$-close to $X$; - $Y^t(x)=X^t(x)$, $\forall t \in \mathbb{R}$; - all the eigenvalues of $P_Y^{\pi(x)}$ are real, and - $Y$ is equal to $X$ outside a small neighbourhood of the orbit of $x$. Let us fix a small $\delta >0$. Let $R_{\theta}$ denote the rotation of angle $\theta$ in the plane. Lemma 6.6 of [@BC] assures that there exists $N=N(\epsilon) \in \mathbb{N}$ satisfying the following: for any $k>N$ and for any $C_1$, $C_2$,...,$C_k \in SL(2,\mathbb{R})$ there are rotations $R_{\theta_1}$, $R_{\theta_2}$,...,$R_{\theta_k}$, with $|\theta_j|< \delta$ for all $j \in \{1,2,...,k\}$, such that the linear map $$C_k \circ R_{\theta_k} \circ C_{k-1} \circ R_{\theta_{k-1}} \circ ... \circ C_1 \circ R_{\theta_1}$$ has real eigenvalues. Let us fix a periodic orbit $\gamma$ and $x \in \gamma$ with $\pi(x) \geq N$. We assume that $P_X^{\pi(x)}(x)$ has a complex eigenvalue associated to a two dimensional invariant subspace $V_x \subset N_x$. Assuming that $\pi(x)=k \in \mathbb{N}$, we consider the linear maps $C_j \colon V_{X^{j-1}(x)} \rightarrow V_{X^{j}(x)}$ defined by $$C_j=\frac {1}{\det(P_X^1(X^{j-1}(x))|_{V_{j-1}})}P_X^1(X^{j-1}(x))|_{V_{j-1}},$$ where $V_j=P_X^j(V_x)$ and $j \in \{1,2,...k\}$. If $\pi(x) \notin \mathbb{N}$ we take $k=[\pi(x)]$, consider $C_1,...,C_{k-1}$ as before and define $C_k \colon V_{X^{k-1}(x)} \rightarrow V_{X^{\pi(x)}(x)}=V_{x}$ by $$C_k=\frac {1}{\det(P_X^{1+\pi(x)-k}(X^{k-1}(x))|_{V_{k-1}})}P_X^{1+\pi(x)-k}(X^{k-1}(x))|_{V_{k-1}}.$$ In what follows, without loss of generality, we assume that $\pi(x)=k \in \mathbb{N}$. We observe that each $C_j$ can be identified with a linear map of $SL(2, \mathbb{R})$ and that $P_X^{\pi(x)}(x)=C_k \circ C_{k-1}\circ...\circ C_1$. Therefore, Lemma 6.6 of [@BC] gives a family of rotations $R_{\theta_1}$, $R_{\theta_2}$,...,$R_{\theta_k}$ with the properties described above. Now we want to apply Lemma \[mpl\] to each arc $\Gamma(X^{j-1}(x),1)$ and to the maps $C_j$ and $R_{\theta_j}$. For that we consider $V=V_{j-1}$ and choose an appropriate $ V_{j-1}^{\prime}$ using the Jordan canonical form so that the perturbation we will construct do not change the other eigenvalues of $P_X^{\pi(x)}$. Then, for each $j \in \{1,2,...,k\}$, we define the one-parameter linear family $\{R_{j,t}\}_{t\in \mathbb{R}}$ associated to $\Gamma(X^{j-1}(x), 1)$ and $V_j$ by - $R_{j,t} \colon N_{X^{j-1}(x)} \rightarrow N_{X^{j-1}(x)}$ is a linear map, for all $t\in \mathbb{R}$, - $R_{j,t}=Id$, for all $t\leq 0$, and $R_{j,t}=R_{j,1}$, for all $t\geq 1$, - $R_{j,t}|_{V_j^{\prime}}\equiv Id$, $\forall t \in [0,1]$, and - $R_{j,t}|_{V_j}=R_{\alpha_j(t)\theta_j}$, where $\alpha_j$ is a $C^{\infty}$ bump function with $\alpha_j(t)=0$ for $t \leq 0$, $\alpha_j(t)=1$ for $t \geq 1$, and $0 \leq \alpha_j^{\prime}(t) \leq 2$, for all $ t \in \mathbb{R}$. A direct computation gives that $\|R_{j,t}^\prime \circ R_{j,t}^{-1}\|=\alpha^{\prime}(t) \theta_j \leq 2 \epsilon$. Therefore we fix $\delta \leq \frac{1}{2}\xi_0(\epsilon_0 /n, X)$, where $n$ is the dimension of $M$ and $\xi_0(\epsilon_0 /n, X)$ is given by Lemma \[mpl\]; thus, applying this lemma we get divergence-free vector fields $Y_1$,...,$Y_{k}$, each one $\epsilon_0/n$-$C^1$-close to $X$ and such that $P_{Y_j}^1(X^{j-1}(x))= P_X^1(X^{j-1}(x)) \circ R_{\theta_j}$, for $j \in \{1,2,...,k\}$. It follows from this construction that these vector fields *glue* together defining a $C^1$ vector field $Y \in \mathfrak{X}_\mu^1(M)$, $\epsilon_0/n$-$C^1$-close to $X$, and such that $$\begin{aligned} P_Y^{\pi(x)}(x)|_{V_x}&=&(P_X^1(X^{k-1}(x)) \circ R_{\theta_k}) \circ ... \circ (P_X^1(x) \circ R_{\theta_1})\\ &=&(C_{k} \circ R_{\theta_k}) \circ (C_{k-1} \circ R_{\theta_{k-1}}) \circ ... \circ (C_1 \circ R_{\theta_1}), \end{aligned}$$ therefore this linear map has real eigenvalues. Finally, we apply these arguments at most $[\frac{n}{2}]$ times (corresponding to the maximal number of two-dimensional eigenspaces associated to complex eigenvalues) to get the vector field $Y$ satisfying the conditions of the lemma. In the previous lemma we can assure that the eigenvalues of $P_Y^{\pi(x)}$ are all real and with different modulus. In fact this can be achieved by adding small directional homotheties in the two-dimensional vector spaces $V_x$. [Proof of Proposition \[VF\]]{} We first note that it is not difficult to see that once we obtain the conclusions of the proposition for a robustly transitive vector field $X$ then they also hold for $Y$ in a small neighbourhood $\mathcal{U}$ of $X$ with the same $\ell$ and $\varrho$. So let us fix a robustly transitive $X \in \mathfrak{X}_\mu^4(M)$ and a positive number $\epsilon$ such that if $Y \in \mathfrak{X}_\mu^1(M)$ and is $\epsilon$-$C^1$-close to $X$ then it is robustly transitive. We assume that the proposition does not hold for $X$. Therefore for each $\ell \in \mathbb{N}$ there exists a periodic point $x_{\ell}$, with period $\pi_{\ell}=\pi(x_{\ell})\geq \ell $ and such that the orbit $\Gamma(x_{\ell}, \pi_{\ell})$ does not admit an $\ell$-dominated splitting. Define $$\Sigma=\bigcup_{\ell \in \mathbb{N}} \Gamma(x_{\ell}, \pi_{\ell}),$$ and consider the linear differential system $\mathcal{A}=(\Sigma, X^t, N_{\Sigma}, \Pi \circ DX)$, that is $A^{\star}=\Pi \circ DX$ . As $M$ is compact and $X$ is $C^1$ $\mathcal{A}$ is bounded and, by construction, it is conservative and a large period system. Now we show that $\mathcal{A}$ is strictly without dominated decomposition. In fact let us assume that there exists an $X^t$-invariant and not finite set $\Sigma_0 \subset \Sigma$ such that $P_X^t$ admits an $L_0$ dominated splitting over $\Sigma_0$, say $N_{\Sigma_0}=E \oplus F$. It follows that there is $L_1>L_0$ such that $E \oplus F$ is an $L$-dominated splitting for the linear Poincaré map, for any $L \geq L_1$. As $\Sigma_0$ is not finite there exists $\ell_0>L_1$ such that $\Sigma_1=\Gamma(x_{\ell_0}, \pi_{\ell_0}) \subset \Sigma_0$. The set $\Sigma_1$ is an $X^t$-invariant subset of $\Sigma_0$ therefore, by choice of $L_1$, it admits an $L$- dominated splitting for any $L \geq L_1$; by other side as $\pi_{\ell_0} \geq \ell_0 >L_1$ it follows that $\Sigma_1$ does not admit an $\ell_0$-dominated splitting, which is a contradiction. Now we can apply Theorem \[2.2\] in order to get a conservative perturbation $\mathcal{B}$ of $\mathcal{A}$ and an infinite set $\Sigma^{\prime} \subset \Sigma$ which is $X^t$-invariant such that for every $x \in \Sigma^{\prime}$ the linear map $\Phi_{B^{\star}}^{\pi(x)}(x)$ as all eigenvalues real and with the same modulus (thus equal to $1$ or to $-1$). As $\mathcal{B}$ is a perturbation of $\mathcal{A}$, for any small $\eta_0>0$ there exits $\overline{x} \in \Sigma^{\prime}$ such that $\|B^{\star}_{X_t(\overline{x})}-A^{\star}_{X_t(\overline{x})}\|<\eta_0$, for all $t \in \mathbb{R}$. We observe that the period of $\overline{x}$, $\pi(\overline{x})$ tends to infinity as $\eta_0$ goes to zero. Now we construct a new vector field $Y \in \mathfrak{X}_\mu^1(M)$ such $Y$ is $\epsilon$-$C^1$-close to $X$, $Y^t(\overline{x})=X^t(\overline{x})$, for all $t \in \mathbb{R}$ and $P_Y^{\pi(\overline{x})}(\overline{x})=\Phi_{B^{\star}}^{\pi(\overline{x})}(\overline{x})$, in particular the linear Poincaré map $P_Y^{\pi(\overline{x})}$ has only eigenvalues equal to $1$ or $-1$. Once we get this vector field we apply Lemma \[linear2\] to get a new vector field $Z \in \mathfrak{X}_\mu^1(M)$, arbitrarily $C^1$-close to $Y$ and having a $Z^t$-invariant tubular neighbourhood of the orbit of $\overline{x}$, which contradicts the fact that $Y$ is robustly transitive thus ending the proof of Proposition \[VF\]. Therefore it remains to explain how we obtain the mentioned vector field $Y$. Consider the arcs of trajectory $$\Gamma_j=\bigcup_{t \in[0,1]}X^t(X^j(\overline{x})), \text{ for } j \in \{0,..., [\pi(x)]-2\},$$ and $$\Gamma_{[\pi(x)]-1}=\bigcup_{t \in[0,1+\pi(\overline{x})-[\pi(\overline{x}]]}X^t(X^j(\overline{x})).$$ For $j \in \{0,..., [\pi(x)]-1\}$ we write $\Phi_{B^{\star}}^1(X^j(\overline{x}))=P_X^1(X^j(\overline{x})) \circ A^j$, where $A^j={(P_X^1(X^j(\overline{x})))}^{-1} \circ \Phi_{B^{\star}}^1(X^j(\overline{x}))$, and consider a continuous one-parameter family $\tilde{A}_t^j$ defined by - $\tilde{A}_t^j=Id$ for $t \leq 0$; - $\tilde{A}_t^j=A^j$ for $t \geq 1$ (for $t \geq 1+\pi(\overline{x})-[\pi(\overline{x})]$, if $j=[\pi(\overline{x})]-1$), and - $\tilde{A}_t^j={(P_X^t(X^j(\overline{x})))}^{-1} \circ \Phi_{B^{\star}}^t(X^j(\overline{x}))$, for $0 \leq t \leq 1$ (or for $0 \leq t \leq 1+\pi(\overline{x})-[\pi(\overline{x})]$, if $j=[\pi(\overline{x})]-1$). We fix small $\tilde{\delta}>0$ and consider $A_t^j$ a one-parameter linear family (see Section \[pl\]), arbitrarily close to $\tilde{A}_t^j$, such that $A_t^j=Id$ for $t \leq \tilde{\delta}$ and $A_t^j=\tilde{A}_t^j$ for $t \geq 1$ (or $t \geq 1+\pi(\overline{x})-[\pi(\overline{x})]$, if $j=[\pi(\overline{x})]-1$). Now, we observe that $\| (A_t^j)^{\prime} (A_t^j)^{-1}\|$ is of order $$\|((P_X^{t}(X^j(\overline{x})))^{-1} \circ \Phi_{B^{\star}}^t(X^j(\overline{x})))^{\prime} \circ ((P_X^{t}(X^j(\overline{x})))^{-1} \circ \Phi_{B^{\star}}^t(X^j(\overline{x})))^{-1}\|,$$ therefore of order $$\mathcal{O}(\overline{x}) =\max_{y \in \Gamma(\overline{x}, \pi(\overline{x}))}\|B^{\star}_y - A^{\star}_y \|.$$ As we mention before $\overline{x} \in \Sigma^{\prime}$ can be chosen such that $\mathcal{O}(\overline{x})$ is arbitrarily small. Therefore, fixing $0 < \delta < \tilde{\delta}$ and small $r>0$, to each arc $\Gamma_j$ we apply Lemma \[mpl\] to get a new vector field $Y_j \in \mathfrak{X}_\mu^1(M)$, $\epsilon$-$C^1$-close to $Y$ and such that 1. $Y_j^t(\overline{x})=X_j^t(\overline{x})$, $\forall t \in \mathbb{R}$; 2. $P_{Y_j}^1(X^j(\overline{x}))=\Phi_{B^{\star}}^1(X^j(\overline{x}))$, and 3. $Y_j|_{\mathcal{T}_j^c}=X|_{\mathcal{T}_j^c}$, where $\mathcal{T}_j=\mathcal{T}(X^j(\overline{x}), \tau, r, \delta)$ and $\tau = 1$ (or $\tau=1+\pi(\overline{x})-[\pi(\overline{x})]$ if $j=[\pi(\overline{x})]-1$). Finally, as $ \delta < \tilde{\delta}$, by construction it follows that $$Y_j|_{\mathcal{T}_j \cap \mathcal{T}_{j+1}}=Y_{j+1}|_{\mathcal{T}_j \cap \mathcal{T}_{j+1}},\,\, \forall j \in \{0,1,..., [\pi(\overline{x})]-1\},$$ which, together with item $3.$ above, implies that these vector fields can be *glued* to obtain a $C^1$ vector field $Y \in \mathfrak{X}_\mu^1(M)$, $\epsilon$-$C^1$-close to $X$ and such that $Y|_{\mathcal{T}_j}=Y_j|_{\mathcal{T}_j}$, for all $j \in \{0,1,..., [\pi(\overline{x})]-1\}$. Thus $Y^t(\overline{x})=X^t(\overline{x})$, for all $t \in \mathbb{R}$ and $P_Y^{\pi(\overline{x})}(\overline{x})=\Phi_{B^{\star}}^{\pi(\overline{x})}(\overline{x})$ as required. This ends the proof of Proposition \[VF\]. [ABC]{} A. Arbieto and C. Matheus, *A pasting lemma and some applications for conservative systems*, Preprint arXiv:math/0601433v1 2006, to appear Ergod. Th. & Dynam. Sys.. A. Baraviera and C. Bonatti, *Removing zero Lyapunov exponents*, Ergod. Th. & Dynam. Sys., 23 (2003), 1655–1670. M. Bessa, *Dynamics of generic multidimensional linear differential systems* Preprint CMUP 2007. M. Bessa and J. Rocha, *Removing zero Lyapunov exponents in volume-preserving flows*, Nonlinearity 20 (2007), 1007–1016. J. Bochi and M. Viana, *The Lyapunov exponents of generic volume-preserving and symplectic maps*. Ann. of Math. (2) 161 (2005), no. 3, 1423–1485. C. Bonatti and S. Crovisier, *Récurrence et généricité*, Invent. Math. 158, 1 (2004), 33–104. C. Bonatti, L.J. Díaz and E. Pujals, *A $C^1$-generic dichotomy for diffeomorphisms: Weak forms of hyperbolicity or infinitely many sinks or sources*, Ann. Math, 158 (2005), 355–418. C. Bonatti, L. J. Díaz and M. Viana, *Dynamics beyond uniform hyperbolicity*, EMS 102, Springer 2005. C. Bonatti, N. Gourmelon and T. Vivier, *Perturbations of the derivative along periodic orbits.* Ergod. Th. & Dynam. Sys. 26, 5 (2006), 1307–1337. L.J. Díaz, E. Pujals and R. Ures, *Partial hyperbolicity and robust transitivity*. Acta Math., 183, 1 (1999), 1–43. B. Dacorogna and J. Moser, *On a partial differential equation involving the jacobian determinant*. Ann. Inst. Henri Poincaré, Anal. Non Linéaire, 7 (1990), 1–26. C. Doering, *Persistently transitive vector fields on three-dimensional manifolds*. Proceedings on Dynamical Systems and Bifurcation Theory, Vol. 160 (1987), 59–89, Pitman. J. Franks, *Necessary conditions for the stability of diffeomorphisms*. Trans. Amer. Math. Soc., 158 (1971), 301-308. V. Horita and A. Tahzibi, *Partial hyperbolicity for symplectic diffeomorphisms*. Ann. Inst. Henri Poincaré, Anal. Non Linéaire, 23, 5 (2006), 641–661. R. Mañé, *An ergodic closing lemma*. Ann. Math, 116 (1982), 503–540. R. Mañé, *Oseledec’s theorem from the generic viewpoint*. Proc. Int. Congress of Mathematicians, Vol. 1, 2 (Warsaw, 1983), 1269–1276, PWN, Warsaw, 1984. J. Moser, *On the volume elements on a manifold*. Trans. Amer. Math. Soc., 120 (1965), 286–294. C. Pugh and C. Robinson, *The $C^1$ Closing lemma, including hamiltonians*. Ergod. Th. & Dynam. Sys., 3, 2 (1983), 261–313. T. Vivier, *Projective hyperbolicity and fixed points*. Ergod. Th. & Dynam. Sys., 26 (2006), 923–936. T. Vivier, *Robustly transitive 3-dimensional regular energy surface are Anosov*. Preprint, Dijon (2005). C. Zuppa, *Regularisation $C^{\infty}$ des champs vectoriels qui préservent l’elément de volume*, Bol. Soc. Bras. Mat., 10, 2 (1979), 51–56.    ([email protected])\ CMUP, Rua do Campo Alegre, 687\ 4169-007 Porto\ Portugal\    ([email protected])\ Dep. Matemática Pura-FCUP, Rua do Campo Alegre, 687\ 4169-007 Porto\ Portugal\ [^1]: Supported by FCT-FSE, SFRH/BPD/20890/2004. [^2]: Partially supported by FCT-POCTI/MAT/61237/2004.

ArXiv

--- abstract: 'The acoustic radiation force exerted on a small sphere located at the focus of an ultrasonic beam is measured in a soft gel. It is proved to evolve quadratically with the local amplitude of the acoustic field. Strong oscillations of the local pressure are observed and attributed to an acoustic Fabry-P[é]{}rot effect between the ultrasonic emitter and the sphere. Taking this effect into account with a simple model, a quantitative link between the radiation force and the acoustic pressure is proposed and compared to theoretical predictions in the absence of dissipation. The discrepancy between experiment and theory suggests that dissipative effects should be taken into account for fully modeling the observations.' author: - Pierre Lidon - Louis Villa - Nicolas Taberlet - 'S[é]{}bastien Manneville' date: November 2016 title: Measurement of the acoustic radiation force on a sphere embedded in a soft solid --- *Introduction -* Acoustic manipulation has recently emerged as a versatile tool to displace, sort and organize small objects such as droplets, bubbles or cells in microfluidic devices [@petersson_2004; @wiklund_2006; @evander_2007; @rabaud_2011]. Indeed, when submitted to an acoustic wave, the interface between two materials displaying an acoustic contrast is exposed to a force $F_\text{rad}$, known as the acoustic radiation force, that arises from non-linear acoustic effects[@chen_1996]. According to the sign of the acoustic contrast, it is possible to push or pull objects in a fluid with a traveling wave[@hertz_1939; @schmid_2014; @li_2015a] or to induce the migration of particles towards the nodes or the antinodes of a standing wave [@goldman_1949; @tran_2012; @guo_2016]. Thus $F_\text{rad}$ allows one to move precisely small objects[@laurell], which has opened a path for microfluidic design of structured materials via bottom-up approaches[@chen_2013; @gesellchen_2014; @ma_2015; @caleap_2014; @caleap_2016]. In order to preserve the obtained mesoscopic structure after insonification, it has been proposed to work in photocurable resins[@raeymaekers_2011] or in yield-stress fluids[@xu_2011]. However, while the acoustic radiation force is well characterized in simple inviscid fluids, the effect of viscosity on $F_\text{rad}$ is still under debate[@doinikov_1994a; @settnes_2012; @karlsen_2015; @sepehrirahnama_2015; @karlsen_2016; @sepehrirahnama_2016] and only a few studies have been devoted to the general case of viscoelastic media[@chen_2002; @ilinskii_2005; @aglyamov_2007; @urban_2011; @suomi_2015]. It is thus of growing interest to characterize the acoustic radiation force in complex, soft materials. In this Letter, we show experimentally that the acoustic radiation force exerted on a rigid sphere embedded in a viscoelastic solid and submitted to a focused ultrasonic field is related to the local pressure $P_\text{loc}$ by $F_\text{rad} = \beta P_\text{loc}^2$ and we provide an estimate of the parameter $\beta$. We further show that there exists a partially standing wave between the sphere and the acoustic transducer, so that $F_\text{rad}$ is highly sensitive to the sphere position. This cavity effect is well captured by a simple model. Finally, the measured prefactor $\beta$ is found to be larger by about 40% than the theoretical value in the absence of dissipation, which hints at the necessity to account for viscosity in order to correctly predict $F_\text{rad}$ in soft materials. *Principle of the experiment -* In order to quantify $F_\text{rad}$, the motion of an intruder within a soft gel under the effect of focused ultrasound is recorded, as sketched in Fig. \[fig:transducteur\_setup\]. A similar experiment has been used to assess the mechanical properties of tissue-like gelatin gels by analyzing the transient displacement of the intruder induced by short pulses of duration up to a few milliseconds[@aglyamov_2007]. Here, we propose a different approach: assuming that the rheology of the gel is known, we measure the acoustic radiation force through the steady-state displacement $\delta r_\text{max}$ of a spherical intruder exposed to much longer insonification. Indeed, assuming a no-slip boundary condition at the sphere surface, $F_\text{rad}$ is related to $\delta r_\text{max}$ by[@urban_2011; @ilinskii_2005] $$F_\text{rad} = 6\pi G_0 a \delta r_\text{max}\, , \label{eq:link_displacement_force}$$ where $G_0$ is the elastic modulus of the material and $a$ the radius of the spherical intruder. ![Experimental setup. The cell and the transducer are immersed in a water bath. “A” stands for power amplifier, “G” for function generator and “C” for computer. \[fig:transducteur\_setup\]](fig1a_transducteur_setup.pdf) *Setup -* This study focuses on a gel made of carbopol ETD 2050 dispersed in water at a concentration of $1\%$ wt and prepared following a protocol already described elsewhere [@lidon_2016b]. Such a gel has a yield stress $\sigma_\text{y}=\SI{19.5}{\pascal}$ below which it behaves as a soft viscoelastic solid of elastic modulus $G_0 \simeq \SI{55}{\pascal}$ and notably smaller viscous modulus ($\simeq\SI{4}{\pascal}$) as measured by a stress-controlled rheometer (Anton Paar MCR-301). The gel is first centrifugated to remove air bubbles. It is then progressively poured into a cylindrical cell of inner diameter $\SI{2}{\centi\meter}$ and length $\SI{5}{\centi\meter}$. A single polystyrene sphere of radius $a=\SI{163 \pm 3}{\micro\meter}$ is introduced into the gel. As the buoyancy stress due to the density mismatch between the sphere ($\rho_\text{p} \simeq \SI{1.05}{\kilo\gram\per\meter\cubed}$) and the gel ($\rho_\text{g} \simeq \SI{1.03}{\kilo\gram\per\meter\cubed}$) is smaller than $\sigma_\text{y}$, the sphere does not sediment. Finally, the cell is closed with a thin plastic wrap. During all these steps, great care is given not to trap any air bubbles within the gel, which would result in important distortion of the acoustic waves. The cell is moved using a three-axis translation stage controlled by a computer. The setup is illuminated by a LED panel providing a uniform luminous background. Images of the sphere are recorded at a rate of $300 \, \text{fps}$ with a CCD camera (Baumer HXC20) mounted on a macroscope (Nikon SMZ745T). Images are analyzed with a standard particle detection algorithm allowing us to measure the sphere trajectory with a resolution of about $\SI{1}{\micro\meter}$. Finally, ultrasonic waves are produced by a focused piezoelectric transducer (Imasonic, diameter $\SI{38}{\milli\meter}$, center frequency $f=\SI{2.25}{\mega\hertz}$). The cell and the transducer are placed within a water bath at temperature $T \simeq \SI{20}{\celsius}$ in order to ensure ultrasound propagation. We checked that the gel, the plastic membrane and the cell do not distort nor attenuate the ultrasonic field. The transducer is driven by a power amplifier (Kalmus 150C) controlled by a function generator (Agilent 33522A). It emits bursts of duration $\SI{0.5}{\second}$, hereafter referred to as “pulses,” and it is left to rest for about one minute to allow for full relaxation of the gel. The sphere can be located precisely at the focus of the transducer by maximizing its displacement under a fixed acoustic intensity. ![(a) Pressure field in water and in the absence of a sphere, around the focal spot of the piezoelectric transducer in the transverse plane (left) and in the propagation plane (right). The color scale corresponds to relative acoustic intensity in decibels, i.e. to $20\log{(P/P_0)}$. (b) Pressure profiles $P$ at focus along a perpendicular axis $x$ (left) and along the direction of propagation $z$ (right), normalized by the maximum pressure at focus $P_0$. Red curves are theoretical predictions by Eqs. (\[eq:pressure\_focused\_transducer\_x\]) and (\[eq:pressure\_focused\_transducer\_z\]) with no adjustable parameter. \[fig:transducteur\_champ\]](fig1b_transducteur_champ.pdf) Figure \[fig:transducteur\_champ\](a) shows the pressure field characterization $P(x,y,z)$ obtained by scanning the vicinity of the focal spot with a needle hydrophone (Precision Acoustics 1717SN with active element diameter $\SI{0.2}{\milli\meter}$) in water and in the absence of the experimental cell. The acoustic field is axisymmetric about the direction of propagation and pressure profiles are displayed in Fig. \[fig:transducteur\_champ\](b). They are well fitted by the following theoretical expressions without any adjustable parameter[@kino]: $$\begin{aligned} P(x,0,0) = &P_0 \frac{\lambda \ell}{\pi a x} \mathrm{J}_1 \left(\frac{2\pi a x}{\lambda \ell} \right) \label{eq:pressure_focused_transducer_x} \\ P(0,0,z) = &P_0 \frac{\ell}{z+\ell} \mathrm{sinc} \left(\frac{a^2}{2\lambda \ell} \frac{z}{z+\ell} \right) \label{eq:pressure_focused_transducer_z}\end{aligned}$$ where $J_1$ is the Bessel function of the first kind, $a=\SI{19}{\milli\meter}$ is the radius of the transducer, $\ell=\SI{38}{\milli\meter}$ its focal length and $\lambda = c_\text{w}/f=\SI{670}{\micro\meter}$ is the acoustic wavelength in water for the considered frequency and sound speed $c_\text{w}=\SI{1480}{\meter\per\second}$ at room temperature. The acoustic field is focused in a small spot of diameter $\SI{1.0}{\milli\meter}$ and focal depth at $\SI{-3}{\deci\bel}$ $\ell_\text{z}\simeq\SI{5}{\milli\meter}$. Consequently, it can be assumed to be homogeneous at the scale of the sphere so that $F_\text{rad}$ will be estimated locally at the sphere position. *Scaling of $F_\text{rad}$ with acoustic intensity -* Figure \[fig:creep\](a) displays the sphere displacement $\delta r$ during successive pulses of increasing amplitude $P_0=\SI{0.4}{\mega\pascal}$ to $\SI{2.0}{\mega\pascal}$. The initial position of the sphere corresponds to the focus of the transducer and the maximum value of $P_0$ is chosen such that the acoustic radiation pressure remains below the yield stress in order to avoid any irrecoverable displacement of the sphere. ![(a) Displacement $\delta r$ of the sphere as a function of time $t$ during pulses of duration $\SI{0.5}{\second}$ and amplitude ranging from $P_0=\SI{0.4}{\mega\pascal}$ to $P_0=\SI{2.0}{\mega\pascal}$. (b) Radiation force $F_\text{rad}$ obtained from Eq.  as a function of $P_0^2$. The red line is a linear fit, $F_\text{rad} = \alpha P_0^2$, with $\alpha=\SI{3.7(3)e-18}{\newton\per\pascal\squared}$. \[fig:creep\]](fig2_creep.pdf) For the range of acoustic intensities considered here, the sphere reaches an almost stationary displacement $\delta r_\text{max}$ at the end of the pulse and $\delta r_\text{max}$ is converted into an estimate of $F_\text{rad}$ by using Eq. . As shown in Fig. \[fig:creep\](b), the acoustic radiation force scales as $P_0^2$, with a slope $\alpha=\SI{3.7(3)e-18}{\newton\per\pascal\squared}$. Although such a quadratic scaling is not surprising –since $F_\text{rad}$ results from a non-linear effect[@chen_1996]–, it is nicely confirmed here for a viscoelastic propagation medium. *Oscillations of $F_\text{rad}$ with sphere position -* In order to explore the effect of the sphere position with respect to the transducer on $F_\text{rad}$, the focal area is scanned by moving the cell in horizontal $(xOy)$ and vertical $(xOz)$ planes centered around the focus by steps of size $\SI{100}{\micro\meter}$. At each position, $\delta r_\text{max}$ is recorded for a fixed pressure at focus $P_0=\SI{1.8}{\mega\pascal}$ as displayed in Fig. \[fig:champ\_force\]. It should be noted that the displacement map has the same overall shape and size as $P^2(x,y,z)$. However, striking oscillations are observed in the direction of ultrasound propagation $(Oz)$. ![Maximum displacement $\delta r_\text{max}$ of the sphere during pulses of duration $\SI{0.5}{\second}$ and amplitude $P_0=\SI{1.8}{\mega\pascal}$ for different initial positions across the pressure field, in the transverse plane (left) and in the propagation plane (right). \[fig:champ\_force\]](fig3_champ_force.pdf) ![Acoustic radiation force $F_\text{rad}$ exerted on the sphere as a function of the distance $z$ from focus along the direction of propagation for $P_0=\SI{1.8}{\mega\pascal}$. The red curve corresponds to Eqs.  and with $r=-0.115$, $\rho=-0.8$ and $\lambda=\SI{680}{\micro\meter}$. \[fig:cavite\_exp\]](fig4_cavite_exp.pdf) In order to better characterize these oscillations, Fig. \[fig:cavite\_exp\] shows a scan performed along the line $(Oz)$ with smaller steps of $\SI{2}{\micro\meter}$. The oscillations display a period close to $\lambda / 2$ and are sharper on the upper part. These features are reminiscent of a Fabry-P[é]{}rot or cavity effect. During the pulse, acoustic waves indeed are reflected back and forth on the sphere and on the transducer a large number of times. This results in the formation of an acoustic cavity: the acoustic wave between the sphere and the transducer is the superposition of the incident traveling wave generated by the transducer and of the standing wave due to the Fabry-P[é]{}rot effect. This partially standing wave structure leads to oscillations of the acoustic force when the sphere is displaced with respect to the transducer[@issenmann_2006]. This effect can be accounted for by a simple model sketched in Fig. \[fig:cavite\_modele\](a). We consider the sphere and the transducer as plane reflectors of reflection coefficients in amplitude $r$ and $\rho$ respectively. For a plane wave, computing the local pressure field $P_\text{loc}$ in the cavity from $r$ and $\rho$ is straightforward. Assuming that the case of a focused beam is simply obtained by modulating this field by the pressure profile $P(0,0,z)$ given by Eq. , we get $$\begin{aligned} F_\text{rad}(z) &=& \alpha \Phi(z) P(0,0,z)^2\label{eq:model_cavity1}\\ \hbox{\rm{with~}}\Phi(z)&=&\frac{A+B \cos{[2k(z+\ell)]}}{\left(C + D \cos{[2k(z+\ell)]} \right)^2}, \label{eq:model_cavity2}\end{aligned}$$ where $k=2\pi / \lambda$ and $A$, $B$, $C$ and $D$ depend only on $r$ and $\rho$[@expression_ABCD]. The speed of sound in our carbopol gel was measured to be $c_\text{g}=\SI{1530}{\meter\per\second}$, corresponding to $\lambda=c_\text{g}/f=\SI{680}{\micro\meter}$. We also estimated the reflection coefficient $\rho\simeq -0.8$ by analyzing the decay of successive echoes of short bursts reflected on the transducer and on a flat, perfectly reflecting water–air interface, as shown in Fig. \[fig:cavite\_modele\](b,c). Finally, with the value of $\alpha$ inferred from the previous set of experiments, we obtain a good agreement between the model and experimental data by taking $r=-0.115$ as the sole free parameter (see red curve in Fig. \[fig:cavite\_exp\]). This value of $r$ should only be interpreted as an effective reflection coefficient for the gel–sphere interface since the sphere diameter is comparable to the acoustic wavelength and a diffraction-based model would be required to fully grasp the details of our focused beam impinging on a sphere smaller than $\lambda$. Moreover, we observe that the oscillation period of the force increases past the focal spot, for $z>0$. This effect, which is not captured by the model, is too significant to be caused by a global temperature drift during the scan along $(Oz)$. It could result from a shift in the spectral content to lower frequencies due to nonlinear propagation and/or attenuation, although further work should confirm such an explanation. ![(a) Schematics of the cavity effect: the acoustic wave is successively reflected by the sphere and by the transducer, leading to various pressure contributions $p_n^{\rm i}$ and $p_n^{\rm r}$ that add up and form a partially standing wave in the cavity. (b) Voltage output of the transducer after successive reflections between the transducer and a water–air interface. The red curve corresponds to the electrical input of duration $\SI{10}{\micro\second}$. (c) Amplitude of the successive echoes as a function of the number of reflections $n$. The red line shows the linear fit $\log{(V_n/V_1)}= (n-1)\log{|\rho|}$ with $\rho=-0.80$. \[fig:cavite\_modele\]](fig5_cavite_modele.pdf) *Discussion -* Due to the aforementioned cavity effect, the local pressure field $P_\text{loc}$ varies significantly on a typical scale $\lambda/2 \simeq \SI{350}{\micro\meter}$ so that, for sphere displacements larger than about $\SI{100}{\micro\meter}$, $F_\text{rad}$ cannot be considered as a constant. Therefore, accurate measurements of the acoustic radiation force require both a good control of the sphere position and moderate displacements, as in Fig. \[fig:creep\](a) for which the sphere was carefully located at focus and $\delta r$ did not exceed $\SI{70}{\micro\meter}$. Moreover, $P_\text{loc}$ differs from the bare pressure field generated by the transducer in the absence of the sphere through a modulation by the oscillating function $\Phi$. Thus, in order to recover the “true” prefactor $\beta$ linking $F_\text{rad}$ and $P_\text{loc}^2$, the coefficient $\alpha$ measured in Fig. \[fig:creep\](b) must be corrected for the effect of the cavity: at focus, one has $P_\text{loc}^2=\Phi(z=0) P_0^2$ so that $\beta = \alpha / \Phi(z=0)$, which yields $\beta = \SI{7.5(7)e-18}{\meter\squared\per\pascal}$. In the literature, the prefactor $\beta$ has been calculated analytically for an elastic sphere located at the center of a focused acoustic field and suspended in an inviscid fluid[@chen_1996]. These calculations should also apply to a suspending medium made of an isotropic elastic solid since the propagation equations remain unchanged as long as there is no dissipation. They predict $\beta_\text{th} = \pi a^2 Y / 4\rho_\text{g} c_\text{g}^2$, where $Y$ involves the ratio $a/\lambda$, the densities ($\rho_\text{p}$, $\rho_\text{g}$) and sound speeds ($c_\text{p}$, $c_\text{g}$) of the particle and gel respectively \[see Eq. (63) in Ref. [@chen_1996] for the full expression of $Y$\]. Using the tabulated value for $c_\text{p}$ in polystyrene, we find $\beta_\text{th} = \SI{5.3e-18}{\meter\squared\per\pascal}$, which is of the same order of magnitude as our measurement. The measured $\beta$ is, however, larger than the theoretical prediction by about 40%, suggesting that the theory underestimates $F_\text{rad}$. This discrepancy could be attributed to dissipative effects in the carbopol gel which have been neglected here, both in the theory and in Eq. . For instance, in Newtonian fluids, viscosity has been shown to increase the amplitude of the radiation force[@doinikov_1994a; @settnes_2012; @karlsen_2015; @sepehrirahnama_2015] which could account for an underestimation of $\beta_\text{th}$. Thus, further theoretical work is needed to incorporate dissipative effects both in acoustical and rheological models. *Conclusion and perspectives -* To summarize, we have measured the acoustic radiation force exerted by a focused acoustic beam on a spherical obstacle embedded in a soft gel and we found that the acoustic force scales quadratically with the local pressure. Due to a cavity effect, the local pressure oscillates significantly when the distance between the sphere and the transducer is varied. This implies that an accurate measurement of $F_\text{rad}$ requires a precise positioning of the sphere. The setup described here is not restricted to the case of yield-stress fluids: provided that buoyant motion is slow enough compared to the experiment duration, this method can be employed for any viscoelastic material. Future work shall thus focus on measuring the acoustic radiation pressure for different intruders, surrounding fluids and pressure fields. To this aim, we will use a microscope to resolve smaller displacements and improve accuracy of the sphere positioning. We will moreover use higher acquisition rates to measure and analyze transient displacements with a higher resolution. This will provide data sets to characterize acoustic radiation pressure in complex materials in situations relevant for microfluidics but also for ultrasonic imaging and therapy. To rationalize the observations, this experimental effort will need to be accompanied by numerical and theoretical modeling of acoustic forces in complex media. *Ackowledgements -* This work was funded by the Institut Universitaire de France and by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement No. 258803. The authors are grateful to Y. Forterre, G. Ovarlez and R. Wunenburger for fruitful discussions. Erratum: “Measurement of the acoustic radiation force on a sphere embedded in a soft solid” \[APL 110, 044103 (2017)\] In this erratum, we make corrections to our recent publication \[APL 110, 044103 (2017)\]. First, the densities of polystyrene and gel are in the wrong unit and should read $\rho_\text{p} \simeq \SI{1.05e3}{\kilo\gram\per\meter\cubed}$ and $\rho_\text{g} \simeq \SI{1.03e3}{\kilo\gram\per\meter\cubed}$. Second, Eq. (4) should be written as follows: $$F_\text{rad}(z) = \alpha \Phi(z) P(0,0,z)^2 .$$ This correction does not affect the rest of the article. Third and most importantly, the theoretical value of $\beta$ is wrong by a factor $2$. The model of Ref. 5 indeed predicts $\beta_\text{th} = \pi a^2 Y / 2 \rho_\text{g} c_\text{g}^2$ and thus $\beta_\text{th} = \SI{10.6e-18}{\meter\squared\per\pascal}$, which is twice the value that is given in the article. This does not change our conclusion that the model fails at describing our measurements but it actually overestimates the experimental value of $\beta$ by about $\SI{40}{\percent}$. This invalidates our short subsequent discussion (top of page 4) that invoked viscous effects to justify that $\beta_\text{th}<\beta$. *Ackowledgements -* The authors are grateful to G.T. Silva for bringing these mistakes to their attention. [33]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****,  ()](\doibase 10.1039/b409139f) [****,  ()](\doibase 10.1209/epl/i2003-10191-2) [****,  ()](\doibase 10.1021/ac061576v) [****,  ()](\doibase 10.1063/1.3579263) @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1039/c4lc00588k) [****,  ()](\doibase 10.1073/pnas.1504484112) @noop [****, ()]{} [****,  ()](\doibase 10.1063/1.4751348) [****,  ()](\doibase 10.1073/pnas.1524813113) @noop [**]{} (, ) [****,  ()](\doibase 10.1021/nn4000034) [****, ()](\doibase 10.1039/c4lc00436a) [****,  ()](\doibase 10.1039/c5nr04272k) [****,  ()](\doibase 10.1073/pnas.1323048111) [ (), 10.1016/j.crhy.2016.02.007](\doibase 10.1016/j.crhy.2016.02.007) [****,  ()](\doibase 10.1063/1.3530670) [****,  ()](\doibase 10.1016/j.biomaterials.2011.07.010) @noop [****,  ()]{} [****, ()](\doibase 10.1103/PhysRevE.85.016327) [****, ()](\doibase 10.1103/PhysRevE.92.043010) [****,  ()](\doibase 10.1103/PhysRevE.92.063309) [****, ()](\doibase 10.1103/PhysRevLett.117.114504) [****,  ()](\doibase 10.1103/PhysRevE.93.023307) [****,  ()](\doibase 10.1121/1.1501276) [****,  ()](\doibase 10.1121/1.1863672) [****,  ()](\doibase 10.1121/1.2774754) [****,  ()](\doibase 10.1121/1.3613939) [****,  ()](\doibase 10.1016/j.ultrasmedbio.2015.07.018) [ (), 10.1007/s00397-016-0961-4](\doibase 10.1007/s00397-016-0961-4) @noop [**]{} () [****,  ()](\doibase 10.1103/PhysRevLett.97.074502)

ArXiv

--- abstract: 'We analyzed two twist-3 distribution amplitudes of pion, i.e. pseudoscalar $\phi^P_{3;\pi}(x)$ and pseudotensor $\phi^\sigma_{3;\pi}(x)$, within the LFQM. Our LFQM descriptions both for twist-3 $\phi^P_{3;\pi}$ and $\phi^\sigma_{3;\pi}$ obtained from the Gaussian radial wave function not only satisfy the fundamental constraint required from the isospin symmetry, but also reproduce exactly the asymptotic forms anticipated from QCD’s conformal limit.' author: - 'Ho-Meoyng Choi' - 'Chueng-Ryong Ji' date: 'Received: date / Accepted: date' title: 'Twist-3 Distribution Amplitudes of Pion in the Light-Front Quark Model [^1] ' --- Introduction {#intro} ============ Hadronic light-cone distribution amplitudes (DAs) have been known to play an essential role in the QCD description of hard exclusive processes via the factorization theorem [@BL80]. Taking the transverse separation to zero, the factorization takes the form of a convolution of a perturbatively calculable hard-scattering amplitude and the process-independent nonperturbative DAs. These nonperturbative DAs motivated many theoretical studies to calculate meson DAs using nonperturbative methods such as the QCD sum rule [@BF; @BBL; @MPS10], the chiral-quark model from the instanton vacuum [@Goeke; @NK06], the Dyson-Schwinger equations (DSE) approach [@DSE13; @DSE2015], and the light-front quark model (LFQM) [@CJ_DA; @CJ_V14; @CJ2015; @FB2015]. Among them, the LFQM appears to be one of the most efficient and effective tools in studying hadron physics as it takes advantage of the distinguished features of the light-front dynamics (LFD) [@BPP]. Specifically, the rational energy-momentum dispersion relation of LFD, namely $p^-=({\bf p}^2_\perp + m^2) / p^+$, yields the sign correlation between the LF energy $p^-(=p^0-p^3)$ and the LF longitudinal momentum $p^+(=p^0 + p^3)$ and leads to the suppression of vacuum fluctuations in LFD. This simplification is a remarkable advantage in LFD and facilitates the partonic interpretation of the amplitudes. Based on the advantage of LFD, the LFQM has been quite successful in describing various static and non-static properties of hadrons such as meson mass spectra [@CJ_99], the decay constants [@Choi07], electromagnetic and weak transition form factors [@CJ_Bc] and generalized parton distribution (GPDs) [@CJ_GPD01]. In Ref. [@CJ_DA], we have analyzed twist-2 DAs of pseudoscalar ($\phi^A_{2;M}(x)$) and vector ($\phi^{||}_{2;V}(x)$) mesons using the LFQM [@CJ_99]. In more recent works [@CJ_V14; @CJ2015], we have extended our LFQM to analyze twist-3 pseudoscalar ($\phi^P_{3;M}(x)$) DAs of pseudoscalar mesons [@CJ2015] and chirality-even twist-3 ($\phi^{\perp}_{3;V}(x)$) DAs of vector mesons [@CJ_V14] and discussed the link between the chiral symmetry of QCD and the numerical results of the LFQM. In particular, we have discussed a wave function dependence of the LF zero-mode [@Zero1; @Zero2; @Zero3] contributions to $\phi^P_{3;M}(x)$ and $\phi^{\perp}_{3;V}(x)$ [ not only for the exactly solvable manifestly covariant Bethe-Salpeter (BS) model but also]{} for the more phenomenologically accessible realistic LFQM [@CJ_DA; @CJ_99] using the standard LF (SLF) approach. We also linked the covariant BS model to the standard LFQM [ providing]{} the correspondence relation [@CJ_V14; @CJ2015] between the two models. [ Effectively, we prescribed a consistent substitution for]{} the LF vertex function in the covariant BS model with the more phenomenologically accessible Gaussian wave function provided by the LFQM analysis of meson mass [@CJ_99]. The remarkable finding is that the zero-mode contribution as well as the instantaneous contribution revealed in the covariant BS model become absent in the LFQM. Without involving the zero-mode and instantaneous contributions, our LFQM result of twist-3 DAs $\phi^P_{3;M}(x)$ and $\phi^{\perp}_{3;V}(x)$ provides the consistency with the chiral symmetry anticipated from QCD’s conformal limit [@BF; @Ball98]. The purpose of this work is to extend our previous work [@CJ2015] to analyze the twist-3 pseudotensor DA $\phi^\sigma_{3;\pi}(x)$ of a pion within the LFQM. The paper is organized as follows: In Sec. \[sec:II\], we compute the twist-3 pseudotensor DA $\phi^\sigma_{3;\pi}(x)$ in an exactly solvable model based on the covariant BS model of (3+1)-dimensional fermion field theory. We then linked the covariant BS model to the standard LFQM following the correspondence relation [@CJ2015] between the covariant BS and LFQM models and present the form of $\phi^\sigma_{3;\pi}(x)$ as well as $\phi^P_{3;\pi}(x)$ in our LFQM. In Sec. \[sec:III\], we present our numerical results. Summary and discussion follow in Sec. \[sec:IV\]. Model Description {#sec:II} ================= The $\phi^P_{2;M}$ and $\phi^\sigma_{3;M}$ are defined in terms of the following matrix elements of gauge invariant nonlocal operators in the light-cone gauge [@BF; @BBL]: \[Deq:3.1\] \^P\_[3;M]{}(x) &=& \^\_[-]{} e\^[-i(P)]{} 0|[|q]{}()i\_5 q(-)|M(P),\ \^\_[3;M]{}(x) &=& - \^\_[-]{} \^x\_0 dx’ e\^[-i’(P)]{} 0|[|q]{}() i(/P /- P)\_5 q(-)|M(P), respectively, where $\eta=(1,0,0,-1)$ and $P$ is the four-momentum of the meson ($P^2=m^2_M$) and $x$ corresponds to the longitudinal momentum fraction carried by the quark and $\zeta =2x -1$. The normalization parameter $\mu_M = m^2_M /(m_q + m_{\bar q})$ results from quark condensate. For the pion, $\mu_\pi = -2\la {\bar q}q\ra / f^2_\pi$ from the Gell-Mann-Oakes-Renner relation [@GOR]. The nonlocal matrix elements ${\cal M}_{\alpha} \equiv \la 0|{\bar q}(\tau\eta) i\Gamma_\alpha q(-\tau\eta) |M(P)\ra$ for pseudoscalar ($\Gamma_\alpha=\gamma_5$) and pseudotensor ($\Gamma_\alpha=(\slash\!\!\!\!P \slash\!\!\!\eta - P\cdot\eta)\gamma_5$) channels are given by the following momentum integral in two-point function of the manifestly covariant BS model \[Deq:4\] [M]{}\_= N\_c e\^[-i k]{} e\^[-i (k-P)]{} H\_0, where $N_c$ denotes the number of colors. The quark propagators of mass $m_q$ and $m_{\bar q}$ carry the internal four-momenta $p =P -k$ and $k$, respectively. In order to regularize the covariant loop, we use the usual multipole ansatz [@CJ_V14; @CJ2015] for the $q{\bar q}$ bound-state vertex function $H_0=H_0(p^2,k^2)$ of a meson: $H_0(p^2,k^2) = g/(p^2 - \Lambda^2 +i\varepsilon)^2$, where $g$ and $\Lambda$ are constant parameters. After a little manipulation, we obtain for the pion ($m_q=m_{\bar{q}}$) \[Deq:5.1\] \^P\_[3;]{}(x) &=& M\^2\_0,\ \^\_[3;]{}(x) &=& - \^x\_0 dx’ ([**k**]{}\^2\_+ m\^2\_q), where $\chi(x,{\bf k}_\perp) = \frac{g}{[x (m_M^2 -M^2_0)][x (m_M^2 - M^2_{\Lambda})]^2}$ and $M^2_{0(\Lambda)} = \frac{ {\bf k}^{2}_\perp + m^2_q(\Lambda^2)}{x} + \frac{{\bf k}^{2}_\perp + m^2_{\bar q}}{(1-x)}$. In the standard LFQM, the wave function of a ground state pseudoscalar meson as a $q\bar{q}$ bound state is given by $\Psi_{\lam{\bar\lam}}(x,{\bf k}_{\perp}) ={\Phi_R(x,{\bf k}_{\perp})\cal R}_{\lam{\bar\lam}}(x,{\bf k}_{\perp})$, where $\Phi_R$ is the radial wave function and the spin-orbit wave function ${\cal R}_{\lam{\bar\lam}}$ with the helicity $\lam({\bar\lam})$ of a quark(antiquark) that is obtained by the interaction-independent Melosh transformation [@Melosh; @Mel2] from the ordinary spin-orbit wave function assigned by the quantum numbers $J^{PC}$. For the radial wave function $\Phi_R$, we use both the gaussian or harmonic oscillator (HO) wave function $\Phi_{\rm HO}$ and the power-law (PL) type wave function $\Phi_{\rm PL}$ [@schlumpf] as follows \[QM2\] \_[HO]{}(x,[**k**]{}\_)= e\^[-[k]{}\^2/2\^2]{}, \_[PL]{}(x,[**k**]{}\_)= , where ${\bf k}^2={\bf k}^2_\perp + k^2_z$ and $k_z=(x-1/2)M_0 + (m^2_{\bar q}-m^2_q)/2M_0$ and $\beta$ is the variational parameter fixed by the analysis of meson mass spectra [@CJ_99]. The normalization of $\Phi_R$ is given by $\int\frac{dx d^2{\bf k}_\perp}{16\pi^3} |\Phi_R(x,{\bf k}_{\perp})|^2=1$. In our previous analyses of twist-2 and pseudoscalar twist-3 DAs of a pion [@CJ2015], we have shown that the SLF results in the standard LFQM is obtained by the the replacement of the LF vertex function $\chi$ in BS model with our LFQM wave function $\Phi_R$ as follows (see Eq. (35) in [@CJ2015]) \[Deq:12\] . The correspondence in Eq. (\[Deq:12\]) is valid again in this analysis of a pseudotensor twist-3 DA $\phi^{\sigma}_{3;\pi}(x)$. We now apply this correspondence to both pseudoscalar DA $\phi^{P}_{3;\pi}(x)$ in Eq. (\[Deq:5.1\]) and pseudotensor DA $\phi^{\sigma}_{3;\pi}(x)$ in Eq. (5) to obtain them in our LFQM as follows: \[Deq:13\] \^[P]{}\_[3;]{}(x) &=& M\^2\_0,\ \^\_[3;]{}(x) &=& \^x\_0 dx’ ([**k**]{}\^2\_+ m\^2\_q), respectively. Numerical Results {#sec:III} ================= In the numerical computations, we use $m_q=(0.22, 0.25)$ GeV and $\beta_{q\bar{q}}=(0.3659,0.3194)$ GeV for the linear (HO) confining potential model parameters for the Gaussian wave function, which were obtained from the calculation of meson mass spectra using the variational principle in our LFQM [@CJ_DA; @CJ_99]. For the PL wave function, we use $m_q=0.25$ GeV and $\beta_{q\bar{q}}=0.335$ GeV adopted from [@schlumpf]. We first compute the second transverse moments $\la {\bf k}^2_\perp \ra^{P(\sigma)}_\pi$ for both pseudoscalar ($P$) and pseudotensor ($\sigma$) channels, and the results obtained from the linear \[HO\] parameters are given by $\la {\bf k}^2_\perp \ra^P_\pi = (553~{\rm MeV})^2 [(480~{\rm MeV})^2]$ and $\la {\bf k}^2_\perp \ra^\sigma_\pi = (481~{\rm MeV})^2 [(394~{\rm MeV})^2]$, respectively. ![\[fig1\] The twist-3 DAs $\phi^P_{3;\pi}(x)$ (left panel) and $\phi^\sigma_{3;\pi}(x)$ (right panel) of pion.](Fig1a.eps "fig:"){height="7cm" width="7cm"} ![\[fig1\] The twist-3 DAs $\phi^P_{3;\pi}(x)$ (left panel) and $\phi^\sigma_{3;\pi}(x)$ (right panel) of pion.](Fig1b.eps "fig:"){height="7cm" width="7cm"} Fig. \[fig1\] shows the two-particle twist-3 pion $\phi^P_{3;\pi}(x)$ (left panel) and $\phi^\sigma_{3;\pi}(x)$ (right panel) obtained from the nonzero constituent quark masses using Gaussian wave functions with HO (solid lines) model parameters and PL wave functions (dashed lines). We also plot our results in the chiral symmetry $(m_{u(d)}\to 0)$ limit for both Gaussian (dotted lines) and PL (dot-dashed lines) wave functions and compare them with the chiral-limit prediction of DSE approach employing the dynamical chiral symmetry breaking-improved kernels [@DSE2015] (double-dot-dashed line) as well as the asymptotic result $6x(1-x)$ (circled data) for the case $\phi^\sigma_{3;\pi}(x)$. Our results for both $\phi^P_{3;\pi}(x)$ and $\phi^\sigma_{3;\pi}(x)$ are normalized without the momentum cutoff (i.e. $|{\bf k}_\perp|\to\infty$). For $\phi^P_{3;\pi}(x)$ case, our results with nonzero constituent quark masses show rather convex shapes for both Gaussian and PL wave functions but they show quite different end point behaviors, i.e. the end points are more enhanced for the PL wave function than the Gaussian wave function. The difference between the two wave functions are more drastic in the chiral symmetry limit, where the result of Gaussian wave function reproduces the result $\phi^P_{3;\pi}(x)\to 1$ anticipated from the QCD’s conformal limit [@BF] but the result of PL wave function shows the concave shape similar to the result of DSE approach [@DSE2015], in which the following asymptotic form $\phi^P_{3;\pi}(x)\to 1 + (1/2) C^{(1/2)}_2(2x-1)$ was obtained. While the authors in [@DSE2015] explained that the difference, i.e. $(1/2)C^{(1/2)}_2(2x-1)$ term in chiral symmetry limit may come from the mixing effect between the two- and three-particle twist-3 amplitudes, we observe that this difference may come from the cutoff scale of the transverse momentum scale associated with the different choice of LF wave functions. Especially, as on can see from Fig. 1, while the end-point suppressed form of $\phi^P_{3;\pi}(x)$ obtained from HO wave function shows constant shape (dotted line) in the chiral symmetry limit, the end-point enhanced form of $\phi^P_{3;\pi}(x)$ obtained from PL wave function shows concave shape (dot-dashed line) in the chiral symmetry limit. The cutoff dependent behaviors of $\phi^P_{3;\pi}(x)$ obtained from both Gaussian and PL wave functions are also presented in [@FB2015], where the concave shape for the Gaussian wave function can also be seen with the cutoff scale $\mu=1$ GeV or less being taken but the cutoff dependence was shown to be more sensitive for the PL wave function than the Gaussian one. For $\phi^\sigma_{3;\pi}(x)$ case, our results with nonzero constituent quark masses for both Gaussian (solid line) and PL (dashed line) show again different end point behaviors, i.e. the end points are more enhanced for the PL wave function than the Gaussian wave function. However, in the chiral symmetry limit, Gaussian (dotted line) and PL (dot-dashed line) wave functions show very similar shapes each other. Furthermore, the result from Gaussian wave function reproduces exactly the asymptotic form $6x(1-x)$. The same chiral-limit behavior was also obtained from the DSE approach [@DSE2015]. As one can see from Fig. \[fig3\], the twist-3 pseudoscalar $\phi^P_{3;\pi}(x)$ is more sensitive to the shape of the model wave functions (Gaussian vs. PL) than the twist-3 pseudotensor $\phi^\sigma_{3;\pi}(x)$. It is quite interesting to note in the chiral symmetry limit that while $\phi^P_{3;\pi}(x)$ is sensitive to the shapes of model wave functions, $\phi^\sigma_{3;\pi}(x)$ is insensitive to them. The twist-3 pseudoscalar DA $\phi^P_{3;M}(x)$ and pseudotensor DA $\phi^\sigma_{3;M}(x)$ are usually expanded in terms of the Gegenbauer polynomials $C^{1/2}_n$ and $C^{3/2}_n$, respectively, as follows [@NK06]: $\phi^P_{3;M} = \sum^{\infty}_{n=0} a^{P}_{n,M}C^{1/2}_n(2x-1)$, and $\phi^\sigma_{3;M} = 6 x(1-x) \sum^{\infty}_{n=0} a^{\sigma}_{n,M}C^{3/2}_n(2x-1)$. The coefficients $a^{P(\sigma)}_{n,M}$ are called the Gegenbauer moments, which describe how much the DAs deviate from the asymptotic one. In addition to the Gegenbauer moments, one can also define the expectation value of the longitudinal momentum, so-called $\xi$-moments, as follows: $\la\xi^n\ra^{P(\sigma)}_M = \int^1_0 dx \xi^n \phi^{P(\sigma)}_{3;M}(x)$. The Gegenbauer- and $\xi$- moments of the pseudoscalar twist-3 DAs $\phi^P_{3;\pi}(x)$ and twist-2 DA $\phi^A_{2;\pi}(x)$ can be found in our previous works [@CJ_DA; @CJ2015]. [@lcccccc]{} Models & $a^{\sigma}_{2,\pi}$ & $a^{\sigma}_{4,\pi}$ & $a^{\sigma}_{6,\pi}$ & $\la\xi^2\ra^{\sigma}_\pi$ & $\la\xi^4\ra^{\sigma}_\pi$ & $\la\xi^6\ra^{\sigma}_\pi$\ HO & -0.1155 & -0.0268 & -0.0046 & 0.1604 & 0.0565 & 0.0263\ Linear & -0.0803 & -0.0256 & -0.0082 & 0.1725 & 0.0647 & 0.0318\ PL & -0.0375 & -0.0092 & -0.0031 & 0.1871 & 0.0762 & 0.0406\ SR [@BBL] & 0.0979& -0.0016 & -0.0011 & 0.2325 & 0.1075 & 0.0624\ DSE [@DSE2015] & $\cdots$ & $\cdots$ & $\cdots$ & 0.20 & 0.085 & 0.047\ $\chi$QM [@NK06] & -0.0984 & -0.0192 & -0.0037 & 0.1663 & 0.0612 & -0.0015\ $6x(1-x)$ & $\cdots$ & $\cdots$ & $\cdots$ & 0.20 & 0.086 & 0.048\ In Table \[t1\], we list the calculated Gegenbauer- and $\xi$- moments of pseudotensor twist-3 pion DA $\phi^\sigma_{3;\pi}(x)$ obtained from the Gaussian wave functions with linear and HO potential models and PL wave function. We also compare our results with other model predictions, e.g. QCD sum rules (SR) [@BBL], DSE approach [@DSE2015] and the chiral quark model ($\chi$QM) [@NK06]. As expected from the isospin symmetry, all odd Gegenbauer and $\xi$ moments are all zero. It is interesting to note that the sign of $a^{\sigma}_{2,\pi}$ is negative from our LFQM and $\chi$QM predictions but is positive for QCDSR prediction. Larger positive value of $a^{\sigma}_{2,\pi}$ leads to more flat shape of DA but the larger negative value leads to more narrower shape of DA. Knowing our LFQM results from the HO model are exact to the asymptotic result in the chiral-symmetry limit, i.e. $[\la\xi^2\ra^{\sigma}_\pi]_{\rm HO}=[\la\xi^2\ra^{\sigma}_\pi]_{\rm asy}$ in $m_q\to 0$ limit, one can see that the $\xi$-moments are reduced when the chiral symmetry is broken. We also should note for the same reason that our LFQM results are in good agreement with DSE results in the chiral-symmetry limit of $\phi^\sigma_{3;\pi}(x)$. Summary {#sec:IV} ======= We analyzed the two twist-3 DAs of pion, i.e. pseudoscalar $\phi^P_{3;\pi}(x)$ and pseudotensor $\phi^\sigma_{3;\pi}(x)$, within the LFQM. We also investigated the discrepancy of the asymptotic forms of $\phi^P_{3;\pi}(x)$ between DSE approach [@DSE2015] and QCD’s conformal limit expression [@BF] from the perspective of dependence of DA on the form of LF wave functions, e.g. Gaussian vs. PL wave functions. In order to compute the twist-3 pseudotensor DA $\phi^\sigma_{3;\pi}(x)$, we utilized the same manifestly covariant BS model used in [@CJ_V14; @CJ2015; @FB2015] and then substituted the LF vertex function in the covariant BS model with the more phenomenologically accessible Gaussian and PL wave functions. Linking the covariant BS model to the standard LFQM, we used the same correspondence relation Eq. (\[Deq:12\]) between the two models as the one found in [@CJ2015; @CJ_V14]. The remarkable finding in linking the covariant BS model to the standard LFQM is that the treacherous points such as the zero-mode contributions and the instantaneous ones existed in the covariant BS model become absent in the LFQM with the Gaussian or PL wave function. Our LFQM descriptions for both twist-3 $\phi^P_{3;\pi}$ and $\phi^\sigma_{3;\pi}$ satisfy the fundamental constraint(i.e. symmetric form with respect to $x$) anticipated from the isospin symmetry. For the $\phi^P_{3;\pi}(x)$ case, our results with nonzero constituent quark masses show rather convex shapes for both Gaussian and PL wave functions but they show quite different end point behaviors, i.e. the end points are more enhanced for the PL wave function than the Gaussian wave function. The difference between the two wave functions are more drastic in the chiral symmetry limit, where the result of Gaussian wave function reproduces the result $\phi^P_{3;\pi}(x)\to 1$ anticipated from the QCD’s conformal limit [@BF] but the result of PL wave function shows the concave shape similar to the result of DSE approach [@DSE2015]. For the $\phi^\sigma_{3;\pi}(x)$ case, our results from both Gaussian and PL wave functions in the chiral symmetry limit, show very similar shapes each other. Especially, the result from Gaussian wave function reproduces exactly the asymptotic form $6x(1-x)$ anticipated from QCD’s conformal limit. The same chiral-limit behavior was also obtained from the DSE approach [@DSE2015]. The remarkable thing is that our predictions for the two twist-3 DAs of $\pi$ and chirality-even twist-2 DA $\phi^{||}_{2;\rho}(x)$ and twist-3 DA $\phi^{\perp}_{3;\rho}(x)$ of $\rho$ [@CJ_V14] obtained from the Gaussian wave function in the chiral limit exactly reproduce the forms anticipated from QCD’s conformal limit. We summarize in Table \[t2\] the asymptotic forms for the two twist-3 DAs of $\pi$ and chirality-even twist-2 and twist-3 DAs of rho meson compared with DSE approach [@DSE2015] and QCD’s conformal limit expression [@BF; @Ball98]. \[t2\] [lllll]{} Model in $m_q\to 0$ & $\phi^P_{3;\pi}$ & $\phi^\sigma_{3;\pi}$ & $\phi^{||}_{2;\rho}(x)$ & $\phi^{\perp}_{3;\rho}(x)$\ LFQM (HO model) [@CJ_V14; @CJ2015] & 1 & $6x(1-x)$ & $6x(1-x)$ & $\frac{3}{4}[1+ (2x-1)^2]$\ DSE approach [@DSE2015] & $1+ \frac{1}{2} C^{1/2}{2}(2x-1)$ & $6x(1-x)$ & $-$ & $-$\ QCD’s conformal limit  [@BF; @Ball98]& 1& $6x(1-x)$ & $6x(1-x)$ & $\frac{3}{4}[1+ (2x-1)^2]$\ The idea of our LFQM is to provide the nonperturbative wave functions at the momentum scale consistent with the use of constituent quark mass. The DAs determined from this nonperturbative wave functions can be fed into the QCD evolution equation to provide the shorter distance information of the corresponding hadrons. While our results for both $\phi^P_{3;\pi}$ and $\phi^\sigma_{3;\pi}$ in Fig. 1 were obtained without the momentum cutoffs (i.e. $|{\bf k}_\perp|\to\infty$), the scales of both DAs obtained from the Gaussian wave functions and that of $\phi^\sigma_{3;\pi}$ obtained from PL wave function are estimated to be $|{\bf k}_\perp|\simeq 1$ GeV. However, the scale of $\phi^P_{3;\pi}$ obtained from PL wave function are estimated to $|{\bf k}_\perp|\simeq 2$ GeV due to the high momentum tail compared to other wave functions. The DAs obtained without the cutoff should not be regarded as the fully evolved DAs but still be nonperturbative as they just mean that the cutoff dependence becomes marginal beyond a certain nonperturbative cutoff scale. [99]{} Lepage, G.P., Brodsky, S. J: Exclusive Processes in Perturbative Quantum Chromodynamics. . Braun, V.M., Filyanov, I.E.: Conformal Invariance and Pion Wave Functions of Nonleading Twist. Z. Phys. C [**48**]{}, 239 (1990). Ball, P., Braun, V.M., Lenz, A.: Higher-twist distribution amplitudes of the K meson in QCD. J. High Energy Phys. 05, 004 (2006). Mikhailov, S.V., Pimikov, A.V., Stefanis, N.G.: Endpoint behavior of the pion distribution amplitude in QCD sum rules with nonlocal condensates. . Petrov, V.Y., Polyakov, M.V., Ruskov, R., Weiss, C., Goeke, K.: Pion and photon light-cone wave functions from the instanton vacuum. . Nam, S.I., Kim, H.-Ch.: Twist-3 pion and kaon distribution amplitudes from the instanton vacuum with flavor SU(3) symmetry breaking. . Chang, L., Roberts, C.D., Schmidt, S.M.: Light front distribution of the chiral condensate. . Shi, C., Chen, C., Chang, L., Roberts, C.D., Schmidt, S.M., Zong, H.-S.: Kaon and pion parton distribution amplitudes to twist three. . Choi, H.-M., Ji, C.-R.: Distribution amplitudes and decay constants for $(\pi, K, \rho, K^*)$ mesons in the light-front quark model. . Choi, H.-M., Ji, C.-R.: Self-consistent covariant description of vector meson decay constants and chirality-even quark-antiquark distribution amplitudes up to twist 3 in the light-front quark model. . Choi, H.-M., Ji, C.-R.: Consistency of the light-front quark model with chiral symmetry in the pseudoscalar meson analysis. . Choi, H.-M., Ji, C.-R.: Light-Front Quark Model Phenomenology Consistent with Chiral Symmetry. EPJ Web of Conference [**113**]{}, 05009 (2016). Brodsky, S. J., Pauli, H. -C. , Pinsky, S.: Quantum chromodynamics and other field theories on the light cone. Phys. Rep. [**301**]{}, 299 (1998). Choi, H.-M., Ji, C.-R.: Mixing angles and electromagnetic properties of ground state pseudoscalar and vector meson nonets in the light-cone quark model. . Choi, H.-M.: Decay constants and radiative decays of heavy mesons in light-front quark model. . Choi, H.-M., Ji, C.-R.: Semileptonic and radiative decays of the B(c) meson in light-front quark model. . Choi, H.-M., Ji, C.-R., Kisslinger, L.S.: Skewed quark distribution of the pion in the light front quark model. . Brodsky, S.J., Hwang, D.S.: Exact light cone wave function representation of matrix elements of electroweak currents. . de Melo, J.P.B.C., Sales, J.H.O., Frederico, T., Sauer, P.U.: Pairs in the light front and covariance. . Choi, H.-M., Ji, C.-R.: Nonvanishing zero modes in the light front current. . Ball, P., Braun, V.M., Koike, Y., Tanaka, K.: Higher twist distribution amplitudes of vector mesons in QCD: Formalism and twist - three distributions. . Gell-Mann, M., Oakes, R., Renner, B.: Behavior of current divergences under SU(3) $\times$ SU(3). Phys. Rev. [**175**]{}, 2195 (1968). Melosh, H.J.: Quarks: Currents and constituents ; Chung, P.L., Coester, F., Keister, B.D., Polyzou, W.N.: Hamiltonian Light Front Dynamics of Elastic electron Deuteron Scattering. . Schlumpf, F.: Charge form factors of pseudoscalar mesons. . [^1]: This work was supported by the Korean Research Foundation Grant funded by the Korean Government (No. NRF-2014R1A1A2057457).

ArXiv

--- abstract: | The crossing property is perhaps the most subtle aspect of the particle-field relation. Although it is not difficult to state its content in terms of certain analytic properties relating different matrixelements of the S-matrix or formfactors, its relation to the localization- and positive energy spectral principles requires a level of insight into the inner workings of QFT which goes beyond anything which can be found in typical textbooks on QFT. This paper presents a recent account based on new ideas derived from “modular localization” including a mathematic appendix on this subject. Its main novel achievement is the proof of the crossing property of formfactors from a two-algebra generalization of the KMS condition. The main content of this article is the presentation of the derailments of particle theory during more than 4 decades: the S-matrix bootstrap, the dual model and its string theoretic extension. Rather than being related to crossing, string theory is the (only known) realization of a dynamic infinite component one-particle wave function space and its associated infinite component field. Here “dynamic” means that, unlike a mere collection of infinitely many irreducible unitary Poincaré group representation or free fields, the formalism contains also operators which communicate between the different irreducible Poincaré represenations (the levels of the “infinite tower”) and set the mass/spin spectrum. Wheras in pre-string times there were unsuccessful attempts to achieve this in analogy to the O(4,2) hydrogen spectrum by the use of higher noncompact groups, the superstring in d=9+1, which uses instead (bosonic/fermionic) oscillators obtained from multicomponent chiral currents is the only known unitary positive energy solution of the dynamical infinite component pointlike localized field project. When the first version of this paper was submitted to hep-th it was immediately removed by the moderator and placed on phys. gen without any possibility to cross list, even though its content is foundational QFT. With the intervention of another member of the advisory comitee at least the cross-listing seems now to be possible. author: - | Bert Schroer\ CBPF, Rua Dr. Xavier Sigaud 150\ 22290-180 Rio de Janeiro, Brazil\ and Institut fuer Theoretische Physik der FU Berlin, Germany date: December 2009 title: | A critical look at 50 years particle theory from the perspective of the crossing property\ [Dedicated to Ivan Todorov on the occasion of his 75th birthday]{}\ [to be published]{} [in Foundations of Physics]{} --- The increasing gap between foundational work and particle theory ================================================================ There has always existed a tendency to romanticize the past when criticizing the present. But the importance of interpretational and philosophical ideas for the development of quantum theory (QT) in the first three decades of particle theory, starting in quantum mechanics (QM) and escorting the beginnings of quantum field theory (QFT), as compared to their superficial role or absence in the ongoing particle theory is hard to be overlooked. Most of the foundational concepts in relativistic QT can be traced back to developments before 1980. One can hardly think of any other branch of physics in which the correct interpretation of observational results was that much dependent on the outcome of a delicate balance between speculative innovations being followed by critical foundational work in which questions of conceptual aspects and philosophal consistency were the main driving force. The strength of this connection between descriptive and conceptual aspects in the beginning of quantum physics was a result of the protagonist’s (Bohr, Heisenberg, Schroedinger,..) intense interests in conceptual and philosophical questions of of quantum theory. Almost the entire arsenal of foundational concepts, including those iconized Gedankenexperiments as *Schroedinger’s cat* and *Fermi’s two-atom experiment* in QED (arguing that $c$ as the classical velocity of light remains the maximal speed after the quantization of electrodynamics), were introduced in order to highlight the philosophical consequences of their discoveries and to facilitate a critical engagement with the new theory for others. But this does not mean that all this impressive grand design was an inevitable outcome of the innovative potential of the protagonists. Even the greatest intellectual brilliance is no insurance for finding the “diretissima” for scientific progress; already a slight change in the chronological ordering of important discoveries could have led to a time-consuming detour or a blind allay. Just imagine that Feynman’s path integral would have entered the scene before matrix mechanics, Schrödinger QM and transformation theory; as a result of the conceptual proximity of an integral over classical orbits with the Bohr-Sommerfeld framework of the largely quasi-classical old quantum theory there is hardly anything more natural than to contemplate such a direct connection. The resulting formalism would have unified all the quasi-classical results of the old quantum theory and lifted it to a new level. It would have streamlined most previous calculations and presented an elegant way how to do computations around quantum oscillators, but it would have missed the important dichotomy between observables and states. Even worse, the elaboration of the Hilbert space formalism and operators acting in it, as well as all the understanding of those important integrable systems as the hydrogen atom (which even with all the present hindsight about path integrals remaines a nontrivial endeavour) without whose operator presentation a course on QM is unthinkable, all these important contributions would have appeared much later and in a very different (and probably more involved) form. Fortunately this was not the way things unfolded; by the time Feynman proposed his path representation in the setting of QM, the conceptual level of operator QT was mature enough to resist the temptation of a fallacious short-sighted interpretation of this elegant, but often conceptual and computational unsafe formalism. In this way many years of confusion in quantum physics were avoided and the path integral could be explored for those purposes for which it is powerful, namely quasiclassical approximations, keeping track of combinatorial aspects of renormalized perturbation theory[^1] and for presenting a flexible metaphoric top soil on which innovative ideas can sprout and specific computational problems be formulated. Many results based on operator formalism on the other hand are either out of reach of the path integral, or can only be obtained by imposing rules which do not follow from its measure theoretic foundation and are therefore less trustworthy than direct operator methods. The conceptual-mathematical control is limited to QM and certain (superrenormalizable) models in low dimensional QFTs, but this does not diminish its value as an intuitive guide and a social cohesion-creating construct in discussions among particle physicists with different backgrounds. Taking into account that progress in particle physics is not only the result of the intellectual capacity and the originality of the involved actors, but also requires an element of good fortune in taking the right turns at the right time on important cross roads, there is ample reason for considering the first three decades of particle physics in retrospect as the “good old days”. The aim of this essay is to shed light on later developments, when innovation, critical analysis and luck began to drift apart. The best way to do this is to revisit the chain of events which started from the S-matrix bootstrap approach and culminated in string theory. It is not difficult to localize the point of no return, from where the present less fortunate direction in particle theory research took its beginning, by following the events in the aftermath of the enormous successful perturbative renormalized quantum electrodynamics (QED). The emerging difficulties to treat the nuclear interactions with the same methods led to a revival of S-matrix based ideas. This time the connection between relativistic local fields and asymptotic in/out particles was better understood than in Heisenberg’s ill-fated first attempt [@Heisen] a decade before the S-matrix bootstrap. Instead of investigating a concrete hadronic model, for which there existed at that time no computational framework, the most reasonable approach was to look for some experimentally accessible consequences of the general causal locality principles underlying QFT. This led to the derivation of a form of the Kramers-Kronig dispersion relations known from optics, but now ingeniously adapted to particle physics. The derivation of these relations from first principles and their subsequent experimental verification in high energy collisions was the main aim in which many of the best brains of the 50s participated. According to the best of my knowledge this was the only topic in post QED particle theory which can be characterized by the words “mission accomplished”; several years of dedicated work led to the solution of the problem, so that one could move on to other problems in an upbeat spirit without being obliged to revisit the problems in order to patch up conceptual holes left behind. It was in the wake of dispersion theory that the notion of the crossing property appeared; first as a property in Feynman graph perturbation theory and soon afterwards as a consequence of the same principles which already led to the dispersion relation. Bros Epstein Glaser and Martin **[@BEG][@E-G-M]** succeeded to proof the validity of crossing property by showing that the two particle elastic scattering amplitude is analytically connected to its crossed[^2] version. The analytic connection between these processes establishes the existence of a “masterfunction” which analytically links all these different processes. Its existence in turn suggested that the asymptotic high energy behavior of the different processes may not be independent, an idea which was confirmed in [@Martin]. There exist also proofs of “asymptotic crossing” for $2\rightarrow3~$scattering and indications about how to generalize this to $2\rightarrow n$ scattering [@Br]. Some comments on the ideas used in this derivation can be found in the next section. Since causal localization is the only foundational property which distinguishes QFT within quantum theory (for this reason often referred to as LQP i.e. local quantum physics [@Haag]), the fact that the wealth of different models with their distinct physical manifestations are in some way related to different realizations of causal localization is to be expected. What is however highly nontrivial is the chain of arguments and the richness of additional concepts which are needed in order to establish this connection. In the present work one of the oldest and most mysterious properties on the border between particle and fields, namely the crossing property, is generalized to formfactors and general scattering amplitudes. The modular localization methods used in that derivation reveal that the conceptual setting is a *two-algebra generalization of the thermal KMS property* (section 5). Although this KMS property is, as the Bros Epstein Glaser arguments which are based on analytic completions of expectation values, derived from locality and localization, the former is easier, furthergoing and more physical. Continuing the S-matrix history, in the subsequent revival of S-matrix theory the newly discovered crossing played an essential role. It was the main distinctive new feature with respect to Heisenberg’s ill-fated prior S-matrix proposal of the 40’s. The S-matrix bootstrap program attracted the attention of many particle theorists for almost a decade before it disappeared from journal publications and conference topics[^3]. The apparent reason was “physical anemia” i.e. its inability to produce any credible calculation from its underlying principles. There was certainly nothing wrong with its S-matrix principles of unitarity, Poincaré invariance and crossing, except that the “maximal analyticity” postulate resulted from a misunderstanding of the role of analyticity in physics since different from the times of the S-matrix bootstrap, it not represent a physical principle but rather results from one which unfortunately was not clearly identified. The connection between the physical causal locality principles and their analytic consequences are subtle and long winding, but there is no way to sidestep these subtleties by turning the logic on its head. What was however grossly misleading was the claim that the *nuclear democracy*[^4] behind the bootstrap principles has at most one solution (the possibility of having no solution was admitted) which describes the entire world of strong interaction. Such sweeping ultra-reductionist uniqueness claims arose occasionally in particle physics, usually in connection with certain nonlinear structures to which it was difficult to find any solution at all (e.g. the Schwinger-Dyson equation). Reasonable formulations as QFT “defuse” such nonlinear structures as e.g. the unitarity of the S-matrix by showing that they result from linear asymptotic properties of fields.The belief in the uniqueness of the bootstrap mechanism contained already germs of modern ideological thinking, which in more recent times took the extreme form of a theory of everything (TOE). Several years after the disappearance of the S-matrix bootstrap, the principles which underlie the construction of so-called factorizing two-dimensional models were discovered [@KTTW] which kick-started a still ongoing stream of results about a family of new interesting soluble models[^5]. These rich results came from the observation that factorizing two-dimensional elastic S-matrices can indeed be classified and constructed by the bootstrap principles of the meanwhile abandoned S-matrix bootstrap approach. Factorization in conjunction with dispersion theoretic analyticity led to meromorphy in terms of the rapidity variables which gave a precise meaning and a physical interpretation to “maximal analyticity” at least in this special case. The protagonists of the old bootstrap program never took notice of these astonishing new observations about an interesting subset of nontrivial QFTs which have become a theoretical laboratory to test ideas of QFT [@Sch][@Lech1]; in this way they spared themselves the confrontation with their earlier premature apodictic statements on this matter. The two-dimensional bootstrap project has infinitely many solutions and serves as the starting point of a new infinitely large family of genuine nontrivial two-dimensional QFTs. These constructions did not only show that the claimed unicity was wishful imagination, but also revealed that the idea that all QFT can be described in a Lagrangian setting was too narrow: the bootstrap classification of all two-dimensional factorizing S-matrices had infinitely many more solutions than those which can be described by pointlike Lagrangian couplings between free fields. There were many ad hoc concepts invented in the wake of the S-matrix bootstrap, the most prominent (used in many later papers) was the Mandelstam spectral representation [@Vech]. At that point the philosophy underlying physical research had significantly changed as compared to the era of dispersion relation[^6]. For the latter it was essential to be a *rigorous* consequence of spectral representation (the Jost-Lehmann-Dyson representations) which in turn were derived from the locality and spectral principles of QFT. Without this strong connection with the underlying principles, the experimental verification of dispersion relations would have remained without much significance since they represented a check of the locality principles of QFT and not of the validity of a particular model. The aim of the work of Mandelstam, as well as that of the later work of Veneziano leading up to string theory, was very different from that of dispersion relations; although it originated with a phenomenological entitlement, it soon turned into a rather freewheeling attempt to explore an imagined area beyond QFT with yet unknown principles. In other words these attempts were excursions into the “blue yonder”, but certainly not from a firm platform of departure to which one could return in case of failure which in particle theory is more common than success. As soon as the phenomenological basis was lost as a result of new experiments which turned out to be incompatible with the Regge trajectory phenomenology, the dual model and string theory became free-floating mathematical ideas without any conceptual basis to which they could safely return. The main part of the paper will be concerned with a critical look at post S-matrix bootstrap ideas as the phenomenological dual model and the closely related string theory, which the protagonists of these models and others thought of as particular implementations of the crossing property. Following [@Mack] it will be shown that the dual model properties are identical to the analytic properties of Mellin transforms of conformal correlation; they have nothing in common with the correctly understood crossing property of formfactors and scattering amplitudes which belong to a very different conceptual setting. Since the *crossing property is one of the most subtle relations between particles and fields*, part of our task consists in presenting an up to date account of a derivation of crossing from the causality and covariance principles of QFT. The full depth of the crisis in contemporary particle physics cannot be perceived, and its causes cannot be understood, without a careful conceptual and mathematical analysis based on a critical first hand historical knowledge. Commemorative articles as [@Vech] are interesting and certainly contain a lot of important background material, but one should not expect to find a critical view in them. If one asks a particle theorist of sufficient age to point at an important difference between the scientific discourse in the old days and the one in more recent decades, he will probably agree that, whereas the intellectual potential has remained the same or even increased, there has been a remarkable reduction of critical contributions and public controversies. The great conceptual discourse of the early years of QT gave way to a new style in which metaphorical arguments were allowed a more permanent position and the appreciation of the pivotal role of the delicate equilibrium between innovative speculations and their critical evaluation (which made particle physics such a success story) was declining. At the time of Pauli, Lehmann, Källén, Feynman, Landau, Jost, Schwinger and others it was the critical analysis of new ideas which kept particle theory on a good track. Although controversies became sometimes abrasive for the persons directly involved, particle physics profited from them. Since the time of Jost’s criticism [@Jost] of the S-matrix bootstrap idea in the 60s, there has not appeared any profound critical article of essay about the ideas leading from S-matrix theory to string theory[^7]. Less than ever was string theory itself subjected to critical evaluation about its conceptual-mathematical structure, the critique of its sociological and philosophical epiphenomena is not sufficient. Those prestigious physicists, who in previous times would have considered it as their privilege, if not their moral duty, to give a critical account, became string theories fiercest defenders, if not to say its propagandists. For a historical and foundational interested researcher with textbook knowledge of QFT, the 40 year lasting dominance of this theory is surrounded by a nearly impenetrable mathematical conceptual cordon which makes it difficult to extract relevant foundational aspects. The present article can not change a situation which has been going on for 40 years and in this way became immunized against conceptual objections [^8], but it does present some unknown facts which may become useful in a not so far future, when historians and philosophers finally become curious about what really went on in particle physics for almost half a century and in particular what happened to all those noisy promises of a TOE. The content of the various sections is as follows. The next section explains the formal aspects of the crossing property. It contains in addition to mathematical facts also philosophical aspects. The third section presents the dual resonance model and its derivatives and explains why the absence of a critical evaluation of this interesting class of models which result from the suitably normalized Mellin transformation of (any) conformal QFT (and have nothing to do with properties of the S-matrix) prepared the ground which led into the metaphoric landscape on which string theory subsequently flourished. The fourth section shows that string theory is, despite its name, not about objects which have a string-like spacetime localization; rather the objects of string theory define a *“dynamical” infinite component pointlike field* [^9]; this section therefore constitutes the core of the critical part of the presentation. Section 5 and 6 present the modern view of the crossing property which to a certain extent explains why it led to so many misunderstandings and metaphoric ideas. Despite the highly mathematical level of these sections, the presentation of the mathematical state of art on crossing is not the principle motivation. But a critical exposition of ideas which historically emanated from an incompletely or even incorrectly understood crossing property would itself be incomplete without giving the modern viewpoint on this subtle property. The conclusions present a resumé and additional critical remarks. The crossing property and the S-matrix bootstrap approach ========================================================= In contrast to QM where particles play the role of stable quanta which keep their identity in the presence of interactions, QFT comes with a much more fleeting particle concept. Even in theories without interactions, where relativistic particles are synonymous with free fields, composite operators as e.g. the important conserved currents exhibit the phenomenon of (finite) vacuum polarization. This makes such an object rather singular (an operator-valued distribution with no equal time restriction) and renders the definition of a *partial charge* corresponding to a finite volume a delicate problem with the help of which Heisenberg [@Hei] discovered the property of vacuum polarization at the beginning of QFT. The full subtlety of this problem only became manifest in the presence of interactions; this is the situation in which Furry and Oppenheimer [@F-O] observed that even the basic Lagrangian fields, which without interactions were linear in the particle creation/annihilation operators, cannot create one-particle states without an admixed infinite cloud[^10] of particle/antiparticle pairs. Re-interpreted in a modern setting, this observation permits the following generalization*: in an interacting QFT there exists no operator localized in a compact spacetime region which, if applied to the vacuum, creates a one-particle state without an infinite vacuum polarization cloud.* Or using recent terminology: a model which contains among its operators a compactly localized PFG (vacuum-**p**olarization-**f**ree **g**enerator) is generated by a free field [@BBS][^11]. The “shape” of the locally generated vacuum polarization cloud depends on the kind of interaction, but its infinite particle content is a characteristic property shared by all interacting theories; a finite number of particle-antiparticle polarization pairs created by “banging” on the vacuum with a local (composite) operator can only happen in a free theory. The sharpness of the localization boundary (horizon) accounts for the unboundedness of the energy/entropy content [@BMS]. The subtlety of the particle/field problem (not to be confused with the particle/wave dualism of QM which was already solved by the transformation theory of the 20s) was confirmed in the discovery of perturbative renormalization and the time-dependent scattering theory[^12]. The main conceptual message was that in interacting QFT the notion of particles at finite spacetime had no intrinsic covariant (reference system-independent) meaning. Only the asymptotic particle states are intrinsic and unique, whereas the fields (basic or composites within the chosen description) form an infinite set of frame independent objects whose physical nature is however somewhat fleeting since observationally they carry a large amount of redundancy (infinitely many different fields in the same local equivalence class lead to the same asymptotic particle and scattering amplitudes. The situation resembles the use of coordinates in geometry; the redundancy inherent in the use of different coordinate systems corresponds to the use of different field coordinatizations generating the same system of local operator algebras which correspond to the intrinsic (coordinate-free) way of doing geometry. This view is reflected in the terminology of the 50s, when fields were referred to as “interpolating” fields, thus highlighting that they should be considered as mediators of events involving particles without acquiring direct observational relevance by themselves. In fact the algebraic approach, which started shortly after the LSZ scattering theory, had as its main aim the establishment of a setting in which the infinite plurality of fields is encoded into the infinite ways of coordinatizing a *unique* *system of spacetime-localized algebras*. In this way the setting of a spacetime-indexed net of operator algebras represents a compromise between an extreme on mass-shell/S-matrix point of view and a formulation in terms of the infinitely many ways of generating the same unique net of spacetime-indexed algebras using different covariant “field coordinatizations”.   This particle-field problem has again become a controversially debated issue in the setting of QFT in curved space time (CST) [@Wa] when the Poincaré symmetry including the notion of the vacuum and particle states is lost. There are many results of QFT which are consistent with the Lagrangian quantization setting (with which QFT is often incorrectly identified), but which cannot be derived by textbook Lagrangian methods and rather require operator algebraic methods. In this case it may be helpful for the reader to replace the standard terminology QFT by local quantum physics (LQP). The main difference is methodological and consists in the use of field-coordinatization-independent algebraic methods wherever this is possible. There exists an important area of QFT for which up to this day the use of pointlike covariant field coordinates cannot be avoided namely *renormalized perturbation theory*. But even there the causal perturbation theory a la Epstein-Glaser [@E-G] in terms of an iterated lowest order input (in the form of an invariant polynomial pointlike coupling between free fields) contains some of the LQP spirit. The coupling of free fields to invariant interaction polynomals has hardly any direct relation to Lagrangian quantization[^13]. The method is based on the iterative application of the causality and spectral principles of QFT; it does not follow the quantum mechanical logic of defining formal operator as e.g. Hamiltonians via momentum space cutoffs as unbounded non-covariant operators whose cutoff dependence must then be removed in order to be formally consistent with the principles. But even when the E-G formalism would reach its limits in the problem of separating the infrared dibergencies from short distance problems of perturbative renormalization of nonabelian gauge couplings, there is still the possibility of a saving grace by separating the issue of states from operators and operator algebras and in this way arrive at an infrared finite local algebraic structure and leave the infrared problems to the construction of states [@Fre]. Such a operator-state dichotomy is impossible in the Lagrangian or functional integral formulation. There was however one seemingly mysterious property in the particle-field relation which, even using the advanced conceptional tool box of LQP, did not reveal its mystery. This is the *crossing property* (often called misleadingly “crossing symmetry”). Only recently this property has lifted some of its secrets (see last two sections). Since this property and other ideas which resulted from it constitute the central subject of the present essay, a clear definition is paramount. Fortunately this is not difficult since the problem is not in its presentation, but rather its connection with the principles of QFT. Its formal aspects in Feynman’s perturbative setting was obtained by combining two observations: the invariance of certain families of subgraphs in the same perturbative order under the conjugate interchange of incoming with outgoing lines (the graphical crossing), and the less trivial mass shell *projection* of the connecting analytical path resulting in an analytic relation *onto the complex mass shell* between amplitudes describing two different scattering processes. It is this second step of demonstrating the existence of an analytic path on the complex mass shell linking the backward mass shell defined by analytic continuation in formfactors with the interchange $in\longleftrightarrow out$ and $particle\longleftrightarrow antiparticle$ which (even in the setting of renormalized perturbation theory) remains somewhat nontrivial. According to the LSZ scattering theory collision amplitudes can be obtained from formfactors, hence it is natural to formulate the crossing identity first in this context. A *formfactor* is a matrix elements of a field between “bra” states, consisting of say n-k outgoing particles, and k incoming particles in a “ket” state. Taking the simplest case of a scalar field $A(x)$ between spinless states of one species it reads$$\begin{aligned} & ^{out}\left\langle p_{k+1},...p_{n}\left\vert A(0)\right\vert p_{1}...,p_{k-1},p_{k}\right\rangle ^{in}\label{cross}\\ & =~^{out}\left\langle -p_{k}^{c},p_{k+1},...p_{n}\left\vert A(0)\right\vert p_{1}...,p_{k-1}\right\rangle _{c.o}^{in}\nonumber\end{aligned}$$ in words: the incoming 4-momentum on the mass ahell $p_{k}$ is “crossed” into the outgoing $-p_{k}^{c},$ where the $c$ over the momentum indicates that the particle has been crossed into its antiparticle and the -sign refers to the fact that the formfactor is not between physical states but rather involves the analytic continuation of one. The subscript $c.o$ (contractions omitted) indicated that contraction terms of $p_{k}~$and the other $p^{\prime}s$ (inner products) which are absent in the uncrossed configuration must be excluded after the crossing. Since their structure is different from that of the uncontracted leading terms, they can be easily separated from the main term. This notational complication can be avoided if one formulates the crossing relation directly in terms of free incoming/outgoing fields instead of particles (section 5). The relation (\[cross\]) would be physically void if it would not come with an assertion of analyticity which connects the unphysical backward mass shell momentum with its physical counterpart. The (still unphysical) crossing identity (\[cross\]) together with the analyticity which connects backward to forward momenta constitute the crossing property; there is no direct identity between physical formfactors, only the affirmation that they are related by analytic continuation. The proof is provided by modular localization which will be the central issue in section 5. The iterative application of the crossing relations permits to compute general formfactors from the vacuum polarization components of $A(x)$$$\left\langle 0\left\vert A(0)\right\vert p_{1},p_{2},...,p_{n}\right\rangle ^{in}=~^{out}\left\langle -p_{k+1}^{c},...-p_{n}^{c}\left\vert A(0)\right\vert p_{1}...,p_{k-1},p_{k}\right\rangle _{c.o}^{in}$$ where the charge conservation forces particles to be crossed into antiparticles. Only the vacuum polarization matrixelement does not need the subscript $c.o~$since contraction terms occur solely between bra and ket momenta. The identity only holds for unphysical momenta. By analytic continuation one can get to any formfactor with the same total number of particles starting from the vacuum polarization component i.e. a local “bang” on the vacuum $A\Omega$ determins all formfactors. The S-matrix elements result from the formfactors by choosing for $A(x)$ the unit operator[^14]; since the latter cannot absorb energy-momentum, the incoming momenta are bound to the outgoing by the energy-momentum conserving delta function which leads to some peculiarities. The analyticity in the momentum space representation can only be valid for the function which remains after extracting the delta function. Hence by crossing one particle it is not possible to return to a physical scattering process. One needs to cross simultaneously an incoming and outgoing pair in order to preserve the energy-momentum delta function for physical momenta. This is particularly obvious if the crossing starts from a two-particle state so that crossing only one particle will not lead to an analytic continuation of a physical process. The formfactor of the identity operator with the vacuum or with the one particle state on one side vanishes. In order to come to a relation whose analytic continuation has a nontrivial relation to elastic 2-particle scattering one must simultaneously cross a particle from the opposite side i.e. cross a pair in exactly the way in which crossing was first observed for two particle scattering in the setting of Feynman graphs. We will return to this case in section 5. Crossing looks as being closely related to TCP. Although both properties can be derived from modular localization, the derivation of crossing turns out to be much more subtle than that of the TCP theorem. **** The main conceptual role of crossing is that it relates the various n-particle matrixelements of a local operator which belong to different distributions of n particle momenta into incoming ket and outgoing bra states of an analytic *master function*. This is of course much more than the tautological statement that these matrix elements can be computed once a concrete model has been selected; it really means that once one process has been computed, the others are uniquely determined in a model-independent way without doing another QFT computation. Since this essay also addresses readers with interests in philosophical aspects, the occasional use of metaphoric arguments as a rapid vehicle to convey a mathematically difficult property which places LQP into sharp contrast with QM (even in its relativistic form [@Co-Po][@interface][^15]) should not cause problems. In any case this will be limited to solved problems whose mathematical presentation can be found in the existing literature. The crossing properties of formfactors point at the most important consequence of causal localization in the presence of interactions: the ability to couple all multi-particle channels with the same superselected quantum numbers with each other, in particular the non-orthogonality of corresponding localized states. This is a double edged knife, it makes QFT much more foundational than QM but it also renders many operator techniques one learns in QM unusable. In the case of formfactors the analytic properties of crossing prevent that there are special matrix elements which vanish, leading to the absence of certain processes. Crossing is a special illustration of a general property of LQP which often is expressed in an intuitive way as a kind of benign form of “Murphy’s law”: *particle states which (by charge superselection rules) are allowed to communicate (via formfactors), actually do communicate i.e. their coupling cannot be prevented, it rather constitutes a structural property of any QFT.* It is this property which is behind the interaction-induced (infinite) vacuum polarization clouds resulting from “banging” with a local operator $A$ on the vacuum; but it is certainly less metaphoric than the standard textbook presentation of the vacuum as a “broiling soup of virtual particles” which is allowed to violate the energy-momentum conservation for short times thanks to the uncertainty relation. A special but important case of Murphy’s law governing the coupling of channels is the principle of *nuclear democracy*. It states that QFT cannot distinguish between elementary and bound particles, the only hierarchy consistent with nuclear democracy is the already mentioned one between basic and fused *charges*. This means in particular that it is consistent to view any particle as the result of a fusion of a cluster of other particles whose fused joint charge is contained in the reduction of the fused charge spectrum of the cluster under consideration. Nuclear democracy is certainly a principle which contradicts the boundstate hierarchy of QM in a very radical way; if even the charge-carrying “elementary” particle can be interpreted as resulting from the collective fusion of its own charge with that of a local cluster of suitably chosen other local charges, then the strict hierarchy between elementary (fundamental) and composite certainly breaks down. Hence regarding the formation of “bound particles” i.e. eigenstates of the mass operator with a fused charge, the situation is *radically different from that in QM* because there is nothing which will prevent this particle from coupling in a formfactor to all other states which the superselection rules permit. The crossing property transfers the validity of *Murphy’s Law* and the resulting principle of nuclear democracy for formfactors to the phenomenon of vacuum polarization where there exist theorems showing that no vacuum polarization component can vanish in a theory with nontrivial interaction[^16].   Let us now sketch the ideas which were used in the original proof of crossing [@BEG]. The elastic 2-particle amplitude is a function of the 3 Mandelstam variables s,t,u which are not independent but obey the relation $s+t+u=~m_{1}^{2}+m_{2}^{2}+m_{3}^{2}+m_{4}^{2}.$ There are 3 physical processes (and their TCP conjugates) which can be reached if one knows the amplitude as a function of the full range of the Mandelstam variables s, t. Bros, Epstein and Glaser started from the LSZ representation in terms of Fourier transforms of time ordered functions and used known analytic properties of the latter in order to show that the physical region in terms of Mandelstam variables s$>$0, t$<$0 can be connected to the two other possible physical regions by an analytic path. The proof is somewhat involved because it is not the primitive analyticity domain of the starting correlation function but rather its holomorphy envelop which leads to the desired result. Although the analytic prerequisites for the continuation between forward and backward mass shell through the complex mass shell are proven, the crossing identity does not explicitely appear in that work. These papers are an illustration of the profound mathematical knowledge which physicists acquired in the pursuit of structural problems in QFT of the 60s. The proof of crossing and later generalizations only addressed special cases of scattering. However the interesting connection with thermal KMS properties and a more general proof only came into the open in connection with formfactor crossing as will be explained in the sequel. Only decades later it became clear that localization in QFT (restriction of the vacuum state to the local subalgebra) converts the vacuum state to a thermal KMS state **** $$\Omega_{vac}\upharpoonright\mathcal{A(O})\text{ }is\text{ }\Omega_{KMS}$$ where the Hamiltonian is canonically determined in terms of $(\mathcal{A(O}),\Omega).$ The mathematical theory behind this is *modular theory*. This theory exists in two interconnected versions, the operator algebraic Tomita-Takesaki theory (of which important physical aspects were discovered independently by Haag, Hugenholtz and Winnink [@Haag]) and the modular localization of relativistic wave functions and states [@BGL][@Sch]. These ideas led to a closer connection of the thermal aspects of event horizons in QFT in CST with thermal aspects caused by restricting the vacuum state of global QFTs to localized algebras. A much discussed case is the restriction of the vacuum to the wedge-localized algebra $\mathcal{A}(W)$ which leads to the Unruh effect and an interesting formula for the entropy near the horizon $H(W)$ (the entropy of a light-sheet [@BMS]). The $W$-Rindler word with its lightfront horizons is created in form of a Gedankenexperiment involving a family of uniformely accelerated observers. Black holes with their event horizons would lead to more real astrophysical illustrations of thermal effects resulting from modular localization. Two additional facts finally led to the somewhat surprising result that also the crossing relation belongs to those phenomena which are related to thermal aspects of localization. The first perception was the observation that at least formally the KMS relation written for formfactors of free fields. For free fields and their composites restricted to a wedge region (with the test functions alway having support in W) one has[^17]$$\begin{aligned} \left\langle A(f_{l+1})..A(f_{n})C(h)A(f_{l})..A(f_{1})\right\rangle & =\left\langle A(f_{1})\Delta A(f_{l+1})..A(f_{n})C(h)A(f_{l})..A(f_{2})\right\rangle \label{id}\\ \left\langle p_{n}..p_{l+1}|C(h)|p_{l}..p_{1}\right\rangle & =\left\langle p_{n}..p_{l+1},-p_{1}|C(h)|p_{l}..p_{2}\right\rangle _{c.o}\nonumber\end{aligned}$$ Here $C(h)$ is a h-smeared composite of a free field. For the validity the KMS relation with respect to the modular Hamiltonian $\Delta=e^{-2\pi K}$ with $K$ the Lorentz boost the smearing functions must be localized in $W$. Since the mass-shell restriction of wedge-localized smearing functions form a dense set of wave functions, the momentum space relation in the second line is a consequence. The negative sign of the first momentum is a result of the analytic continuation implied by the imaginary $2\pi$ Lorentz rotation together with the Hermitian adjoint from passing from ket to bra states (\[bra\]); for obvious reasons the backward mass shell momenta are referring to particles with the opposite charge i.e. “anti” with respect to the original one before the cyclic permutation. To obtain the particle states from field states one must Wick-order the $A$-field states on the left hand side and remember that the cyclic permuted $A(f_{1})$ has no contractions with the fields on the right from where it was coming. In the transcription of this relation to particles, the absence of left backward momentum states contracted with right forwards ones is indicated by $c.o.$ (contractions omitted). Hence the crossing relation in the interaction-free case is noting else than the thermal KMS relation of wedge localization (featuring in the Unruh effect) rewritten as a relation between particle matrix-elements (formfactor). In section 5 it will be shown that the interacting case is the particle transcription of a *new modular theory-based field relation which extends the KMS relation*. The correlation functions have analyticity properties in the Lorentz boost parameter, they are analytic in the multi-strip [@Araki]$$\begin{aligned} 0 & \leq\tau_{1}\leq\tau_{2}\leq..\leq\tau_{l}\leq\tau\leq\tau_{l+1}..\leq\tau_{n}\leq1\label{Ara}\\ A(f_{i}) & \rightarrow e^{2\pi\tau_{i}K}A(f_{i})e^{-2\pi\tau_{i}K},~C(h)\rightarrow e^{2\pi\tau K}A(f_{i})e^{-2\pi\tau K}\nonumber\end{aligned}$$ The support properties in x space of wave functions are equivalent to analyticity properties. In particular they imply that certain complex Lorentz transformations which act on the Fourier transformed operators can be absorbed in the analytic continuation of test functions and vice versa. **** Looking only at the contribution of (\[id\]) without contractions among the free fields and using the density of Fourier transformed wedge-supported smearing functions on-massshell, one obtains the crossing relation for the free formfactor $$\left\langle p_{n}..p_{l+1}\left\vert C(0)\right\vert p_{l}..p_{1}\right\rangle =a.c._{q^{c}\rightarrow-p_{1}^{c}}\left\langle q,p_{n}..p_{l+1}\left\vert C(0)\right\vert p_{l}..p_{1}\right\rangle _{c.o} \label{test}$$ where the subscript $c.o$ has the same meaning as before[^18]. This is an identity between a particle matrixelement of C and an a crossed formfactor at an analytically continued momentum; (the notation $-p^{c}$ instead of simply $-p$ indicates that the momentum on the backward shell is that of an antiparticle relative to what it was before the crossing on the ket side. The only somewhat tricky part of rewriting the KMS relation (\[id\]) into the crossing form (\[test\]) is taking the operator $A(f_{1})\Delta$ as its conjugate to the bra vacuum and using modular theory to bring the resulting bra state into the desired form (for the notation see appendix) $$\begin{aligned} & \Delta A(f_{1})^{\ast}\Omega=\Delta SA(f_{1})\Omega=\Delta^{\frac{1}{2}}JA(f_{1})\Omega=A^{c}(\check{f}_{1})\Omega=\int\frac{d^{3}p}{2p_{0}}\left\vert p^{c}\right\rangle \bar{f}(-p)\label{bra}\\ & S=J\Delta^{\frac{1}{2}},~SA\Omega=A^{\ast}\Omega,~A\in\mathcal{A}(W),\mathbf{~~~}\check{f}(p)=\bar{f}(-p)\nonumber\end{aligned}$$ More details can be looked up in section V.4 of [@Haag][^19]. The application of the unbounded modular operators $\Delta^{\frac{1}{2}}=e^{-\pi K},K=W$-associated Lorentz boost generator requires precisely that analytic continuability which is guarantied by the wedge localization. With respect to analyticity there is no difference between the KMS setting and its two-algebra generalization needed for the derivation of crossing in section 5. Instead of invoking modular theory, the free field relation (\[test\]) can also be checked by explicit computation, but this privilege does not exist in the presence of interactions. There is in fact a serious obstacle against applying this argument to interacting formfactors in order to establish the identity (\[cross\]). The reason is obvious since there are 3 different algebras involved $A_{in}(W),A_{out}(W),A(W)$ and the modular operators of interacting operator algebras are different from those generated by their asymptotic free fields. But there is a fortunate circumstance which comes to one’s rescue: at least the domains of the unbounded Tomita S operators $S_{in},S_{out},S$ are identical i.e. the $\Delta^{\prime}s$ coalesce and hence the dense subspace of localized states are the same. The consequences of the identity of the domains are the subtle ingredients in the proof of crossing. We will return to this problem in section 5 and 6 and show that this suffices in order to derive crossing in the formfactor- as well as in the scattering- form. It is interesting to compare the old derivation [@BEG][@E-G-M] which uses holomorphy properties of correlation functions in several variables, including the sophisticated tool of computing holomorphy envelopes (cutting “noses”), with the present one. The wedge localization approach is quite different, even though both rely on analyticity properties coming from locality. Its analytic underpinning is that of Araki’s KMS analyticity for correlation functions and states [@Araki]. The modular approach is more economical in the sense that only the analyticity which is really necessary for the derivation of crossing is used, and analytic completion techniques, whose physical interpretation is not known, are avoided. In this way the important role of crossing in the construction of factorizing models becomes clearer [@Sch][@Lech1]. Finally crossing becomes part of a structural problem of wedge algebras whose thermal manifestations are important in the Unruh effect associated with a (Rindler) wedge and its causal horizon as well as in thermal aspects related to event horizons, including vacuum polarization-induced entropy near null-horizons [@BMT]. This connection between properties from the center of particle theory, with properties which at least historically come from black hole physics, is the real surprise here. At the time of the Bros-Epstein-Glaser work on crossing some quantum field theorists pinned high hopes on the use of new analytic methods for functions of several complex variables for a nonperturbative understanding of QFT. Källén and Wightman [@Ka-Wi] tried for many years to construct a representation of the 3-point function which fulfilled all linear requirements of QFT. They never reached their goal, and this kind of technique subsequently fell out of favor. **** Whether is returns one day together with different problems, who knows? **** The dual resonance model, superseded phenomenology or progenitor of a new fundamental theory? ============================================================================================= The history of the crossing property starting in the early 60s is the key for understanding the direction into which a good part of particle physics research developed afterwards. It began by more or less accidentally stumbling across a property whose importance in particular for an S-matrix-based approach to particle physics was apparent, but whose foundational aspects remained hidden. The necessary conceptual and mathematical tools for its understanding only appeared at the end of the century in the form of modular localization (appendix and sections 5,6). Direct numerical attempts to find approximate solutions of the extreme nonlinear properties of the the S-matrix bootstrap “axioms” ended in failure but unfortunately only strengthened the misleading belief of the existence of a unique non-Lagrangian theory of strong interactions. This was neither the first nor the last time that an ultra reductionist “theories of everything” (TOE) entered the particle theory discourse. As mentioned before, after the completion of the dispersion theory project the underlying philosophy of research began to change. The new strategy was most clearly formulated by Mandelstam. In analogy to the rigorously established Jost-Lehmann-Dyson spectral constructions for matrixelements of field commutators [@Tod] (generalizations of the simpler Källen-Lehmann representation for the two-point function) which became a seminal tool in the derivation of the dispersion relations, Mandelstam proposed an spectral representation for the two-particle scattering amplitude [@Vech] in the hope that the crossing property may be simpler accessed in terms of spectral functions. This representation was never proven and the hope about its use did not materialize, but taken together with ideas about the use of Regge pole trajectories in strong interaction phenomenology it led Veneziano to the mathematical construction of the dual resonance model for elastic two-particle scattering[^20] [@Vech] which was later generalized to an arbitrary number of particles. In terms of Feynman graph terminology it represented the tree approximation for a process of two incoming particles which couple via trilinear interaction vertices to an infinite tower of intermediate particles with ever increasing masses and spins. The decrease of the coupling strengths is carefully tuned in such a way that the sum of all these contributions from the infinite mass/spin tower of the interaction mediating particle poles not only converge in the s-channel (using the canonical terminology introduced by Mandelstam), but represents a function which allows a t-channel interpretation in terms of another sum of infinitely many exchanges via particles from the same mass/spin tower. To find such function in a pedestrian manner, without an operational backup, just by using known properties of gamma and beta functions, is an astonishing achievement which even nowadays commands respect [@Vech]. In hindsight it is somewhat surprising that it was not realized that the dual model and its Nambu-Goto Lagrangian string theory analog which share the same particle/spin tower were the first nontrivial realizations of an object which less than one decade earlier was searched for under the label *infinite component fields*. The motivation came from a completely different corner, namely from the analogy to the $O(4,2)$ “dynamic symmetry” of the hydrogen atom. Infinite component fields in the sense of Fronsdal, Barut, Kleinert and other authors [@Tod] were not just infinitely many fields of varying mass and spin put together as a direct sum, but there was a “dynamic” content consisting in the existence of operators which “vertically” communicate between the different tower levels and set the mass/spin spectrum. This dynamic aspect was expected to arise from noncompact group representations which extend those of the Lorentz group, but this hope did not materialize and the cited authors remained empty-handed. This dynamic requirement makes the construction of an infinite component field a difficult problem. In fact up to date the 10-dimensional superstring field has remained the only dynamical infinite component pointlike solution in which the representation of the Poincaré group is a positive energy unitary ray representation. String theory owed its social success as an infinite component field theory only 6 years after the ill-fated infinite component program to the replacement of higher noncompact groups by the infinite degrees of freedom inherent in multicomponent chiral conformal currents or equivalently in a canonical quantization of the bilinearized (square roots removed) Nambu-Goto Lagrangian. It is one of the missed chances in history that even though the followers of the infinite component field program and those of the later dual model community (which afterwards became incorporated into the string community) had both strong phenomenological roots, they never noticed the proximity of their ideas. It certainly would have been very interesting to be informed that the duality requirement imposed on the vertices of a pole approximation for a scattering amplitude and the mass/spin tower of the Nambu-Goto description of string theory can be encoded into an infinite component field containing operators which intertwine between the levels of the infinite tower. The word “string” would never have appeared and there would have been no danger in misreading string theory as having something to do with string-like objects. Since one cannot change 50 year old customs, the word string theory will always refer to this infinite component pointlike field, whenever we talk about real string (e.g. in gauge theories) we will use the terminology “string-localized”. The duality idea arose from consistency arguments between the low energy resonance contributions and the expected high energy Regge behavior. Veneziano’s first implementation led to several generalizations [@Vech]. The formulation which is most suitable for an in-depth critical analysis is the operational setting of Fubini et al. [@Fu-Ve] which uses multi-component conformal currents and their potential. It may be helpful for the reader to recall at this point some results about conformal currents [@Vech]. The simplest situation is that of a one-component current which, similar to a free field, is determined by its commutation relation$$\begin{aligned} & \left[ j(x),j(y)\right] =-\delta^{\prime}(x-y)\\ Q & =\int j(x)dx,~~\psi(x)=~"e^{i\alpha\Phi(x)}",~~\Phi(x)=\int_{-\infty }^{x}j(x)dx \label{int}$$ Despite its simplicity it leads to a very rich representation theory. There are continuously many representations (labeled by $\alpha)~$as a consequence of the continuous spectrum of the charge Q [@BMT], so in the jargon of chiral QFT these models are extremely “non-rational”. Formally such charged fields are written as exponentials of “potentials” i.e. half space integrals over the current. The quotation marks are meant to indicate that such formulas are conceptually not quite correct since the charge $\alpha$ carrying field $\psi$ does not live in the vacuum sector as the naive reading of this formula would indicate[^21]. This observation is inexorably linked with the infrared divergence of the integral representation, which is the way in which the exponential announces that it is not a quantum field like the others in the Hilbert space generated by the currents. Unfortunately the extended algebra which incorporates *all* charge-carrying fields lives in an inseparable Hilbert space. In order to use currents as a two-dimensional theoretical laboratory following the intrinsic logic of QFT, Buchholz, Mack and Todorov [@BMT] introduced the concept of *maximal local extension* of the algebra of currents. The extension is done by adding certain fields of the form $\psi_{\alpha}(x),$ whose dimension $d_{\alpha}\sim\alpha^{2}$ is integer (and hence which for different localization points commute among each other) to the algebra of currents and view the resulting larger bosonic algebra as the *extended observable algebra*. This reduces the number of charge sectors in a drastic way, their number is now not only countable but even finite (“rational chiral theories”). It turns out that the denumerable set of maximal extension can be explicitly constructed. These do not commute among themselves or with each other but rather obey (abelian) braid group commutation relation. The multi-component generalization of the representation theory of a current turned out to lead to a theory of remarkable richness [@Stas][@Longo]. In this case the maximal extensions are classified by *even lattices* $L$ in $\mathbb{R}^{n},~L:$ $\left( \alpha,\beta\right) =2\mathbb{Z}$.$~$The sectors are then classified by equivalence classes of the dual lattice $L^{\ast}/L$ of which there exist finitely many. The cases with $L=L^{\ast}$ are particularly interesting$.$ These constitute a finite number of models which only exist in their vacuum representation. They are related to finite exceptional groups, among them the famous “moonshine model”. This is just to mention that the following operator dual model construction happens in a fascinating neighborhood. Besides this use of multicomponent current models following the intrinsic logic of LQP, these currents have also been used in an operational approach to the dual model in the work of Fubini at al. [@Fu-Ve] which is somewhat different from the conformal field theoretic logic. Their interest was in the direct use of the potentials $\Phi_{i}$ of the multi-component currents as some quantum mechanical objects$$\begin{aligned} \Phi_{i}(x) & =\int_{-\infty}^{x}j_{i}(x)dx\rightarrow X_{i}(z),~i=1....d\\ Q_{i} & \rightarrow P_{i},~\alpha_{i}\rightarrow p_{i},\text{ }V(z,p)=e^{iP\cdot X(z)}\nonumber\end{aligned}$$ this symbolic formulas are in need of some detailed explanation. The first line indicates a passage from the noncompact to the compact picture $x\rightarrow z,$ and the notation $X_{i}(z)$ anticipates that the potentials are now going to be interpreted as quantum coordinates which classically would trace out a curve in a d-dimensional spacetime. The second line expresses the fact that one really wants to take this reinterpretation into a different direction by adding the identification of the d-component charge operator with the momentum operator and writing the charge-carrying exponential of an would-be n-component “spacetime” as a vertex (or in more recent generalizations *a chiral sigma field*) operator $V$ which carries a noncompact spacetime symmetry (which from the chiral conformal viewpoint of the source theory would be called an inner symmetry). This is the famous source-target relation which later led to the notion of world sheets. But is this strange interpretation of multicomponent charge values as momenta with the operator dimensions of the charge-carrying operators passing to particle masses and the current potential $\Phi_{i}$ becoming a kind of position operator in an multicomponent internal symmetry space consistent? Is such a magic of defining a “target” spacetime in a source-target relation supported by any physical concept? This picture would suggest that the potentials of the conformal current theory define an embedding of a line/circle into spacetime which in turn is the origin of the worldsheets (in analogy to Feynman’s worldlines) in the string theoretic extension of the dual model. Admittedly the identification of an internal symmetry space with noncompact physical spacetime is one of the strangest ideas which entered particle physics [^22], but is it consistent? As oftern, the idea is abscured by the very inappropriated terminology “field space” which somehow insinuates that the space in which classical fields take their values continues to make sense in QFT. We will use instead “internal symmetry space” which is probably what is meant. We will show in the following that the worldsheet picture is incorrect and that instead the localization is pointlike as in standard QFT, which leads to worldlines and is consistent with what was said before about ST really dealing with infinite component pointlike fields. Fact is that one cannot embed any lower dimensional QFT into a higher dimensional one. The frame-independent modular localization of QFT as opposed to frame-dependent Born-Newton-Wigner localization of (relativistic) QM, is a totally holistic concept which does not allow such embeddings which in QM would be a triviality. The degrees of freedom of the theory to be embedded never go into a stringlike extension of localization, they rather form their own space which manifests itself as an internal infinite component space. A quantum mechanical construction does not know where it is going to live, the acting designer has to tell where he wants it. One can only restrict a higher dimensional QFT to lower dimensional part of spacetime. But physically this is generally not a very useful procedure because the lower dimensional theory obtained in this way will inevitably have too many phase-space degrees of freedom for being a candidate for a physical QFT in that lower dimension. The only exception to this rule is the holographic projection onto a null-surface resulting from a causal or event horizon. A AdS$_{5}$-CFT$_{4}$ correspondence or a restriction of a QFT to a brane is mathematically possible but only with an unphysical abundance of phase space degrees of freedom on the side with the lower dimensions or a unphysical anemia in the opposite direction. The unphysical consequences have been studied by people who have studied QFT beyond its Lagrangian confines since in that case it is impoetant to know which property guaranties the causal propagation (this is much more than the spacelike commutativity) and the existence of temperature states for arbitrary temperatures. In order to discuss problems of unitarity of Poincaré representations on inner symmetry space of chiral theories it is inconvenient to use the dual model setting. The reason is that even in case of the 10 dimensional superstring for which Hilbert space- as well as energy- positivity can be satisfied, the supersymmetric unitary representation is only obtained after passing to a subspace and dividing out zero norm-states. This blurrs the picture of the target spacetime resulting from an inner symmetry of a multicomponent potential of a current, and the presentation in terms of the bilinearized Nambu-Goto Lagrangian (see next section) is more convenient. As indicated before one replaces the superselected charges by superposable momenta and the potentials by operators $X_{i}(z)$ whose lowest Fourier mode (which includes a logarithmic contribution) $X_{i}(z)_{0}=x_{i}^{op}+icp_{i}^{op}\ln z$ defines quantum mechanical $x^{op},~p^{op}$ operators[^23]. In this way the inseparable Hilbert space which describes charged representations for a continuum of charges is avoided and the continuous direct sum becomes a quantum mechanical direct integral in the sense of spectral decomposition theory. Although the presence of these quantum mechanical degrees of freedom prevent the conformal covariance of the zero dimensional $X_{i}(z)$ field$,$ there is no problem with the covariance of the exponential vertex operators which carry an anomalous dimension proportional the square of charges which in the new reading corresponds to the square of momenta i.e. of masses$$d_{\psi}\sim\alpha\cdot\alpha,~d_{V}\sim p\cdot p=m^{2}$$ So in the Fubini et al. formalism [@Fu-Ve] Veneziano’s rather involved gamma function setting is replaced by a formalism using the conformal invariant part (the part which depends only on the anharmonic ratios) of the 4-point function of the vertex operator. The higher point function dual model amplitude results from the invariant part of the higher correlations; in this way one arrives at a dual model representation for $n\rightarrow m$ particle scattering. It is hard to criticize a proposal which is phenomenological in nature, apart from expressing some unease about putting together raw phenomenology ideas (which were later contradicted by new experiments) with subtle mathematical concepts which already had a different very precise conceptual position. It is probably the attractive mathematical aspect which explains why this proposal did not disappear completely together with the Regge phenomenology when the latter came to an end. Being a somewhat too ambitious setting for a mere phenomenological description, the theory had its later comeback in the form of string theory; but whereas its mathematical entitlement was natural, the same cannot be said about its physical interpretation. It finally became acclaimed as the millennium TOE which, different from the S-matrix bootstrap, allegedly also includes gravity. The conceptual distinction resulting from the of apparent uniqueness of mathematically ambitious projects as the implementation of the highly nonlinear duality structure has often mislead people[^24]. In the beginning there was only Veneziano’s version of the dual model which was constructed by a clever use of properties of gamma functions. But now we know that there are myriads of functions of the Mandelstam variables $s_{ij}$ which are meromorphic with an infinite tower of particle poles in the position of duality. They are constructed by starting from any conformal theory in any spacetime dimension. As explained in detail in a beautiful paper of Mack [@Mack], one only has to write the connected part of a conformal n-point function as a Mellin transform $M$ $$G^{c}(x_{1},...x_{n})=\left( \frac{1}{2\pi i}\right) ^{n/2}\int..\int d\delta M^{c}(\left\{ \delta_{ij}\right\} ){\displaystyle\prod\limits_{ij}} \Gamma(\delta_{ij})\left( -\frac{1}{2}x_{ij}\right) ^{-\delta_{ij}}$$ There are as many integration variables as there are independent conformal invariant anharmonic ratios. The aim is to show that by identifying the operator dimensions of the conformal fields with the masses of particles and the Mellin variables $\delta_{ij}$ to the Mandelstam variables $s_{ij}$ one obtains a meromorphic Mellin transform which has the correct poles as required by the duality property. The reduced Mellin transform $M^{c}$ can be defined in such a way that the spacetime dimensionality does not enter[^25] i.e. one can obtain dual models in a fixed spacetime dimension from conformal theories in *any* dimension, not only from chiral conformal theories. The convergence of the infinite sums over poles as well as certain positivity properties of the associated residues follow from the established validity of *global operator expansions in conformal theories* [@Mack]. At this level, there is however no claim that the Mandelstam variables are related to momenta on which a unitary representation of the Poincaré group acts. This problem was not part of the dual model program since the only positivity requirement in the Mandelstam setting of scattering amplitudes are conditions on the correct sign of residua of poles. It however became a pressing problem after the original phenomenological purpose of the formalism was abandoned and the setting was allowed to become the driving force of a free-roaming TOE under some mathematical, but practically no conceptual control. The rallying point for this development was the observation that the only unitary positive energy representation of a Poincaré group which can act on the index space of a multi-component current and its potentials or on the oscillator space of the Nambu-Goto model is the 10-dimensional superstring representation. In this case the Mandelstam invariants result from a unitary momentum space representation of the Poincaré group. In the present context the Mellin formalism demystifies Veneziano’s observation to some extent in that it shows that the duality structure, far from being a lucky discovery of a special way to implement (an approximated form of) the crossing property, is in reality a kinematical aspect of a certain transformation property of conformal correlation functions. Unlike the Fourier transform of correlation functions it cannot be expressed in terms of single operators but needs the entire correlation function for its definition[^26]. The operator version of the Veneziano dual model [@Fu-Ve], which starts from a chiral conformal current model, turns out to be a special case of Mack’s conformal Mellin transformation formalism. But whereas in the former the momenta enter explicitly via the continuous charge spectrum, the appearance of momenta in Mack’s setting is less overt; they only enter in parametrizing a relation which links the anomalous dimension of the conformal theory to the independent variable in the Mellin transform [^27]. As mentioned before the existence of a unitary positive energy representation of the Poincaré group behind the Mandelstam variables is not part of the Mellin transformation formalism. The verification of its existence in d=10 (the superstring theory) is certainly an unexpected curiosity since there was no reason at the beginning to expect a chiral conformal theory to support a noncompact inner symmetry as a Lorentz group representation start[^28]. But to take such a property of a two-dimensional conformal theory as a hint of leading to a new understanding about spacetime is far-fetched if not a step into mysticism. As mentioned before, the infinite component field of superstring theory in d=10 is the first and only nontrivial realization of a dynamic pointlike irreducible infinite component theory in the before explained sense [@Tod]. The protagonists of the infinite component field idea (if some of them are still around) would perhaps notice with satisfaction that by allowing quantum mechanical oscillators to connect the levels and to generate the mass/spin spectrum one obtains the first illustration of what they had in mind; perhaps they would have been less than happy about the high spacetime dimension of this unique realization and its resulting metaphoric epiphenomenon. A relation between masses and operator dimensions which is not related to Mellin transformation occurs in a more intrinsic physical context of the AdS-CFT correspondence. This correspondence will appear in a different context in the concluding remarks. Whatever one wants to make out of the operator setting of the dual model or the Mellin formalism, there is certainly no intrinsic physical reason why one should re-interpret charges as momenta and inner symmetry spaces of chiral theories as spacetime arenas for physical events. And why should one follow somebody who claims that the generating objects of ST are stringlike, in blatant contradiction to the pointlike computational results which lead to worldsheets on such an incorrect metaphoric path? The systematic construction of dual model amplitudes via conformal QFTs has nothing to do with the physical picture of an imagined approximation to the conjectured Mandelstam representation, nothing can conceptually be farther apart than scattering theory of particles and conformal QFT. Why mystify the different 10 dimensional pointlike superstrings and their presumed connection via M-theory as revealing deep secrets of physical spacetime when there is the autonomous possibility of explaining these properties as peculiarities of inner symmetries of chiral models which are known not to have to follow the standard inner symmetry pattern in terms of compact group representation of higher dimensional symmetries? Behind all this is the general question: is particle physics only interesting after, following the modern Zeitgeist, it has been sexed up or mystified? String theory, a TOE or a tower of Babel within particle theory? ================================================================ String theory addresses some of the questions which the dual model left open or could not handle convincingly as: can one really obtain a unitary representation of the Poincaré group on the internal symmetry space of a chiral current theory and if yes, what is the covariant localization concept in such a source-target relation and in particular does it really lead, as claimed, to a notion of world sheets? Last not least one would like to know whether the use of special exponentials of potentials (in the operator duality approach) can be replaced by a more general setting in which, similar to the Wigner approach to particles, a representation space is defined in terms of generating wave functions with clear localization properties, which are then used to pass to an (interaction-free) operator field formalism. For this purpose it has turned out to be covenient to start from a slightly more general point of view which prepares the desired unitary representation theory more directly in terms of the current potentials $X^{\mu}(z).$ But before going into these technicalities some general remarks are in order. There exist operator algebras and state spaces which have no pointlike but rather semiinfinite string-like generators; Wigner’s massless *infinite spin* representation family presents the only noninteracting illustration [@MSY] and it shows that string localization is incompatible with a Lagrangian description. In this case one may speak of world sheets being traced out in spacetime. But the generating wave function of string theory and their second quantized counterparts are pointlike generated. Originally the string world sheets were not part of the dual model of old, they appeared in a later stage when it was incorrectly claimed that the source-target relation can be understood as an embedding of the one-dimensional chiral theory as a one-dimensional submanifold into a 10-component target space representing spacetime. To support such a picture string theorists invented a Lagrangian description of relativistic particles [@Pol]. Compared with Wigner’s clear representation-theoretical classification, the functional integral representation in terms of relativistic particle mechanics falls short of a convincing attempt to support string theory; it is mathematically ill-defined[^29], does not describe all irreducible positive energy representations, and was never used by particle physicists who characterize particles following Wigner. Such ad hoc inventions whose only purpose is just to make one point in an analogy sometimes backfire instead of lending support. Before going into ST details, it is helpful to start with a theorem from unitary representation theory which limits the localization of states (appendix). The causal localization (modular localization, see appendix) inherent in unitary positive energy representations of the covering of the Poincaré group is pointlike generated apart from Wigner’s massless infinite spin representation whose optimally localized generators are semiinfinite spacelike strings [@MSY]. Some comments are in order. Unitary positive energy representations are canonically related to free fields or (in case of reducible representations) to direct sums of free fields. The bilinear Nambu-Goto Lagrangian is interaction free, hence the localization is completely determined by the representation theoretical content. One only has to show the absense of the Wigner infinite spin representation from the positive energy unitary 10-dimensional superstring representation (which is obvious) in order to secure that it is pointlike generated where pointlike generated means that there is a collection of infinite component singular functions[^30] (wave function-valued distributions) $\psi(x)$ whose smearing with test functions generate the one string space. In the Fock space extension this corresponds to a collection of infinite component pointlike fields whose one-string projection leads to the singular wave functions. This theorem also covers the localization in string theory, since the Lagrangian which underlies the quantum string is bilinear and hence the (graded) commutator must be a c-number. This Lagrangian supplies the operator formalism acting in the Hilbert space of the string wave functions. This one-string representation space is an analog of the Wigner one particle space apart from the fact that there is a severe restriction from the unitarity of the action of the Poincaré group. This is because the central issue is the quantization of a Lagrangian and the unitarity problem is an additional restriction The situation resembles vaguely that of the vector potentials in QED in that one has to form sub- and factor- spaces in order to get rid of the negative and zero norm states. But whereas in QED this idea is independent of the spacetime dimension and certainly does not effect the noninteracting theory (where it only appears if one uses potentials instead of field strengths), the origin in string theory is quite different. It can be traced back to the unmotivated (i.e. not physically justifiable) demand that one wants a *unitary positive energy representation of the covering of the Poincaré group on an internal symmetry space of quantum theory.* . Nature could have answered this extravagant requirement by providing the same negative response which has been known in higher dimensional QFT namely: any inner symmetry is necessarily described by a compact group; noncompact groups as spacetime symmetries would be in contradiction with the localization principles of LQP [@Haag]. But surprisingly there are exceptions in chiral QFT where besides “rational” models (which are in many ways similar to the inner symmetry structure of higher dimensional models) and models with countably many superselection sectors, there are also quite different “irrational” internal symmetries. Models in which the observable algebras are defined by multicomponent abelian currents belong to the latter. They have a continuum of charged representations and there is indeed the possibility to have (in an appropriate sense) a positive energy representation of the covering of the Poincaré group on a 10 dimensional internal symmetry space of a chiral current model. But from the context in which this somewhat surprising observation arises it is clear that it has nothing to do with a new mysterious insight into foundational problems of spacetime but rather with an unexpected property of the particular chiral model (other surprising properties of maximal extended current algebras were mentioned in the previous section). Whatever one’s position is towards spacetime symmetries appearing on the inner symmetry space of chiral currents, there can be no doubt about the fact that the one string space (or the uniquely associated string string field theory) is pointlike generated. This is the unavoidable conclusion from the previously stated theorem as well as from the below mentioned concrete calculations. At this point it is important not to equate the localization of states with that of operators beyond the setting of free fields. Whereas only the family of massless infinite spin Wigner representations is semiinfinite string-like generated [@BGL][@MSY], the absence of pointlike algebraic generators in certain charged subalgebras is quite common. The best known case is that of electrically charged fields in QED [@Jor], it is impossible to localize a charge-carrying operator in a compact spacetime region. Within massive theories the possibility of such a situation was investigated by Buchholz and Fredenhagen [@Bu-Fr], but since in this case there would be no infrared manifestation of string localization in Lagrangian perturbation theory, there are presently no illustrative models of string localization in theories with mass gaps [^31]. A B-F stringlike or an electrically charged field applied to the vacuum decomposes into pointlike generated wave functions, but this decomposition process has no counterpart in the local algebras. By leaving the issue of localization in string theory to be settled as a special consequence of a powerful structural theorem in local quantum physics as above, one deprives oneself of some interesting insight into one of the most fascinating episodes in 20th century particle physics namely a more detailed understanding of where did the arguments leading up to string theory fail. For this reason we will now follow this more interesting path. The formal starting point is the bilinear Lagranian form in which the Nambu-Goto Lagrangian [@Nambu][@Goto] is used in string theory$$\begin{aligned} L & ={\displaystyle\iint} (\partial_{\tau}X_{\mu}\partial_{\tau}X^{\mu}-\partial_{\sigma}X_{\mu}\partial_{\sigma}X^{\mu})d\tau d\sigma\\ & \left( \partial_{\tau}^{2}-\partial_{\sigma}^{2}\right) X^{\mu }(z)=0\nonumber\end{aligned}$$ In the simplest case the $\tau,\sigma$ dependent “zero scale dimension position field” $X_{\mu}(\tau,\sigma)$ (the string analog of the Fubini... potential) is considered to be defined on $R\times(0,\pi)$ with appropriate (Neumann) boundary conditions. The equation of motion is a two-dimensional wave equation which together with the boundary conditions leads to the Fourier representation $$X^{\mu}(\tau,\sigma)=x^{\mu}+p^{\mu}\tau+i\sum_{n\neq0}\alpha_{n}^{\mu }e^{-in\tau}\frac{\cos n\sigma}{n}$$ The $\alpha_{n}^{\mu}$ are oscillator-type creation and annihilation operators which by Lorentz covariance are forced to act in an indefinite metric space. Denoting these chiral current potentials by $X^{\mu}$ may create the delusion that one is describing a path in target space; with the conservative notation $\Phi^{\mu}$ such an association is less automatic. In the present form there is yet no free go for a unitary Poincaré on target space, such a move must be more carefully prepared. Imposing subsidary conditions$$\left( \partial_{\sigma}X\pm\partial_{\tau}X\right) ^{2}=0$$ does the job, after they have been adjusted to the quantum setting (valid only on states). The Klein Gordon equation on target space with a mass operator of an integral spaced spectrum. These conditions express reparametrization invariance and they would have been a consequence of the true Nambu-Goto Lagrangian which is a nonlinear expression in the $\partial X.$ Clasically the true N-G Lagrangian is equivalent to its bilinearized version plus constraints. The reparametrization invariance trivializes part of the infinite dimensional conformal covariance. All these aspects are subordinated to the construction of a unitary Poincaré group representation on the appropriately defined target space of a multicomponent current potential; they do not have any intrinsic physical meaning. Any unitary representation of the Poincaré group acts in a Hilbert space can be obtained by a two-step process from a formal covariant representation in a negative metric space of the form $$H_{sub}\subset L^{2}(\mathbb{R}^{n},\rho(\kappa)d\kappa d\mu(p,\kappa))\otimes H_{QM}$$ where the first factor is a spinless relativistic particle representation space with a continuous mass distribution and $H_{QM}$ contains vector-valued or spinor-valued quantum mechanical variables (as the $\alpha_{n}^{\mu}$) which strictly speaking are prevented by Lorentz covariance to be genuine “quantum” (acting in a Hilbert space). In the simplest case of finite dimensional massive representations, the representation space is the n-dimensional (nonunitary) vector representation space of the Lorentz group $H_{QM}=V^{(n)}.$ To get to a unitary massive $s=1~$representation of the Poincaré group one uses Wigner’s idea of the little group and obtains a *unitary p-dependent Lorentz transformation law* which results from the original non-unitary covariant law through an *intertwiner* (a 4-component function on the forward mass shell) between the original n=4 vector representation with and its manifestly unitary form which acts covariantly on a positive metric subspace $H_{phys}=H_{sub}\subset L^{2}(R^{n})\times V^{(n)}.$ In the N-G case at hand the selection of the mass specrum is done by imposing the Klein-Gordon equation with the mass operator, its spectrum then leads to a direct sum over equally spaced mass eigenstates including a “lowest” tachyonic contribution$$\sum_{\kappa=-2,0,2,...}L^{2}(\mathbb{R}^{n},d\mu(p,\kappa))\otimes H_{Osc}$$ And one has to still implement the complete set of subsidery conditions. For this purpose one uses the vector-valued oscillators belonging to the higher Fourier components of the current potential whose Lorentz invariant inner product is indefinite. There is no chance to find a subspace through subsidiary conditions which is positive semidefinite with one exception. Only for the multicurrent model with 26 components does one arrive at a semidefinite metric [@Brower]. The last step is canonical, having arrived at a semidefinite situation, the positive definite situation is gratis. Details can be found in many articles [@Dim]. The obtained 26 dimensional representation is not of positive energy as a result of the presence of a tachyon. However admitting spinorial-valued chiral current components (which would require a spinorial change of the N-G Lagrangian) one arrives at the 10 dimensional positive energy superstring representation. The transition from unitary to covariant representations is done with the help of so called $u,v$ intertwiners. This step is in complete analogy to what Weinberg presented in a group theoretic setting in the first volume of his well-known textbook [@Weinberg]. It also can be obtained by applying modular localization to the Wigner representation theory [@MSY] (see appendix). Although for each irreducible unitary representation there is only one wave function space, there are infinitely many different looking covariant wave functions and free fields (see appendix). Since a unitary representation of (necessarily noncompact) spacetime symmetry group on an internal symmetry space of a current algebra is a strange requirement from a viewpoint of local quantum physics[^32], it would be very natural to have received a negative answer to the target space issue. But inner symmetries in low dimensional QFT are different from their standard realization and lo and behold there is precisely one exception namely the positive energy “superstring” representation in 10 spacetime dimension. But does the existence of this exception indicate some mysterious new insight into spacetime? Certainly not, it does however reveal some unexpected property of the potentials $\Phi^{\mu}$ (and their charge-carrying exponentials) of multicomponent chiral currents. Actually the solution is not completely unique since there is a finite number of 10 dimensional superstrings and there exists even a conjecture (M-theory) about their possible relations. It would be interesting to present these observations (in analogy to Mack’s Mellin formalism) solely in terms of multicomponent currents and their potentials, leaving spacetime metaphors aside. The correct reading of the string as a dynamic infinite component field[^33] shares the inner symmetry$~\rightarrow~$spacetime symmetry reinterpretation with that of string theory. But there is less temptation to elevate the construction of a (possibly unique) dynamic infinite component field to a new foundational insight into spacetime or to interpret its near unicity as the indicating a TOE. Every correct investigation of localization by string theorists led to the pointlike result. The safest calculation is that via the commutator of two string fields. All these calculations led to one result: a pointlike localized spacelike (graded) c-number commutator, whose explicit form still depends on the choice of the internal part (the “vertically” acting oscillators) of the smearing function [@Dim]. With other words the infinite mass/spin tower spectrum is a general characteristic property of the theory, but the strength with which these levels contribute to a particular point-localized wave function or second quantized field analog can be manipulated with operators acting between the levels. String theorist in their correct calculations of (graded) commutators of string fields of course do not obtain anything but a pointlike localization [@Mar][@Lowe]. But being members of a globalized community of string theory they do not present their result in this form. Their subconcious desire to serve the community course leads them to describe a conceptual chimera: some sort of extended but at the same time hidden object, a kind of invisible string of which only the c.m. point is sticks out. This shows to what extend the critical power, which was keeping particle theory strong during healthy times, was lost in the sociology of globalized communities and their guru like leaders This phenomenon is one of the strangest I ever came across. Remembering that the conquest of quantum theory is inexorably linked with a clear exposition of quantum reality and localization in particular, one wonders why string theory leads people to mystical regressions. The cited papers constitute an interesting historical document for a time in which clear calculations could not prevent their metaphoric interpretation. The tower of Babel in particle theory is erected on the difference between computations and prevailing ideology. It is of course important that the calculations are correct, and it is not plausibe that the interpretation which fails to match the calculation was distorted on purpose. The tower of Babel effect is rather the result of the Zeitgeist in the service of a dominating TOE. Perhaps the path into a self-defeating metaphoric world started already with such innocent looking choice of notation which feigns string-like target space localization as writing $X^{\mu}(\tau)$ for the current potentials $\Phi^{\mu }(\tau)$ or even before by introducing the notion of target space (field space) which has no place in QFT and is at best a metaphor for “arena of action for inner symmetries”. With the loss of conceptual knowledge about local quantum physics, the idea of a stringlike target space localization may have received a helping hand from an unlucky notation which could have exacerbated already present misunderstandings. String theory unlike QFT has no built-in operational way of introducing interactions. Whereas the spacetime principles underlying QFT are strong enough to not only determine the form of interactions consistent with the locality principle but also to rigorously derive scattering theory, all these ideas of deriving global properties from local principles are lost in a pure S-matrix approach. Its principles of unitarity, Poincaré invariance and possibly crossing are the only guides and every additionally imposed structure has to justify itself a posteriori by its phenomenological success. Hence it is not surprising that interactions are defined “by hand” via highlighting certain operators which already played that role in the dual model. Being deprived of large time asymptots which relates the S-matrix with a Lagrangian via interpolating fields, string theorists simply define the lowest order (tree approximation) of the string S-matrix by functional formulas which are equivalent to the the Fubini-Virasoro exponential expressions. Already in the setting of the dual model, attempts were made to find reasonably looking recipes to imitate the loop corrections of QFT by adapting Feynman’s rules for world lines to world sheets. String theorists introduced computational recipes in form of graphical descriptions in terms of rules for combining and splitting tubes which are supposed to represent the world sheet traced out by closed strings, but what does this mean for pointlike objects whose spacetime string extension is metaphoric and not real? Whereas such recipes in QFT can be shown to be a graphical illustration of operator relations, their quantum meaning in string theory remain unclear. The characteristic feature of a relation formula or idea in quantum theory is that it can be expressed in terms of operators and states; if this is not the case it is not part of QT, this was the leimotiv of Bohr and Heisenberg which led them to the notion of “observables”. Despite a search over more than 4 decades for an operator formulation behind those recipes for perturbative string S-matrix amplitudes by the best minds in the string community, no such quantum theoretical formulation was ever found. In this context it is interesting to remind oneself that Stuekelberg discovered Feynman rules precisely in this graphical recipe form. In his studies of macrocausality properties of an S-matrix he realized that, whereas the spacelike macrocausality amounts to the cluster factorization of the S-matrix, there was a finer macrocausality property for asymptotic timelike separation. A 3$\rightarrow$3 particle scattering for example should contain the possibility that first 2 particles interact in form of a 2-particle scattering and afterwards the third particle enters the causal future of the first process and meets and interacts with one of the outgoing particles. He showed that the timelike trajectory between the two local scattering centers corresponds to a propagator with (what later became known as) Feynman’s $\varepsilon$-prescription, expressing the fact that the second interaction happened later. By assuming that interaction regions can be idealized as pointlike vertices he obtained the Feynman rules. Of course nobody, including himself, paid much attention to such an ad hoc recipe. The general acceptance came only with the derivation in terms of operators and states which started with Feynman and found its most concise expression in the work of Dyson. In the string case not only is there no operator formulation for the world sheet picture, such a formulation would create a clash with the pointlike nature of the free string. There remains of course the possibility that an infinite collection of pointlike fields offers a new kind of pointlike interaction which has no counterpart in the standard setting of polynomial (possibly infinite degree) interactions. But even if such a possibility exists, any quantum interaction must allow a formulation beyond recipes and prescriptions in terms of the quantum setting of operators and states. String theory, either in its factual infinite component pointlike setting, or its metaphoric guise of a “invisible string” is markedly different from (finite component) QFT if it comes to the notion of *degrees of freedom*. QFT has more phase space degrees of freedom than QM; whereas in QM there is a finite number of degrees of freedom in a finite phase space volume, the cardinality in QFT is described by a mild form of infinity (the compactness or nuclearity property of QFT [@Haag]). This is precisely what guaranties the existence of thermal states at any temperature and the causal shadow property which states that the algebra of a spacetime region equals that of its causal completion [@Swie] (the quantum counterpart of the Cauchy wave propagation). Both properties are lost in string theory. In view of its importance for the problem of holographic relations of QFTs in different dimensions this issue will reappear (last section) in a different context. Modular localization, the KMS condition and the crossing property =================================================================  In order to become aware of the significant conceptual differences between the crossing property and duality it is necessary to have a profound understanding of crossing. In the sequel I will for the first time present some recent insight on this problem within the setting of modular localization theory (appendix). The important concept from modular theory which relates to the crossing property is *localization equivalence with respect to the wedge W spacetime region[^34] (*which will be denoted denote by ** $\overset {W}{\sim})\ $ ** between operators affiliated to ($\prec$) different wedge algebras $\mathcal{A}(W)$ and $\mathcal{B}(W)$ which live in the same Hilbert space and share the same positive energy representation of the Poincaré group.$$B\overset{W}{\sim}A:B\Omega=A\Omega,~~B\prec\mathcal{B}(W),~A\prec \mathcal{A}(W) \label{bi}$$ Since under such conditions modular theory identifies the dense subspaces generated by applying the two wedge algebras[^35] to the vacuum, it brings about a one to one relation between generally unbounded operators which does not respect the algebraic multiplication structure. Hence the $\overset{W}{\sim}$ relation is a *bijection* between the individual operators affiliated to two wedge-localized operator algebras which both live in the same Hilbert space and share the same unitary representation of the Poincaré group, but may be very nonlocal relative to each other. The situation which is relevant for the derivation of crossing is that in which $\mathcal{B}(W)$ is the wedge-localized algebra from an interacting net of local algebras which admits a complete asymptotic interpretation and $\mathcal{A}(W)=\mathcal{A}_{in}(W)$ is the wedge algebra generated by its asymptotic incoming fields. It is convenient for the following to introduce a flexible notation. If we want to refer everything to the algebra $\mathcal{A}$ we will use the notation $B_{\mathcal{A}}$ when we want to substitute a $B\in\mathcal{B}$ by its bijectively related operator (\[bi\]) $B_{\mathcal{A}}\in\mathcal{A};$ conversely we write $A_{\mathcal{B}}$ if our aim is the characterization of an operator in the algebra $B$ which is bijectively related to $A\in\mathcal{A}.$ Returning to our situation of interest of an asymptotically complete theory and $\mathcal{B}=\mathcal{B}(W),$ $\mathcal{A}=\mathcal{A}_{in}(W)$ we can picture a $A_{\mathcal{B(}W\mathcal{)}}$ in a more concrete fashion: it is an operator in $\mathcal{B}(W)~$whose creation component (the one involving only creation operators $a^{\ast}(p)$) is identical to the creation component of $A\in\mathcal{A}(W).$ The other components are uniquely determined by the requirement that the operator belongs to a particular W-localized algebra. The underlying idea resembles in some sense the algebraic notion of *relatively local fields* which led to the concept of Borchers equivalence class [@STW]. But since there is no furthergoing algebraic connection beyond the bijection between the operators of two local algebras which only share the same localized states and the same representation of the Poincaré group (and as a consequence, the same Reeh-Schlieder subspace [@Haag]), the two algebras may be quite different in the algebraic sense as exemplified by an interacting wedge algebra and its (via scattering theory) associated asymptotic incoming algebra. The existence of this bijection is a straightforward generalization of an argument about modular theory in [@BBS]. In that work the interacting representation of a wedge-localized one-particle state[^36] was considered. Such vacuum polarization-free objects are not available in interacting theories for compact localization region, in fact the wedge region is in passing from compact to non-compact localized causally closed spacetime regions the first for which such interacting one-particle generators exist. In a more intuitive formulation: *wedge regions lead to the best compromise between particles and fields* in the presence of interactions. Only algebras generated by free fields have vacuum polarization free generators for *any* localization region. Hence localization-caused vacuum polarization clouds offer an autonomous criterion for the presence of interactions. In a journal on foundations of physics it may be appropriate to mention that these dense subspaces have attracted the attention of renown philosophical and foundational motivated physicists. [@Cl-Hal][@Ea-Rue][@Fra][@Rob][@Summers]; in fact this has been a small window of intense communication between physics and philosophy to which the critical remarks in the introduction do not apply. In fact the existence of these subspaces was a surprise for those who obtained her/his physical intuition from QM; they constitute one of the most characteristic features of QFT. Although the domains $domS$ only depend on the representation of the Poincaré group (the mass/spin spectrum), the way how the different $S$ act on this domain carries the informations about the interaction (appendix). In the case the algebras generated by the cyclically acting fields are identical $\mathcal{A}(W)=\mathcal{B}(W),$the bijection $\overset{W}{\sim}$ leads back to the trivial relation $A=B.~$Hence it is a generalization of the algebraic notion of local equivalence which is closely related to the notion of the *Borchers class* of relative local fields. Both concepts are also related (but not identical) to *weak locality [@STW]*. The bijection concept comes with a prize. If the operator $A(f)\prec \mathcal{A}_{in}(W)$  is a $f~$- smeared covariant field with $suppf\subset W,~$having the standard Wightman domain properties, the existence of $B^{\prime}s$ is paid for by unwieldy domain properties. Although acting on the vacuum they do induce the same dense space of states; their domain properties are weaker than those of smeared Wightman fields. Their generally smaller domain is not translational invariant i.e. the translated domain of an operator $B\prec\mathcal{B}(W)$ is outside $domB~$[@BBS]$.$ The translation invariance of the domain (the *temperateness* of $B$) would imply $S=\mathbf{I}$ if $d>1+1,$ whereas in case d=1+1 the model has only elastic scattering [@BBS]. This shows that modular theory does not only reveal deep connections between spacetime geometry and the mathematics of operator algebras, but also sheds new light on connections between domain properties of unbounded operators and the presence of interactions. For $S_{scat}=\mathbf{1}$ to occur it is enough if such a Poincaré invariant dense domain exists for a particular $B(f)\prec\mathcal{B(W)}$ which is in bijection with $A_{in}(f),$ supp$f\in W$ $$B(f)\Omega=A_{in}(f)\Omega$$ i.e. a temperate B which **g**enerates a vacuum **p**olarization **f**ree *one particle state* (such a B is called a **PFG**[^37]) [@BBS]. The triviality of the scattering matrix $S_{scat}=1,$ and therefore the equality of the Tomita operator $S_{Tomita}=S_{free}$ with that of a free field follows (as long as one avoids low dimensions ($d=1+1$) and and 3-dimensional models with plektonic statistic). The interesting question to what extend this implies the absence of interaction in the stronger sense of $\mathcal{B=A}_{in}$ will be commented on later. The case of factorizing models, for which the S-matrix is nontrivial but has a rather simple structure, will be presented in detail in the next section. The important relation which leads to the derivation of the crossing property is [@BBS]$$\begin{aligned} B\Omega & =\Phi,\text{ }\Phi=A\Omega,~i.e.~\Phi\in domS_{\mathcal{A}},~B\in B(W)\label{rep}\\ & \curvearrowright B^{\ast}\Omega=S_{\mathcal{B}}\Phi\nonumber\end{aligned}$$ This is a formula for the computation of the action of the conjugate of an operator on the vacuum if the operator itself is unknown except that its action on the vacuum should result in a state vector $\Phi$ which has no direct relation to the $\mathcal{B}(W)$ algebra apart from its membership to the space $domS_{\mathcal{A}}=domS_{\mathcal{B}}$. The theorem tells one how to compute $B^{\ast}\Omega$ from these data. These prerequisites are always met if the two algebras share the same representation of the Poincaré group i.e. have the same mass/spin particle content. The crossing relation in its simplest field theoretic formulation (selfconjugate spinless fields, only incoming fields in the uncrossed configuration) reads$$\left\langle B(A_{in}^{(1)})_{\mathcal{B}}A_{in}^{(2)}\right\rangle =\left\langle A_{out}^{(2)}{}\Delta BA_{in}^{(1)}\right\rangle \label{orig}$$ It is important to note that $A_{in}^{(1)}$ on the left hand side appears as its bijectively related counterpart $(A_{in}^{(1)})_{\mathcal{B}}$ which off vacuum represents a different operator. $A_{in}^{(1)}$ and $A_{in}^{(2)}$ may be products of W-smeared free fields$$A_{in}^{(1)}=:A_{in}(g_{1}),..A_{in}(g_{k}):,~~A_{in}^{(2)}=:A_{in}(f_{1})...A_{in}(f_{l}):$$ The proof of this relation is reminiscent of the modular derivation of the KMS relation, in fact if $A_{in}^{(1)}=1=(A_{in}^{(1)})_{\mathcal{B}}$ it reduces to the above extension of the KMS property. In the present case there are not only operators from different algebras to start with, but the action of $S_{B}$ on $A_{in}^{(2)}$ brings a third algebra into the game namely $\mathcal{A}_{out}(W).$ As in the case of KMS, the presence of the unbounded analytically continued operator $\Delta$ leads precisely to the same analytic properties as those found by Araki. The formfactor crossing$$\left\langle B|p_{1}..p_{k}q_{1},.q_{l}\right\rangle _{in}=~_{out}^{a.c.}\left\langle -q_{1}..-q_{l}|B|p_{1}...p_{k}\right\rangle _{in}^{c.o} \label{cro}$$ results from the previous field theoretic crossing if one takes instead the over all Wick product $:(A_{in}^{(1)})_{\mathcal{B}}A_{in}^{(2)}:.$ But what are the precise conditions under which the subsript ()$_{\mathcal{B}}$ can be omitted? This will be explained below. The W-localized wave functions which were still present in the field theoretic crossing (\[orig\]) have been removed in (\[cro\]) as the result of their on-shell denseness so that the crossing identity only involves momenta[^38]. The notation is as follows, the $a.c.$ refers to the analytic continuation from the positive mass shell to the backward shell (using the momentum space analyticity of wedge localized mass-shell reduced test functions) and the $c.o.$ indicates the omission of contractions between the $p^{\prime}s$ and $q^{\prime}s$ which reflects the fact that the l+k particle state on the right hand results from a Wick-ordered product of in-fields; since there are no contractions between in-particle on the right hand side, there can be none after crossing to the left hand side either. The negative momenta $-q$ are a result of the combined action of $S^{\ast}=\Delta^{\frac{1}{2}}J$ where $\Delta^{\frac{1}{2}}$ has the geometric interpretation of an imaginary $\pi$ Lorentz rotation $\Delta ^{\frac{1}{2}}=e^{-\pi iK};$ these particles are outgoing and they would be antiparticles in case the particles carry a superselected charge (the fields are not selfadjoint). For the proof one uses the formula (\[rep\]) (first line, second equation) $$\begin{aligned} \left\langle B(A_{in}^{(1)})_{B}A_{in}^{(2)}\right\rangle & =((B(A_{in}^{(1)})_{\mathcal{B}})^{\ast}\Omega,A_{in}^{(2)}\Omega)=(S_{\mathcal{B}}(BA_{in}^{(1)})\Omega,S_{\mathcal{B}}A_{out}^{(2)\ast}\Omega)=\label{rel}\\ & =\left\langle A_{out}^{(2)}\Delta BA_{in}^{(1)}\right\rangle \nonumber\end{aligned}$$ where in the the last line the antilinearity of $S_{\mathcal{B}}$ as well as the relation $S^{\ast}S=\Delta$ was used.  Apart from the involvment of different algebras, the derivation of the crossing relation resembles strongly the modular derivation of the KMS property of localized algebras and may be seen as a *generalization of the KMS setting*. In fact the relation (\[rel\]) is a KMS relation in the presence of two different wedge-localized algebras$~\mathcal{B}(W),\mathcal{A}(W)$ which share the same representation of the modular group (the Lorentz boost). In the case at hand, the joint modular group results from the sharing of the same Poincaré group representation between the interacting theory and its asymptotes. It is a special case of an *extended KMS relation for two algebras* which are standard with respect to the same vector state and have the same modular group but different modular reflections $J_{\mathcal{A}}\neq J_{\mathcal{B}}$ $$\left\langle BA\right\rangle =\left\langle ((A^{\ast})_{\mathcal{B}})^{\ast }\Delta B\right\rangle ,~(A_{\mathcal{B}})^{\ast}\Omega\equiv S_{\mathcal{B}}A\Omega$$ with $S_{\mathcal{B}}$ being the modular Tomita operator for the algebra $\mathcal{B}(W)$[^39], the two-algebra generalization of the KMS situation evidently reduces to the one-algebra case for $\mathcal{B}(W)=\mathcal{A}(W)~$and hence$~A_{\mathcal{B}}=A_{\mathcal{A}}\equiv A.$ In the interacting case only the creation components coalesce; the statement that all other (model-dependent) contribution to the formfactor contain momentum space delta functions and are removed under the c.o operation will be deferred to a separate publication[^40]. This two-algebra extension of KMS offers a wealth of new applications whose presentation would go beyond the theme of this essay. The formfactor relation which follows from (\[rel\]) has the follwing form in momentum rapidity space$$\left\langle B\left( a^{\ast}(\theta_{1})..a^{\ast}(\theta_{k})\right) _{\mathcal{B}}a^{\ast}(\vartheta_{1})..a^{\ast}(\vartheta_{l})\right\rangle =\left\langle a(\vartheta_{1})..a(\vartheta_{l})Ba^{\ast}(\theta_{1})..a^{\ast}(\theta_{k})\right\rangle$$ For simplicity of notation we specialized to two spacetime dimensions (omission of transverse momenta components) and assumed that the particles are selfconjugate. Wick-products of Bose fields are symmetric in their arguments and this symmetry property is inherited by the (..)$_{\mathcal{B}}.$ It is custumary in formfactor constructions to get rid of this redundancy and make the convention that n-particle states are always $\theta-$ordered i.e. $$Ta^{\ast}(\theta_{1})..a^{\ast}(\theta_{k})=a^{\ast}(\theta_{i_{1}})..a^{\ast }(\theta_{i_{k}})~~if\text{ }\theta_{i_{1}}>\theta_{i_{2}}>...>\theta_{i_{k}}$$ But since $a^{\ast}(\theta)_{\mathcal{B}}\equiv b^{\ast}(\theta)$ is generally a very complicated operator whose only simplicity consists in that it defines a wedge-localized PFG (i.e. its application to the vacuum created a one-particle state without vacuum polarization), the problem is to define a symmetric product which agrees for ordered arguments with the operator product and which does not require any special commutation relations as those which characterize free fields. The only answer is the *theta-ordered product[^41]*. One then obtains the desired composition rule$$\begin{aligned} (Tb^{\ast}(\theta_{1})..b^{\ast}(\theta_{k}))a^{\ast}(\theta_{1})..a^{\ast }(\theta_{k})\Omega & =a^{\ast}(\theta_{1})..a^{\ast}(\theta_{k})a^{\ast }(\vartheta_{1})..a^{\ast}(\vartheta_{l})\Omega\\ if\text{ }\theta_{i} & >\vartheta_{j}~\forall~i,j\text{ }\nonumber\end{aligned}$$ i.e. the interacting ($a^{\ast}(\theta_{1})..a^{\ast}(\theta_{k}))_{\mathcal{B}}\equiv$ $Tb^{\ast}(\theta_{1})..b^{\ast}(\theta_{k})$ operators in our bijection act as products of interaction-free operators only in special orderings between T-ordered clusters. Only in the well-studied case of factorizing models the $b$-operators have simple properties, namely they have a translation invariant domain and obey Zamolodchikov-Faddeev commutation relations (see next section). In that case one also knows the analytic continuation properties under exchange of rapidities which are not part of the crossing relation (which only refers to the cyclic KMS permutations). More on the properties of these bijectively related operators will be contained in forthcoming joint work [@M-S]. Since the analyticity properties result from the domain properties of $\Delta,$ it is helpful to remind the readers of the standard analytic KMS properties as Araki [@Araki] first established them. Let $\mathfrak{C}$ be a C$^{\ast}$algebra on which $\alpha_{t}$ acts as a one parameter automorphism group. Then $\omega$ is called a KMS state with respect to $\alpha_{t}$ at temperature $\beta>0$ if for each pair of operators $A,B$ $\in~\mathfrak{C}$ there exists a function $F_{A,B}(z),$ analytic on the open strip $\left\{ z\in C,~0<\operatorname{Im}z<\beta\right\} ,$ continuous and bounded on its closure, such that$$F_{A,B}(t)=\omega_{\beta}(Aa_{t}(B)),~~F_{A,B}(t+i\beta)=\omega_{\beta}(a_{t}(B)A) \label{KMS}$$ Araki showed that the n-point correlation functions in a KMS state are boundary values of analytic functions in the strip $C_{\beta,<}^{(n)}$ given by $$\begin{aligned} & \omega_{\beta}(\alpha_{t_{1}}(B_{1})....\alpha_{t_{n}}(B_{n}))=\lim _{\operatorname{Im}z\rightarrow0}\omega_{\beta}(\alpha_{z_{1}}(B_{1})....\alpha_{z_{n}}(B_{n}))\label{Araki}\\ & C_{\beta,<}^{(n)}:0<\operatorname{Im}z_{1}<.....<\operatorname{Im}z_{n}<\beta\nonumber\end{aligned}$$  and $\omega_{\beta}$ exists under rather general conditions for all $\beta>0.$ There are similar analytic properties of KMS states which come with only half the strip region [@Araki]$$\begin{aligned} & \alpha_{t_{1}}(B_{1})....\alpha_{t_{n}}(B_{n})\Omega_{KMS}=\lim _{\operatorname{Im}z\rightarrow0}\alpha_{z_{1}}(B_{1})....\alpha_{z_{n}}(B_{n})\Omega_{KMS}\\ & C_{\beta/2,<}^{(n)}:0<\operatorname{Im}z_{1}<.....<\operatorname{Im}z_{n}<\beta/2\nonumber\end{aligned}$$ Note that there is no statement on whether different orderings can be related by analytic continuation; in general this is not possible. In the case of Wightman functions however this follows from spacelike (graded) commutativity, and for the so called temperate PFGs of d=1+1 factorizing theories this is a consequence of the Zamolodchikov-Faddeev commutation relations for generators of wedge localized algebras (next section). In the case at hand the thermal aspect does not arise in the standard way i.e. by subjecting a global algebra of QFT to a heat bath which converts its ground state into a KMS state. It rather originates by restricting a global vacuum state to a wedge-localized subalgebra. With the conventions from modular theory we have$$\begin{aligned} \Delta^{it} & =~e^{-2\pi itK},~K=generator~of\ W-Lorentz~boost\\ \beta_{\operatorname{mod}} & =-1~corresponds~to~\beta=2\pi\nonumber\end{aligned}$$ Whereas the Hamiltonian and the temperature $kT=\beta^{-1}$ are dimensionful quantities, the modular temperature and the modular Hamiltonian are dimensionless since they arise in a geometric context. This raises a very fundamental question: is the KMS analytic aspect of crossing with real thermal physics only a parallelism in the mathematical formalism or does it extend to the physical content. The question how the basic quantities of a heat bath situation, as energy and entropy, are related to their counterparts arising from localization is a fundamental problem of quantum theory, in view of its astrophysical applications perhaps the most fundamental problem of our times. Although its understanding does not contribute anything directly to the crossing property, some general comments on such a pivotal problem are in order. Both problems are related to KMS states on the same algebra namely the hyperfinite type III$_{1}~$factor algebra. In [@BMS] the reader finds rather tight arguments that the thermodynamic infinite volume limit of a heat bath system corresponds to a certain “funnel” approximation of a localized algebra by a family of slightly larger algebras defined in terms of “the split property” in such a way that the divergent volume limit for the entropy can be placed in direct correspondence with a logarithmically corrected divergent area law. It is to be expected that such a method, even if ingeniously applied as in [@BEG], is too bulky for a general solution of the crossing problem, in particular in view of the fact that it does not point to the relevant physical setting (KMS from localization). Its exploration came to an end already in the 70s after Källén and Wightman tried for many years in vain to derive a general representation of a 3-point function on the basis of computations of natural muti-variable analyticity domains. Historically thermal properties of localization entered QFT through the Hawking radiation of quantum matter behind an event horizon. For some time this was thought of as a separate issue of QFT in curved spacetime. But the main difference between event horizons in curved spacetime and causal horizons in Minkowski spacetime QFT is that the former are objective locations given by the external metric, whereas the latter are Gedanken-constructs whose physical realization depends on non-inertial observers (viz. the Unruh effect). The fleeting existence (i.e. not experimentally realizable) of causal horizons does not at all mean that they are unimportant for a structural comprehension of QFT. The fact that the insufficiently understood crossing property of particle physics reveals it full physical significance in the setting of thermal manifestations of modular localization confirms this. This confluence of particle physics concepts with concepts coming from black hole physics is a very exciting process of ongoing conceptual unification which promises to bring a wealth of new insights. There are interesting structural consequences of the crossing property, e.g. the Aks theorem [@Aks] stating that d$>$1+1 quantum fields cannot lead to elastic scattering without the presence of inelastic scattering processes. The factorizing models in d=1+1 are an exceptional case; such models carry the full infinite vacuum polarization, but its S-matrices are certain combinatorial products of two-particle $S_{scat}^{(2)}(\theta_{1}-\theta_{2}).$ Another expected consequence of localization equivalence and crossing is that $S_{scat}=1$ implies that the theory is that of a free field[^42], but the arguments given in [@inverse] can presently only be made rigorous for factorizing models. Crossing is a consequence of the specific field theoretic (modular) localization and not a general property of relativistic QT. There exists a relativistic particle quantum mechanics, the DPI (direct particle interaction theory) [@interface] which is based on the non-covariant Born-Newton-Wigner localization [@N-W] resulting from the spectral decomposition of the selfadjoint position operator. The DPI Hilbert space carries an interacting multiparticle representation of the Poincaré group which fulfills the cluster factorization property. However it contains no covariantly localized objects at finite times, the only covariant object is the (global) S matrix which is invariant and has the cluster decomposition property for spacelike directions (macrocausality). In fact it fulfills all properties which one is able to formulate in terms of particles [@interface]. On the other hand the properties presented in this section need the causal relativistic localization which, although leading to important consequences for particle scattering (as crossing), cannot be understood in a *pure* particle setting. The velocity of light in DPI setting, similar to the velocity of acoustic waves, comes about through quantum mechanical state averaging at large times; it refers to the center of a wave packet whereas in QFT it is a microscopic property of the observable algebra which is not related to the c.m. movements of wave packets. The good news is however that in QFT the *BNW localization becomes asymptotically covariant* and thus consistent with modular localization. In particular the asymptotic interpretation of QFT inherits the BNW probability without which one could not obtain invariant scattering cross section. An exceptional case of localization equivalence: d=1+1 factorizing models ========================================================================= In the generic setting of formfactor crossing there is no property by which one can *interchange* the position of the rapidities by the process of analytic continuation[^43]. For the d=1+1 factorizing models this is however possible i.e. there is only one analytic masterfunction which relates *all* rapidity orderings. This additional analytic structure makes it possible to use the extended analytic setting as the start of a classification and explicit construction of models through the S-matrix and the formfactors of its fields: the bootstrap-formfactor project [@Ba-Ka]. From the viewpoint of modular localization based construction favored in the present paper, these properties turn into powerful tools of model constructions. These models are distinguished by the fact that their wedge algebra contains what has been referred to as “temperate PFGs” (vacuum **p**olarization-**f**ree **g**enerators) [@BBS]. PFGs are operators operators which applied to the vacuum have translation invariant domains and, as a consequence, well behaved Fourier transforms. With other words the d=1+1 $B$-fields which are bijectively related via their shared wedge localized state space to the wedge-localized state space generated by incoming/outgoing free fields have now translational invariant domains (i.e. are temperate). It turns out that all so-called factorizing models [@Ba-Ka] are in this class and it appears that temperate PFG always lead to factorizing models. The covariant domain properties result in the existence of a wedge-independent on-shell Fourier transformation leading to a free field like representation [@BBS] which for the simplest family of models (the Sinh-Gordon model) [@AOP][@Lech1] are of the form $$\begin{aligned} & \Phi(x)=\frac{1}{(2\pi)^{\frac{3}{2}}}\int\left( Z^{\ast}(\theta )e^{ipx}+h.c.\right) \frac{d\theta}{2},~~p=m(ch\theta,sh\theta)\label{ZF}\\ & Z(\theta_{1})Z(\theta_{2})=s(\theta_{1}-\theta_{2})Z(\theta_{2})Z(\theta_{1}),~Z(\theta_{1})Z^{\ast}(\theta_{2})=s(\theta_{1}-\theta _{2})Z^{\ast}(\theta_{2})Z(\theta_{2})+\delta(\theta_{1}-\theta_{2})\nonumber\end{aligned}$$ where it is convenient to use the mass shell rapidity instead of the mass shell momentum. Here $s$ is the two-particle scattering function of the Sinh-Gordon model; in the general case of factorizing models the $Z$-operators are multi-component creation/annihilation operators and the scattering function becomes a scattering matrix. The Z-commutation relations are a special special case of the Zamolodchikov-Faddeev algebra structure, but in contrast to their original use as pure algebraic calculational devices, the Z’s in the present wedge localized approach have a spacetime interpretation. Although the affiliated field $\Phi$ for $s\neq1$ lacks pointlike localization, it can be shown to be at least wedge-like localized [@Sch]. When these properties became clear during the 90s [@AOP], the guiding idea that the larger the spacetime region one has at one’s disposal, the easier the control of vacuum fluctuation andr to find simple generators of the localized algebra[^44]. Pointlike local fields are of course generators for algebras localized in arbitrary spacetime regions, but in the new constructive approach they appear at the end as the cherry on the cake. The best localization region below the full algebra associated with the Minkowski spacetime which still admits a particle structure is the (noncompact) wedge region. this algebra is the “smallest” which contains for the first time PFG operators i.e. operators which once applied to the vacuum behave like a free field but have a complicated action on other states; i.e. although far more involved than free fields, in their application to the vacuum they behave precisely like a free field. This was the beginning of a new construction principle which I applied to factorizing models [@Sch] before Gandalf Lechner [@Lech1] used it to proof the first existence theorem of the strictly renormalizable (short distance singularities involve powers worse than those of free fields), but not superrenormalizable models. The Fourier transforms of the wedge generating fields were the Z-F operators of the above form. In the standard terminology $\Phi$ is a *nonlocal* on-mass-shell covariant field, but an application of modular theory shows that it is far from being completely nonlocal since it is wedge localized [@Lech1] in the sense that smeared with W-supported test functions $\Phi(f)\prec \mathcal{B}(W).$ Contrary to free fields for which the localization is entirely governed by the support of the test function, the use of compact localized test function inside W does not improve the situation. The possibility of “localizing in momentum space” in d=1+1 i.e. to work with operators $Z(\theta)$ (\[ZF\]) with Wightman-like domain simplifies the discussion and permits to arrive at more detailed results than the crossing of the previous section where algebraic properties of the operators $B(f,..),$ which are comparable to those of the temperate PFG generators $Z,$ are not available. There exists a very simple-minded almost kinematical argument why in d=1+1 the temperateness of wedge localized PFGs does not exclude interactions. It so happens that the two-dimensional energy-momentum conserving delta function coalesces with the tensor product of two particle mass shell delta functions which appear in the inner product of a two-particle state. This has as a consequence that the *cluster factorization argument* for the S-matrix cannot distinguish between an elastic $S^{(2)}$ and a trivial $S^{(2)}=\mathbf{1}$ i.e. clustering in d=1+1 cannot remove a two particle interaction and arrive at a trivial scattering amplitude. In this sense the models stay close to non-interacting situations. Nevertheless the *off-shell structure of these models is surprisingly rich*, in particular they possess the full vacuum polarization struture for compact spacetime localization, although they have no on-shell particle creation through scattering. Their mathematical and conceptual structure has been the object of many studies and they continue to play the role of a theoretical laboratory in which quantum field theoretical ideas can be tested and studied under full mathematical control. The states obtained by the iterative application of the $Z$ have a very simple structure$$\begin{aligned} & TZ^{+}(\theta_{1})....Z^{+}(\theta_{n})\Omega=a_{in}^{\ast}(\theta _{1})...a_{in}^{\ast}(\theta_{n})\Omega\\ & \bar{T}Z^{+}(\theta_{1})....Z^{+}(\theta_{n})=a_{out}^{\ast}(\theta _{1})...a_{out}^{\ast}(\theta_{n})\Omega\nonumber\end{aligned}$$ where $T$ is the $\theta$-ordering (same symbol as for time-ordering) and $\bar{T}$ denotes the opposite ordering and the right hand side only involves symmetric Bose operators. The analytic properties of the vacuum polarization component for a fixed order$$F(\mathcal{O},\theta_{1}...\theta_{n})\equiv\left\langle 0\right. |\mathcal{O}\left\vert Z^{\ast}(\theta_{1})...Z^{\ast}(\theta_{n})\right\rangle ^{in},\text{ ~~}\theta_{1}>...>\theta_{n} \label{vac}$$ are those expected from the the previous section. But now, as mentioned before, the analytic properties go beyond those coming from the cyclic KMS property since the $Z$ commutation relations also encode what happens when the order is analytically interchanged. This is similar to the extension of the primitive tube domain of Wightman functions by the use of locality. On should not confuse this commutation with (graded) bosonic statistics. The latter has been already absorbed by encoding states which coincide after applying particle statistics into one ordered master-state $\theta_{i_{1}}>..>\theta_{i_{n}\text{ }}$written as $\left\vert Z^{\ast}(\theta_{i_{1}})...Z^{\ast}(\theta_{i_{n}})\right\rangle $; it couples the $\theta$-order with the operator order in products whereas the analytic change of order is dynamic and extends Watson’s observation that the boundary values of the two-particle formfactor in the elastic region are determined by the elastic part of the scattering matrix [@Ba-Ka]. Without this analytic interchange it is not possible to understand the algebraic aspects of the work on the bootstrap formfactor construction. The knowledge about commutation properties is not availabe in the general case; in the derivation crossing in the previous section we only used the extended Araki KMS analyticity. Crossing does not tell anything about an analytic exchange of two $\theta^{\prime}s,$ i.e. the analyticity which permits to change the order of rapidities comes from the algebraic commutation structure of the $Z$ generators. The crucial property which permits the explicit computation of formfactors of fields is this analytic exchange of rapidities (not to be confused with the exchange property due to particle statistics) in which the factorizing S-matrix shows up. The Z-F commutation relations result from the algebraization of this analytic structure. Its higher dimensional generalization is the so called Watson theorem: the difference between the upper and lower branch of the elastic scattering cut of the two-particle formfactor is given by the elastic part of the S-matrix. The introduction of rapidities “unfold” this cut, but since in non-factorizing theories there exist all the higher inelastic cuts, a uniformization in terms of rapidities which leads to a meromorphic function in the plane is not possible. Hence the constructive power of factorizing models does not only come from the general crossing property but rather results also from the powerful analytic exchange property in conjunction with crossing. The factorizing models confirm again that crossing has no conceptual relation to duality. One-particle bound states which are poles in scattering processes have no special place in crossing; models without or with boundstates fulfill crossing and in case there are bound states present, they are mixed via crossing with the scattering continuum in a complicated way. Even in perturbative crossing the one-particle direct or exchange contributions do not play any special role, there is no crossing in which only one-particle states participate. 40 years of research on S-matrix based particle theory (duality, ST) have been founded on misunderstandings of the crossing property. Mathematically all the dual model constructions are synonymous with Mellin transforms of conformal QFTs, a topic which physically could not be more removed from crossing. So the discovery of dual models is nothing else than the discovery of the infinite pole-structure of the Mellin transform of conformal QFTs which in turn reflects the properties of the global operator expansions [@Mack]. It is hard to imagine anything further from the quantum field theoretical collision theory. The full analytic setting of the so-called bootstrap-formfactor program (which resulted from a correct understanding of crossing) was already formulated at the late 70s [@Kar]; since that time there has been a steady stream of novel models and new insights based on the analytic properties of their formfactors [@Ba-Ka]. In all cases the calculated formfactors were not only meromorphic functions in the multi-strip regions (where their poles have a direct interpretation in terms of bound states), but they were even meromorphic in the full complex $\theta$-plane (the infinitely many different sheets in the Mandelstam s-t variables). The conceptual basis of this approach received a significant boost when it was observed that the analytic rules for the construction of formfactors permit a formal algebraic encoding. What was first introduced as a trick without any apparent intrinsic physical meaning [@Za] in the 90s acquired the spacetime meaning of being closely related to wedge localization [@AOP] which finally led to the first existence proof for factorizing models [@Lech1]. The interesting problem is to find an higher dimensional counterpart of these observation. In the present context one certainly does not expect a simple analog  of $Z^{\#}$ operators which relate the different $\theta~$orderings in the sense that the connection between the different $\theta$-orders in the vacuum polarization formfactor (\[vac\]) can be encoded into the operator positions. The difficulty is that outside the temperate setting there is no known mechanism by which one could get to an analytic exchange of two repidities. To look for an algebraic interpretation of analytic continuation in terms of an auxiliary QFT is not so absurd as it appears at first sight. The analogy with Wightman theory is worth exploring. Wightman functions are distributions whose primitive analytic properties come from the energy positivity. The analytic tube regions for different spacetime orderings are related by the algebraic properties of covariance and local commutativity. This gets quite complicaed in case of d=1+2 braid group commutation structures where the analytic continuation leads to multivalued functions. The formfactors in factorizing theories are also multi-valued in the Mandelstam variables and by rewriting this in the temperate case into the uniformizing $\theta$ variables one finds an algebraic structure. The crucial question is whether the analyticity properties of formfactors in the general case also permits to encode the analytic change of $\theta$-orders into the position of generalized Z-operators; such a property would go beyond crossing and could play an important role in nonperturbative model constructions beyond factorization. There are many more d=1+1 unitary elastic S-matrices satisfying crossing than there are pointlike Lagrangian couplings i.e. most of the existing factorizing models do not have a Lagrangian name. There is no reason to believe that this is in any way different in higher dimensions, so there is a strong suggestion that even outside factorizing models the Lagrangian formalism only covers a tiny area. As often with physical ideas, the best insight into their inner workings may have little resemblance with the history which led to their discovery. Indeed the original observation leading eventually to factorizing models had little to do with what was presented in this section, in fact it was not even related with factorizing S-matrices but rather with integrable looking quasiclassical mass spectra of certain field theories (notable Sine-Gordon). In analogy to integrable systems of QM as the hydrogen atom, it was natural to look for higher conservation laws. But historically the first hints came from mass shell restriction of perturbative correlation functions leading to scattering amplitudes which were expected to show the absence of on-shell creation as an indication of their integrability [^45]. From such confidence-building calculations sprung the first suspicion that behind these observation there was the S-matrix bootstrap, but this time without the old ideological bombast [@Weiss][@KTTW]. The first structural arguments pointing into the direction of the S-matrix bootstrap approach set off a frenzy of model classifications and construction according to the bootstrap S-matrix program. It soon became part of a new *bootstrap-formfactor approach to factorizing models* (for more on the history see [@Swie]). Resumé, some personal observations and a somewhat downbeat outlook ================================================================== The era of post renormalization QFT began at the end of the 50s with a return of the incompletely understood age-old *particle-field problem*. The formulation of the LSZ scattering theory and its rigorous derivation by Haag, Ruelle and Hepp are important landmarks in this conquest. Another more recent important step is the partial resolution of the apparent contradiction between the noncovariant Born-Newton-Wigner localization, which brings the indispensable probabilistic concept of QM[^46] into QFT, and the modular localization, which is intrinsic to QFT but does not lead to the probability of finding a particle in a specified spacetime region [@interface]. It is deeply satisfying that in the large time scattering limit both localizations match; hence in particular the noncovariant BNW localization becomes covariant[^47] and the modular localization becomes consistent with a probability concept which in turn is the prerequisite for an invariance S-matrix and the probabilistic interpretation of the associated cross sections. This large time asymptotic coexistence between particles and fields or their generated localized operator algebras is crucial for our understanding of QFT and the crossing property is the (perhaps most subtle) manifestation of the particle-field relation. The first successful test of scattering theory consisted in the derivation of the experimentally verified Kramers-Kronig dispersion relations from analytic properties of field theoretic locality. This was important for strengthening the confidence in the locality and spectrum principes of QFT. It was in this context that the crossing relation arose in form of the existence of an analytic masterfunction which connects different processes with different distribution between incoming and outgoing particles. This was a crossing identity in which the crossed in/out particles were in an unphysical position. One still needed analytic continuation properties which the LSZ scattering formalism by its own theory did not provide. For certain scattering configurations this analytic argument was supplied in [@BEG]. In the S-matrix bootstrap approach the crossing analyticity was simply assumed under the heading “maximal analyticity”, it was treated as a basic postulate together with the other physical principles as Poincaré invariance and unitarity. This way of looking at a problem by elevating a mathematical property as analyticity to be on par with physical properties foreclosed the chance to understand crossing in terms of localization and ensuing thermal KMS properties; in particular the KMS-like cyclic permutation property (\[rel\]) of scattering amplitudes and formfactors remained unnoticed. Historically the next step was the successful use of the crossing relations within the bootsstrap-formfactor program [@Ka-Wei] for factorizing models. This did not involve a structural understanding; rather crossing was one of the assumptions in the constrution of these models and the fact that at the end one had constructed a nontrivial model meant that crossing is really a property of this particular class of models. No connection to modular localization properties and their thermal manifestations was noticed. This changed with the realization that behind the Zamolodchikov-Faddeev algebraic reformulation of factorizing models there are nonlocal wedge-localized generators [@Sch]. Only then the construction in terms of recipes for formfactors finally became a classification and construction of factorizing models according to the underlying principles of QFT without the inference of additional recipes [@Lech1]. The derivation of crossing for formfactors presented in this paper is according to my best knowledge the first one outside the narrow setting of two-dimensional factorizing models. Since the context of this paper is a rather broad one, a more detailed specific account of crossing from modular localization theory and implications thereof will be given in a separate publication [@M-S]. The history of crossing shows also that an early flare-up of ideas, before their conceptual-mathematical understanding is available, may under certain sociological conditions cause disarray[^48]. The dual model and string theory and with it that strange idea of a millennium TOE would not have come about without a certain amount of conceptual confusion. As we know nowadays the properties of Mellin transform of conformal QFTs are synonymous with dual models, including those first discovered by Veneziano and others; for their construction one does not have to know anything about the crossing property but only how to construct functions with a certain pole structure in different Mellin variables. The pedestrian construction which used properties of Gamma function in a very clever way was physically interpreted as a one-particle approximations of the conjectured Mandelstam representation for scattering amplitudes. It led to the belief that one had discovered a deep and mysterious new area of particle physics outside of QFT, whereas in reality it was the entrance into a physical no man’s land. The string theoretic extension of the dual model aggravated the problem of its conceptual positioning, in particular since its pointlike localized nature was overlooked as the result of confounding the presence of oscillators of a quantum mechanical string with the presence of a string localized in spacetime. The decisive factor which cemented this confusion through all those decades up to the present was however the sociological impact of the enthusiastic support by renown members of the physics community. Who will deny the impact of statements about string theory as “a present of the 21st century to the 20th”, “there is no other game in town” or the citation of Churchills famous die-hard slogan “never, never,...never give up” is living in an ivory tower. At this point the difference to particle physics before the 80s becomes clear: the fragile equilibrium between the innovative and speculative side of particle physics and the critical counterweight had broken down[^49]. The historically grown pre-electronic basic knowledge about QFT appears now, in the presence of a millennium TOE increasingly irrelevant. This is accompanied by a growing split between applied QFT, where the main aim is to find computational recipes about a subject which is thought of as having been basically understood, and foundational research in LQP which is expected to lead to profound structural discoveries by following the inner logic of the theory but often at the prize of loosing contact with the actual reality of particle physics. There is hardly any cross fertilization; the one side fails to penetrate the conceptual-mathematical barrier to comprehend new structural insights into QFT (and often thinks it is not even worth a try), whereas the other side has distanced itself so much from the phenomena that even when one of their findings can be connected to observational particle physics, it would probably go unnoticed. Speculative proposals with little conceptual support but popular appeal were made at all times; particle theory is by its own nature a highly speculative science where it is sometimes necessary to take a dive into the “blue yonder” of the unknown. What was however different during the decades of dominance of string theory is that the critical counterweight, which had quite a tradition on the old continent, was not available after the 70s when it was most needed. The leading figures in mathematical physics and (algebraic) quantum field theory who had the conceptual insight to play this indispensible critical role unfortunately did not enter the fray, and thus the old “Streitkultur” was lost. In the beginning of this disengagement the phenomenological proposals of Regge-trajectories were far removed from any structure which one could relate with known principles of relativistic quantum theory; but when the sudden transition to a fundamental TOE took place[^50], the uncritical acceptance of the new string theory as a TOE happened with such a speed that a critical discourse was hardly possible. The string protagonists occupied research and university positions within a short time. Often their main credentials were that they are working on the allegedly most important millennium theory. After some of the leading High Energy laboratories began to hire string theorists, it was a matter of national and scientific pride to have a representative of string theory as a kind of signboard of participation in the new millennium project. In order to avoid misunderstandings, the derailment of parts of particle physics caused by string theory did not come about because it is mathematically nonsensical. As an infinite component pointlike QFT which contains operators which communicate between the different floors of an infinite particle/spin tower it is well-defined and the problem that there exists a finite number of such objects in 10 spacetime dimensions related to each other only becomes problematic if it used as the starting point for making claims about the structure of the universe. The point where the conceptual confusion starts is that in order to introduce interactions one uses pictures *as if the pointlike localized infinite component field[^51] would be stringlike,* replacing the lines in Feynman graphs by the tubelike world-heets traced out by closed strings. Therefore the recognition that the localization is pointlike does not put an end to the confusion but rather creates new problems. It is important to note that, different from Feynman rules, these tube (worldsheet) rules, despite an intense search by the creme of string theorists over many decades, did not permit a presentation in terms of operators and states. Having bungled the localization properties of infinite component fields, at the latest this lack of presentability of would be hifger orders in terms of operators and states should have set off alarm bells. The string theoretic way of metaphoric thinking can perfectly exist outside string theory. One look at the recent paper [@Verlinde] supports this point; here a string theoretician sets out to colonize territories outside string theory by treating them with similar Zeitgeist-compatible metaphors. Perhaps the most spectacular episode triggered by string theory is the fray which developed around the anti-De Sitter–conformal field theory (AdS-CFT) correspondence, an issue which, although not directly related to string theory, suddenly obtained prominence as its alledged consequence. Within a short time string theorists managed to convert this subject into something mystical, if not to say surreal. The subtle point of this correspondence is the radical change of the spacetime localization involved in the *spacetime reordering of quantum matter* passing from AdS spacetime to a lower dimensional CFT. Since physics is not only determined by the abstract quantum matter (e.g. CCR or CAR or any other matter characterized by its abstract spacetime independent properties), but also by its spacetime ordering, some physical properties do change with the spacetime reordering in passing from $AdS_{5}$ to$~CFT_{4}.$ The relevant question is how much can they change if the abstract matter which is ordered according to the causal locality in different spacetimes with different dimensions remains the same? The answer is, that although there is no correspondence (isomorphism) between pointlike fields, there is one between tertain operator algebras which are generated by pointlike fields[^52]. This coarser than pointlike correspondence is sufficient to fix one side of the correspondence in terms of the other [@Rehren]. The naive expectation about any isomorphismus (correspondence) is that when one starts from a theory with a *physically acceptable cardinality of degrees of freedom* (intuitively speaking, one coming from Lagrangian quantization) and spatially reorders them in such a way that there remains a local algebraic isomorphism for certain regions[^53], then there will be *too many degrees of freedom* in case that the reordering leads to a spacetime of lower dimension as in the AdS$_{5}$–$>$CFT$_{4}~$correspondence. Although perfectly consistent from a mathematical viewpoint, this causes serious physical pathologies (Hagedorn temperature or no thermal states at all, anomalies in the causal propagation etc.) [@Swie]. In the opposite direction CFT$_{4}$–$>$AdS$_{5}$ the resulting AdS theory obtained from a physical CFT model will be too “anemic” concerning its degrees of freedom in order to be of any direct physical interest (the degrees of freedom hover near the boundary). This is the content of a rigorous mathematical theorem [@Rehren] and can be explicitly illustrated in terms of a free field AdS model [@Du-Re]. For example the CFT theory one obtains as an image under the correspondence from a free massive AdS model is a generalized free CFT theory with an increasing Kallén-Lehmann spectral function (a power law, depending on the AdS mass) which violates the causal shadow property and has no physical thermal states. Far from being a disease of this particular model, it is a structural property of the correspondence itself. The Maldacena conjecture [@Mal] is more specifically places a concrete supersymmetric Yang-Mills theory[^54] on the CFT side of the correspondence and expects a supersymmetric gravity model on the AdS side (suggested by string theory). The prerequiste for conformal invariance is the vanishing of the beta function. Rigorous proofs for the absence of radiative corrections and in particular the vanishing of the Beta functions in certain models were given in the 70s by combining Callen-Symanzik equations with Ward identities [@Lo1][@Lo2]. Apparently the knowledge about these techniques has been lost, the new order by order or lightcone quantization attempts applied to the supsersymmetric N=4 Yang Mills model are unconvincing. But our main criticism is independent of these weaknesses and concerns the phase space degree of freedom issue which stands in contradiction to the underlying tacit assumption that both sides represent physical theories. The above theorem says that this is structurally impossible; if one side is physical, the other is a purely mathematical chimera which however still may be useful in order to study certain physical properties of the physical side which in the original description were not easily accessible. Since the Maldacena statement is only a conjecture as compared to Rehren’s theorem, there is no paradox here. What renders the whole situation delicate from a sociological viewpoint however is the fact that meanwhile more than 6000 papers have been written in support of Maldacena’s conjecture (but, as expected, without any conclusion about its validity) and the saying that so many people cannot err is, as well-known, one of the most accepted vernaculars. It is hard to think of a more convincing illustration about the loss of solid scientific knowledge than this episode around the Maldacena correspondence. The discovery that instead of the finite phase space degrees of freedom in QM (one per unit phase space cell of size $\hbar$), the cardinality of degrees of freedom in QFT is different, namely “mildly infinite” (compact, nuclear) was made in the 60s [@Ha-Sw][@Swie]. In the spirit of this article it is important to emphasize that this difference is a consequence of the different concepts of localization [@interface]. If one compresses the O(4,2) symmetric degrees of freedom from a physical density in a five-dimensional spacetime into four dimensions, then there are “too many phase space degrees” in order to sustain the causal propagation property which is the LQP version of the classical causal Cauchy propagation. With too many phase space degrees of freedom the quantum causal shadow property $\mathcal{A(O})=\mathcal{A(O}^{\prime\prime}),$ where $\mathcal{O}^{\prime\prime}$ is the causal completion of the spacetime region $\mathcal{O}$ (the causal complement taken twice), is being violated; the right hand side is bigger. From the viewpoint of somebody whose intuitive understanding of QFT comes from Lagrangian quantization which formally obey this property, the violation may appears mysterious. The only way he can uphold his picture of propagation is by using a metaphor that some degrees of freedom enter “sideways” from an extra dimension or from another universe (“poltergeist degrees of freedom”). Within the present Zeitgeist inspired by string theory, where metaphoric arguments are en vogue and extra dimensions and multiverses are concepts on which articles are written, this only sounds like a harmless addition. The problem is that deep concepts as the cardinality of degrees of freedom [@Ha-Sw][@Bu-Wi] and their preservation in correspondences between QFT in different spacetimes have vanished from the conceptual screen of the 80s so that especially those who work on holographic problems are not aware of their existence. The notion that metaphoric arguments should at most be tolerated as placeholders in a conceptual emergency for a limited time has been lost. A more detailed recent presentation of this phase space degrees of freedom issue can be found in [@Swie]. Part of the problem of holographic spacetime reordering of quantum matter is that it is too radical in order to allow a formulation in terms of the standard setting of QFT using individual pointlike fields; there is however no problem to express this in terms of operator algebras associated with suitable causally closed regions [@Rehren]. The only kind of holography which complies with the thinning out of phase space degrees of freedom is the holography onto nullsurfaces i.e. the holographic projection of bulk QFT onto causal or event horizons. In that case the reduction of degrees of freedom goes hand in hand with a reduction of symmetry: the symmetry of a lightfront is a 7-parameter subgroup of the Poincaré group and the problem of “filling up” the degrees of freedom to their orginal strength is equivalent to knowing the action of the remaining Poincaré transformations on the lightfront degrees of freedom. Equivalently it would suffice to know the lightfront theory in a “GPS manner” in different positions; in d=1+3 not more than three different positions are necessary [@interface]. The problems which led to a derailment of a large part of particle theory can however not fully explain why the comperativly healthy standard model, after impressive initial gains, entered a period of stagnation. For almost 4 decades there has been not a single conceptual addition to the age-old central problems of gluon and quark confinement and the Schwinger-Higgs screening mechanism. Such a situation is certainly unique in the more then 8 decades lasting history of particle physics. In some cases there was even a regress in that earlier promising ideas have been lost in the maelstrom of time [@Swie]. If there was any influence of S-matrix approach on the standard model research, it certainly was not of a hepful kind. Rather the perilous charm, which a TOE supported by prominent community members, exerts on intelligent and zealous newcomers could have been one reason why the standard model research may not have attracted the brightest minds; not to mention the considerable material support enjoyed by string-related research; a closely related argument is the prediction of the leading string theorists that the standard model has to appear anyhow as a “low energy effective theory” of a TOE. Finally there is a widespread but misleading opinion that the remaining theoretical problems of the standard model are basically of a computational nature; this is strengthened by the credo that QFT is a reasonably well understood low energy footnote of string theory. These beliefs have eroded the enthusiasm for new conceptual investments. A serious obstacle against a conceptual renewal is the fact that the teaching of QFT has fallen back behind what can be found in books written before 1980 e.g. in the book of Itzykson and Zuber. More recent books often appear as a kind of QFT filtered through string theory glasses. It is nearly impossible to start research on important conceptual problems (as the problem of the crossing property in this paper) on the basis of contemporary books on QFT. This has led to a situation in which the number of people who know QFT sufficiently well in order to contribute to a conceptual progress of QFT has shrunk to a few individuals in an advanced age. Speculative proposals with little conceptual support but a lot of public attraction were of course made at all times; particle theory by its very nature is a highly speculative science where it is necessary (at least once in a while) to take a dive into the “blue yonder”. What was however different during the last 4 decades of dominance of string theory, is that the critical counterweight, which had quite a tradition in the Streitkultur of the old continent, was not available when it was most needed. The leading figures in mathematical physics and (algebraic) quantum field theory who are in the possession of the necessary conceptual insight to play this indispensible critical role did not enter the fray. At the beginning the phenomenonological proposals (the Regge trajectory setting) were far removed from any structure which one could relate with known principles of relativistic quantum theory, and when the sudden transition to a pretended fundamental TOE took place[^55], the uncritical spread of the new string theory was too rapid, so that there was hardly time for a critical discourse. The string protagonists occupied research and university positions within a short time, and often their only credentials were that they are working on the most important millennium theory. There are of course others who understand more or less the causes behind the derailment. In some of their articles one even finds the statements that strings are, contrary to their name, really point-localized objects. But since no critical conclusions are drawn; such articles do not create frictions with their string theory colleagues. They are tolerated, even when they contribute jointly to the same book [@FRS]), as the kind of critical remarks which show that string theory is a living science. As long as they do not lead to a serious conceptual encounter whose outcome could threaten the continued existence of a more than 40 years lasting development in particle physics, the present stalemate will continue. The fruitful Streitkultur belonged to the bygone “golden age” of critical engagement in particle theory. Similar arguments apply to the sociological [@Woit][@Smolin] and philosophical [@He] critique of string theory. Whereas scientific critique may have the power to erode metaphoric constructs, sociological and philosophical arguments do not constitute any danger to the popularity of string theory and certainly have nothing in common with a critical engagement within a scientific Streitkultur; to the contrary they lead to a profitable symbiosis between string propagandists and their critics, with the latter running the risk of loosing their subject without the presence of the former. Reading the books and articles of the aforementioned authors, the following questions comes to one’s mind. Why can’t a theory which has strong conceptual credentials be explored for whatever time is necessary to get to its limits, and isn’t a consistent theory which, as claimed by string theorists, incorporates the existing one as a limiting case an interesting goal even if it does not describe reality? And is observational agreement the only criterion for evaluating a new theory? The old (pre-oxigen) phlogiston theory of burning which dominated for many decades shows that a wrong theory may be able to live for a long time in reasonable agreement with observational facts, especially if it explains sufficienty many observed phenomena. *The only kind of critique which a theoretician must take serious in the long run is one which, as presented in this paper, demonstrates that a theory is conceptually flawed.* All these observations show that the adventurous journey that started more than 4 decades ago with some misunderstandings in the particle-field relation around the crossing property, has grown into a profound crisis of particle physics. The resulting metaphoric discourse of placing superficial conclusions based on calculations done outside any conceptual control above profound critical evaluations is not any more confined to ST; the concomittant sociological phenomenon around the AdS-CFT issue is a clear indication of the spread of the crisis beyond the borders of string theory. The disappearance of criticism has led to a new culture of establishing a scientific truth starting from a conjecture and ending after several reformulations and turns with the acceptance within a community at the level of a theorem. This process has been insightfully described in a series of essays by a young string theorist [@Zapata]. The author, Oswaldo Zapata, has an ambivalent position with respect to string theory; having been raised with string theory and being aware about his limitations with respect to QFT, he knows that he cannot confront it on its *scientific* truth content. Instead he carefully analyzes the sociological aspects of its discourse and comes to remarkable conclusions. His aim is to understand how his fellow string theorists, having disposed of classical methods of establishing theorems, arrive at what they consider as truths, and how they present their results without becoming subjectively dishonest within the community and to the outside world. He does this by studying changes in the string communities discourse from conjectures to theorems during a time in which there was no change in the facts. Interestingly enough he gives the strongest argument for his thesis about the relation of the string community to facts involuntary by not referring to the aforementioned rigorous theorems about AdS-CFT. They are all in the public domain, but their conceptual mathematical content [@Rehren] is not known by the community members because most of them are not on a level on which they can understand structural theorems on local quantum physics. This shows that the control of the community over facts does not end at what is coming from the inside (which Zapata as an insider of this community is well aware of), but it extends also to shielding inconvenient theorems from the outside in the most possible honest manner, namely by ignorance about large parts of QFT. In this way even Zapata remains uninformed that his critical sociological observations about the discourse of the string community have a profound scientific counterpart. Reading Zapata’s essay may not help to learn about conceptual errors of string theory. But his method is very successful in exposing the surreal aspect which accompanies the string community’s almost messianic “end of the millennium belief” in a TOE. His account of how a metaphoric conjecture ends after several sweeps through the community as a community-accepted fact is truely remarkable. It shows that some individuals of the string generation, having been deprived of a critical conceptual scientific basis, can still make fascinating critical observations about the logic and sociology of the discourse within the string community. It is quite revealing that Zapata takes a dim view on some missing arguments in two books by Lee Smolin and Peter Woit [@Smolin][@Woit]. These authors take a critical look at the dominant position of string theory and explain very well the sociological reasons why younger people uncritically internalize the catechism of string theory. *But they never explain why respectable older people, who are under no such career pressures (especially those who are the main string proselytizers mentioned before) believe in the validity of the theory.* It is of course common practice to blame the foot-soldiers (in the present context, the young partisans of string theory) and the propaganda division (Brian Green and others), but spare the generals; there should be no place for this attitude in particle physics. It would be wishful thinking that articles as the present one or the essay of Zapata could have an influence on the tide of events. But they provide a valuable help for historians and philosophers of science to analyze what went on in particle theory during a substantial part of the 20$^{th}$ and the beginning of the 21st century. Since readers need some encouragement in the conclusions, the present essay should not end in a downbeat mood. There are some interesting new developments around higher spin field, in particular massless fields. They start from the observation made in the appendix in (\[line\]) where it was mentioned that the reduced possibilities for $(m=0,s)$ with $s=1,2$ which exclude covariant vector potentials and $g_{\mu\nu}$ tensors, can be complemented to the full spinorial formalism (so that the massless situation is on par with massive case) if one permits semiinfinite string localization [@charge]. This leads to a new way of looking at the problems behind gauge theory. Already in the abelian case of QED for which it has been known for a long time that electrically charged states are semiinfinite string-localized (associated to infraparticles), the new setting incorporates the perturbative aspects of these physical charge-carrying fields into the formalism i.e. they do not have to be defined by hand outside the perturbation formalism as in the famous stringlike formulas of Dirac-Jordan-Mandelstam [@Swie][@Jor]. The Higgs model results as a Schwinger-Higgs screening of the electric charge of a scalar fields and leads to a theory in which the massive matter field is neutral (real) and pointlike localized [@Swie]. The new conceptual frame of modular localization promises to lead to a significant enlargement of the range of renormalizability [@charge]. Appendix: a sketch of modular localization ========================================== Modular localization of states ------------------------------ The simplest context for a presentation of the idea of modular localization is the Wigner representation theory of the Poincaré group. It has been realized by Brunetti, Guido and Longo [@BGL] [^56] there is a natural localization structure on the Wigner representation space for any positive energy representation of the proper Poincaré group. Upon second quantization this representation theoretically determined localization theory gives rise to a local net of operator algebras on the Wigner-Fock space over the Wigner representation space. The starting point is an irreducible representation $U_{1}~$of the Poincaré group on a Hilbert space $H_{1}$ that after “second quantization” becomes the single-particle subspace of the Hilbert space (Wigner-Fock-space) $H_{WF}$ of the field[^57]. The construction proceeds according to the following steps [@BGL][@Fa-Sc][@MSY]. To maintain simplicity, we limit our presentation to the spinless bosonic situation. One first fixes a reference wedge region, e.g. $W_{0}=\{x\in\mathbb{R}^{d},x^{d-1}>\left\vert x^{0}\right\vert \}$ and considers the one-parametric L-boost group (the hyperbolic rotation by $\chi$ in the $x^{d-1}-x^{0}$ plane) which leaves $W_{0}$ invariant; one also needs the reflection $j_{W_{0}}$ across the edge of the wedge which is apart from a $\pi$-rotation in the transverse plane identical to the TCP transformation. The Wigner representation is then used to define two commuting wedge-affiliated operators $$\mathfrak{\delta}_{W_{0}}^{it}=\mathfrak{u}(0,\Lambda_{W_{0}}(\chi=-2\pi t)),~\mathfrak{j}_{W_{0}}=\mathfrak{u}(0,j_{W_{0}})$$ where attention should be paid to the fact that in a positive energy representation any operator which inverts time is necessarily antilinear[^58]. A unitary one- parametric strongly continuous subgroup as $\delta_{W_{0}}^{it}$ can be written in terms of a selfadjoint generator as $\delta_{W_{0}}^{it}=e^{-itK_{W_{0}}}$ and therefore permits an “analytic continuation” in $t$ to an unbounded densely defined positive operators $\delta_{W_{0}}^{s}$. Poincaré covariance allows to extend these definitions to wedges in general position, and intersections of wedges lead to the definitions for general localization regions (see later). Since the localization is clear from the context, a generic notation without subscripts will be used. With the help of this operator one defines the unbounded antilinear operator $\mathfrak{s}$ which has the same dense domain.$$\begin{aligned} & \mathfrak{s}=\mathfrak{j\delta}^{\frac{1}{2}},~dom\mathfrak{s}=dom\mathfrak{\delta}^{\frac{1}{2}}\\ & \mathfrak{j\delta}^{\frac{1}{2}}\mathfrak{j}\mathfrak{=\delta}^{-\frac {1}{2}}$$ Whereas the unitary operator $\delta^{it}$ commutes with the reflection, the antiunitarity of the reflection causes a change of sign in the analytic continuation as written in the second line. This leads to the involutivity of the s-operator as well as the identity of its range with its domain $$\begin{aligned} \mathfrak{s}^{2} & \subset\mathbf{1}\\ dom~\mathfrak{s} & =ran~\mathfrak{s}$$ $\mathbf{.}$ Such operators which are unbounded and yet involutive on their domain are quite unusual; according to my best knowledge they only appear in modular theory and it is precisely these unusual aspects which are capable to encode geometric localization properties into domain properties of abstract quantum operators. The more general algebraic context in which Tomita discovered modular theory will be mentioned later. The idempotency means that the s-operator has $\pm1$ eigenspaces; since it is antilinear the +space multiplied with $i$ changes the sign and becomes the - space; hence it suffices to introduce a notation for just one of the two eigenspaces$$\begin{aligned} \mathfrak{K}(W) & =\{domain~of~\Delta_{W}^{\frac{1}{2}},~\mathfrak{s}_{W}\psi=\psi\}\\ \mathfrak{j}_{W}\mathfrak{K}(W) & =\mathfrak{K}(W^{\prime})=\mathfrak{K}(W)^{\prime},\text{ }duality\nonumber\\ \overline{\mathfrak{K}(W)+i\mathfrak{K}(W)} & =H_{1},\text{ }\mathfrak{K}(W)\cap i\mathfrak{K}(W)=0\nonumber\end{aligned}$$ It is important to be aware that, unlike QM, we are dealing here with real (closed) subspaces $\mathfrak{K}$ of the complex one-particle Wigner representation space $H_{1}$. An alternative which avoids the use of real subspaces is to directly work with complex dense subspaces as in the third line. Introducing the graph norm of the dense space, the complex subspace in the third line becomes a Hilbert space in its own right. The upper dash on regions in the second line denotes the causal disjoint (which is the opposite wedge) whereas the dash on real subspaces means the simplectic complement with respect to the simplectic form $Im(\cdot,\cdot)$ on $H_{1}.$ The two equations in the third line are the defining property of what is called the *standardness* of a subspace[^59]; any standard K-space permits to define an abstract s-operator as follows$$\begin{aligned} \mathfrak{s}(\psi+i\varphi) & =\psi-i\varphi\\ \mathfrak{s} & =\mathfrak{j}\delta^{\frac{1}{2}}\nonumber\end{aligned}$$ whose polar decomposition (written in the second line) returns the two modular objects $\delta^{it}$ and $\mathfrak{j}$ which outside the context of the Poincaré group has in general no geometric significance. The domain of the Tomita $s$-operator is the same as the domain of $\delta^{\frac{1}{2}}$ namely the real sum of the K space and its imaginary multiple. Note that in the present context this domain is determined solely by Wigner’s group representation theory. It is easy to obtain a net of K-spaces by $U(a,\Lambda)$-transforming the K-space for the distinguished $W_{0}.$ A bit more tricky is the construction of sharper localized subspaces via intersections $$\mathfrak{K}(\mathcal{O})={\displaystyle\bigcap\limits_{W\supset\mathcal{O}}} \mathfrak{K}(W)$$ where $\mathcal{O}$ denotes a causally complete smaller region (noncompact spacelike cone, compact double cone). Intersection may not be standard, in fact they may be zero in which case the theory allows localization in $W$ (it always does) but not in $\mathcal{O}.$ Such a theory is still causal but not local in the sense that its associated free fields are pointlike. There are three classes of irreducible positive energy representation, the family of massive representations $(m>0,s)$ with half-integer spin $s$ and the family of massless representation which consists of two subfamilies with quite different properties namely the $(0,h),$ $h$ half-integer class (the neutrino, photon class), and the rather large class of $(0,\kappa>0)$ infinite helicity representations parametrized by a continuous-valued Casimir invariant $\kappa$ [@MSY]$.$ For the first two classes the $\mathfrak{K}$-space is standard for arbitrarily small $\mathcal{O}$, but this is definitely not the case for the infinite helicity family for which the compact localization spaces turn out to be trivial[^60]. Their tightest localization, which still permits nontrivial (in fact standard) $\mathfrak{K}$-spaces for *all* positive energy representations, is that of a *spacelike cone* [@BGL] with an arbitrary small opening angle whose core is a *semiinfinite string* [@MSY]; after “second quantization” (see next subsection) these strings become the localization region of string-like localized covariant generating fields[^61]. The modular localization of states, which is governed by the unitary representation theory of the Poincaré group, has only two kind of generators: pointlike state and semiinfinite stringlike states; generating states of higher dimensionality (“brane states”) are not needed. Although the observation that the third Wigner representation class is not pointlike generated was made many decades ago, the statement that it is semiinfinite string-generated and that this is the worst possible case of state localization is of a more recent vintage [@BGL] since it needs the application of the modular theory. There is a very subtle aspect of modular localization which one encounters in the second Wigner representation class of *massless finite helicity representations* (the photon, graviton..class). Whereas in the massive case all spinorial fields $\Psi^{(A,\dot{B})}$ the relation of the physical spin $s$ with the two spinorial indices follows the naive angular momentum composition rules [@Weinberg][@charge]$$\begin{aligned} \left\vert A-\dot{B}\right\vert & \leq s\leq\left\vert A+\dot{B}\right\vert ,\text{ }m>0\label{line}\\ s & =\left\vert A-\dot{B}\right\vert ,~m=0\nonumber\end{aligned}$$ the second line contains the considerably reduced but still infinite number of spinorial descriptions for zero mass and finite helicity [@MSY]. By using the recourse of string-localized generators $\Psi^{(A,\dot{B})}(x,e)$ one can *restore the full spinorial spectrum* for a given $s$ i. e. one can move from the second line to the first line in (\[line\]) by relaxing the localization. Even in the massive situation where pointlike generators exist but have short distance singularities which increase with spin. there may be good reasons (lowering of short distance dimension down to sdd=1) to use string-like generators. In all cases these generators are covariant and “string-local” $$\begin{aligned} U(\Lambda)\Psi^{(A,\dot{B})}(x,e)U(\Lambda) & =D^{(A,\dot{B})}(\Lambda ^{-1})\Psi^{(A,\dot{B})}(\Lambda x,\Lambda e)\\ \left[ \Psi^{(A,\dot{B})}(x,e),\Psi^{(A^{\prime},\dot{B}^{\prime})}(x^{\prime},e^{\prime}\right] _{\pm} & =0,~x+\mathbb{R}_{+}e><x^{\prime }+\mathbb{R}_{+}e^{\prime}\nonumber\end{aligned}$$ Here the unit vector $e$ is the spacelike direction of the semiinfinite string and the last line expresses the spacelike fermionic/bosonic spacelike commutation. The best known illustration is the ($m=0,s=1$) representation; in this case it is well-known that although a generating pointlike field strength exists, there is no pointlike vectorpotential. The modular localization approach offers as a substitute a stringlike vector potential $A_{\mu}(x,e).$ In the case ($m=0,s=2$) the “field strength” is a fourth degree tensor which has the symmetry properties of the Riemann tensor; in fact it is often referred to as the linearized Riemann tensor. In this case the string-localized potential is of the form $g_{\mu\nu}(x,e)$ i.e. resembles the metric tensor of general relativity. The consequences of this localization for a reformulation of gauge theory will be taken up in a separate subsection. The most radical form of string localization occurs in the massless infinite spin representation family. In that case the representation space does not contain any pointlike localized generators which play the role of field strength, hence such a theory is without any local observables. A different kind of spacelike string-localization arises in d=1+2 Wigner representations with anomalous spin [@Mu1]. The amazing power of this modular localization approach is that it preempts the spin-statistics connection already in the one-particle setting, namely if s is the spin of the particle (which in d=1+2 may take on any real value) then one finds for the connection of the simplectic complement with the causal complement the generalized duality relation $$\mathfrak{K}(\mathcal{O}^{\prime})=Z\mathfrak{K}(\mathcal{O})^{\prime}$$ where the square of the twist operator $Z=e^{\pi is}~$is easily seen (by the connection of Wigner representation theory with the two-point function) to lead to the statistics phase: $Z^{2}=$ statistics phase [@Mu1]. The one-particle modular theory also leads to a relation which may be considered as the proto-form of crossing in the one-particle space$$\mathfrak{\Delta}^{\frac{1}{2}}\mathfrak{j\psi(}p\mathfrak{)=}\overline {\mathfrak{\psi(-}p\mathfrak{)}}$$ in words the $\mathfrak{\Delta}^{\frac{1}{2}}\mathfrak{j=s}^{\ast}~$transformed wave function is equal to the complex conjugate (antiparticle) and from forward to backward mass shell analytically continued (through the connecting complex mass shell) wave function. That one never has to go beyond string localized wave functions (and in fact, apart from those mentioned cases, even never beyond point localization) in order to obtain the generating fields for a QFT is remarkable in view of the many attempts to introduce extended objects into QFT. It should be clear that modular localization, which is formulated in terms of either real or dense complex subspaces, cannot be connected with probabilities and projectors. It is rather related to causal localization aspects and the standardness of the K-space for a compact region is nothing else then the one-particle version of the Reeh-Schlieder property. Fortunately one needs the probability and the projectors from the BNW localization only for asymptotic timelike scattering distances in which case they become frame-independent and the discrepancy with modular localization disappears. Localized subalgebras --------------------- A net of real subspaces $\mathfrak{K}(\mathcal{O})$ $\subset$ $H_{1}$ for an finite spin (helicity) Wigner representation can be “second quantized”[^62] via the CCR (Weyl) respectively CAR quantization functor; in this way one obtains a covariant $\mathcal{O}$-indexed net of von Neumann algebras $\mathcal{A(O)}$ acting on the bosonic or fermionic Fock space $H=Fock(H_{1})$ built over the one-particle Wigner space $H_{1}.$ For integer spin/helicity values the modular localization in Wigner space implies the identification of the simplectic complement with the geometric complement in the sense of relativistic causality, i.e. $\mathfrak{K}(\mathcal{O})^{\prime}=\mathfrak{K}(\mathcal{O}^{\prime})$ (spatial Haag duality in $H_{1}$). The Weyl functor takes the spatial version of Haag duality into its algebraic counterpart. One proceeds as follows: for each Wigner wave function $\varphi\in H_{1}$ the associated (unitary) Weyl operator is defined as$$\begin{aligned} Weyl(\varphi) & :=expi\{a^{\ast}(\varphi)+a(\varphi)\},Weyl(\varphi)\in B(H)\\ \mathcal{A(O}) & :=alg\{Weyl(\varphi)|\varphi\in\mathfrak{K}(\mathcal{O})\}^{^{\prime\prime}},~~\mathcal{A(O})^{\prime}=\mathcal{A(O}^{\prime })\nonumber\end{aligned}$$ where $a^{\#}(\varphi)$ are the usual Fock space creation and annihilation operators of a Wigner particle in the wave function $\varphi$. We then define the von Neumann algebra corresponding to the localization region $\mathcal{O}$ in terms of the operator algebra generated by the functorial image of the modular constructed localized subspace $\mathfrak{K}(\mathcal{O})$ as written in the second line. By the von Neumann double commutant theorem, our generated operator algebra is weakly closed by definition. The functorial relation between real subspaces and von Neumann algebras via the Weyl functor preserves the causal localization structure, hence the spatial duality passes to its algebraic counterpart. The functor also commutes with the process of sharpening localization through intersections $\cap$ according to $K(\mathcal{O})=\cap_{W\supset O}K(W),~\mathcal{A(O})=\cap_{W\supset O}\mathcal{A}(W)$ as expressed in the commuting diagram$$\begin{aligned} & \left\{ K(W)\right\} _{W}\longrightarrow\left\{ \mathcal{A}(W)\right\} _{W}\\ & \ \ \downarrow\cap~~~\ \ \ \ \ \ \ \ \ \ ~\ ~\downarrow\cap\nonumber\\ ~~ & \ \ \ K(\mathcal{O})\ \ \ \longrightarrow\ \ ~\mathcal{A(O})\nonumber\end{aligned}$$ Here the vertical arrows denote the tightening of localization by intersection, whereas the horizontal ones denote the action of the Weyl functor. The case of half-integer spin representations is analogous [@Fa-Sc], apart from the fact that there is a mismatch between the causal and simplectic complements to be taken care of by a *twist operator* $\mathcal{Z}$ and as a result one arrives at the CAR functor instead of the Weyl functor. In case of the large family of irreducible zero mass infinite spin representations for which the lightlike little group, different from the finite helicity representations, is faithfully represented, the finitely localized K-spaces are trivial $\mathfrak{K}(\mathcal{O})=\{0\}$ and the *most tightly localized nontrivial spaces* *are of the form* $\mathfrak{K}(\mathcal{C})$ for $\mathcal{C}$ a *spacelike cone*. As a double cone contracts to its pointlike core, the core of a spacelike cone $\mathcal{C}$ is a *covariant spacelike semiinfinite string*. The above functorial construction works the same way for the Wigner infinite spin representation, except that there are no nontrivial compactly localized algebras with a smaller localization than $\mathcal{A(C})$, and there are no generating fields which are sharper localized than a semiinfinite spacelike string. Point- (or string-) like covariant fields are singular generators of these algebras i.e. operator-valued distributions. Stringlike generators, which are also available in the pointlike case, turn out to have an improved short distance behavior; whereas e.g. the short distance dimension of a free pointlike vectorfield is $sddA_{\mu}(x)=2,$ its stringlike counterpart has $sddA_{\mu}(x,e)=1~$[@MSY] thanks to the fact that the vacuum fluctuations are spread into $e$ as well. Covariant representations are constructed from the unique Wigner representation by so called intertwiners between the canonical and the many possible covariant (dotted-undotted spinorial representations of the L-group) representations. Whereas for pointlike generators this is done by group theoretic methods as in [@Weinberg], the construction of string-like intertwiners require the use of modular localization [@MSY]. The Euler-Lagrange formalism plays no role in these construction since the causal aspect of hyperbolic differential propagation are fully taken care of by modular localization. A basis of local covariant field coordinatizations is defined by Wick composites of the free fields. The string-like fields do not follow the classical behavior; already before introducing composites one has a continuous family of non-classical intertwiners between the unique Wigner infinite spin representation and the *continuously many covariant string interwiners*. These non-classical aspects, in particular the absence of a Lagrangian, are the reason why their spacetime description in terms of semiinfinite string fields has been discovered only recently and not at the time of Jordan’s field quantization nor at the time of Wigner’s representation theory. Using the standard notation $\Gamma$ for the second quantization functor which maps real localized (one-particle) subspaces into localized von Neumann algebras, and extending this functor in a natural way to include the functorial images of the $\mathfrak{K}(\mathcal{O})$-associated objects $s,\delta,j$ (denoted by $S,\Delta,J),$ one arrives at the Tomita Takesaki theory of the interaction-free local algebra ($\mathcal{A(O}),\Omega$) in standard position[^63]$$\begin{aligned} & H_{Fock}=\Gamma(H_{1})=e^{H_{1}},~\left( e^{h},e^{k}\right) =e^{(h,k)}\label{mod}\\ & \Delta=\Gamma(\delta),~J=\Gamma(j),~S=\Gamma(s)\nonumber\\ & SA\Omega=A^{\ast}\Omega,~A\in\mathcal{A}(O),~S=J\Delta^{\frac{1}{2}}\nonumber\end{aligned}$$ With this result we got to the core statement of the Tomita-Takesaki theorem which is a statement about the action of the two modular objects $\Delta^{it}$ and $J$ on the algebra$$\begin{aligned} \sigma_{t}(\mathcal{A(O})) & \equiv\Delta^{it}\mathcal{A(O})\Delta ^{-it}=\mathcal{A(O})\\ J\mathcal{A(O})J & =\mathcal{A(O})^{\prime}=\mathcal{A(O}^{\prime})\nonumber\end{aligned}$$ in words: the reflection $J$ maps an algebra (in standard position) into its von Neumann commutant and the unitary group $\Delta^{it}$ defines an one-parametric automorphism-group $\sigma_{t}$ of the algebra. In this form (but without the last statement involving the geometrical causal complement $\mathcal{O}^{\prime})$ the theorem hold in complete mathematical generality for standard pairs ($\mathcal{A},\Omega$). The free fields and their Wick composites are “coordinatizing” singular generators of this $\mathcal{O}$-indexed net of algebras in the sense that the smeared fields $A(f)$ with $suppf\subset\mathcal{O}$ are (unbounded operators) affiliated with $\mathcal{A(O}).$ In the above second quantization context the origin of the T-T theorem and its proof is clear: the spatial symplectic disjoint passes via the functorial operation to the operator algebraic commutant and the spatial one-particle modular automorphism goes into its algebraic counterpart. The definition of the Tomita involution $S$ through its action on the dense set of states (guarantied by the standardness of $\mathcal{A}$) as $SA\Omega=A^{\ast}\Omega$ and the action of the two modular objects $\Delta,J$ (\[mod\]) is part of the general setting of the modular Tomita-Takesaki theory; standardness is the mathematical terminology for the Reeh-Schlieder property [@Haag] i.e. the existence[^64] of a vector $\Omega\in H$ with respect to which the algebra acts cyclic and has no “annihilators” of $\Omega.$ Naturally the proof of the abstract T-T theorem in the general setting of operator algebras or even in the more restricted context of interacting QFT is more involved [@Haag]. The important property which renders this formalism useful beyond free fields as a new constructive tool in the presence of interactions, is that for $\left( \mathcal{A}(W),\Omega\right) ~$ the antiunitary involution $J$ depends on the interaction, whereas $\Delta^{it}$ continues to be uniquely fixed by the representation of the Poincaré group i.e. by the particle content. In fact it has been known for some [@Sch] time that $J$ is related with its free counterpart $J_{0}$ through the scattering matrix$$J=J_{0}S_{scat} \label{scat}$$ This modular role of the scattering matrix as a relative modular invariant between an interacting theory and its free counterpart comes as a surprise. It is precisely this role which opens the way for an inverse scattering construction [@inverse] and the constructive approach to factorizing models [@Lech1]. Closely related to this observation is the realization that the wedge region leads to a coexistence of one particle states in interacting theories (section 6) with modular localization; namely there is a dense set of wedge-localized one particle states and their multiparticle in/out extensions in the interacting theory. With other words the wedge region is the “smallest” region for which PFGs (vacuum **p**olarization **f**ree **g**enerators) and their multiparticle generalizations are available. This is the origin of the crossing property as explained in section 5. For the construction of a QFT it suffices to specify wedge algebra $\mathcal{A}(W)$ for one particular wedge $W$ as well as the action of the Poincaré group on $\mathcal{A}(W)$ which results in a net of wedge algebras $\left\{ \mathcal{A}(W)\right\} _{W\in\mathfrak{W}}.$Knowing a wedge algebra means knowing its position in the global algebra $\mathcal{A}(W)\subset B(H);$ in practice this is achieved by describing $\mathcal{A}(W)$ in terms of generators as explained before in the special case of factorizing models. By taking suitable intersections of wedge algebras one obtains (in case the double cone intersections are nontrivial) a net of local observables i.e. a nontrivial local QFT or (if they are trivial) there is no local QFT associated with the system of wedge algebras. In this way one is able to separate the existence proof for a local QFT from the harder problem of the construction of its pointlike fields[^65] via their correlation functions or formfactors. Hence the construction of a QFT may be seen as a generalization of those ideas which lead to a proof of the crossing property. An “observable net” is a spacetime-indexed family of operator algebras consisting of chargeless operators. By definition these operators fulfill spacelike commutativity and have, as the vacuum, vanishing charge. There exists a very deep theory which intrinsically constructs all charged sectors and combines them to a generally quite large “field-algebra” which in a way defines the maximal extension of the observable algebra; this is the famous Doplicher-Haag-Roberts (DHR) superselection theory [@Haag]. It explains statistics and inner symmetries in terms of spacetime localization properties of the observable net[^66]. From a point of view of principles of QFT one can show that in more than 3 dimensions all compact groups can appear. What does not appear in this classification is supersymmetry. A slight reformulation of this algebraic setting which leads to a (philosophically) quite spectacular new view of the core nature of local quantum physics. Namely it is possible to encode the entire content of QFT i.e. the net of local observables *as well* as all its superselected charge sectors and their interpolating charged fields including the representation of the Poincaré group acting on it, into a finite set of copies of the monad (physically interpreted as $\mathcal{A}(W)s$) carefully positioned in a joint Hilbert space with the help of modular theory, using concepts of “modular inclusion” and “modular intersection ” within a joint Hilbert space [@interface]. The representation theory of the Poincaré group and therefore of spacetime itself arises from the joint action of the individual modular groups in the form of unitary operators in the shared Hilbert space. This is as close as one can get to how Leibniz envisaged reality as emerging from *relations* between monads, the monads (here copies of the unique hyperfinite Type III$_{1}\operatorname{factor}$ algebra) themselves being structureless[^67].   It is an interesting open question whether  a characterization of a QFT in terms of positioning of a finite number of monads can be extended to curved spacetime. The recent successful quantum formulation of the principle of local covariance [@BFV] nourishes some hope that this may be the case. Acknowledgement: I am indepted to Jens Mund who on several occasions gave me advice on matters of modular localization. [99]{} W. Heisenberg, Z. Naturforschung **1**, (1946) 608 J. Bros, H. Epstein and V. Glaser, Com. Math. Phys. **1**, (1965) 240 H. Epstein, V. Glaser, A.Martin, Commun. Math.Phys. **13**, (1969) 257 A.Martin, *Scattering Theory, Unitarity, Analyticity and Crossing*, Springer-Verlag, 1969 J. Bros, Physics Report **134**, (1986) M. Karowski, H.-J. Thun, T.T. Truoung and P. Weiss, Phys. Rev. Lett. B **67**, (1977) 321 H. Babujian, A. Förster and M. Karowski, Nucl. Phys. B **825**, (2010) 396 P. Di Vecchia, *The birth of string theory*, arXiv 0704.0101 R. J. Jost, Helvetica Physica Acta **36**, (1963) 77 W. Heisenberg, Verh. d. Sächs. Akad. **86**, (1934) 317 W. H. Furry and J. R. Oppenheimer, Phys. Rev. **45**, (1934) 245 H. J. Borchers, D. Buchholz and B. Schroer, Commun.Math.Phys. **219** (2001) 125 R. F. Streater and A. S. Wightman, PCT, Spin and Statistics and all that, New York, Benjamin 1964 R. Haag, *Local Quantum Physics*, Springer, second edition, 1996 R. E. Wald, *The History and Present Status of Quantum Field Theory in Curved Spacetime*, arXiv: gr-qc 0608018 H. Epstein and V. Glaser, Ann. Inst. Henri Poincare A **XIX**, (1973) 211 M. Duetsch and K. Fredenhagen, Commun. Math. Phys. **203**, (1999), 71, hep-th/9807078. F. Coester and W. N. Polyzou, Phys. Rev. D **26**, (1982) 1348 and references therein B. Schroer, *Localization and the interface between quantum mechanics, quantum field theory and quantum gravity I (the two antagonistic localizations and their asymptotic compatibility)* arXiv:0912.2874,  *Localization* .....*gravity II* ((The search of the interface between QFT and QG) arXiv:0912.2886 R. Brunetti, D. Guido and R. Longo, *Modular localization and Wigner particles*, Rev.Math.Phys. **14**, (2002) 759 B. Schroer, Nucl. Phys. B **499**, (1997) 447 G. Lechner, *An Existence Proof for Interacting Quantum Field Theories with a Factorizing S-Matrix*, Commun. Mat. Phys. **227**, (2008) 821, arXiv.org/abs/math-ph/0601022 D. Buchholz, G. Mack and I. Todorov, Nucl.Phys. B, Proc. Suppl. **5B**, (1988) 20 G. Källén and A. S. Wightman, Mat. Fys. Skrifter Kongl. Dansk. Vid. Selsk. **1**, (1958) 6 N. N. Bogoliubov, A. Logunov, A. I. Oksak and I. T. Todorov, General principles of quantum field theory, Dordrecht Kluwer S. Fubini and G. Veneziano, Annals of Physics 63, 12 (1971),  E. Del Giudice, P. Di Vecchia and S. Fubini, Annals of Physics, 70, (1972) 378 C. P. Staszkiewicz, Die lokale Struktur abelscher Stromalgebren auf dem Kreis, Freie Universitaet Thesis Berlin 1995 S. Doplicher and J. E. Roberts, Commun. Math. Phys. **131**, (1990) 51 R. Longo and Y. Kawahigashi, Adv. Math. **206**, (2006) 729, and references therein G. Mack, *D-dimensional Conformal Field Theories with anomalous dimensions as Dual Resonance Models*, arXiv:0909.1024, *D-independent representations of conformal field theories in D dimensions via transformations to auxiliary dual resonance models. The scalar case*, arXiv:0907:2407 J. Mund, B. Schroer and J. Yngvason, *String-localized quantum fields and modular localization*, CMP** 268** (2006) 621, math-ph/0511042 J. Polchinski, String Theory, Cambridge University Press (1998) B. Schroer, *Pascual Jordan’s legacy and the ongoing research in quantum field theory*, in preparation D. Buchholz and K. Fredenhagen, Commun. Math. Phys. **84**, (1982) 1 Y. Nambu, Lectures at the Copenhagen Symposium, 1970, unpublished T. Goto, Progr. Theor. Phys. **46** (1971) 1560 R.C. Brower, Phys. Rev. D **6**, (1972) 1655,  P. Goddard and C.B. Thorn, Phys. Lett. B **40**, (1972) 235 S. Weinberg, *The Quantum Theory of Fields I*, Cambridge University Press 1995 J. Dimock, *Locality in Free String Field Theory II*, Annales Henri Poincaré **3** (2002) 613, math-ph/0102027 E. Martinec, Class. Quant. Grav. **10**, (1993) 187 D. A. Lowe, Phys. Lett. B 326, (1994) 223 J. Scherk anf J. Schwarz, Nucl. Phys. B **81**, (1974) 118 R. Clifton and H. Halvorson, Brit. J. Phil.Sci. **52**, (2001) 417 Arageorgis, J. Earman and L. Rutsche, Philosophy of Science **70**, (2003) D. Fraser, SHPMP **19**, (2008) 841 R. Clifton and H. Halvorson, Stud. Hist. Philos. Mod. Phys. **32** (2001) 1, arXiv:quant-ph/0001107 S. J. Summers, *Yet More Ado About Nothing: The Remarkable Relativistic Vacuum State*, arXiv:0802.1854 H. Araki, *Multiple Time Analyticity of a Quantum Statistical State Satisfying the KMS Boundary Condition*, Publ. RIMS, Kyoto Univ. Ser. A, **4**, (1968) 361 B. Schroer, *BMS symmetry, holography on null-surfaces and area proportionality of “light-slice” entropy*, arXiv:0905.4435 D. Buchholz and S. J. Summers, *Scattering in Relativistic Quantum Field Theory: Fundamental Concepts and Tools*,  arXiv: math-phys/0509047 S. Aks, J. Math. Phys. **6**, (1965) 516 B. Schroer, Annals Phys. 307 (2003) 421, arXiv:hep-th/0106066 T. D. Newton and E. P. Wigner, Rev. Mod. Phys. **21**, (1949) 400 A. B. Zamolodchikov and A. Zamolodchikov, AOP **120**, (1979) 253 B. Schroer Annals Phys.**321**, (2006) 435, and references to previous publications of the author H. Babujian, A. Foerster and M. Karowski, Nucl. Phys.B **736**, (2006) 169 H. Babujian and M. Karowski, Int. J. Mod. Phys. **A1952**, (2004) 34, and references therein to the beginnings of the bootstrap-formfactor program B. Schroer, T.T. Truong and P. Weiss, Phys. Lett. B **63**, (1976) 422 M. Born, Zeitschr. für Physik **38**, (1926) 803 J. H. Lowenstein and B. Schroer, Phys. Rev. D **7**, (1973) 1929 M. Gomes and J.H. Lowenstein, Nucl. Phys. B **45**, (1972) 252 *The crossing property derived from an extended KMS condition*, in preparation K.-H. Rehren, *QFT Lectures on AdS-CFT*, hep-th/0411086 E. P Verlinde, *On the Origin of Gravity and the Laws of Newton*, arXiv:1001.0785 M. Duetsch, K.-H. Rehren, *A comment on the dual field in the AdS-CFT correspondence*, Lett.Math.Phys. 62 (2002) 171 J. A. Maldacena, Adv. Theor. Math. Phys. **2**, (1998) 231 R. Haag abd J. A. Swieca, Commun. Math. Phys. **1**, (1965) 308 D. Buchholz and E. Wichmann, Commun. Math. Phys. **106**, (1986) 321 R. Haag and B. Schroer,J. Math. Phys. **3**, (1962) 248 J. Schwinger, Phys. Rev. **128**, (1962) 2425 J. Schwinger, Trieste Lectures, 1962, p. 89, I.A.E.A., Vienna 1963 J. A. Swieca, Phys. Rev. D **13**, (1976) 312 B. Schroer, *Particle physics in the 60s and 70s and the legacy of contributions by J. A. Swieca*, arXiv:0712.0371 M- Karowski and P. Weisz, Nucl. Phys. B **139**, (1978) 445 O. Zapata Marin, *On Facts in Superstring Theory. A Case Study: The AdS/CFT Correspondence*, arXiv:0905.1439, see also: Spinning the superweb, essays on the history of string theory, http://www.spinningthesuperweb.blogspot.com/ L. Smolin, *The Trouble With Physics: The Rise of String Theory, the Fall of a Science, and What Comes Next*, Sept. 2006 P. Woit, *Not Even Wrong, the failure of string theory and the continuing challenge to unify the laws of physics*, Jonathan Cape London 2006 R. Hedrich, *The Internal and External Problems of String Theory - A Philosophical View*, physics/0610168 K. Fredenhagen, K.-H. Rehren and E. Seiler, Springer Lecture Notes Phys. **721** (2007) 61,  arXiv:hep-th/0603155 L. Fassarella and B. Schroer, J. Phys. **A 35**, (2002) 9123 J. Mund, Commun. Math. Phys. **286**, (2009) 1159 R. Brunetti, K. Fredenhagen and R. Verch, Commun. Math. Phys. **237**, (2003) 31 B. Schroer, *Unexplored regions in QFT and the conceptual foundations of the Standard Model*, arXiv:1006.3543 [^1]: Despite the fact that it is not valid in interacting QFT, its is an ideal intuitive starting point and with some hinsight about the nature of renormalization it carries its user right into the renormalized perturbation formalism. Even though the result does not satisfy the path representation from which everything started, it was a valuable guide to arrive at the correct result. [^2]: An incoming particle changes its position with an outgoing one and, as required by charge conservation, both particles become anti-particles. [^3]: The fate which the S-matrix bootstrap community in conference publications predicted for QFT was “to fade away like a mortally wounded soldier on a battle field”, but little did they know that this would become its own fate shortly after. [^4]: The quantum mechanical hierarchy between elementary and bound particles cannot be maintained in QFT; the only hierarchy which is consistent with interaction caused vacuum polarization clouds is that between basic and fused superselected charges. [^5]: See the most recent one [@Foer] and the references quoted therein. [^6]: I recall warnings by Källén, Lehmann, Jost, Martin and others. [^7]: By this I mean primarily an inner theoretical critical discourse clarifying the conceptual position with respect to the principles underlying previous successful theories. [^8]: In the words of Feynman: “string theory has no arguments instead it uses excuses”, [^9]: The “dynamical” has been added in order to distinguish the intended meaning from the trivial case of an infinite direct sum of irreducible free field representations. In addition to such an infinite mass/spin tower there are also *intertwiners between these representations* without which one cannot generate a mass/spin spectrum. [^10]: In the sequel “cloud” is intended to automatically imply an infinite number of particles. [^11]: The theorem is the algebraic version of the Jost-Schroer theorem, see [@STW]. The latter shows that the existence of a local covariant field which acts on the vacuum as PFG implies that it is a free field whereas the former replaces the pointlike covariance with the affiliation to a compact localized algebra. [^12]: The elegant formulation leading to the well-known useful expressions in terms of correlation functions are due to Lehmann, Symanzik and Zimmermann (LSZ formalism) whereas the proof of the asymptotic convergence towards free fields was supplied by Haag and Ruelle [@Haag]. [^13]: The covariantization of Wigner’s unique representation theoretical classification leads to infinitely many covariant (spinorial) fields (appendix), but most of them do not result from an Euler-Lagrange principle. The latter may be necessary in starting from euclidean functional integral representations, but they are not required in the E-G setting. [^14]: This could be achieved by cluster factorization of $A(x),$ assuming that $A$ has a nontrivial vacuum component. [^15]: The “direct particle interaction” (DPI) [@Co-Po] is a relativistic particle setting which fulfills all properties of relativistic particles which one can formulate in terms of particles only including macro-causality (cluster factorization). Crossing can however not be implemented in such a setting. [^16]: The strongest result is a forthcoming theorem by Jens Mund (private communication) which generalizes the old Jost-Schroer theorem (see [@STW]). [^17]: The reader should pay attention to the changes of notation between expectation values and matrixelements of operators between states. [^18]: Instead of omitting certain contration terms one might as well use the unmodified formfactor and subtract terms of the form $c.t.=\sum_{r=l+1}\delta(p_{1}^{c}-p_{r})\cdot lower~formfactors$ [^19]: Especially recommended to philosophically motivated readers who prefer conceptual clarity over mathematical rigor. [^20]: The added “resonance” expressed the wish to unitarize the model so that it could pass as an S-matrix Ansatz. [^21]: If one uses such formulas outside of the theory of superselected charges one must add the charge conservation by hand; only then does one obtain a Wightman theory in a Hilbert space. [^22]: In higher dimensions it has been shown from first principles that all inner symmetries are described by compact groups [@Do-Ro]. [^23]: The appearance of the logaithmic term is a mark of the formal infrared divergence of the potentials which by themselves (outside their exponential form) are not conformal fields. The label op distinguishes quantum mechanical operators from the numerical momentua (alias chage values) [^24]: The nonlinear S-matrix bootstrap and the Schwinger-Dyson illustrate this point. [^25]: The properly reduced Mellin amplitudes are independent of spacetime dimensions; this is similar (actually closely related) to the invariant part of conformal correlation which only depends on dimension-independent conformally invariant harmonic ratios. [^26]: One needs the conformally invariant part of the correlation, a step which permits no operator formulation. [^27]: The interpretation of the (appropriately defined) Mellin transform as a 4-dimensional dual model is idependent of the spacetime dimensionality of the associated conformal model. For the Fubini et al. [@Fu-Ve] model it is a multi-component abelian chiral current. [^28]: The use of inner symmetry indices of a QFT as an arena for representations of spacetime symmetries is one of the strangest proposals ever made in particle physics. Once accepted, it opened the flood gates for other metaphoric ideas as e.g. the conversion of unwanted spacetime dimensions via “compactification” into inner inner symmetries. [^29]: It requires to pass through apparently unavoidable infinite intermediate steps resulting from the necessity to extract infinite factors coming from reparametrization invariance which have nothing to do with intrinsic properties of particles. [^30]: The different wave functions are distinguished by different relative strength with which the different irreducible components contribute to the mass/spin tower. String theory provides operators which change this decomposition. [^31]: The presence of zero mass photons with an infrared-strong coupling to charged particles results in a weakening in localization of the latter. The optimal (sharpest) localization of the latter is semiinfinite stringlike as described by the well-known Dirac-Jordan-Mandelstam line integral representations. Charged fields interpolate “infraparticles” instead of Wigner particles [^32]: The unresisted acceptance of identifying inner symmetries of conformal symmetries with actual spacetime and its opposite of mutating spacetime dimension into inner symmetries by “rolling them up” (compactification) is an indicator for how much the conceptual framework of QFT principles has been lost and replaced by a collection of computational recipes. [^33]: One has less problems with looking at the source  –$>$ target embedding as a purely formal device. [^34]: Although localization equivalence can be defined between operator algebras which share the same Poincaré representation theory in the same Hilbert space, only the wedge situation leads to the crossing relation. [^35]: More precisely modular theory identifies the range of the two algebras after closing it in the graph norm of shared $\Delta^{\frac{1}{2}}$ which defines the same dense subspace. This domain of Wightman fields is believed to include that subspace but the range of those $B(f,..)$ which are l.e. in the expained sense is smaller (see later). [^36]: Such operators were called PFGs (vacuum-polarization-free-generators). They allow to generalize the Jost-Schroer theorem (saying essentially that interacting theories cannot have compact localizable PFGs) and play a crucial role in the modular construction of factorizing models (see next section). [^37]: PFGs do not exist for causally complete subwedge regions unless the theory is generated by a free field. (stronger than the triviality of scattering). The wedge is the “smallest” causally closed region for which PFGs exist, though generally only at the prize of nontranslational invariant domains. Well behaved (“temperate”) PFGs for Ws only exist in d=1+1. [^38]: In the case of non selfconjugate particles the $q$-momenta refer to antiparticles and it would be better to use the notation $\bar{q}.$ [^39]: The explicit computation of the action of a modular $S$- operator on a state generated by a $\Delta$-related operator algebra is generally a difficult problem. [^40]: J. Mund and B. Schroer: “A generalized KMS condition and its relation to the crossing property” in preparation. [^41]: Note that not even in the standard perturbative setting the time ordered product is simply the time-ordering of the unordered product in any naive sense. [^42]: I am indepted to Jens Mund who informed me about a forthcoming paper by him on this generalization of the Jost-Schroer theorem. [^43]: The particle statistics (Bosons, Fermion) is used to bring the rapidities into the natural order. The n! natural orders are generally belonging to n! analytic functions which are not analytic continuations of each other. [^44]: The knowledge about the system however decreases with increaing localization size. [^45]: As a curiosity I remember how one of my Ph.D students (Bernd Berg) in the beginning of the 70s demonstrated such statements numerically on one of the old Hewlett-Packard pocket calculators. [^46]: Born [@Born] introduced this probability concept first in the setting of scattering theory (the Born approximation for the cross section) before it was extended to x-space wave functions. [^47]: In the literature one sometimes encounters an “effective” version stating that covariance is attained for distances beyond the Compton wave length.. [^48]: Usually premature observations disappear and return often in a different context when the understanding of their conceptual-mathematical struture is in place [@Jor]. [^49]: The first version of the present paper was uploaded to arXiev:hep-th when a moderator placed it to the general physics setion with a built-in barrier to prevent any crosslisting of the paper. There is no more fitting description of the present sociological state of particle theory, any commentary about this episode is superfluous. [^50]: The begin of modern string theory has a date, it is the week in Paris in 1974 when Scherk and Schwarz [@S-S] wrote up their famous paper. Underlining the rapidity of change one may call it the Bartholomew-like massacre of the old string theory which started with phenomenology of Regge trajectories. [^51]: The attempts to construct infinite component irreducible (in the described sense) pointlike fields based on higher noncompact group representations (similar to the O(4,2) hydrogen spectrum) are described in [@Tod]. Unfortunately there was no communication between the two groups of which only the string construction was successful (a success certainly not appreciated by string theorists). [^52]: Although there is no theorem that a net of local algebras is always generated by local fields, the lack of any counterexample suggest that even if this does not hold for all local nets, it is valid for a large subset. [^53]: Neither in the case of the AdS-CFT correspondence, nor in the case of holographic projections on the horizon (a nullsurface) of a bulk region, the dimension-changing holographic map can be expressed in terms of pointlike fields. [^54]: If the supersymmetric N=4 Yang-Mills theory would be the only 4-dimensional CFT, then the correspondence would be unique. [^55]: The begin of modern string theory has a date, it is the week in Paris in 1974 when Scherk and Schwarz [@S-S] wrote up their famous paper. Underlining the rapidity of change one may call it the Bartholomew-like massacre of the old string theory which started with phenomenology of Regge trajectories. [^56]: With somewhat different motivations and lesser mathematical rigor see also [@Sch]. [^57]: The construction works for arbitrary positive energy representations, not only irreducible ones. [^58]: The wedge reflection $\mathfrak{j}_{W_{0}}$ differs from the TCP operator only by a $\pi$-rotation around the W$_{0}$ axis. [^59]: According to the Reeh-Schlieder theorem a local algebra $\mathcal{A(O})$ in QFT is in standard position with respect to the vacuum i.e. it acts on the vacuum in a cyclic and separating manner. The spatial standardness, which follows directly from Wigner representation theory, is just the one-particle projection of the Reeh-Schlieder property. [^60]: It is quite easy to prove the standardness for spacelike cone localization (leading to singular stringlike generating fields) just from the positive energy property which is shared by all three families [@BGL]. [^61]: The epithet “generating” refers to the tightest localized singular field (operator-valued distribution) which generates the spacetime-indexed net of algebras in a QFT. In the case of localization of states the generators are state-valued distributions. [^62]: The terminology 2$^{nd}$ quantization is a misdemeanor since one is dealing with a rigorously defined functor within QT which has little in common with the artful use of that parallellism to classical theory called “quantization”. In Edward Nelson’s words: (first) quantization is a mystery, but second quantization is a functor. [^63]: The functor $\Gamma$ preserves the standardness i.e. maps the spatial one-particle standardness into its algebraic counterpart. [^64]: In QFT any finite energy vector (which of course includes the vacuum) has this property as well as any nondegenerated KMS state. In the mathematical setting it is shown that standard vectors are “$\delta-$dense” in $H$. [^65]: The necessarily singular pointlike fields are universal generators for algebras of arbitrary (small) localization. [^66]: There is a complication in low-dimensional theories in which braid group statistics may occur in which case there is no sharp separation between inner and spacetime symmetries. Nevertheless these representations appear in the DHR theory [@Haag]. [^67]: In this sense they are like points in geometry except that monads can be mutually included and intersected.

ArXiv

--- abstract: 'Extended main-sequence turnoffs (eMSTO) have been observed in the color-magnitude diagram (CMD) of intermediate-age and young star clusters. The origin of the eMSTO phenomenon is still highly debated. Calculations show that the blue and faint (BF) stars in the CMD of NGC 1866 are hydrogen main sequence (MS) + naked He star systems. The He star derives from the massive star of a binary system. The BF stars and the red and faint MSTO stars belong to the same stellar population. The values of $m_{F336W}$ and $m_{F336W}-m_{F814W}$ of the BF stars are mainly determined by the masses of He stars and H-MS stars, respectively. The behaviors of the BF stars in the CMD are well explained by the H-MS + He-star systems. The BF stars provide a strict restriction on the age of the stellar population. Moreover, the bimodal MS of NGC 1866 can also be reproduced by a younger binary population. The calculations show that part of the blue and bright (BB) MS stars of NGC 1866 are H-MS + He-star systems, H-MS + white dwarf systems, and merged stars in a binary scenario. The H-MS stars of the H-MS + He-star systems for the BB stars are significantly more massive than those of the BF stars. Once the H-MS + He-star systems and their membership in NGC 1866 are confirmed, the extended star-formation histories and the effects of binaries can be confirmed in the young star cluster.' author: - Wuming Yang title: 'The Effects of Binary Stars on the Color-Magnitude Diagrams of Young-Age Massive Star Clusters' --- INTRODUCTION ============ In the classical theory of star formation, a star cluster is considered to be composed of stars belonging to a simple, single stellar population (SSP) with a uniform age and chemical composition. However, the discoveries of double or extended main-sequence turnoffs (eMSTO) [@mack07; @glat08; @gira09; @goud09; @milo09] in the color-magnitude diagram (CMD) of intermediate-age star clusters in the Large Magellanic Clouds (LMC) are challenging the classical hypothesis. The phenomenon of eMSTOs has been interpreted as meaning that the clusters have experienced extended star-formation histories (eSFH) with durations of $\sim$100-700 Myr [@glat08; @mack08; @milo09; @gira09; @rube10; @goud11; @goud14; @corr14]. As an alternative interpretation, eMSTOs are thought to be due to coeval populations with different rotation rates [@bast09; @yang13; @lic14; @bran15; @dant15; @nied15a] or due to interacting binaries [@yang11; @liz12; @liz15]. Moreover, stellar variability [@sali16; @degr17] and a variable overshoot of the convective core of stars [@yang17] may play a potential role in shaping the eMSTO regions as well. The nature of the eMSTO phenomenon is still highly debated. The problem of the origin of eMSTOs becomes more complicated when they and split main sequences (MS) are found in the CMD of young clusters (less than $\sim 400$ Myr) in the LMC. The split MS was first discovered in the CMD of young cluster NGC 1844 [@milo13]. However, the eMSTOs of young clusters were first observed in NGC 1856 by [@corr15] and [@milo15]. Young cluster NGC 1856 hosts an eMSTO and a double MS, which changes our understanding of young clusters in the LMC. The eMSTO of NGC 1856 can be interpreted as the superposition of two main populations having the same age but different rotation rates [@dant15] or as the effects of a variable overshoot of the convective core of stars [@yang17]. [@bast17] have inferred the existence of rapidly rotating stars in NGC 1856 and NGC 1850 from H-alpha excess fluxes likely being due to so-called Be candidate stars. Moreover, eMSTOs and bimodal MS have been observed in young clusters NGC 1755 [@milo16], NGC 1850 [@bast16; @corr17], and NGC 1866 [@milo17]. Neither stellar populations with different ages only, nor coeval stellar populations featuring a distribution of stellar rotation rates, properly reproduce the observed split MS and eMSTO [@milo17; @corr17]. eMSTOs were found in young clusters NGC 330, NGC 1805, NGC 1818, and NGC 2164 as well [@lic17]. [@lic17] also show that the observed eMSTOs cannot be explained by stellar rotation alone. Similar cases have been found in intermediate-age star clusters NGC 1987 and NGC 2249, whose eMSTOs cannot be explained solely by a distribution of stellar rotation rates [@goud17]. For these clusters, a combination of rotation and an age spread seems to be required to explain observational results [@milo16; @milo17; @corr17; @lic17; @goud17]. [@piat17] analyzed the data of young cluster NGC 1971 and found that NGC 1971 exhibits an eMSTO originated mostly by a real age spread. Moreover, [@dupr17] obtained the spectra of 29 eMSTO stars in NGC 1866. The direct spectroscopic measures clearly demonstrate the presence of rapidly rotating stars that are cooler than a population of slowly rotating objects, arguing for an actual spread in age of NGC 1866. However, [@lema17] studied the chemical composition of several Cepheids located in NGC 1866 and found that six Cepheids have a homogeneous chemical composition and are consistent with the red giant branch in the cluster. Their analysis shows that the Cepheids belong to the same stellar population. In line with the comment of [@milo17] on their observational results of NGC 1866, these observations raise many more questions than they solve. In addition to the main characteristics of the eMSTO and bimodal MS described by [@milo17], Figure \[fig1\] shows that NGC 1866 has two main MSTOs and that there are many blue and bright (BB) stars and blue and faint (BF) stars in the CMD of NGC 1866. There is a gap between the BB stars and the blue or red MS stars and an upper limit of luminosity for the BF stars; that is, the value of $m_{F336W}$ for most of the BF stars is larger than $\sim 21$. An age spread or the effects of rotation cannot produce such BF stars. The behaviors of the BF and BB stars in the CMD might result from binaries. If the BF and BB stars are members of NGC 1866, they provide different perspectives on the populations of NGC 1866 and aid us in understanding the nature of the eMSTO phenomenon in young clusters. In the present work, we mainly focus on whether the characteristics of NGC 1866 can be reproduced by binaries. The paper is organized as follows. Some initial assumptions are given in Section 2, calculation results are shown in Section 3, and the results are discussed and summarized in Section 4. STELLAR MODELS AND POPULATION SYNTHESIS ======================================= For a binary system, the mass of the primary star, $M_{1}$, is generated in terms of the lognormal initial mass function (IMF) of [@chab01]. The mass of the secondary star is then determined by $qM_{1}$, where the $q$ is the ratio of the mass of the secondary to that of the primary and is generated according to an assumed distribution. The distribution of separations ($a$) between the primary and the secondary stars is assumed to be constant in $\log a$ [@han95]: $$an(a)=\alpha_{a},$$ where $\alpha_{a} \approx 0.12328$. The eccentricity ($e$) of each binary system is assumed to be a uniform distribution within $0-1$. In this study, the initial metallicity $Z$ was fixed at $0.008$. Once the initial distributions of the masses $M_{1}$ and $M_{2}$ ($qM_{1}$), separation $a$, and eccentricity $e$ were given by Monte Carlo simulation, the sample was evolved to a given age by using the Hurley rapid single/binary evolution codes [@hurl00; @hurl02] to obtain the luminosities and effective temperatures of the stellar population with the given age. The metallicity $Z$ was converted into \[Fe/H\] by \[Fe/H\] $= \lg(Z/X)- \lg(Z/X)_{\odot} \simeq \lg(Z/Z_{\odot})$. There is a bitter controversy between helioseismology and observation about solar metallicity [@yang16 and the references therein], but the value of $Z_{\odot}$ in our calculations is $0.02$. The quantities (\[Fe/H\], $T_{\rm eff}$, $\log g$, $\log L$) of evolutionary models were then transformed into colors and magnitudes using the color transformation tables of [@leje98]. In computing their colors and magnitudes, the binaries with $a \le 10^{4} R_{\odot}$ were treated as a single point-source object according to the formulas in [@zhan04]. A distance modulus $(m-M)_{0}=18.3$ and reddening $E(B-V)=-0.01$ were adopted in our calculations. CALCULATION RESULTS =================== Results of a Uniform $q$ Distribution ------------------------------------- There are about $9900$ objects with $15.0\leq m_{F336W}\leq 23.0$ and $-2.0\leq m_{F336W}-m_{F814W}\leq 1.0$ in the observed data of NGC 1866 [@milo17]. With the assumption that the mass ratio $q$ is a uniform distribution within $0-1$, a sample was evolved to given ages. The CMDs of the simulated populations with different ages are shown in Figure \[fig2\]. In each panel of Figure \[fig2\], there are about $4000$ simulated objects with $15.0\leq m_{F336W}\leq 23.0$ and $-2.0\leq m_{F336W}-m_{F814W}\leq 1.0$, where the merged stars make up about $13\%$. Others are binaries. In our synthesized populations, we included observational errors taken to be a Gaussian distribution with a mean value of $0$ and a standard deviation of $0.025$ in magnitude and color. The calculations show that the eMSTO and bimodal MS of NGC 1866 cannot be reproduced by the effects of the binaries alone. In order to reproduce the eMSTO region, an age spread of about $150$ Myr (from 190 to $340$ Myr) is required in the simulation. The simulation produced a few BF stars and minimal BB stars. The results cannot exclude the effects of binaries. The mass-ratio $q$ is the key parameter determining the evolutions of binaries, so the calculated results can be affected by the distribution of $q$. Results for Binary Populations with Different $q$ Distributions --------------------------------------------------------------- ### The binary population reproducing the BF stars In order to study whether the BB and BF stars can be reproduced by the evolutions of binaries, we computed the evolutions of a sample of binaries with an uncorrelated $q$. A sample was first generated at random from the IMF as primary stars. Then secondary stars were generated at random from the same IMF. We obtained about $3000$ objects with $15.0\leq m_{F336W}\leq 23.0$ and $-2.0\leq m_{F336W}-m_{F814W}\leq 1.0$ in each simulation, with the merged stars making up around $13\%$. The CMDs of the simulated population are shown in Figure \[fig3\]. For clarity, when the mass of the secondary star evolved to the given age is larger than that of the primary star, the value of the mass ratio is redefined as $1/q$ in the Figures. The simulations cannot reproduce enough BB stars, but they can generate some BF stars (see Figure \[fig3\]), which indicates that the BF stars are relevant to the evolutions of binaries. The initial value of $q$ of the simulated BF stars is mainly between $3$ and $4$, and the values of initial $M_{1}+M_{2}$ of the stars are mainly between about $4$ and $5.5$ for the population with the age of $340$ Myr. Not all binary systems with the special $q$ and initial $M_{1}+M_{2}$ between $4$ and $5.5$ can evolve into BF stars; only some of the binaries can do so. According to the IMF, the more massive the stars, the smaller their number. If the secondary mass is determined by $qM_{1}$ and the initial $q$ is a uniform distribution within $0-1$, the BF stars could hardly be reproduced in a cluster because there are not enough binary systems with the initial $q$ between $0.25$ and $0.33$ and $M_{1}+M_{2}$ between $4$ and $5.5$ . The presence of a large number of BF stars indicates that the initial $q$ for the population of NGC 1866 may follow other distributions or that the BF stars do not belong to NGC 1866. To reproduce the BF stars, the distribution of $q$ is assumed to be $$n(q)=\frac{2q}{\mathbf{\beta}},$$ where $\mathbf{\beta}$ is a free parameter. Then mass-ratio $q$ is generated at random by $$q=\sqrt{\mathbf{\beta}}\times\sqrt{r_{i}},$$ where $r_{i}$ is a random number within $0-1$. Figure \[fig4\] represents the CMDs of simulated populations with different values of $\mathbf{\beta}$ but with the same other parameters, which shows that the BB and BF stars of NGC 1866 are reproduced well by the simulations with $\sqrt{\mathbf{\beta}}= 4$ or $5$ (see panels $b$ and $c$ of Figure \[fig4\]). When $\sqrt{\beta} \lesssim 4$, the number of simulated BB and BF stars increases with the increase in $\sqrt{\beta}$. The smaller the $\sqrt{\beta}$, the redder the produced BF stars. The BF stars of NGC 1866 cannot be reproduced by the simulations with $\sqrt{\beta} < 4$. The larger the value of $\sqrt{\beta}$, the more massive the star produced by $qM_{1}$. When the value of $qM_{1}$ is larger than a certain value, the star evolves to a later stage rather than H-MS or He-MS stage at the age of $340$ Myr. Thus, when $\sqrt{\beta} \gtrsim 5$, the number of simulated BB, BF, and MS stars decreases with an increase in $\sqrt{\beta}$. In order to produce the same number of BF stars, the number of initial models of simulation with $\sqrt{\beta} = 6$ is $1.5$ times as large as that with $\sqrt{\beta} = 5$. The main difference between a simulated population with $\sqrt{\beta} = 4$ and that with $\sqrt{\beta} = 5$ is that a simulation with $\sqrt{\beta} = 5$ can produce some BF systems with $-1.5 \lesssim m_{F336W}-m_{F814W} \lesssim -1.0$ but a simulation with $\sqrt{\beta} = 4$ cannot produce BF systems with $-1.5 \lesssim m_{F336W}-m_{F814W} \lesssim -1.0$ (see Figure \[fig5\]). Moreover, the number of BB and MS stars of the population with $\sqrt{\beta} = 5$ is slightly lower than that of the population with $\sqrt{\beta} = 4$. Furthermore, the simulation with $\sqrt{\beta} = 5$ produces a larger number of bright stars that deviate from observation than the simulation with $\sqrt{\beta} = 4$ (see panels $c$ and $d$ of Figure \[fig5\]). The BF stars of NGC 1866 seem to have a gap between $m_{F336W}-m_{F814W} \approx -1.0$ and $ m_{F336W}-m_{F814W} \approx -1.4$ (see panel $a$ of Figure \[fig5\]), which could be used to limit the value of $\beta$. Fixing the value of $\sqrt{\mathbf{\beta}}$ at $4$, we computed stellar populations with different ages. The results are represented in Figure \[fig6\]. The calculations show that the BF stars and the red and faint MSTO of NGC 1866 are reproduced well by the populations with an age between about $320$ and $340$ Myr. The BF stars and the red and faint MSTO belong to the same population. There are almost no BF stars with $m_{F336W} < \sim21$ for the stellar populations (see panels $c$ and $d$ of Figure \[fig6\]). A large number of BB stars were reproduced by the calculations as well. ### The blue and faint stars For both the simulated and observed populations, there are only a few BF stars brighter than $m_{F336W} \sim 21$ (see Figure \[fig5\] and panels $c$ and $d$ of Figure \[fig6\]), which indicates that the luminosity of $m_{F336W}$ for BF stars with the given age has an upper limit. Our calculations show that a BF star is a binary system consisting of a hydrogen MS star and a naked helium star. The initial value of $q$ of the system is mainly between about $2.5$ and $4$. The He stars evolve from the massive stars of the systems. The initial masses of the massive stars are mainly in the range of $\sim 3.0 - 3.5$ for the population of the age of $340$ Myr. More massive stars have evolved into later phases, but lower mass stars have not yet evolved into He stars. The masses of the He stars of BF systems are mainly between about $0.45$ and $0.54$ , but the masses of the H-MS stars are mainly in the range of $\sim 0.7 - 1.3$ . The value of $m_{F336W}$ of a BF system is mainly dependent on its He star because the value of $m_{F336W}$ of its H-MS star is much larger than that of the He star when the mass of the H-MS star is lower than $1.4$ (see Figure \[fig7\]). The luminosity of a He star is determined by its mass. The value of $m_{F336W}$ of the He-MS stars with masses in the range of $\sim 0.50 - 0.54$ is around $21$. More massive He stars have evolved into the Hertzsprung Gap or later phases. The lifetimes of the phases are very short, which leads to the fact that there are only a few BF stars with $m_{F336W} <$ $\sim21$ in our simulations. Therefore, the BF candidates cannot blend with BB stars in the CMD unless the number of H-MS + He-star systems is large enough to contain many more massive He stars at the given age. The luminosities of $m_{F814W}$ of the BF systems are mainly dependent on the H-MS stars of the systems because the values of $m_{F814W}$ of the H-MS stars are much lower than those of the He stars (see Figure \[fig7\]). But the values of $m_{F336W}$ of the systems only slightly decrease with an increase in mass of the H-MS stars. Thus the variation in mass of the H-MS stars mainly affects the values of $m_{F336W}-m_{F814W}$. For the BF stars, the more massive the H-MS stars, the lower the values of $m_{F814W}$, and the larger the values of $m_{F336W}-m_{F814W}$; but the more massive the He stars, the lower the values of $m_{F336W}$. The characteristics of the BF stars of NGC 1866 are consistent with those of H-MS + He-star systems (see Figure \[fig8\]). However, when the H-MS star is massive than about $1.4$ , the luminosity of the H-MS + He-star system can be significantly affected by the H-MS star. When the H-MS star is massive than about $1.5$ , the values of $m_{F336W}$ and $m_{F814W}$ of the system are mainly determined by the H-MS star. The system appears as a MS star rather than a BF star. As a consequence, there is an upper limit of $m_{F336W}$ for BF stars with the age of $340$ Myr at $m_{F336W} \sim 21$. The increase in mass of the H-MS stars of BF systems cannot lead to the fact that BF candidates blend with the BB stars in the CMD. The BF stars with an approximate $m_{F336W}$ are almost in line in the CMD due to the difference in mass of H-MS stars (see panels $c$ and $d$ of Figure \[fig6\]). These characteristics can be used to restrict the age of young star clusters. If the age of the BF stars is younger than about $320-340$ Myr, the value of the upper limit of $m_{F336W}$ will be smaller than $21$ due to the presence of more massive He stars (see panel $a$ of Figure \[fig6\]). In the observed sample, there are only several BF stars with $m_{F336W}$ between about $20$ and $21$, but a large number between around $21$ and $23$. Both the BF stars and the red and faint MSTO stars belong to the same population. This indicates that there is a stellar population as old as $\sim320-340$ Myr in NGC 1866. Moreover, BF stars can only derive from binary systems with an initial mass ratio in a narrow range. Thus there should be many binaries in NGC 1866. The orange dots in the lower-left corners of the panels of Figure \[fig8\] represent He-star + white dwarf (WD) systems. The magnitudes of these systems are almost entirely determined by their He stars. Thus they look like a single He star and have an approximately equal $m_{F336W}-m_{F814W}$ or $m_{F555W}-m_{F814W}$. Figure \[fig7\] shows that the longer the effective wavelength or the more massive the H-MS star, the more easily is the magnitude of H-MS + He-star system affected by the H-MS star. Therefore, the H-MS + He-star systems are more easily separated from MS by $m_{F336W}-m_{F814W}$ rather than by $m_{F555W}-m_{F814W}$. The lower the mass of a H-MS star, the smaller is the contribution of the H-MS star to the $m_{F336W}$ and $m_{F555W}$ of the H-MS + He-star system; so the BF systems with lower-mass H-MS stars are closer to WD + He-star systems in $m_{F336W}-m_{F555W}$ (see panel $c$ of Figure \[fig8\]). This indicates that BF systems are more easily displayed on $m_{F336W}-m_{FxxxW}$, where the $xxx$ represents the effective wavelength of other filters, such as $450$, $555$, or $814$. It indicates, too, that BF systems have different behaviors in different colors, which can aid us in confirming the BF systems. The simulation with a larger $\beta$ can produce the H-MS + He-star systems with a lower-mass H-MS star, i.e. bluer BF stars when $\sqrt{\beta} \leq 5$, which leads to the fact that simulations with $\sqrt{\beta} =4$ cannot produce the BF stars with $m_{F336W}-m_{F814W} \lesssim -1.0$. ### The blue and bright stars The simulations also produce many BB stars that are mainly merged stars, H-MS + He-star systems, and H-MS + WD systems. Such stars are also called blue stragglers [@stro70; @pols94]. The H-MS stars had accreted mass from their companions. The value of initial $q$ of the systems is mainly in the range of around $1.5 - 2.5$. As with BF stars, the initial masses of the massive stars that evolved into the naked He-MS stars are mainly between about $3.0$ and $3.5$ ; but the initial masses of the H-MS stars are higher than those of BF systems. The masses of the H-MS stars of BB systems are mainly between about $2.5$ and $4$ , which are obviously more massive than those of BF systems. The luminosities of the BB systems are more dependent on their H-MS stars than on their He stars. The BB stars thus look like MS stars and are more luminous than BF systems. The accretion and merging make the stars bluer and brighter than the MSTO stars. The value of mass accretion rates can affect the luminosity of stars, but the mass accretion rate and the process of merging are not definitely known in the theory of stellar evolution. The more the mass accreted by a star, the larger is the star’s luminosity. The uncertainty of the rate could give rise to the fact that the simulated BB stars are more scattered than the observed ones in the CMD (see Figures \[fig6\] and \[fig8\]). The contribution of the He stars to the $m_{F336W}$ of the systems is larger than to the $m_{F814W}$ of the systems. The radius of the He stars is of the order of $0.1$ . The He stars of the systems could be outshone by their companions or disks and be difficult to observe. This may lead to the fact that the observed BB stars are slightly fainter and redder than the theoretical models in the CMD. Some BB stars are H-MS + WD systems. The masses of the H-MS stars are mainly between about $3$ and $4$ , but those of the WDs are mainly in the range of $\sim 0.55 - 0.7$ . The luminosities of the systems are mainly determined by the H-MS stars. Thus they look like blue MS stars rather than WDs. ### The blue MS Panel $a$ of Figure \[fig6\] shows that the simulated stellar population with the $q$ distribution and age of $190$ Myr cannot reproduce the blue MS of NGC 1866. Moreover, the population contains many BF stars with $m_{F336W}$ between about $20$ and $21$. The absence of BF stars with $m_{F336W}$ between about $20$ and $21$ in the observed sample indicates that properties of the blue MS stars of NGC 1866 should be different from those of the simulated population. We computed the evolutions of binary populations with a Gaussian mass-ratio distribution. The mean value and standard deviation of the Gaussian distribution is $0.6$ and $0.11$, respectively. Figures \[fig9\] and \[fig10\] show that the bimodal MS and blue MS of NGC 1866 can be reproduced by the binary population with an age of about $190$ Myr. The blue stragglers with $m_{F336W} \sim 16$ and $m_{F336W}-m_{F814W}< -1.0$ of NGC 1866 are reproduced as well. As we have just said, however, the BF stars do not appear in this simulation; so this young population is not incompatible with the constraints of the observed BF stars. The blue MS is mainly composed of merged stars and binaries with a $q$ less than about $0.5$ (see Figure \[fig11\]). Most of the simulated systems have an initial $q$ larger than $0.5$, which leads to the fact that the red MS is denser than the blue MS. Mass accretion and merging can lead to the fact that the number of stars with masses in a certain range can increase, but those with masses in another range decreases ( see Figure \[fig11\]). This results in the fact that the blue MS is discontinuous in the CMD. The simulated blue stragglers are composed of merged MS stars and H-MS + He-star systems. The masses of the He-MS stars are mainly between about $0.6$ and $0.7$ . As we have noted, the H-MS stars had accreted mass from their companions. The masses of these stars and the merged MS stars are mainly in the range of $\sim 5.0$ and $6.5$ , which is much larger than the masses of the MSTO stars of the SSP with the age of $190$ Myr. The luminosity of the H-MS + He-star system is mainly dependent on the H-MS star rather than the He star. Thus the system is a blue and bright star rather than a blue and faint star. Moreover, part of the simulated blue MS ($m_{F336W} < 19$) are merged MS stars, H-MS + He-star systems, and H-MS + WD systems. The masses of the He-MS stars are mainly between about $0.6$ and $0.7$ ; but those of the WDs mainly between the range of $\sim 0.9 - 1.0$ . For this young stellar population, the masses of the He stars of H-MS + He-star systems are mainly between about $0.6$ and $0.7$ , which are larger than those of the stellar population with the age of $340$ Myr. Figure \[fig12\] presents the CMD of simulated multiple populations that are the mixture of the binary population characterized by Gaussian mass-ratio distribution and an age of $190$ Myr with the population with $\sqrt{\beta}=5$ and an age of $340$ Myr. It shows that the blue and the red MS gradually merge when the value of $m_{F336W}$ is larger than $20$. When the value of $m_{F336W}$ is less than $20$, however, the separation between the blue and the red MS is almost not affected by the old population. ### Other Clusters The CMDs of NGC 1806 and NGC 1856 have been given by [@milo09; @milo15]. There are almost no BF stars in the observed CMDs. With the same initial parameters and $\sqrt{\beta}=5$, we computed stellar populations with different ages. The blue stragglers and the blue and bright MSTO of NGC 1806 can be reproduced by a population with the age of $1.35$ Gyr (see Figure \[fig13\]); but there are only a few BF stars in the simulated population. The faint and red MSTO of NGC 1856 can be reproduced by a population with the age of 600 Myr, and the few BF stars of NGC 1856 also are reproduced by the simulated population. The observations and simulations show that BF stars are correlated with the age of clusters. They appear more easily in a young cluster. DISCUSSION AND SUMMARY ====================== Discussion ---------- We noticed that the ages of binary populations are larger than those given by [@milo17], which may be due to the effects of binaries. Similar to [@milo17], we obtained an age of about $150$ Myr for NGC 1866 when we used a SSP to fit the blue MS of NGC 1866 (see Figure \[fig14\]). Moreover, the simulated age of a cluster is related to \[Fe/H\]. The uncertainty of \[Fe/H\] can affect the age. Assuming a constant mass-ratio distribution, [@milo17] found that the fraction of binaries in NGC 1866 is about $0.28$. In our simulations, in the primary stage of evolutions all models are members of binaries. When the models are evolved to the age of $190$ or $340$ Myr, the merged stars make up only about $12\%-14\%$ of the obtained objects; but all the others are still binaries. Moreover, all the simulated BF stars are binaries. The ratio of merged He stars to He-star + WD systems in Figures \[fig8\] and \[fig15\] is around $0.08-0.12$. Our binary fraction is much larger than that given by [@milo17]. The IMF does not affect our results. Basing on [@salp55] IMF, we obtained almost the same results (see Figure \[fig15\]). The difference between the results obtained from [@salp55] IMF and those based on [@chab01] IMF can be neglected. The value of $E(B-V)$ is $-0.01$ for NGC 1866, $0.08$ for NGC 1856, and $0.16$ for NGC 1806, which may be related to the mixing-length parameter. That parameter is unadjustable in the [@hurl02] codes and is calibrated to a solar model. The larger the parameter, the smaller the radius of stellar models, and the higher their effective temperature. If the value of the mixing-length parameter increases with an increase in the mass of stars but is fixed at the value calibrated to a low-mass star, one could find that for clusters with the same reddening the value of theoretical $E(B-V)$ decreases with a decrease in age of the clusters because the masses of MSTO stars increase with the decrease in age of the clusters. The rotation velocity of merged stars is hard to estimate. If they lost angular momentum as they merged, they are slowly rotating stars; but if their angular momentum was conservative, they are fast rotators. In order to distinguish the merged stars from binaries in Figure \[fig16\], their velocities are assumed to be $150$ km s$^{-1}$. We calculated rotational velocities and orbital velocities with the assumption that the rotation rate of a star is equal to the revolution rate of the binary system. The rotational and orbital velocities of many stars are of the order of $100$ and $200$ km s$^{-1}$, respectively; but orbital velocities can be as high as about $300$ km s$^{-1}$. The fraction of MSTO stars with orbital velocities higher than $200$ km s$^{-1}$ in the old population is larger than that in the young population (see panels $b$ and $d$ of Figure \[fig16\]). The eMSTO and bimodal MS of NGC 1866 can be explained by a combination of rotation and age spread [@milo17; @corr17]; but the eMSTO can also be explained by the effects of a large overshoot of the convective core of stars [@yang17]. A single He star, WD, or MS star cannot appear as a BF star, BB star, or blue straggler. Thus the H-MS + He-star systems and H-MS + WD systems are the main characteristics that can distinguish the binary scenario from the rotation scenario and the overshoot scenario. The calculations show that the BF stars are H-MS + He-star systems. The H-MS star is a low mass star, and it hardly accretes mass from its companion. The BF stars and the red and faint MSTO belong to the same population. The BF stars can provide a constraint on the age of the stellar population. Once the BF stars are determined to be the members of the cluster, the existence of an old stellar population in the cluster will be confirmed. The initial $q$ of the H-MS + He-star systems is in a narrow range. The observed BF stars could be used to estimate the fraction of binaries in the cluster. The simulated bimodal MS is sensitively dependent on the mean value of the Gaussian mass-ratio ($q$) distribution and can be reproduced when the mean value is in the range of about $0.6-0.67$. For a population with an age of $340$ Myr, the masses of He stars and WDs are mainly within the range of $\sim0.46-0.54$ and $\sim0.55-0.7$ , respectively; but for that with an age of $190$ Myr, their masses are mainly within the range of $\sim0.6-0.7$ and $\sim0.9-1.0$ , respectively. The fundamental parameters of binary stars are more easily determined from light curves than those of single stars. If the young cluster consists of two main populations, there would be a bimodal distribution of the mass of He stars. Furthermore, part of the blue stragglers are massive H-MS + He-star systems. If they derive from a young population with an age of about $190$ Myr, the masses of the He stars are between about $0.6$ and $0.7$ , but those of the H-MS stars are between around $5$ and $6.5$ . The masses of the H-MS and He stars of the BF systems of the stellar sub-population with the age of $340$ Myr are virtually all lower than $1.4$ and $0.54$ , respectively. If one can determine that part of the blue stragglers are the massive H-MS + He-star systems and that the BF stars are the low mass H-MS + He-star systems, that will confirm the existence of the eSFH and the role of binaries. **Photometric errors for bright MS stars are around $0.02$ mag in color [@corr14; @milo16]. The errors could increase with an increase in magnitude. The characteristics of the BF population are mainly dependent on He stars. The values of $m_{F336W}-m_{F814W}$ of the BF population are mainly between about $-2.0$ and $-0.3$, i.e. the change of $|m_{F336W}-m_{F814W}|$ is large than $0.1$. Even photometric errors increase from $0.02$ to $0.1$ mag, our results for the BF population are not changed (see Figure \[fig17\]); but the simulated MS broadens.** Summary ------- In this work, we calculated different binary populations. The calculations show that the BF stars of NGC 1866 are H-MS + He-star systems, which derive from the evolutions of binaries with the initial $q$ in the range of about $2.5-4$. The He star comes into being from the evolution of the massive star of the system. The value of $m_{F336W}$ of a BF star is mainly dependent on the mass of the He star. The mass of most of the He stars is lower than $0.54$ for the population with the age of $340$ Myr, which leads to the fact that the $m_{F336W}$ of most of the BF stars is larger than $\sim 21$. The value of $m_{F336W}-m_{F814W}$ of a BF star mainly depends on the mass of the H-MS star that lies mainly within the range of about $0.7-1.3$ . The more massive the H-MS star, the larger the value of $m_{F336W}-m_{F814W}$ for a BF star. However, when the mass of the H-MS star is larger than $1.5$ , the system looks like an MS star rather than a BF star. If the population of BF stars can be confirmed to be members of NGC 1866, this would directly show that NGC 1866 hosts a population older than the blue MS, otherwise a mechanism making some stars bluer and brighter, such as variable overshoot, is required to explain the blue MS. The bimodal MS of NGC 1866 can be produced by the binary population with the Gaussian $q$ distribution and the age of $190$ Myr. The calculations show that the BB stars of NGC 1866 are mainly merged stars, H-MS + He-star systems, and H-MS + WD systems, which leads to the discontinuities between the BB stars and the blue MS. The H-MS + He-star systems and H-MS + WD systems are the main characteristic of the binary models, and this can be used to confirm or exclude the binary scenario. Moreover, the calculations show that the blue stragglers of NGC 1866 and the blue MS belong to the same population. The blue stragglers consist of the merged stars and H-MS + He-star systems. The masses of the He stars are mainly between about $0.6$ and $0.7$ , which are larger than those of the BF stars. Therefore, the existence of the blue stragglers and the BF stars can confirm the existence of the eSFH. To explain the eMSTO and bimodal MS of NGC 1866, a combination of an age spread and binary population is required. The role of binaries in the formation of the eMSTO and bimodal MS can be tested by whether part of the BB stars are H-MS + He-star systems or H-MS + WD systems. Moreover, the eSFH can be confirmed by whether the BF stars are members of NGC 1866, because the BF stars belong to an older population in the theoretical model. The author thanks the anonymous referee for helpful comments that helped the author improve this work, A. P. Milone for providing the observed data of NGC 1866, and Daniel Kister for help in improving the English, and acknowledges the support from the NSFC 11773005, U1631236, and 11273012. Bastian, N., Cabrera-Ziri, I., Niederhofer, F., et al. 2017, , 465, 4795 Bastian, N., & de Mink, S. E. 2009, MNRAS, 398, L11 Bastian, N., Niederhofer, F., Kozhurina-Platais, V., et al. 2016, , 460, L20 Brandt, T. D., Huang, C. X. 2015, , 807, 25 Chabrier, G. 2001, ApJ, 554, 1274 Correnti, M., Goudfrooij, P., Kalirai, J. S., et al. 2014, , 793, 121 Correnti, M., Goudfrooij, P., Puzia, T. H., de Mink, S. E. 2015, , 450, 3054 Correnti, M., Goudfrooij, P., Bellini, A., Kalirai, J. S. Puzia, T. H. 2017, , 467, 3628 D’Antona, F., Di Criscienzo, M., Decressin, T., et al. 2015, , 453, 2637 de Grijs, R. 2017, NatAs., 1, 0011 Dupree, A. K., Dotter, A., Johnson, C. I., et al. 2017, ApJL, 846, L1 Girardi, L., Rubele, S., & Kerber, L. 2009, MNRAS, 394, L74 Glatt, K., Grebel, E. K., & Sabbi, E. et al. 2008, AJ, 136, 1703 Goudfrooij, P., Girardi, L., Kozhurina-Platais, V., et al. 2014, , 797, 35 Goudfrooij, P., Puzia, T. H., Chandar, R., & Kozhurina-Platais, V. 2011, , 737, 4 Goudfrooij, P., Puzia, T. H., Kozhurina-Platais, V., & Chandar, R. 2009, AJ, 137, 4988 Goudfrooij, P., Girardi, L., Correnti, M. 2017, ApJ, 846, 22 Han, Z., Podsiadlowski, P., & Eggleton, P. P. 1995, MNRAS, 272, 800 Hurley, J. R., Pols, O. R., & Tout, C. A. 2000, MNRAS, 315, 543 Hurley, J. R., Tout, C. A., & Pols, O. R. 2002, MNRAS, 329, 897 Lejeune, T., Cuisinier, F., & Buser, R. 1998, , 130, 65 Lemasle, B., Groenewegen, M. A. T., Grebel, E. K., et al. 2017, A&A, 608, A85 Li, C., de Grijs, R., Deng, L. 2014, Nature, 516, 367 Li, C., de Grijs, R., Deng, L., & Milone, A. P. 2017, ApJ, 844, 119 Li, Z., Mao, C., Chen, L. 2015, , 802, 44 Li, Z., Mao, C., Chen, L., Zhang, Q. 2012, , 761, 22 Mackey, A. D., & Broby Nielsen, P. 2007, MNRAS, 379, 151 Mackey, A. D., Broby Nielsen, P., Ferguson, A. M. N., & Richardson, J. C. 2008, ApJ, 681, L17 Milone, A. P., Bedin, L. R., Piotto, G., & Anderson, J. 2009, A&A, 497, 755 Milone, A. P., Bedin, L. R., Cassisi, S., Piotto, G., Anderson, J., Pietrinferni, A., Buonanno, R. 2013, A&A, 555, A143 Milone, A. P., Bedin, L. R., Piotto, G., et al. 2015, , 450, 3750 Milone, A. P., Marino, A. F., D’Antona, F., et al. 2016, , 458, 4368 Milone, A. P., Marino, A. F., D’Antona, F., et al. 2017, MNRAS, 465, 4363 Niederhofer, F., Georgy, C., Bastian, N., Ekström, S. 2015a, , 453, 2070 Piatti, A. E., & Cole, A. 2017, MNRAS, 470, L77 Pols, O. R., & Marinus, M. 1994, A&A, 288, 475 Rubele, S., Kerber, L., & Girardi, L. 2010, MNRAS, 403, 1156 Salinas, R., Pajkos, M. A., Strader, J., Vivas, A. K., Contreras Ramos, R. 2016, ApJL, 832, L14 Salpeter, E. E. 1955, ApJ, 121, 161 Strom, K. M., & Strom, S. E. 1970, ApJ, 162, 523 Yang, W. 2016, , 821, 108 Yang, W. M., Bi, S. L., Meng, X. C., Liu, Z. 2013, , 776, 112 Yang, W. M., Meng, X. C., Bi, S. L., et al. 2011, , 731, L37 Yang, W., & Tian, Z. 2017, ApJ, 836, 102 Zhang, F., Han, Z., Li, L., Hurley, J. R. 2004, A&A, 414, 117

ArXiv

--- abstract: | A new statistical ensemble is examined using the example of classical one-component simple fluid. It’s logical to call it an open ensemble, because its peculiarity is the inclusion in the consideration some surrounding area. Calculations point to the necessity of taking into account the restricting surface, exactly when the system is not separated by anything from the bath, and the whole medium is uniform. The “surface tension coefficient”, included in the partition function corresponds to the interface of the fluid and hard solid, due to the strict compliance of probability and potential limitations. The number of surface particles corresponds exactly to near surface number density distortions (oscillations) arising in the neighborhood of fluctuation cavities. In contrast to grand canonical ensemble, an open statistical ensemble satisfies the scale invariance requirement: general term of the included subsystem distribution corresponds to that of the original system. It is this ensemble which should be used where consideration of a truly open system is required, since it properly integrates the surface terms. Furthermore, this ensemble may be employed in studies of small systems, since it has no lower limits for the volume of the system. Finally, it is useful in the investigation of fluctuations. For example, it demonstrates that the variance (the mean square deviation) of the number of particles is divided into the bulk and surface terms. author: - 'V. M. Zaskulnikov' title: 'Open statistical ensemble: new properties (scale invariance, application to small systems, meaning of surface particles, etc.)' --- \[sec:01\]Introduction ====================== The Gibbs grand canonical distribution is one of the foundations of statistical physics and, in its basic parts, is not doubt. However, it will be interesting to check the grand canonical ensemble (GCE) relations for correspondence to some criteria. One of such criteria is a peculiar kind of scaling, which lies in the fact that the distribution for the subsystem (calculated from the source) should functionally coincide with the original. Such testing seemed reasonable from the very beginning. Indeed, the familiar treatment of nucleation dating back to J.W.Gibbs [@gibbs1961 p.242] involves nucleus surface into consideration. The grand canonical distribution itself describes the fluctuations of the particles number in a given volume. Thus, initially, at the GCE level, for example, the probability of the formation of the hole must include the surface terms. The presence of such a term in the partition function of GCE was established earlier [@Bellemans1962], however, it is easy to see (sections \[subsec:03a\_1\], \[subsec:04f\], \[subsec:06b\]), that it is included with the wrong sign. Consequently, the conventional procedure of passing to the thermodynamic limit is in doubt. The thing is that the increase in the GCE volume does not eliminate the boundary distortions because they remain “glued” to its borders. Correct thermodynamic limit must fix the volume of observation, and take away the boundary distortions to infinity (section \[subsec:06a\]). Introduction of terms defining the interaction over the surrounding surface in statistical distribution calls for the exploration of the medium beyond the limits of the system. Thus we arrive at the idea of an open statistical ensemble (OSE). It can be constructed with the use of the scaling check procedure. Though OSE construction itself is connected with the violation of GCE scale invariance, it is preserved during the subsequent breakdown of the system. In the case of one-component system liquid/gas the surface tension coefficient depends solely on one variable, either pressure or temperature, due to the phase equilibrium curve. The situation under discussion is beyond the limits of this condition, and “surface tension coefficient” depends on two variables. As a result, a non-zero number of surface particles arises for the system boundary. It is necessary to establish the physical meaning of this parameter. Another issue complicating and limiting the work with the GCE is the impossibility of direct application of this distribution to small systems. This limitation is habitual and, at the first glance, seems reasonable but it has no deep foundation. There is no reason not to build the distribution, which will be valid for arbitrarily large and arbitrarily small volumes. In particular, this would allow one to calculate the number density fluctuations for small volumes. In addition, it is interesting to simplify an extremely complicated method of calculations of surface terms, based on the diagram technique [@Bellemans1962]. It is impossible to work with a series if you do not have the general term. It should be emphasized that in the given case the surface terms appear in the description of a homogeneous medium involving no real surface but just a hypothetical one, contouring the system under study. \[sec:02\]Primary definitions ============================= To simplify operations, in most cases we shall define the integration domain by characteristic functions $$\psi^v(\bm{r}_i) = \psi^v_i = \left\{ \begin{array}{ll} 1 & (\bm{r}_i \in v)\\ 0 & (\bm{r}_i \notin v) \end{array} \right. \label{equ:01}$$ and $$\chi^v_i = 1 - \psi^v_i = \left\{ \begin{array}{ll} 0 & (\bm{r}_i \in v)\\ 1 & (\bm{r}_i \notin v), \end{array} \right. \label{equ:02}$$ where $v$ denotes the volume of the domain, and integration in all the integrals is considered to be performed over the infinite space unless otherwise specified. Note that the algebra of characteristic functions plays an important part in the approach at hand. Besides, as it will be seen further, the role of these functions is significant from the standpoint of distribution functions determination. \[subsec:02a\]Canonical ensemble -------------------------------- The probability density to find the given spatial configuration of a specific set of particles [@hillstatmeh1987 p.181] is defined by the expression $$P^{(k)}_{1...k} = \frac{1}{Z_N^V}\int \left [ \prod_{l = k+1}^N \psi^V_l \right ] \exp(-\beta U^N_{1...N})d\bm{r}_{k+1}...d\bm{r}_N, \label{equ:03}$$ where $N$ is the number of particles in the system, $\beta = 1/k_BT$, $k_B$ is the Boltzmann constant, $T$ is the temperature, $U^N_{1...N}$ is the interaction energy of particles between each other, $V$ is the system volume. Integration is performed over the coordinates of the ensemble particles $\bm{r}_{k+1}...\bm{r}_N$. $Z_N^V$ is the configuration integral $$Z_N^V = \int \left [ \prod_{k = 1}^N \psi^V_k \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N. \label{equ:04}$$ Passing to the distribution functions for arbitrary set of particles, we have $$\varrho^{(k)}_{C,1...k} = \frac{N!}{(N-k)!}P^{(k)}_{1...k}, \label{equ:05}$$ where $\varrho^{(k)}_{C,1...k}$ gives the probability density to find a given configuration of $k$ arbitrary particles for the canonical ensemble. \[subsec:02b\]Grand canonical ensemble -------------------------------------- Average equality (\[equ:05\]) over fluctuations of the number of particles, i.e., apply the operation $\sum_{N=0}^\infty P_N^V$ to both its sides, where $$P_N^V = \frac{z^N Z_N^V}{N!\Xi_V} \label{equ:06}$$ is the probability for GCE to have a definite number of particles $N$ inside the volume $V$. Here $z$ is the activity $$z = \frac{e^{\mu/k_BT}}{\Lambda^3}, \label{equ:07}$$ where $\mu$ is the chemical potential, $\Lambda = h/\sqrt[]{2 \pi mk_BT}$, $h$ is the Planck constant, $m$ is the particle mass, and $\Xi_V$ is a large partition function of the system of the volume $V$ $$\Xi_V = 1 + \sum_{N=1}^\infty \frac{z^N Z_N^V}{N!}. \label{equ:08}$$ Thus $$\varrho^{(k)}_{G,1...k} = \sum_{N=k}^\infty \varrho^{(k)}_{C,1...k} P_N^V, \label{equ:09}$$ or $$\begin{aligned} &&\varrho^{(k)}_{G,1...k} (\psi^V) = \frac{z^k}{\Xi_V} \left \{ \exp(-\beta U^{k}_{1...k}) + \sum_{N=1}^\infty \frac{z^N}{N!} \right . \label{equ:10} \\ && ~~~~~~ \left . \times \int \left [ \prod_{l = k+1}^{k+N} \psi^V_l \right ] \exp(-\beta U^{N+k}_{1...N+k})d\bm{r}_{k+1}...d\bm{r}_{k+N} \right \}, \nonumber\end{aligned}$$ where $\varrho^{(k)}_{G,1...k}(\psi^V)$ is the $k$-particle distribution function for GCE. As already mentioned, these functions specify the probability density to find a certain configuration of arbitrary particles. For ideal gas $\varrho^{(k)}_{G,1...k} = \varrho^{k}$, where $\varrho = \overline{N}/V$ is the number density. Here $\psi^V$ cannot be treated as the definition domain of the function in the ordinary sense, since, as it will be seen further, the coordinates of free particles can fall outside the limits it defines. For brevity, we shall call it the assignment domain. For example, in equation (\[equ:10\]) $\psi^V$ is the assignment domain of the function $\varrho^{(k)}_{G,1...k}$. Characteristic functions appear both in the nominator and in the denominator of (\[equ:10\]); in the last case through $\Xi_V$ (\[equ:08\]). This will be discussed in detail below. Further the significance of the assignment domain $\psi^V$ will become obvious, so we shall give it in the explicit form, and the fact that $\varrho^{(k)}_{G,1...k}(\psi^V)$ belongs to GCE type will be denoted by the index $G$. Commonly, the assignment domain $\psi^V$ is noted and controlled only slightly, however, from the standpoint of the present paper it is of crucial importance. Complete integrals $\varrho^{(k)}_{G,1...k}(\psi^V)$ satisfy the relation $$\int \left [ \prod_{i = 1}^k \psi^V_i \right ] {\varrho}^{(k)}_{G,1...k} (\psi^V) d\bm{r}_1... d\bm{r}_k = \left \langle \frac{N!}{(N-k)!} \right \rangle, \label{equ:11}$$ following from the definition. Here the triangular brackets denote averaging over the number of particles of type (\[equ:09\]). Note that for GCE integration and assignment domains in (\[equ:11\]) rigorously coincide. \[subsec:02c\]Presence of an external field ------------------------------------------- The configuration integral of the inhomogeneous closed system is given by the expression $$Z^U_N = \int\limits_{V}^{} \exp(-\beta\sum_{i=1}^N u_i-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N, \\ \label{equ:12}$$ where $u_i$ is the energy of the interaction of the $i$-th particle with the force field. For the GCE, we introduce the quantity $$\Xi^U_V = 1+ \sum_{N=1}^\infty \frac{z^N Z^U_N}{N!}, \label{equ:13}$$ which is obviously the large partition function of the system in the presence of an external field. \[sec:03\]OSE construction ========================== The idea of OSE construction consists in the separation of a certain subsystem rather far removed from the boundaries of the whole GCE system. In practice it is sufficient if the subsystem is smaller than the system just by several atomic layers. Lower bounds of the system size are absent, so however small open ensembles may be considered. \[subsec:03a\]OSE partition function ------------------------------------ The probability of a hole formation is an important parameter that defines the ensemble partition function. Thus we start with the examination of the first and the most significant term of this distribution, and then pass to other terms. First, the canonical ensemble will be considered, then GCE, and finally OSE definition. Consider the probability of the fluctuation formation of the cavity inside a uniform statistical system - the region of the volume $v$ containing no particles. Let us take that the cavity is rather far away from the system boundaries. Evidently, the desired probability for the closed system is defined as $$p^v_{C,0} = \frac{1}{Z_N^V}\int \left [ \prod_{k = 1}^N(\psi^V_k-\psi^v_k) \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N \label{equ:14}$$ or $$p^v_{C,0} = \frac{Z_N^{V-v}}{Z_N^V}. \label{equ:15}$$ Here the index $C$ on $p^v_{C,0}$ shows that the probability is defined in the canonical ensemble on condition that $N$ particles are in the volume $V$. Passing to the probability for GCE $$p^v_{G,0} = \sum_{N=0}^\infty p^v_{C,0} P_N^V, \label{equ:16}$$ and substituting (\[equ:06\]) in (\[equ:16\]) we obtain $$p^v_{G,0} = \frac{\Xi_{V-v}}{\Xi_V}. \label{equ:17}$$ As is seen from the derivation procedure, both large partition functions are considered at one and the same activity. Introduce the quantity $$\Upsilon_v = \frac{\Xi_V}{\Xi_{V-v}}, \label{equ:18}$$ by definition it is the partition function of OSE system of the volume $v$. Strictly speaking, it would be reasonable to retain the index $V$ in the partition function $\Upsilon_v$ as well, or to define it via passing to the limit $V\rightarrow \infty$. This passage will be done in examining the expansions in powers of the activity. However, we shall see that the dependence on the system volume $V$ is so slight that (\[equ:18\]) is a rather good definition even when the boundaries of two systems are separated just by several atomic layers. Essentially, the difference between OSE and GCE is in surface effects; this is clear from (\[equ:18\]). When the subsystem does not interact with the medium, $\Xi_V = \Xi_v \Xi_{V-v}$, and this equation reduces to a banal one $$\Upsilon_v = \frac{\Xi_v \Xi_{V-v}}{\Xi_{V-v}} = \Xi_v, \label{equ:19}$$ thus demonstrating the identity of the ensembles. \[subsec:03a\_1\]Equivalence of the potential and statistical limitations ------------------------------------------------------------------------- Potential and statistical restrictions closely correspond to each other for the potential of hard solid. This is evident from the fact that the Bolzmann factor corresponds to the characteristic function of (\[equ:01\]) type in this case. Consequently, the partition function (\[equ:12\]) coincides exactly with $Z_N^{V-v}$, appearing in (\[equ:15\]), and (\[equ:13\]) - with $\Xi_{V-v}$ from (\[equ:17\]). Thus, equations (\[equ:15\]), (\[equ:17\]) may be interpreted as the ratios of partition functions where the denominator involves the partition function of the initial system, and the nominator - that of the system with hard solid immersed in it. Therefore the logarithm of the ratio of partition functions acquires (accurate to the factor) the meaning of the change in $\Omega$ - potential (grand potential) $$\begin{aligned} \Omega^u - \Omega &=& [-P(V-v) + \sigma a] - [-PV] \nonumber \\ &=& vP(z,T) +a \sigma (z,T). \label{equ:20} \end{aligned}$$ Here the index $u$ denotes the system with the immersed solid, $P(z,T)$ is the pressure, $a$ - the surface restricting the subsystem (hard solid), and $\sigma (z,T)$ - “surface tension coefficient” at the hard solid/fluid interface. Using (\[equ:20\]), obtain $$\Upsilon_v = \exp{\beta [ vP(z,T) + a\sigma (z,T) ]}. \label{equ:21}$$ This expression is valid, obviously, for rather smooth surface only, i.e., for not very small volume $v$. Naturally, $$p^v_0 = \frac{1}{\Upsilon_v}, \label{equ:22}$$ where $p^v_0$ is the probability of a hole of the volume $v$ formation for OSE. So, the “surface tension coefficient” included in the OSE always corresponds to the boundary between fluid and hard solid. The near surface number density oscillations close to the above fluctuation cavities also correspond exactly to the distortions at the boundary of hard solid and fluid. This also agrees with the Bolzmann idea concerning fluctuation thermodynamics (entropy) definition in terms of appropriate quantities for the system in a such field, where a given fluctuation is an equilibrium configuration. Below we shall show the separation of surface terms in OSE, here we also refer to the results of the work [@Bellemans1962] from which it is clear that for GCE $$\Xi_V = \exp{\beta [ VP(z,T) - A\sigma (z,T) ]}, \label{equ:23}$$ where $A$ is the surface restricting the system. Substituting (\[equ:23\]) in (\[equ:18\]) it is easy to see that the expressions (\[equ:21\]) and (\[equ:23\]) correspond to each other. Quite interesting are the opposite signs of surface terms in GCE and OSE. In section \[subsec:06b\] a comparison of these ensembles will be made. \[subsec:03b\]OSE distribution ------------------------------ Let us calculate the general term of OSE distribution. Thus, consider the probability for a uniform system to find exactly $m$ particles in the volume $v$ - $p^v_m$. In essence, calculations will correspond to the above calculation of a zero term. We begin with the probability for the canonical ensemble $$\begin{aligned} p^v_{C,m} = \binom{N}{m} && \frac{1}{Z_N^V}\int\exp(-\beta U^N_{1...N})\label{equ:24} \\ \times && \left [\prod_{i = 1}^m\prod_{j = m+1}^N\psi^v_i (\psi^V_j-\psi^v_j) \right ] d\bm{r}_1...d\bm{r}_N. \nonumber\end{aligned}$$ Then we pass to analogous probability for GCE $$p^v_{G,m} = \sum_{N=m}^\infty p^v_{C,m} P_N^V \label{equ:25},$$ and substituting (\[equ:06\]) gives $$\begin{aligned} p^v_{G,m}& = & \frac{z^m}{m!\Xi_V}\sum_{N=0}^\infty \frac{z^N}{N!}\int\exp(-\beta U^{N+m}_{1...N+m}) \nonumber \\ &\times& \left [\prod_{i = 1}^m\prod_{j = m+1}^{m+N}\psi^v_i (\psi^V_j-\psi^v_j) \right ] d\bm{r}_1...d\bm{r}_{N+m}. \label{equ:26}\end{aligned}$$ Eventually, by multiplication and division of (\[equ:26\]) by $\Xi_{V-v}$ and taking the sum of the series, we get $$p^v_m = \frac{1}{m! \Upsilon_v} \int \left [ \prod_{i = 1}^m \psi^v_i \right ] \varrho^{(m)}_{G,1...m}(\psi^V-\psi^v) d\bm{r}_1...d\bm{r}_m. \label{equ:27}$$ This is just the OSE distribution general term we are interested in. The sense of this expression is clear: the probability of finding $m$ particles in the volume $v$ is determined by the probability of finding the hole of the same size, $1/\Upsilon_v$, multiplied by the probability density of revealing some configuration of $m$ particles (on condition that other particles are beyond the limits of the volume) integrated over all configurations. The factor $1/m!$ arises due to the symmetry (indistinguishability) under particles permutations. (Here we deal exactly with the case where the coordinates of particles $1,...m$ are beyond the limits of the distribution function assignment domain defined by the functions $\psi^V-\psi^v$.) First note the similarity between expression (\[equ:27\]) and the general term of GCE distribution (\[equ:06\]). They coincide accurate to the normalizing factor in the approximation $$\varrho^{(m)}_{G,1...m} \approx z^m \exp(-\beta U^{m}_{1...m}), \label{equ:28}$$ i.e., in the low density limit when the first term plays the main role in the distribution function expansion [@hillstatmeh1987 p.410], (\[equ:a05\]). In particular, this means that for low densities the role of surface tension is less important. As will be seen later the reason is that in this case the expansion of surface terms into a series in activity begins with a quadratic term. Expressions (\[equ:18\]), (\[equ:21\]) and (\[equ:27\]) solve the problem of OSE construction, and now we can turn to its properties. \[sec:04\]Some properties of OSE ================================ \[subsec:04f\]$p^v_0$ as the fluctuation probability ---------------------------------------------------- The equation for the cavity formation probability which has the form $$p^v_0 = \exp{- \beta [vP(z,T) + a \sigma (z,T)]} \label{equ:29}$$ (as follows from (\[equ:21\]), (\[equ:22\])) agrees with the general expression for fluctuation probability $$p_f \propto \exp{( -\beta R_{min})}, \label{equ:30}$$ [@landaulifshitz1985 p.339], where $R_{min}$ is the minimum work of fluctuation removal. Indeed, since the chemical potential of particles outside the field is not changed under local imposition of the field, we can take that in this case the process of creation/removal of fluctuation occurs at constant chemical potential and temperature, i.e., natural variables of $\Omega$-potential (grand potential). So minimum cavity formation work is equal to the change in $\Omega$-potential (\[equ:20\]) $$R_{min} = \Omega^U - \Omega = vP(z,T) +a \sigma (z,T). \label{equ:31}$$ As we are interested in the fluctuation at which the density changes in a stepwise way, this fluctuation creation process must be performed by hard solid (see Section \[subsec:03a\_1\]). \[subsec:04a\]Normalization condition ------------------------------------- Normalization condition $$\sum_{m=0}^{\infty} p^v_{G,m} = 1 \label{equ:32}$$ gives the expression $$\begin{aligned} &&\sum_{m=0}^{\infty} p^v_{G,m} = \frac{1}{\Xi_V} \sum_{m=0}^\infty \frac{1}{m!} \sum_{N=m}^\infty \frac{z^N}{(N-m)!}\label{equ:33} \\ &&\times \int \left [\prod_{i = 1}^m\prod_{j = m+1}^N\psi^v_i (\psi^V_j-\psi^v_j) \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N. \nonumber\end{aligned}$$ Changing the order of summation, we have $$\begin{aligned} &&\sum_{m=0}^{\infty} p^v_{G,m} = \frac{1}{\Xi_V} \sum_{N=0}^\infty \frac{z^N}{N!}\int \sum_{m=0}^N \binom{N}{m}\label{equ:34} \\ &&~~~\times \left [\prod_{i = 1}^m\prod_{j = m+1}^N\psi^v_i (\psi^V_j-\psi^v_j) \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N. \nonumber\end{aligned}$$ Using the Bolzmann factor symmetry under permutations of particles and binomial formula, we arrive at $$\begin{aligned} \sum_{m=0}^{\infty} p^v_{G,m} &=& \frac{1}{\Xi_V} \sum_{N=0}^\infty \frac{z^N}{N!} \label{equ:35} \\ &\times & \int\left [\prod_{i = 1}^N\psi^V_i \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N = 1. \nonumber\end{aligned}$$ \[subsec:04b\]The mean number of particles ------------------------------------------ With the aid of (\[equ:27\]), one can find the mean number of particles in a certain volume $v$ $$\begin{aligned} &&\overline m = \frac{1}{\Xi_V} \sum_{m=1}^\infty \frac{1}{(m-1)!} \sum_{N=m}^\infty \frac{z^N}{(N-m)!}\label{equ:36} \\ &&\times \int \left [\prod_{i = 1}^m\prod_{j = m+1}^N\psi^v_i (\psi^V_j-\psi^v_j) \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N. \nonumber\end{aligned}$$ Changing the order of summation and applying binomial distribution and symmetry under permutations again, we have $$\overline m = \int \psi^v_1 \varrho^{(1)}_{G,1}(\psi_V) d\bm{r}_1. \label{equ:37}$$ Though relation (\[equ:37\]) seems banal, it has a non-trivial sense. Since in the given case integration is performed far from the system boundaries, the function $\varrho^{(1)}_{G,1}$ involves no boundary distortions, thus, unlike GCE, the mean number of particles for OSE is exactly equal to the number of volume particles. So, $$\overline m = m_b, \label{equ:38}$$ where $$m_b = v\frac{\partial P}{\partial \mu}, \label{equ:39}$$ since in the limits of the volume $v$ the following equation holds $$\varrho^{(1)}_{G,1}(\psi_V) = \varrho = \frac{\partial P}{\partial \mu}. \label{equ:40}$$ This will be the subject of further discussion in connection with the comparison between OSE and GCE. \[subsec:04c\]Fluctuations of the number of particles ----------------------------------------------------- Averaging the expression $m(m-1)(m-2)...(m-k+1)$ with the use of (\[equ:27\]), we obtain $$\begin{aligned} &&\left \langle \frac{m!}{(m-k)!} \right \rangle = \frac{1}{\Xi_V} \sum_{m=k}^\infty \frac{1}{(m-k)!} \sum_{N=m}^\infty \frac{z^N}{(N-m)!}\label{equ:41} \\ &&\times \int \left [\prod_{i = 1}^m\prod_{j = m+1}^N\psi^v_i (\psi^V_j-\psi^v_j) \right ] \exp(-\beta U^N_{1...N}) d\bm{r}_1...d\bm{r}_N. \nonumber\end{aligned}$$ Proceeding as in the case with (\[equ:33\]) $\rightarrow$ (\[equ:34\]) $\rightarrow$ (\[equ:35\]), we have $$\int \left [ \prod_{i = 1}^k \psi^v_i \right ] \varrho^{(k)}_{G,1...k}(\psi^V) d\bm{r}_1...d\bm{r}_k = \left \langle \frac{m!}{(m-k)!} \right \rangle. \label{equ:42}$$ Though this expression seems similar to (\[equ:11\]), there is an essential distinction: integration and assignment domains of the functions $\varrho^{(k)}_{G,1...k}$ do not coincide in this case. This results in the following. First, in contrast to the case with GCE, the surface terms enter in (\[equ:42\]) in the proper way, since here we are not concerned with the region of near surface distortions of $\varrho^{(k)}_{G,1...k}$ at the outer boundary of GCE. Second, unlike (\[equ:11\]), in (\[equ:42\]) the integration domain may be however small, and this allows one to employ this expression to study fluctuations in small volumes. For example, quadratic fluctuation in the volume $v$ may be found from the equation $$\int \psi^v_1 \psi^v_2 \left ( \varrho^{(2)}_{G,1,2} - \varrho^{(1)}_{G,1} \varrho^{(1)}_{G,2} \right ) d\bm{r}_1 d\bm{r}_2 = \overline {m^2} - {\overline m}^2 - \overline {m}. \label{equ:43}$$ Further the fluctuations will be discussed in greater detail. Note that theorems (\[equ:32\]), (\[equ:37\]), (\[equ:42\]) can be proved in a different way, namely, with the help of a series in powers of the activity [@zaskulnikov200911]. \[subsec:04d\]Small systems --------------------------- It is clear from the derivation of OSE expressions that this distribution may be applied to however small volumes including those less than a molecule size. However, as already mentioned, the expression for partition function (\[equ:21\]) does not work for such sizes. For very small $v$ the partition function may be calculated as follows. Just as from physical considerations, it follows from (\[equ:27\]) that at sizes of $v$ less than the volume of a hard core of particles, of the whole distribution $p_m^v$ only two terms, $p_0^v$ and $p_1^v$ differ from zero. Since (\[equ:37\]) gives $\overline m = \varrho v$, we have $$\overline m = 0 \times p_0^v + 1 \times p_1^v = p_1^v = \varrho v. \label{equ:44}$$ Then $p_0^v = 1 - \varrho v$ and $$\Upsilon_v = \frac{1}{1 - \varrho v}, \label{equ:45}$$ which is a universal expression for OSE partition function at the system sizes tending to zero. Consideration similar to the above one may be done for the volume containing not more than $2, 3,$ etc., atoms. It is clear that no analog may be given (except formal expressions) for GCE, since in this case the derivation of basic relations makes use of the assumption of large sizes of the system. However, if we do it, we shall obtain the analogs of (\[equ:44\]), (\[equ:45\]) with the replacement of $\varrho \rightarrow z$ which is obviously absurd. Now consider fluctuations in a small volume. With this aim we return to equation (\[equ:43\]). Let us use local character of the Ursell distribution function $${\cal F}^{(2)}_{1,2} = \varrho^{(2)}_{G,1,2} - \varrho^{(1)}_{G,1} \varrho^{(1)}_{G,2}. \label{equ:46}$$ Transforming (\[equ:43\]),and in view of $$\int {\cal F}^{(2)}_{1,2} d\bm{r}_2 = \varrho ( \varrho k_B T \varkappa_T - 1), \label{equ:47}$$ where $\varkappa_T$ is the isothermal compressibility, we obtain $$\frac{\overline {m^2} - {\overline m}^2}{\overline {m}} = \varrho k_B T \varkappa_T - \frac{1}{\overline {m}} \int \psi^v_1 \chi^v_2 {\cal F}^{(2)}_{1,2} d\bm{r}_1 d\bm{r}_2. \label{equ:48}$$ This expression, just as the original one (\[equ:43\]), is exact and holds for any volumes. We see that, along with ordinary term $\varrho k_B T \varkappa_T$, an additional term appears in the right-hand side. For volumes restricted by rather smooth surfaces, and hence for rather large volumes (compared with the molecular size) the second term in the right-hand side of (\[equ:48\]) is proportional to the surface (see Section \[subsec:05a\]). Performing integration over the surface in (\[equ:48\]), one has $$\frac{\overline {m^2} - {\overline m}^2}{\overline {m}} = \varrho k_B T \varkappa_T - \frac{a}{\overline {m}} \int \psi^v_1 \chi^v_2 {\cal F}^{(2)}_{1,2} dx_1 d\bm{r}_2, \label{equ:49}$$ where $x_1$ is the coordinate perpendicular to the system surface. Since ${\cal F}^{(2)}_{1,2}$ depends only on the relative configuration of particles, we can perform integration over $x_1$ in (\[equ:49\]). Making the change of variables $\bm{r}_2' = \bm{r}_2 - \bm{r}_1$ and performing some other simple transformations, we get $$\frac{\overline {m^2} - {\overline m}^2}{\overline {m}} = \varrho k_B T \varkappa_T - \frac{a}{\overline {m}} \int_{x>0} x {\cal F}^{(2)}(r) d\bm{r}. \label{equ:50}$$ Equations (\[equ:49\]), (\[equ:50\]) give the expressions for the mean square relative deviation with allowance for surface corrections. For absolute form we have $$\overline {m^2} - {\overline m}^2 - \overline {m} =v \int {\cal F}^{(2)}(r) d\bm{r} - a \int_{x>0} x {\cal F}^{(2)}(r) d\bm{r}. \label{equ:51}$$ For even more small volumes where separation of surface terms is impossible other expressions are valid. Let us use the fact that it is impossible to place two atoms into the volume less than that of a particle hard core. For such a volume, it follows from (\[equ:42\]) that for $k > 1$ $$\left \langle \frac{m!}{(m-k)!} \right \rangle = 0. \label{equ:52}$$ For example, for $k=2$, we obtain $$\overline {m^2} - \overline {m} = 0 \label{equ:53}$$ and $$\frac{\overline {m^2} - {\overline m}^2}{\overline {m}} = 1 - \varrho v. \label{equ:54}$$ Using (\[equ:43\]), we can also get the exact expressions. Performing a procedure similar to the transition from (\[equ:49\]) to (\[equ:50\]), we obtain $$\int f(\bm{r}) {\cal F}^{(2)}(r) d\bm{r} = \overline {m^2} - {\overline m}^2 - \overline {m}, \label{equ:55}$$ where $$f(\bm{r}) = \int \psi^v(\bm{r}_1) \psi^v(\bm{r}_1 + \bm{r}) d\bm{r}_1. \label{equ:56}$$ Recall that $\psi^v$ is a characteristic function, specifying a certain volume. Equality (\[equ:55\]) holds for the volume of any shape and size. For example, in the case of the sphere, we have $$f(\bm{r}) = \left\{ \begin{array}{ll} \pi(\frac{4}{3}R^3 - R^2 r + \frac{1}{12} r^3) & (r < 2R)\\ 0 & (r \geq 2R) \end{array} \right. \label{equ:57},$$ where $R$ is the radius of the sphere. Substituting $f$ in (\[equ:55\]), we obtain the exact expression $$\begin{aligned} && \overline {m^2} - {\overline m}^2 - \overline {m} = \frac{4}{3}\pi R^3 \int \limits_{0}^{2R} {\cal F}^{(2)}(r) d\bm{r} \nonumber \\ && -\pi R^2 \int \limits_{0}^{2R} r {\cal F}^{(2)}(r) d\bm{r} + \frac{\pi}{12} \int \limits_{0}^{2R} r^3 {\cal F}^{(2)}(r) d\bm{r}. \label{equ:58}\end{aligned}$$ When the sphere radius $R$ exceeds the correlation radius (the radius of ${\cal F}^{(2)}(r)$ oscillations decay), the first term in the right side of (\[equ:58\]) coincides with that in the right-hand side of (\[equ:51\]), and the second - with the second one, respectively, as can be verified by performing the integration over the angles in the last term of equation (\[equ:51\]). Thus, the surface terms of the mean-square deviations are determined by the first moments of the pair Ursell function or, eventually, by the first moments of the radial distribution function. If $R \rightarrow 0$, ${\cal F}^{(2)}(r) \rightarrow - \varrho^2$ and using (\[equ:58\]) we obtain $$\overline {m^2} - {\overline m}^2 - \overline {m} = -\frac{16}{9} \pi^2 R^6 \varrho^2, \label{equ:59}$$ which coincides with (\[equ:53\]). Higher order fluctuations and fluctuations for the volumes of greater sizes may be calculated in a similar way. \[subsec:04e\]Scale invariance ------------------------------ Probabilistic distributions for two different volumes are obviously connected by complete probability formula $$w^v_m = \sum_{N=m}^\infty P(m|N) W^V_N, \label{equ:60}$$ where $w^v_m$ is the unconditional probability to find $m$ particles in the volume $v$, $W^V_N$ is the unconditional probability to find $N$ particles in the volume $V$, and $P(m|N)$ is the probability to find $m$ particles in the volume $v$ on condition that the volume $V$ contains $N$ particles. Naturally the identity of intensive parameters of the medium and its homogeneity are assumed. Consider the case where the corresponding volumes are inserted in one another, and their boundaries are separated by a sufficient distance. So statistical distribution scale invariance is to mean that the following equality holds $$w^V_N = W^V_N. \label{equ:61}$$ In other words, functional dependences of unconditional probabilities are to be the same for different volumes. Comparing (\[equ:60\]) with (\[equ:25\]), we arrive at the conclusion that scale invariance is absent in GCE, since, though the character of relation (\[equ:25\]) is similar to (\[equ:60\]), scaling condition (\[equ:61\]) is violated; this is seen from the comparison between (\[equ:06\]) and (\[equ:27\]). Note that the volume term retains its form; therefore, this part of GCE is scaled. To prove OSE scale invariance, first it is necessary to construct the conditional probability $P(m|N)$. We cannot use $p^v_{C,m}$ (\[equ:24\]) because of the additional condition: there is a surface at the boundary of the volume $V$. To find $P(m|N)$, employ multiplication rule $$P(m \cap N) = P(m|N) p^V_N, \label{equ:62}$$ where $P(m \cap N)$ is the probability of a simultaneous finding of $m$ particles in the volume $v$ and $N$ particles in the volume $V$. Calculation of this quantity by analogy with transformations (\[equ:24\]) $\rightarrow$ (\[equ:26\]) yields $$P(m \cap N) = \frac{z^N}{m!(N-m)!\Xi_{V'}}\sum_{M=0}^\infty \frac{z^M}{M!}\int\exp(-\beta U^{N+M}_{1...N+M}) \left [\prod_{i = 1}^m\prod_{j = m+1}^N\prod_{k = N+1}^{N+M}\psi^v_i (\psi^V_j-\psi^v_j)(\psi^{V'}_k-\psi^V_k) \right ] d\bm{r}_1...d\bm{r}_{N+M}, \label{equ:63}$$ where $V'$ is some volume exceeding $V$. Using the binomial formula and the Boltzmann factor symmetry under particle permutation again, it can readily be seen that $$p^v_m = \sum_{N=m}^\infty P(m \cap N). \label{equ:64}$$ To make calculations we change the summation order again, and $p^v_m$ is defined by the expression of (\[equ:27\]) type with the replacement of $V \rightarrow V'$. Besides, it is easily established that $$p^V_N = \sum_{m=0}^N P(m \cap N). \label{equ:65}$$ For the desired conditional probability, we obtain from (\[equ:62\]) $$P(m|N) =\binom{N}{m}\frac{\displaystyle \sum_{M=0}^\infty \frac{z^M}{M!}\int\exp(-\beta U^{N+M}_{1...N+M}) \left [\prod_{i = 1}^m\prod_{j = m+1}^N\prod_{k = N+1}^{N+M}\psi^v_i (\psi^V_j-\psi^v_j)(\psi^{V'}_k-\psi^V_k) \right ] d\bm{r}_1...d\bm{r}_{N+M}} {\displaystyle \sum_{M=0}^\infty \frac{z^M}{M!}\int\exp(-\beta U^{N+M}_{1...N+M}) \left [\prod_{i = 1}^N\prod_{k = N+1}^{N+M}\psi^V_i (\psi^{V'}_k-\psi^V_k) \right ] d\bm{r}_1...d\bm{r}_{N+M}}. \label{equ:66}$$ As is seen, after division the first term of series in powers of $z$ (\[equ:66\]) coincides with expression (\[equ:24\]). Eventually, based on properties (\[equ:64\]), (\[equ:65\]) and the form of expression (\[equ:66\]) one can write $$p^v_m = \sum_{N=m}^\infty P(m|N) p^V_N \label{equ:67}$$ and thus OSE scale invariance is proved. \[sec:05\]Series in activity ============================ In this section we give the expressions for GCE and OSE objects as the series in powers of activity $z$, which are analogues to the Mayer series for dense gases. The first subsection presents the separation procedure for volume and surface terms. Finally, the convergence of series will be discussed. Here we assume that the systems under discussion are restricted by rather smooth surfaces, i.e., they are deprived of asperities of molecular sizes. Consequently, this condition imposes the constraints on the systems size as well, in the cases where it is necessary - for OSE distributions it must essentially exceed molecular sizes. \[subsec:05a\]The series for GCE and the procedure of separating surface terms ------------------------------------------------------------------------------ Taking the logarithm of equality (\[equ:08\]), in view of (\[equ:04\]) and the logarithmic form of generating function for Ursell factors ${\cal U}^{(t)}_{1...t}$ (see Appendix \[subsec:appenda1\]), one has $$\ln{\Xi_V} = \sum_{t=1}^\infty \frac{z^t}{t!} \int \left [ \prod_{i = 1}^{t} \psi^V_i \right ] {\cal U}^{(t)}_{1...t} d\bm{r}_1...d\bm{r}_t. \label{equ:68}$$ Transforming (\[equ:68\]) by adding and subtraction of identical terms, we obtain $$\begin{aligned} \ln{\Xi_V} &=& \sum_{t=1}^\infty \frac{z^t}{t!} \int \psi^V_1 {\cal U}^{(t)}_{1...t} d\bm{r}_1...d\bm{r}_t \label{equ:69} \\ &-& \sum_{t=2}^\infty \frac{z^t}{t!} \int \psi^V_1 \left [1 - \prod_{i = 2}^{t} (1 - \chi^V_i) \right ] {\cal U}^{(t)}_{1...t} d\bm{r}_1...d\bm{r}_t.\nonumber\end{aligned}$$ The first term in the right-hand side of (\[equ:69\]), the volume term, is the Mayer familiar expansion $$Pv = k_BT\sum_{t=1}^\infty \frac{z^t}{t!} \int \psi^v_1{\cal U}^{(t)}_{1...t} d\bm{r}_1...d\bm{r}_t \label{equ:70}$$ or $$P(z,T) =zk_BT + k_BT \sum_{k=2}^\infty \frac{z^k}{k!} \int {\cal U}^{(k)}_{1...k} d\bm{r}_2...d\bm{r}_k. \label{equ:71}$$ However, it has the analog for surface tension in the form given below. The second term in the right-hand side of (\[equ:69\]) is proportional to the surface size. Indeed, its structure is such that at least two particles always reside on different sides of the boundary. If we expand the product in it, all terms will have the form $$\int \psi^V_1 \left [ \prod_{i = 2}^{j} \chi^V_i \right ] {\cal U}^{(t)}_{1...t} d\bm{r}_1...d\bm{r}_t, \label{equ:72}$$ where $2 \leq j \leq t $. For the fixed first particle, by virtue of Ursell factors locality, the integrals of the type of (\[equ:72\]) are defined by local region near it. So $\int ... d\bm{r}_2...d\bm{r}_t$ do not depend on the first particle displacement along the system boundary. When the first particle moves away from the boundary, they decay rapidly due to the fixing factors $\chi^V_i$ and Ursell factors locality. Performing integration over the surface and placing the area outside the summation sign, we arrive at (\[equ:23\]), where $A\sigma (z,T)$ has the form $$\begin{aligned} A\sigma (z,T) &=& k_BT\sum_{t=2}^\infty \frac{z^t}{t!} \label{equ:73} \\ & \times & \int \psi^V_1 \left [1 - \prod_{i = 2}^{t} (1 - \chi^V_i) \right ] {\cal U}^{(t)}_{1...t} d\bm{r}_1 ...d\bm{r}_t,\nonumber\end{aligned}$$ and for $\sigma (z,T)$ we have the expression $$\begin{aligned} \sigma (z,T) &=& k_BT\sum_{t=2}^\infty \frac{z^t}{t!} \label{equ:74} \\ &\times & \int \psi^V_1 \left [1 - \prod_{i = 2}^{t} \psi^V_i \right ] {\cal U}^{(t)}_{1...t} dx_1 d\bm{r}_2 ...d\bm{r}_t,\nonumber\end{aligned}$$ where $x_1$ is the coordinate perpendicular to the surface defined by the boundary $\psi^V_1$. The axis $x_1$ direction is chosen such that $dx_1 > 0$. The first terms of expression (\[equ:74\]) were obtained in paper [@Bellemans1962] by diagram technique; however, construction of the far terms of the series in the same way seems extremely difficult. The translational invariance of ${\cal U}^{(t)}_{1 ... t}$ permits to perform the integration over the $x_1$, just as we did when considering the fluctuations in small systems (section \[subsec:04d\]). Performing the change of variables $\bm{r} _i'= \bm{r}_i - \bm{r}_1$ in (\[equ:74\]), shifting the coordinates of the Ursell factors, and integrating over $ x_1 $, arrive at $$\begin{aligned} \sigma (z,T) =&& k_BT\sum_{t=2}^\infty \frac{z^t}{t(t-2)!} \nonumber \\ && \times \int_{ x_2> \{0, x_3... x_k \} } x_2 {\cal U}^{(t)}_{0,2...t} d\bm{r}_2 ...d\bm{r}_t, \label{equ:75}\end{aligned}$$ where we use the Ursell factors symmetry under particle permutations. Finally, using the change of variables $\bm{r}'_i = \bm{r}_i - \bm{r}_2$, for $i = 3,...t$, we have $$\begin{aligned} \nu (z,T) =&& k_BT\sum_{t=2}^\infty \frac{z^t}{t(t-2)!} \nonumber \\ && \times \int_{ x_2 > 0,\dots x_t > 0} x_2 {\cal U}^{(t)}_{0,2...t} d\bm{r}_2 ...d\bm{r}_t. \label{equ:76}\end{aligned}$$ In this case we used the invariance of ${\cal U}^{(t)}$ under particle permutations and under spatial inversion ($\bm{r} _i'= - \bm{r}_i$). \[subsec:05b\]Series for OSE ---------------------------- Taking the logarithm of both sides of equation (\[equ:18\]) and passing to the limit $V \to \infty$, we obtain $$\ln{\Upsilon_v} = \lim_{V\to \infty} \left ( \ln{\Xi_V} - \ln{\Xi_{V-v}} \right ), \label{equ:77}$$ of course, if the intensive parameters of the environment and the parameters of the volume $v$ are preserved. Substitution of (\[equ:68\]) gives $$\begin{aligned} \ln{\Upsilon_v} &=& \lim_{V\to \infty} \sum_{t=1}^\infty \frac{z^t}{t!}\int {\cal U}^{(t)}_{1...t} \label{equ:78} \\ &&\times \left [ \prod_{i = 1}^{t} \psi^V_i - \prod_{i = 1}^{t} (\psi^V_i - \psi^v_i) \right ] d\bm{r}_1...d\bm{r}_t.\nonumber\end{aligned}$$ Using local character of Ursell factors (\[equ:a02\]), it can easily be shown that for OSE partition function we have the series $$\ln{\Upsilon_v} = \sum_{t=1}^\infty \frac{z^t}{t!} \int \left [1 - \prod_{i = 1}^{t} \chi^v_i \right ] {\cal U}^{(t)}_{1...t} d\bm{r}_1...d\bm{r}_t. \label{equ:79}$$ It is interesting to compare it with the expression for GCE partition function logarithm (\[equ:68\]). Separating volume and surface terms by the above technique, we obtain (\[equ:21\]) and (\[equ:74\]) with the replacement of $\psi^v_i \leftrightarrow \chi^v_i$ which is always possible due to the spatial symmetry of the problem. To derive the series for the general term of OSE distribution we apply the expression for distribution function (\[equ:a05\]). Substituting it in (\[equ:27\]) gives $$\begin{aligned} p^v_m &=& \frac{z^m}{m! \Upsilon_v}\sum_{t=0}^\infty \frac{z^t}{t!} \label{equ:80} \\ &\times& \int \left [ \prod_{i = 1}^m \prod_{j = m+1}^{m+t}\psi^v_i\chi^v_j \right ] {\cal B}^{(m,t)}_{1...m+t} d\bm{r}_1...d\bm{r}_{m+t}, \nonumber\end{aligned}$$ where ${\cal B}^{(m,k)}_{1...m+k}$ are partial localization factors intermediate between Boltzmann and Ursell factors (see Appendix \[subsec:appenda2\]). As is easily seen the first term of series (\[equ:80\]) coincides with GCE distribution (\[equ:06\]) by virtue of the equality ${\cal B}^{(m,0)}_{1...m} = \exp(-\beta U^{m}_{1...m})$ (\[equ:a06\]) accurate to normalizing factors - partition functions. The structure of series (\[equ:80\]) is of interest. The position of delocalized group of $m$ particles is defined by characteristic functions $\psi^v_i$ inside the system volume. The position of the localized group of $t$ particles outside it - by the functions $\chi^v_j$, but in close relation to the volume owing to functions ${\cal B}^{(m,t)}_{1...m+t}$. Summation is made over greater and greater clusters. One can say that volume properties of the distribution are specified by the first $m$ particles - ordinary, delocalized, while surface ones - by $t$ particles - localized. Since the distinctions between GCE and OSE are related to surface terms, therefore, all terms of series (\[equ:80\]) beginning with the 2-d term contain both the volume and the surface of the system in different powers. \[subsec:05c\]Convergence of series ----------------------------------- Convergence of series in powers of activity where the integrals of Ursell factors and partial localization factors are used as the expansion coefficients is studied, in particular, in paper [@ruellestatmeh1969]. References to other original works are also given in this paper. Summarizing the results, one can say that these series converge at least for dense gases where the convergence radius is defined by the interaction potential. In a great number of cases the convergence condition is of the form $$z \int {\cal U}^{(2)}_{1,2} d\bm{r}_2 \leq 1, \label{equ:81}$$ or something like that. The potentials with “hard” core - hard spheres or Lennard-Jones potentials (the so-called stable and regular potentials) satisfy the conditions providing the convergence of such series. Of course, the Coulomb potentials lead to divergence on the infinity. We can conclude that at least for such potentials and conditions of (\[equ:81\]) type the series considered in this section converge. \[sec:06\]Discussion ==================== \[subsec:06a\]OSE-limit ----------------------- In (\[equ:79\]), (\[equ:80\]) passing to OSE-limit is performed $$\begin{minipage}[c]{0.80\linewidth} \centering $V \rightarrow \infty$~~~ \\ $v = const$ \\ $ip = const$, \end{minipage} \label{equ:82}$$ where $ip$ are intensive parameters of the medium. Expression (\[equ:27\]) takes the form $$p^v_m = \frac{1}{m! \Upsilon_v} \int \left [ \prod_{i = 1}^m \psi^v_i \right ] \varrho^{(m)}_{1...m}(\chi^v) d\bm{r}_1...d\bm{r}_m. \label{equ:83}$$ Here $\varrho^{(m)}_{1...m}$ has the meaning of the probability density of finding some configuration of $m$ particles in the given volume $v$, on condition that all other particles are outside it. This probability density is defined for infinite homogeneous medium, and has the series of (\[equ:a05\]) type as one of representations in the cases where the series converges. The thermodynamic limit in the form of (\[equ:82\]) is preferable to the limit commonly used for GCE $$\begin{minipage}[c]{0.80\linewidth} \centering $V \rightarrow \infty$~~~ \\ $ip = const$. \end{minipage} \label{equ:84}$$ The thing is that in the last case, despite the increased size of the system, the walls of the vessel are always “stuck” to the system boundaries; this leads to surface deformation preservation and violation of the relation of the mean number of particles. This will be the subject of discussion in the next section. \[subsec:06b\]The difference between OSE and GCE ------------------------------------------------ For illustration in this section we consider two ensembles - GCE and OSE that are of the same volume $V$. Differentiating (\[equ:23\]), (\[equ:21\]) with respect to chemical potential, we have $$\frac{\partial\ln{\Xi_V}}{\partial \ln{z}} = N_b + N_s = \overline{N} \label{equ:85}$$ and $$\frac{\partial\ln{\Upsilon_V}}{\partial \ln{z}} = N_b - N_s = \overline{N} - N_s, \label{equ:86}$$ where $$N_b = V\frac{\partial P}{\partial \mu} \label{equ:87}$$ - the number of volume particles, and $$N_s = - A\frac{\partial \sigma}{\partial \mu} \label{equ:88}$$ - that of surface ones. As is known, the last equality in (\[equ:85\]) is readily obtained by differentiation (\[equ:08\]) with respect to chemical potential, and in (\[equ:86\]) we used expression (\[equ:38\]). Thus OSE and GCE distributions contain surface terms of different signs. The opposite sign in the surface term in (\[equ:23\]) is not accidental; GCE, as a set of closed systems, fails to completely reproduce an open system. Since each of the terms of series (\[equ:08\]) involves the surface corresponding to a closed system, GCE is a hybrid. This ensemble is open from the standpoint of the volume properties, because it describes adequately the fluctuations of the total number of particles, and is closed from the point of view of the surface properties that is obvious from (\[equ:85\]). Probably, it is not a coincidence that the author of paper [@Bellemans1962] has concluded that GCE describes a “drop”. It should only be added that it is a “drop” restricted by potential barriers. Now we can return to equation (\[equ:11\]). It follows from (\[equ:11\]) and (\[equ:85\]) that $$\int \psi^V_1 \varrho^{(1)}_{G,1} (\psi^V) d\bm{r}_1 = \overline{N} = N_b + N_s. \label{equ:89}$$ So we see that $\varrho^{(k)}_{G,1...k}$ involve surface terms, and (\[equ:89\]) does not fit the conception of the GCE as part of a homogeneous medium. For comparison, we derive from (\[equ:42\]) and (\[equ:86\]) for OSE $$\int \psi^V_1 \varrho^{(1)}_{1} d\bm{r}_1 = \overline{N} = N_b, \label{equ:90}$$ which is related to the absence of near surface density distortions in the given case. \[subsec:06c\]Small fluctuations -------------------------------- It should be emphasized once again that in the case under study the surface terms appear in the description of a homogeneous medium involving no real surface but just a hypothetical one restricting the separated system. This unexpected result becomes clearer taking into account that in statistical distributions only fluctuations within the limits of the separated volume are of interest to us. The states corresponding to these fluctuations already contain the restricting surface, and this gives rise to the above effect. Consequently, the surface terms are maximum for great fluctuations, and are compensated for small values. Really, it would be strange if the probability of finding the mean number of particles in a certain volume contained the surface. Such compensation actually takes place. First this is evident from the equality $m_b = \overline m$ itself. Since at small values of the number of particles in a given volume the probability depends on the value of the restricting surface, as is clear from (\[equ:29\]), and the mean value is independent of it, the only variant is compensation of the surface quantities for distribution terms with large $m$. Second, calculations show that for the first expansion terms in activity we have a rigorous compensation of surface terms near the mean values. Finally, for the general case it can easily be shown by considering OSE distribution near the mean values that the above compensation holds, and distribution terms corresponding to the mean values do not have surface components. \[subsec:06d\]Surface particles ------------------------------- Since $\sigma$, according to (\[equ:74\]), depends on two variables, we obtain the non-zero value of the number of surface particles $$N_S = - A \frac{\partial \sigma}{\partial \mu} = - \beta A z \frac{\partial \sigma}{\partial z}, \label{equ:91}$$ or, using (\[equ:73\]), $$N_S = \sum_{t=2}^\infty \frac{z^t}{(t-1)!} \int \psi^V_1 \left [\prod_{i = 2}^{t} \psi^V_i - 1 \right ] {\cal U}^{(t)}_{1...t} d\bm{r}_1 ...d\bm{r}_t. \label{equ:92}$$ Equality (\[equ:92\]) is almost evident taking into account the well-known formula for density expansion into a series in powers of activity for GCE $$\varrho_G(\bm{r}_{1},z) = z + z\sum_{n=1}^\infty \frac{z^n}{n!} \int \left [ \prod_{i = 2}^{n+1} \psi^V_i \right ]{\cal U}^{(n+1)}_{1...n+1} d\bm{r}_{2}...d\bm{r}_{n+1}, \label{equ:93}$$ following, for example, from (\[equ:85\]) or (\[equ:a05\]) and analogous quantity for OSE $$\varrho(z) = z + z\sum_{n=1}^\infty \frac{z^n}{n!} \int {\cal U}^{(n+1)}_{1...n+1} d\bm{r}_{2}...d\bm{r}_{n+1}, \label{equ:94}$$ corresponding to (\[equ:40\]), (\[equ:71\]). The number of surface particles is identified with the quantity $$N_S = \int \limits_V { \left [\varrho_G(\bm{r},z) - \varrho(z) \right ] d \bm{r}}, \label{equ:95}$$ and the surface number density $\varrho_S = N_S/A$ is defined by the expression $$\varrho_S = \int \limits_{L_t} { \left [\varrho_G(x,z) - \varrho(z) \right ] d x}, \label{equ:96}$$ where $x$ is the coordinate perpendicular to the surface, as before; $L_t$ is the transition region near the surface. In (\[equ:95\]), (\[equ:96\]) we use the expression for the near surface density of particles at the outer GCE boundary. However, it is easily seen that for hard solid immersed in the system the conclusion is the same. Formula (\[equ:95\]) coincides with the well-known expression for the surface number density at a impenetrable wall [@SokolowskiStecki1980; @HendersonSwol1984]. So, “surface tension” corresponds exactly to near surface (at hard solid or cavity) oscillations (deviations) of the number density. \[sec:07\]Summary ================= 1. A new ensemble - OSE (open statistical ensemble) is presented, the main peculiarity of which is a correct account of surface terms for an open system. This results in the replacement of the Bolzmann factors in configuration integrals of a grand canonical ensemble (GCE) by distribution functions of a specific type (\[equ:27\]). 2. OSE partion function involves volume and surface terms (\[equ:21\]), and in the expanded form is given by expression (\[equ:79\]). Similar expression for OSE distribution is given by sum (\[equ:80\]), with GCE being the first term up to a factor. 3. The “surface tension coefficient” involved in OSE partition function corresponds to non-zero number of surface particles which is unambiguously determined by the deviations of the number density from the mean value near fluctuation cavities (\[equ:95\]). 4. The expression for OSE partition function agrees with the thermodynamic approach for fluctuation formation probability (\[equ:30\]). 5. In contrast to GCE, OSE has the property of scale invariance: distributions in the initial and embedded volumes functionally coincide (\[equ:67\]). 6. OSE distribution has no lower bound of the volume, and may be directly applied to the studies of small systems (Section \[subsec:04d\]). 7. Thermodynamic and statistical OSE relations for dense gases are provided by recurrence relations (Appendix \[sec:appendb\]) of a new class of functions ${\cal B}^{(m,k)}_{1...m+k}$ (Appendix \[subsec:appenda2\]). 8. The functions ${\cal B}^{(m,k)}_{1...m+k}$ (\[equ:a04\]) generalize the concepts of Boltzmann and Ursell factors (\[equ:a01\]), and involve them as the extreme values (\[equ:a06\]), (\[equ:a07\]). 9. The basic expressions and properties of OSE may be obtained both on the level of factors, and on the level of distribution functions. Thus problems related to the convergence of series in number density (activity) may be avoided. For example, there are two types of expressions for the general term of OSE distribution (\[equ:80\]), (\[equ:83\]). 10. For a homogeneous medium the mean number of particles for OSE, unlike GCE, involves no surface terms: $\overline m = m_b$. 11. GCE is a hybrid the volume terms of which correspond to an open system, and the surface ones - to a closed system (Section \[subsec:06b\]). In particular, GCE partition function contains surface terms with the sign that is opposite to OSE one (\[equ:21\]), and corresponds to a closed system (\[equ:23\]). \[sec:appenda\]Factors ====================== \[subsec:appenda1\]Ursell factors --------------------------------- Ursell factors ${\cal U}^{(k)}_{1...k}$ are also called cluster functions. Just these factors appear in the Mayer well-known expansion in powers of activity for pressure (\[equ:71\]) [@hillstatmeh1987 p.129], [@landaulifshitz1985 p.232]. (From this it follows that a logarithm serves as a generating function for Ursell factors). First they were introduced in paper [@ursell1927]. They may be defined by the equality $$\begin{aligned} {\cal U}^{(k)}_{1...k} &=& \sum_{\{\bm{n}\}}(-1)^{l-1}(l-1)!\prod_{\alpha = 1}^l \exp(-\beta U^{k_\alpha}(\{\bm{n}_\alpha\})), \nonumber \\ 1 & \leq & ~ k_\alpha \leq k, ~~ \sum_{\alpha = 1}^l k_\alpha = k, ~~ \exp(-\beta U^{1}) = 1, \label{equ:a01}\end{aligned}$$ where $\{\bm{n}\}$ denotes some partition of the given set of $k$ particles with the coordinates $\bm{r}_1,...\bm{r}_k$ into disjoint groups $\{\bm{n}_\alpha\}$, $l$ is the number of groups of a particular partition, $k_\alpha$ is the size of the group with the number $\alpha$, the sum is taken over all possible partitions, and the meaning of the condition $\exp(-\beta U^{1}) = 1$ is apparent: single groups make no contribution into the products. For illustration, several first ${\cal U}^{(k)}_{1...k}$ are given $$\begin{aligned} {\cal U}^{(1)}_{1}\mkern 9mu &=& 1 \label{equ:a02} \\ {\cal U}^{(2)}_{1,2} \mkern 8mu &=& \exp(-\beta U^{2}_{1,2}) - 1 \nonumber \\ {\cal U}^{(3)}_{1,2,3} &=& \exp(-\beta U^{3}_{1,2,3}) - \exp(-\beta U^{2}_{1,2}) \nonumber \\ &-& \exp(-\beta U^{2}_{1,3}) - \exp(-\beta U^{2}_{2,3}) + 2 \nonumber \\ \dotso \nonumber\end{aligned}$$ Ursell factors decay rapidly when any group of particles (including single one) moves away. \[subsec:appenda2\]Partial localization factors ----------------------------------------------- These quantities are a hybrid of Boltzmann and Ursell factors. They play an important role in OSE mathematical formalism. As far as we know, they were introduced for the first time in paper [@ruellestatmeh1969]. A portion of particles appearing in these functions do not cause the decay when they move away (delocalized group), another portion does it (localized one). Introduce the designation for them $${\cal B}^{(m,k)}_{1...m+k}. \label{equ:a03}$$ Here the superscripts $m$ and $k$ define the number of delocalized and localized particles, respectively ($m = 1,2,3,\dots, k = 0,1,2,\dots$). The subscripts denote the coordinates of particles. Let us take that the first $m$ particles are delocalized, and other particles are localized. The structure of these functions is similar to that of Ursell factors of the $k+1$ rank, however, note that in construction by (\[equ:a01\]) type, the first $m$ particles (delocalized) are treated as one compound particle. In other words, define ${\cal B}^{(m,k)}_{1...m+k}$ by the equality $$\begin{aligned} {\cal B}^{(m,k)}_{1...m+k} &=& \sum_{\{\bm{n}\}}(-1)^{l-1}(l-1)!\label{equ:a04}\\ &\times& \prod_{\alpha = 1}^l \exp(-\beta U^{k_\alpha+(m-1)\delta_{\alpha\nu}}(\{\bm{n}_\alpha\})), \nonumber \\ 1 \leq& k_\alpha& \leq k+1; ~~ \sum_{\alpha = 1}^l k_\alpha = k+1; ~~ \exp(-\beta U^{1}) = 1,\nonumber \end{aligned}$$ where the designations are analogous to (\[equ:a01\]) on condition that the sum is taken over all possible partitions of the set of $k+1$ particles including one compound particle. $\delta_{\alpha\nu}$ is the Kroneker delta, and $\nu$ is the group number, which involves a compound particle. The generating function for ${\cal B}^{(m,k)}_{1...m+k}$ is distribution function of GCE type $\varrho^{(m)}_{G,1...m}$. Expanding the partition function $\Xi_V$ in (\[equ:10\]), and dividing the series, we have $$\begin{aligned} &&\varrho^{(m)}_{G,1...m} (\psi^V) = z^m \left \{ {\cal B}^{(m,0)}_{1...m} + \sum_{k=1}^\infty \frac{z^k}{k!} \right . \label{equ:a05} \\ && ~~~~~~ \left . \times \int \left [ \prod_{i = m+1}^{m+k} \psi^V_i \right ] {\cal B}^{(m,k)}_{1...m+k} d\bm{r}_{m+1}...d\bm{r}_{m+k} \right \}. \nonumber\end{aligned}$$ The proof of this relation is given in Appendix \[subsec:appendb1\]. ${\cal B}^{(m,k)}_{1...m+k}$ first in the localized group are $$\begin{aligned} {\cal B}^{(m,0)}_{1...m} \mkern 12mu &=& \exp(-\beta U^{m}_{1...m})\label{equ:a06} \\ {\cal B}^{(m,1)}_{1...m+1} &=& \exp(-\beta U^{m+1}_{1...m+1}) - \exp(-\beta U^{m}_{1...m})\nonumber \\ {\cal B}^{(m,2)}_{1...m+2} & = & \exp(-\beta U^{m+2}_{1...m+2}) - \exp(-\beta U^{m+1}_{1...m+1})\nonumber \\ &-& \exp(-\beta U^{m+1}_{1...m,m+2}) - \exp(-\beta U^{m}_{1...m}) \nonumber \\ &\times& \exp(-\beta U^{2}_{m+1,m+2}) + 2\exp(-\beta U^{m}_{1...m}) \nonumber \\ \dotso \nonumber\end{aligned}$$ and in the delocalized one - $${\cal B}^{(1,k-1)}_{1...k} = {\cal U}^{(k)}_{1...k} \label{equ:a07},$$ including the case where $k=1$ (for homogeneous medium) $${\cal B}^{(1,0)}_{1} = 1 \label{equ:a08}.$$ As is seen from (\[equ:a06\]), (\[equ:a07\]), partial localization factors generalize the notions of Boltzmann and Ursell factors including them as the limiting cases. For the factors ${\cal B}^{(m,k)}_{1...m+k}$ a number of recurrence relations are valid (see Appendix \[sec:appendb\]) that provide the fulfillment of various physical relations. \[sec:appendb\]Recurrence relations for ${\cal B}^{(m,k)}_{1...m+k}$ ==================================================================== So, each operation with OSE distribution is ensured by a definite class of recurrence relations for ${\cal B}^{(m,k)}_{1...m+k}$. For brevity, we shall say that operation generates a recurrence relation or a class. Since operations considered below have already been proved by alternative methods, this means that the recurrence relations generated by them are also proved. Let us examine some of these operations. \[subsec:appendb1\]Correspondence to the definition --------------------------------------------------- Substituting expression (\[equ:10\]) in (\[equ:a05\]), expanding the series for $\Xi_V$ and performing multiplication of the series, we obtain $${\cal B}^{(m,k)}_{1...m+k} = {\cal B}^{(m+k,0)}_{1...m+k} - \sum_{n=1}^k \sum_{perm} {\cal B}^{(n,0)}_{1...n}{\cal B}^{(m,k-n)}_{n+1...m+k}, \label{equ:b01}$$ where $m \geq 1$, $k \geq 1$, and the internal sum is taken over samplings of localized particles only ($n$ from $k$). Relation (\[equ:b01\]) is proved either by direct enumeration of partitions according to (\[equ:a04\]), or by repeat substitution of the expression for ${\cal B}^{(m,k)}_{1...m+k}$ in the right-hand side of (\[equ:b01\]). Paper [@percus1964] gives another recurrence relation for Ursell functions which in terms of ${\cal B}^{(m,k)}_{1...m+k}$ looks like $${\cal B}^{(1,k)}_{1...k+1} = {\cal B}^{(k+1,0)}_{1...k+1} - \sum_{n=1}^{k} \binom{k}{n} \left [ {\cal B}^{(n,0)}_{1...n} {\cal B}^{(1,k-n)}_{n+1...k+1}\right ]_{perm}, \label{equ:b02}$$ where $k \geq 1$, and square brackets mean averaging over particles permutations. In principle, this is the same equation (\[equ:b01\]) at $m=1$ but written in a symmetric form. The question of symmetrization of all recurrence relations for ${\cal B}^{(m,k)}_{1...m+k}$ that will arise further is beyond the scope of the present contribution. However, note that in our case symmetrization is not needed: equation (\[equ:b01\]) holds rigorously in asymmetric form as well. Nevertheless, (\[equ:b01\]) may also be represented in a symmetric form; however, this requires that both the sum and the left-hand side of the equation be averaged over permutations. It is not improbable that all relations given below can also have asymmetric form but we shall employ a symmetric form of them. In some cases the symbol $\dosimm$ will be used. We imply that symmetrization is performed in the right- and left-hand sides of the equality over the indices where it is necessary. In the given case symmetrization will also mean, along with averaging over permutations, the aligning of indices to natural series, if necessary. This operation is due to property (\[equ:a08\]). \[subsec:appendb2\]Normalization condition ------------------------------------------ $\sum p^v_m$ over all $m$ must be equal to 1, and we arrive at the equation $$\begin{aligned} \Upsilon_v &=&1 + \sum_{m=1}^\infty \frac{z^m}{m!}\sum_{t=0}^\infty \frac{z^t}{t!} \label{equ:b03} \\ &\times& \int \left [ \prod_{i = 1}^m \prod_{j = m+1}^{m+t}\psi^v_i\chi^v_j \right ] {\cal B}^{(m,t)}_{1...m+t} d\bm{r}_1...d\bm{r}_{m+t}. \nonumber\end{aligned}$$ This is another form of OSE partition function. Here it is not easy to see expression (\[equ:79\])! To show the identity of these formulae, we consider their logarithmic derivatives. Thus $$\frac{1}{\Upsilon_v}\frac{\partial\Upsilon_v}{\partial z} = \beta \frac{\partial(Pv+\sigma a)}{\partial z}, \label{equ:b04}$$ and $\Upsilon_v$ in (\[equ:b04\]) is to be defined by expression (\[equ:b03\]), while $P v$ and $\sigma a$ - by (\[equ:70\]) and (\[equ:73\]), respectively, where ${\cal U}^{(t)}_{1...t}$ are replaced by ${\cal B}^{(1,t-1)}_{1...t}$, according to (\[equ:a07\]). To simplify calculations, we change the order of summation in the expressions. For (\[equ:b03\]): $$\begin{aligned} \Upsilon_v &=&1 + \sum_{s=1}^\infty \frac{z^s}{s!}\int\sum_{m=1}^s \binom{s}{m} \label{equ:b05} \\ && \times \left [ \prod_{i = 1}^m \prod_{j = m+1}^{s}\psi^v_i\chi^v_j \right ] {\cal B}^{(m,s-m)}_{1...s} d\bm{r}_1...d\bm{r}_{s}. \nonumber\end{aligned}$$ Performing differentiation and multiplication of the series, and equating the expressions at different powers of $z$ and products $\psi^v_i$, we obtain the recurrence relations $$\begin{aligned} {\cal B}^{(m,k)}_{1...m+k} &\dosimm& {\cal B}^{(m-1,k+1)}_{1...m+k} \label{equ:b06} \\ &+& \sum_{n=0}^k \binom{k}{n} {\cal B}^{(m-1,k-n)}_{1...m+k-n-1} {\cal B}^{(1,n)}_{m+k-n...m+k}, \nonumber\end{aligned}$$ valid at $k \geq 0$ and $m \geq 2$, and $${\cal B}^{(1,k)}_{1...k+1} = \sum_{m=1}^{k} (-1)^m \binom{k}{m} \left [ {\cal B}^{(m,k-m)}_{1...k} - {\cal B}^{(m+1,k-m)}_{1...k+1}\right ]_{perm}, \label{equ:b07}$$ where $k \geq 1$, and the class of relations having two free indices and containing low indices of localized particles $$\begin{aligned} \sum_{m=0}^{k}&(-1)^{m}& \binom{k}{m} {\cal B}^{(m+t+1,k-m)}_{1...t+k+1}~ \dosimm ~\sum_{l=0}^{k} \binom{k}{l} \label{equ:b08}\\ \times &(-1)^{l}& \sum_{n=1}^{t} \binom{t}{n} {\cal B}^{(l+n,k-l)}_{1...k+n} {\cal B}^{(1,t-n)}_{k+n+1...k+t+1} \nonumber \\ &+& \sum_{s=1}^{k} (-1)^{s} \binom{k}{s} {\cal B}^{(s,k-s)}_{1...k} {\cal B}^{(1,t)}_{k+1...k+t+1}, \nonumber\end{aligned}$$ where $k \geq 1$, $t\geq 1$ and which at $k = 1$ reduces to (\[equ:b02\]). \[subsec:appendb3\]Calculation of $\overline{m}$ ------------------------------------------------ We proceed from the natural assumption $\overline{m} = m_b$ (\[equ:38\]). The outline of the proof is as follows. According to (\[equ:80\]), $\overline{m}$ has the form $$\begin{aligned} \overline{m} &=& \frac{1}{\Upsilon_v}\sum_{m=1}^\infty\frac{z^m}{(m-1)!} \sum_{t=0}^\infty \frac{z^t}{t!} \label{equ:b09} \\ &\times& \int \left [ \prod_{i = 1}^m \prod_{j = m+1}^{m+t}\psi^v_i\chi^v_j \right ] {\cal B}^{(m,t)}_{1...m+t} d\bm{r}_1...d\bm{r}_{m+t}. \nonumber\end{aligned}$$ Let us change the order of summation $$\begin{aligned} \overline{m} &=& \frac{1}{\Upsilon_v}\sum_{s=1}^\infty\frac{z^s}{(s-1)!}\int \sum_{m=1}^s \binom{s-1}{m-1} \label{equ:b10} \\ &\times& \left [ \prod_{i = 1}^m \prod_{j = m+1}^{s}\psi^v_i\chi^v_j \right ] {\cal B}^{(m,s-m)}_{1...s} d\bm{r}_1...d\bm{r}_{s}. \nonumber\end{aligned}$$ This expression must be set equal to (\[equ:87\]), where (\[equ:70\]) should be used for $P$. Multiplying the series and equating the expressions at different powers of $z$ and the products $\psi^v_i$, we get a class of recurrence relations containing low indices of delocalized particles $$\begin{aligned} \sum_{m=0}^{k}(-1)^{m}&& \binom{k}{m} {\cal B}^{(m+1,t+k-m)}_{1...t+k+1} ~ \dosimm ~ \sum_{l=1}^{k} (-1)^{l} \label{equ:b11}\\ \times && \binom{k}{l}\sum_{n=0}^{t} \binom{t}{n} {\cal B}^{(l,k+n-l)}_{1...k+n} {\cal B}^{(1,t-n)}_{k+n+1...k+t+1}, \nonumber\end{aligned}$$ where $k \geq 1$, $t\geq 0$, which at $k = 1$ reduces to (\[equ:b06\]) with $m=2$, and at $t = 0$ - to (\[equ:b07\]). [11]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [[**]{}]{}, Vol.  (, , ) @noop [****,  ()]{} @noop [[**]{}]{} (, , ) @noop [[**]{}]{},  ed., Vol.  (, , ) @noop [“,” ]{} (),  @noop [[**]{}]{} (, , ) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} in @noop [[**]{}]{},  (, , ) pp.  VZ, 12.08.2011, v82.

ArXiv

--- abstract: 'Graph neural networks (GNNs) have been attracting increasing popularity due to their simplicity and effectiveness in a variety of fields. However, a large number of labeled data is generally required to train these networks, which could be very expensive to obtain in some domains. In this paper, we study active learning for GNNs, i.e., how to efficiently label the nodes on a graph to reduce the annotation cost of training GNNs. We formulate the problem as a sequential decision process on graphs and train a GNN-based policy network with reinforcement learning to learn the optimal query strategy. By jointly optimizing over several source graphs with full labels, we learn a transferable active learning policy which can directly generalize to unlabeled target graphs under a zero-shot transfer setting. Experimental results on multiple graphs from different domains prove the effectiveness of our proposed approach in both settings of transferring between graphs in the same domain and across different domains.' author: - | Shengding Hu\ Tsinghua University\ `[email protected]`\ Zheng Xiong\ Tsinghua University\ `[email protected]`\ Meng Qu\ MILA\ `[email protected]`\ Xingdi Yuan\ Microsoft Research\ `[email protected]`\ Marc-Alexandre Côté\ Microsoft Research\ `[email protected]`\ Zhiyuan Liu\ Tsinghua University\ `[email protected]`\ Jian Tang\ HEC Montreal & MILA\ `[email protected]`\ bibliography: - 'reference.bib' title: | Graph Policy Network for Transferable\ Active Learning on Graphs ---

ArXiv

--- abstract: 'Most of the routing algorithms for unmanned vehicles, that arise in data gathering and monitoring applications in the literature, rely on the Global Positioning System (GPS) information for localization. However, disruption of GPS signals either intentionally or unintentionally could potentially render these algorithms not applicable. In this article, we present a novel method to address this difficulty by combining methods from cooperative localization and routing. In particular, the article formulates a fundamental combinatorial optimization problem to plan routes for an unmanned vehicle in a GPS-restricted environment while enabling localization for the vehicle. We also develop algorithms to compute optimal paths for the vehicle using the proposed formulation. Extensive simulation results are also presented to corroborate the effectiveness and performance of the proposed formulation and algorithms.' author: - 'Kaarthik Sundar$^{\dagger}$[^1], Sohum Misra$^{*}$[^2], Sivakumar Rathinam$^{\ddagger}$[^3], Rajnikant Sharma$^{*}$' bibliography: - 'references.bib' title: '**Routing Unmanned Vehicles in GPS-Denied Environments** ' --- path planning; localization; mixed-integer linear programming; unmanned aerial vehicles; GPS-denied environments Introduction {#sec:intro} ============ Unmanned Vehicles (UVs), both aerial and ground, have been finding its use in a plethora of civilian [@Johnson2017; @Ou2014] and indoor applications (see [@Nonami2007; @Chen2014; @Tomic2012] and references therein). Knowledge of position and orientation is necessary to ensure decision-making and autonomy for these vehicles. The problem of localization deals with the estimation of position and orientation of the UVs. Localization therefore requires sensing, and correspondingly, localization procedures are dependent of the available sensory measurements. Since sensory measurements are usually contaminated with noise, the problem of localization also requires filtering the noise in order to determine an accurate estimate of location and orientation. The investment in GPS-based vehicle positioning systems that rely on the GPS measurements have garnered a lot of attention in the literature [@Reina2007]. However, most indoor environments and many parts of the urban canyon do not have access to GPS; even if available, the access is intermittent and not reliable. Hence, localization in a GPS-denied or GPS-restricted environment is an active area of research; it also has additional advantages of robustness, efficiency, and flexibility [@Sharma2013]. In this paper, we present a joint route optimization and localization algorithm in a GPS-denied environment to visit a set of targets while enabling localization as the UV traverses its route. The proposed approach uses known objects in the environment referred to as landmarks (LMs) to aid in localization. The UV determines it relative position and orientation with respect to the LMs using exteroceptive sensor like camera, laser, etc. and assess its motion and attitude information such as velocity and turn rate using interoceptive sensors like IMU, encoders, etc. This enables the UV to estimate its states and localize itself in a GPS-denied environment. Conditions under which the UV would be able to estimate its states and localize itself using the LMs are first derived. These conditions are then embedded into an optimization framework where the sequence in which the UV should visit the targets, the route it takes, and the location where the LMs should be placed to enable localization are optimized. In this way, we are ensured that the UV can perform localization using the LMs as it traverses its route and visits the targets. The problem statement is can be summarized as follows: We remark that, it is also easier to think about the above problem as a single vehicle routing problem with additional constraints for LM placement to aid in localization. Henceforth, we shall refer to this problem as a single vehicle routing problem using LMs to aid localization, abbreviated as SVRP-LL. SVRP-LL, being a generalization of the traveling salesman problem (TSP), is NP-hard. Related work {#sec:lit_review} ============ The problem of localization of a vehicle, aerial and ground, in urban and indoor environments where GPS signals are not always reliable is well studied in the literature. In particular, authors in [@Wong2014] used techniques in computer vision to address the problem. Numerous variants of Simultaneous Localization and Mapping (SLAM) techniques have also been developed for high precision vehicle localization [@Levinson2007; @Weiss2011]. Another frequently used method is infrastructure-aided localization for aerial and ground vehicles. In particular, infrastructure capable of providing range measurements to the vehicles are pre-installed in the environment and algorithms are developed to estimate the position and orientation of the vehicles [@Mao2007]. The problem of localization is also of independent interest to infrastructure-aided localization for transportation applications. The idea of infrastructure-aided navigation and control for automated vehicles is not new and has been considered at least since California PATH’s Automated Highway Systems (AHS) program. However, this idea is useful for other applications such as Advanced Traffic Management Systems (ATMS) and Advanced Traveler Information System (ATIS); one can design V2I [@Doan2009] (vehicle to infrastructure) communication protocols by which the type of vehicle is identified along with the time stamp for the communication thereby obviating the need for traffic measurement devices such as loop detectors which are error-prone. Authors in [@Ou2014] proposed a Vehicular Ad-Hoc Network (VANET) based localization approach in which each vehicle estimates its location on the basis of beacon messages broadcast periodically by pairs of Road Side Units (RSUs) deployed on either side of the road. Authors in [@Khattab2015] modified the previous approach by using two-way time of arrival information to localize vehicles based on communication with a single road side unit (RSU or LM) instead of having 2 RSUs/LMs. The first work to consider placement of LMs given a paths for multiple vehicles is [@Rathinam2015]. The authors formulated the landmark placement problem as a multiple vehicle path covering problem and presented an approximation algorithm using geometric arguments to address the problem. This article is an extension of the work in [@Rathinam2015] on three fronts: (1) we formulate the joint vehicle routing and landmark placement problem as a mixed-integer linear program (MILP), (2) we present an algorithm to compute an optimal solution to the MILP, and finally, (3) we present extensive computational and simulation results showing the effectiveness of the proposed approach with a suitable estimation algorithm. The rest of the paper is organized as follows: in Sec. \[sec:vm\], the vehicle model and the algorithm to localize the vehicle when the position of the LMs are known a priori are detailed; conditions under which the localization algorithm would provide accurate position and orientation estimates are also discussed. Sec. \[sec:defn\], \[sec:formulation\], and \[sec:bandc\] present the optimization problem definition, formulation, and the branch-and-cut algorithm to solve the SVRP-LL to optimality, respectively. Finally, the controller architecture and the computational results are detailed in Sec. \[sec:arch\] and \[sec:results\], respectively. Vehicle model & Localization algorithm {#sec:vm} ====================================== For the localization problem, an unmanned ground vehicle or an unmanned aerial vehicle flying at a constant altitude and traveling with a constant velocity $v$ is considered; the kinematic constraints of the vehicle in state-space form is as follows: = f(x, u) v ()\ v ()\ where, $v$ is the velocity of the vehicle, $\psi$ is the heading of the vehicle and $\omega$ is the rate of change of heading with respect to time. The vector $\bm x$ is the vector of state variables given by $(x, y, \psi)^T$; here $x$ and $y$ denote the position of the vehicle and $\psi$ denotes the heading information. The control input vector, $\bm u$, for the vehicle consists of $v$ and $\omega$. The interoceptive sensors detect the velocity, $v$, and angular velocity, $\omega$, of the vehicle. The exteroceptive sensors are used to detect the relative bearing measurement of the vehicle with respect to the known LMs. The vehicle is assumed to have a complete $360^{\circ}$ field of view. Without loss of generality, it is considered that the vehicle cannot move backwards, *i.e.*, $v \geqslant 0$. Furthermore, the sensing range of the vehicle’s exteroceptive sensor, denoted by $\rho_s$, is assumed to be constant. To route the UVs, the precise knowledge of position and heading of the vehicles is necessary. In GPS-denied environments, relative position measurements using range or bearing angle sensors to known landmarks can be used to localize the vehicle. An Extended Kalman Filter (EKF) or its information form given by the Extended Information Filter (EIF) can be used to estimate the vehicle’s states $\bm x$. This article uses the EIF instead of the EKF for to estimate the vehicle’s state $\bm x$ as the EIF is more favorable from a computational perspective. The estimation algorithm will provide meaningful localization estimates (consistent and bounded) if and only if the sensors provide enough information for localization, or in other words, if the system is observable. It has been shown that the bound on uncertainty is related with the eigenvalues of the observability gramian [@Song1992]. In our previous work [@Sharma2012], we have shown that the UV needs bearing angle measurements from 2 different landmarks in order for the system to be observable using the Hermann-Krener criterion [@Hermann1977] for local observability. This technique of state estimation using the bearing information alone is referred to as *bearing-only localization* [@Sharma2012; @Sharma2013]. The condition for enabling accurate estimation of the states of the vehicle can also be illustrated using a relative position measurement graph (RPMG). A RPMG is a graph that is used to represent the interaction and information flow between the vehicle and LMs. The RPMG consists of two types of nodes: the vehicle position at different time instants and the landmarks. An edge between a landmark $i$ and the vehicle at a particular time instant $t$ indicates that the vehicle obtains bearing measurement from the landmark $i$. An example of RPMG with single vehicle and multiple LMs with edges between the vehicle at time $t_1$ and $t_2$ and the landmarks is illustrated in Fig. \[fig:RPMG\]. The main result of [@Sharma2012] that will be used in the forthcoming sections is that for the estimation algorithm, using an EIF, to provide accurate localization estimates for the vehicle at any given time $t$, the RPMG should contain at least two edges from the node that represents the vehicle at time $t$ to two different LMs. We also remark that having path to more than 2 LMs provides more information to the vehicles, thereby quickening the convergence rate of the estimation algorithm. Optimization problem definition {#sec:defn} =============================== In the previous section, we elucidated the condition that enables localization of the vehicle from bearing measurements which is that, at any given instant of time the vehicle requires bearing information from at least two LMs for observability and for the estimation algorithm to converge to accurate state estimates. We shall now enforce this condition explicitly in the optimization problem to jointly optimize the routes and landmark placement. To that end, let $V$ denote the set of targets $\{v_1, \dots, v_n\}$ and let the vehicle be initially stationed at $v_1$. Let $K$ denote the set of potential locations where a LM can be placed. Associated with each vertex $v_i$ is a subset of locations, $K_{v_i}$; if a LM is placed in a location in $K_{v_i}$, then it can provide bearing measurement for the vehicle when the vehicle is at the target $v_i$. We note that since the vehicle has a $360^{\circ}$ field of view and the sensing range of the exteroceptive sensor on the vehicle is $\rho_s$, a location $k$ is in the set $K_{v_i}$ if and only if the distance between the location and the target $v_i$ is less than $\rho_s$, the sensing range of the exteroceptive sensor. The vehicle can perform localization along an edge $e \in \{(v_i,v_j): i<j\}$ if it has bearing measurements from at least 2 LMs as it traverses that edge. Hence, associated with each edge $e \in \{(v_i,v_j): i<j\}$ is a subset of potential LM locations, $K_e \subseteq K$; for a given edge $e \in \{(v_i,v_j): i<j\}$, $k \in K_e$ if and only if $k \in K_{v_i} \cap K_{v_j}$. For the vehicle to be able to localize itself as it traverses the edge $e$, LMs have to placed at at least two locations in $K_e$. Now, the problem is defined on an undirected graph $G=(V, E)$ where $E$ is the set of edges $\{e=(v_i, v_j):i<j\}$. We say that a set of LMs “cover” an edge $e$ if they are installed in a subset of locations $K_1$ such that $|K_1 \cap K_e| \geq 2$. The set of LMs that cover an edge can provide bearing measurements for the vehicle as it traverses that edge and thereby enabling the vehicle localization. Each edge $(v_i, v_j) \in E$ is associated with a non-negative cost $c_{ij}$ required to travel from target $i$ to target $j$. Also associated with each location $k\in K$ is a non-negative cost $d_k$ that denotes the cost for installing a LM at location $k$; if we wish to minimize the number of LMs placed then each $d_k$ takes a value $1$. The SVRP-LL consists for finding a path for the vehicle and placing/installing LMs in a subset of locations $K$ such that (i) each target in the set $V$ is visited at least once by the vehicle, (ii) each edge traversed by the vehicle is covered by at least two installed LMs , and (iii) the sum of the cost of the path and installation cost of the LMs is a minimum. Mathematical formulation {#sec:formulation} ======================== We now present a mathematical formulation for the SVRP-LL, inspired by the models for the standard routing and covering problems [@Toth2014] (see [@Sundar2016; @Sundar2015; @Manyam2016; @Sundar2016a; @Sundar2016b; @Sundar2016c] for routing and path planning problems concerning UVs). We associate to each feasible solution $\mathcal F$ for the SVRP-LL, a vector $\bm x \in \mathbb{R}^{|E|}$ (a real vector indexed by the elements of the edge set $E$) where each component $x_e$ of $\bm x$ takes a value one if the vehicle traverses the edge and zero otherwise. Similarly, associated to $\mathcal F$, is also a vector $\bm y \in \mathbb{R}^{|K|}$; the value of the component $y_k$ associated with a potential sensor location $k \in K$ is equal to one if a sensor is placed in the location $k$ and zero otherwise. For any $S \subseteq V$, we define $\delta(S) = \{(i,j) \in E: i\in S, j\notin S\}$. If $S = \{i\}$, we simply write $\delta(i)$ instead of $\delta(\{i\})$. Finally for any $\mathcal E \subset E$, we define $x(\mathcal E) = \sum_{e \in \mathcal E} x_e$. Using the notations introduced thus far, the SVRP-LL is formulated as a mixed-integer linear program as follows: & \_[eE]{} c\_e x\_e + \_[kK]{} d\_k y\_k &\[eq:obj\] &\ & &\ & x((v\_i)) = 2 v\_i V, &\[eq:degree\]\ & x((S)) 2 S V, &\[eq:sec\]\ & \_[k K\_e]{} y\_k 2 x\_e e E, &\[eq:sensor\]\ & x\_e {0,1}, y\_k {0,1} eE, k K. & \[eq:integer\] In the above formulation, the constraints are the degree constraints for the targets and they ensure the number of edges incident on any target is $2$. The constraints are the sub-tour elimination constraints and they ensure that any feasible route for the vehicle has no sub-tours of any subset of the targets $V$. The constraints ensure that each edge $e$ that is traversed by the vehicle is covered by a subset of installed LMs to enable vehicle localization as it traverses the edge $e$. Finally, the constraints are the binary restrictions on the $x_e$ and $y_k$ variables. In the next section, we present a branch-and-cut algorithm to solve the mathematical formulation to optimality. Branch-and-cut algorithm {#sec:bandc} ======================== In this section, we briefly present the main ingredients of a branch-and-cut algorithm that is used to solve the two formulations presented in the previous section to optimality. The formulation can be provided to off-the-shelf commercial branch-and-cut solvers like Gurobi or CPLEX to obtain an optimal solution to the SVRP-LL. But, the formulation contains exponential number of sub-tour elimination constraints in and complete enumeration of all the constraints to provide the formulation to the solver would be computationally intractable. To address this issue, we use the following approach: we relax these constraints from the formulation, and whenever the solver obtains an integer solution feasible to this relaxed problem (or a fractional solution with integrality constraints dropped), we check if any of these constraints are violated by the feasible solution, integer or fractional. If so, we add the infeasible constraint and continue solving the original problem; we refer to this technique as a dynamic cut-generation or cutting plane algorithm. This process of adding constraints to the problem sequentially has been observed to be computationally efficient for the TSP and some of its variants [@Sundar2016; @Sundar2015a; @Sundar2016a]. The algorithms used to identify violated constraints are referred to as separation algorithms. For the sake of completeness, a detailed pseudo-code of the branch-and-cut algorithm for the SVRP-LL is given detailed. To that end, let $\bar \tau$ denote the optimal solution to an instance of SVRP-LL.\ S[<span style="font-variant:small-caps;"></span>]{} 1 (Initialization). Set the iteration count $t \gets 1$ and the initial upper bound to the optimal solution $\bar \alpha \gets + \infty$. The initial linear sub-problem is then defined by formulation in Sec. \[sec:formulation\] without the sub-tour elimination constraints in and the binary restrictions on the variables relaxed. The initial sub-problem is solved and inserted in a list $\cal L$. S[<span style="font-variant:small-caps;"></span>]{} 2 (Termination check and sub-problem selection). If the list $\cal L$ is empty, then stop. Otherwise, select a sub-problem from the list $\cal L$ with the smallest objective value. S[<span style="font-variant:small-caps;"></span>]{} 3 (Sub-problem solution). $t \gets t + 1$. Let $\alpha$ denote the objective value of the sub-problem solution. If $\alpha \geqslant \bar \alpha$, then proceed to S[<span style="font-variant:small-caps;"></span>]{} 2. If the solution is feasible for the SVRP-LL, then set $\bar \alpha \gets \alpha$, update $\bar \tau$ and proceed to S[<span style="font-variant:small-caps;"></span>]{} 2. S[<span style="font-variant:small-caps;"></span>]{} 4 (Constraint separation and generation). Using the separation algorithm for the sub-tour elimination constraints, identify the violated constraints . Add the violated constraints to the initial linear sub-problem and proceed to S[<span style="font-variant:small-caps;"></span>]{} 3. If no constraints are generated, then proceed to S[<span style="font-variant:small-caps;"></span>]{} 5. S[<span style="font-variant:small-caps;"></span>]{} 5 (Branching). Create two sub-problems by branching on a fractional $x_e$ or $y_i$ variable. Then insert both the sub-problems in the list $\cal L$ and go to S[<span style="font-variant:small-caps;"></span>]{} 2.\ Now, we detail the separation algorithm used in S[<span style="font-variant:small-caps;"></span>]{} 4 to identify violated sub-tour elimination constraints. To that end, let $G^* = (V^*, E^*)$ denote the support graph associated with a given fractional solution $(\bm x^*, \bm y^*)$ i.e., $V^* = T$ and $E^* := \{e \in E: x_e^* > 0\}$. Here, $\bm x$ and $\bm y$ are the vector of decision variable values in SVRP-LL. Next, we examine the connected components in $G^*$. Each connected component that does not contain all the targets in $T$ generates a violated sub-tour elimination constraint for $S = C$. If the number of connected components is one, then the most violated constraint of the form $x(\delta(S)) \geqslant 2$ can be obtained by computing the global minimum on a capacitated undirected graph $G^*$; let the cut be denoted by $(S, V^* \setminus S)$. $S$ defines a violated sub-tour elimination constraint if the value of the cut is strictly less than 2. System architecture {#sec:arch} =================== The branch-and-cut algorithm in the previous section provides the sequence in which the vehicle should visit the targets and the locations where the LMs should be placed. The constraints ensure that the vehicle can perform localization when the LMs when placed at the specified locations. ![Block diagram showing the system architecture for estimating the states.[]{data-label="fig:arch"}](ControlArch) The Fig. \[fig:arch\] shows the block diagram of the system architecture that is used for estimating the states using the bearing measurements from the LMs. The dashed lines denote that the computation is performed offline, which in this case is the solution to the optimization problem using the branch-and-cut algorithm. The sequence in which the UV visits the targets provides the way points for the paths. The bearing sensors are the exteroceptive sensors on the vehicle which obtain the bearing information from the LMs placed on the locations chosen by the optimization problem. This bearing measurements are in turn provided to the EIF which estimate the states of the system $\bm{\hat x}$. This is provided as input to a proportional controller that computes the corresponding $\omega$, the control input to the vehicle. The effect of choosing different values for gain of the proportional controller is detailed in the forthcoming section on computational and simulation results. Computational and simulation results {#sec:results} ==================================== This section presents extensive computational and simulation results for all the algorithms developed thus far. All the computational experiments were performed on a MacBook Pro with a 2.9 GHz Intel Core i5 processor and 16 GB RAM using CPLEX 12.7 as a mixed-integer linear programming solver. Instance generation {#subsec:instance} ------------------- The number of targets, $|V|$, for all the test instances was chosen from the set $\{15, 20, 25, 30\}$. For each value of $|V|$, $20$ random instances were generated. The targets were randomly placed in a $100 \times 100$ grid. As for the potential LM locations, $5 \cdot |V|$ locations on the $100 \times 100$ grid was chosen at random. In total, we had $80$ instances on which all the experiments were performed. The sensing range for the bearing sensors, $\rho_s$, was fixed at $35$ units. Branch-and-cut algorithm performance {#subsec:bandc} ------------------------------------ The branch-and-cut algorithm with the dynamic cut-generation routine presented in Sec. \[sec:bandc\] was implemented in C++ using the callback functionality of CPLEX. The internal cut-generation routines of CPLEX were switched off and CPLEX was used only to manage the enumeration tree in the branch-and-cut algorithm. All computation times are reported in seconds. The performance of the algorithm was tested on randomly generated test instances. The branch-and-cut algorithm to compute optimal solution to the problem is very effective in computing optimal solutions for instances with less than $30$ targets. The computation time for all the instances was less than a seconds and hence, for all practical purposes it can be converted to an online computation if $|V| \leq 30$. coordinates [(15,10.27) (20,12.21) (25,14.90) (30,20.70) (35, 0)]{}; coordinates [(15,8.89) (20,9.05) (25,9.05) (30,8.50) (35, 0)]{}; The Fig. \[fig:avg\] shows the average number of sub-tour elimination constraints in added by the dynamic cut generation procedure detailed in Sec. \[sec:bandc\] and the average number of LMs required in the optimal solution; this indicates the effectiveness of the dynamic cut generation approach. The average number of LMs remains fairly steady as the number of targets is increased indicating that it is a function of the area of the grid. Simulation results {#subsec:simulation} ------------------ For the simulation *i.e.*, online estimation algorithm using the results of the branch-and-cut algorithm, we consider $3$ cases where the route for the UV and the positions of the landmarks are known a prioi (using the branch-and-cut algorithm). In all cases, the vehicle is required to travel in an optimal path such that it covers all way points and has connection to at least two LMs at all time; this constraint is enforced by the formulation presented in Sec. \[sec:formulation\]. For every run, the controller gain was chosen such that a minimum distance requirement condition to each way point was maintained. This in turn implies that for the vehicle to switch from the current way point to the next one, the vehicle needs to come in close proximity to the current way point meeting at least the minimum distance requirement for the switching to take place. For this simulation, it was considered that the vehicle can have a very high turn rate such that on reaching a way point, it can immediately point towards the next way point. The estimated states for the vehicle were used in the way point controller instead of the true states to show that the vehicle can indeed travel optimally in a GPS-restricted environment provided that the condition for path to at least two LMs is always maintained. The simulations were ran for $3000$ iterations. For the purpose of simulation, the unit for distance and time is chosen as meters (*m*) and seconds (*s*), respectively, without loss of generality. An instance of the simulation has been provided below in Fig. \[fig:C1\_R35L15\_OTG\] denoting the Landmarks, way points, true and estimated position of the vehicle, the associated uncertainty ($3\sigma$) bound ellipse and vehicle’s path to LM(s) within sensing range. ![Plot showing the vehicle motion at an arbitrary point in time during the simulation with sensing range as $35$ units for $|V| = 15$.[]{data-label="fig:C1_R35L15_OTG"}](Exp_pic/Case1/OTG_WP15.pdf) In all cases, there were $8$ LMs and the starting position for the vehicle was chosen to be $(0,35)$, without loss of generality. The plots for each of the $4$ cases and the different conditions in which these instances were ran are provided below. There are 2 distinct categories of plots that are provided for each simulation and they contain the following information: the first set of plots show the true and estimated trajectories for the vehicle and the second set of plots show the error plots along with the $3\sigma$ bounds for each case. ### Case 1 In this instance, $15$ way points (WPs) were provided through which the vehicle needed to route. The first scenario is such that the controller gain is set to $2.0$ and the minimum distance to WPs is $1.0$ unit. On reaching a distance of $1$ unit or closer to the WP, the vehicle can turn sharply and head towards the next WP due to high controller gain value. In the second scenario, the gain is reduced to $1.0$ and the minimum distance to WP is set at a value of $2.0$ units. A third and final scenario is provided for this particular case in which the sensing range was dropped to $20$ units from the required value of $35$ units such that the condition for path to at least 2 LMs are not always maintained. The controller gain was kept at $2.0$ and the minimum distance to WP at $1.0$ unit. The vehicle still routes through the WPs provided, but it does so with larger deviation and higher uncertainty. It can be seen from Fig. \[fig:C1\_R20L15\_M1T\] that there exist a large deviation from desired path during transit from WP-15 to WP-1 and from WP-12 to WP-13 as compared to Fig. \[fig:C1\_R35L15\_M1T\] and Fig. \[fig:C1\_R35L15\_M2T\]. This is because only one or no LM was visible at some points in this path for a reduced sensing range. The errors and $3\sigma$ bounds are also higher in this case (see Fig. \[fig:C1\_R35L15\_M1E\], \[fig:C1\_R20L15\_M1E\], and Fig. \[fig:C1\_R35L15\_M2E\]) This validates the necessity for meeting the path to at least 2 LMs condition. It is evident from the graphs that the error stays within $3\sigma$ bound at all time. The relative orientation of the vehicle with respect to LMs dictates the uncertainty ellipse at any given point. ### Case 2 In this instance, $20$ WPs were provided through which the vehicle needed to route. Here, the simulation was performed for 2 scenarios; one with controller gain as $2.0$ and minimum distance to WPs as $1.0$ unit and the other with controller gain as $0.4$ and minimum distance to WPs as $5.0$ units. While the first scenario is closer to ideal behavior, it requires tighter turn rate and higher vehicle agility. The vehicle gets closer to the given WPs in the first scenario than the next one in general. As a result, the second scenario produced smoother trajectory due to reduced gain and higher deviation from WPs in general, due to relaxed minimum distance requirement, than the first one. It was observed that in the second scenario, reducing controller gain resulted in requirement for a higher turning radius. Therefore, the vehicle overshot the space covered in $100 \times 100$ sq. units at times, especially when the WPs were placed very close to the edge of the square or rectangular area. Comparing Fig. \[fig:C2\_R35L20\_M1T\] and Fig. \[fig:C2\_R35L20\_M5T\], it can be seen that relaxing the minimum distance requirement from 1 unit to 5 units, switching from current to next WP occurs at a much earlier time. As a result, it becomes difficult to distinguish navigation conditions, especially for closely spaced WPs as observed for WP-7, WP-9 and WP-18 particularly in Fig. \[fig:C2\_R35L20\_M5T\] Real life scenarios would have predefined turn radius and turn rate constraints on including vehicle dynamics model. For such cases, way point navigation algorithms such as Dubins path is required to be implemented. Comparing the error plots in Fig. \[fig:C2\_R35L20\_M1E\] and Fig. \[fig:C2\_R35L20\_M5E\], it is observed that the $3\sigma$ bounds are small and comparable for both scenarios. This is so because error is not dependent on the minimum distance to WPs condition. Rather, it depends on path to LMs criteria. ### Case 3 In this instance, $25$ WPs were provided through which the vehicle needed to route. The simulation was performed for 2 scenarios. The first scenario had controller gain set to $2.0$ and the minimum distance to WPs as $1.0$ unit. The second scenario had controller gain tuned to $0.7$ and the minimum distance requirement to WPs as $3.0$ unit. Conclusion {#sec:conclusion} ========== In this paper, a systematic method to address the problem of joint routing and localization for UVs in a GPS-denied or GPS-restricted environment is presented. The optimization problem computes routes and identify the minimal set of locations where landmarks have to be placed to enable vehicle localization. This solution is combined with estimation algorithms to estimate the states of the vehicle. An efficient algorithm to compute an optimal solution to the optimization problem is presented. The proposed system architecture is tested extensively via simulation experiments. Future work can be focused on multiple vehicle versions of the problem and considering more realistic models for the sensors on the vehicles. [^1]: $^{\dagger}$ Center for Non-Linear Studies, Los Alamos National Laboratory, NM. `[email protected]` [^2]: $^{*}$ Dept. of Aerospace Engg., University of Cincinnati, OH. `` [^3]: $^{\ddagger}$ Dept. of Mechanical Engg., Texas A&M University, College Station, TX.

ArXiv

--- abstract: 'We study the geometry of the morphism that sends a smooth hypersurface of degree $d+1$ in $\PP^{n-1}$ to its associated hypersurface of degree $n(d-1)$ in the dual space $\bigl(\PP^{n-1}\bigr)^\vee$.' address: - | Department of Mathematics\ Boston College\ 140 Commonwealth Ave\ Chestnut Hill, MA 02467, USA - | Mathematical Sciences Institute\ Australian National University\ Acton, Canberra, ACT 2601, Australia author: - Maksym Fedorchuk - Alexander Isaev bibliography: - 'associated\_arxiv.bib' title: Associated form morphism --- [^1] Introduction ============ One of the first applications of Geometric Invariant Theory is a construction of the moduli space of smooth degree $m$ hypersurfaces in a fixed projective space $\PP^{n-1}$ [@GIT]. This moduli space is an affine GIT quotient $$U_{m,n}:=\left(\PP\HH^0\bigl(\PP^{n-1}, \cO(m)\bigr) \setminus {\Delta}\right) {/\hspace{-0.25pc}/}\operatorname{PGL}(n),\label{E:smooth-GIT}$$ where $\Delta$ is the discriminant divisor parameterizing singular hypersurfaces. The GIT construction produces a natural compactification $$U_{m,n} \subset V_{m,n}:=\left(\PP\HH^0\bigl(\PP^{n-1}, \cO(m)\bigr)\right)^{ss} {/\hspace{-0.25pc}/}\operatorname{PGL}(n),\label{E:compact-GIT}$$ given by a categorical quotient of the locus of GIT semistable hypersurfaces. We call $V_{m,n}$ the GIT compactification of $U_{m,n}$. The subject of this paper is a certain rational map $V_{m,n} \dashrightarrow V_{n(m-2),n}$, where $n\ge 2$, $m\ge 3$ and where we exclude the (trivial) case $(n,m)=(2,3)$. While this map has a purely algebraic construction, which we shall recall soon, it has several surprising geometric properties that we establish in this paper. In particular, this rational map restricts to a locally closed immersion $\bar A\colon U_{m,n}\to V_{n(m-2),n}$, and often contracts the discriminant divisor in $V_{m,n}$. Consequently, the closure of the image of $\bar A$ in $V_{n(m-2),n}$ is a compactification of the GIT moduli space $U_{m,n}$ that is different from the GIT compactification $V_{m,n}$. To define $\bar A$, we consider the *associated form morphism* defined on the space of smooth homogeneous forms $f\in \CC[x_1,\dots,x_n]$ of fixed degree $m\ge 3$. Given such an $f$, its associated form $A(f)$ is a degree $n(m-2)$ homogeneous form in the graded dual polynomial ring $\CC[z_1,\dots, z_n]$. In our recent paper [@fedorchuk-isaev], we proved that the associated form $A(f)$ is always polystable in the sense of GIT. Consequently, we obtain a morphism $\bar A$ from $U_{m,n}$ to $V_{n(m-2),n}$ sending the image of $f$ in $U_{m,n}$ to the image of $A(f)$ in $V_{n(m-2),n}$. Our first result is that the morphism $\bar A$ is an isomorphism onto its image, a locally closed subvariety in the target. \[T:barA\] The morphism $$\bar A \colon U_{m,n} \to V_{n(m-2),n}$$ is a locally closed immersion. In the process of establishing Theorem \[T:barA\], we generalize results of [@alper-isaev-assoc-binary] to the case of an arbitrary number of variables, and, in particular, prove that the auxiliary gradient morphism sending a semistable form to the span of its partial derivatives gives rise to a closed immersion on the level of quotients (see Theorem \[T:gradient\]). Our second main result is Theorem \[T:barA1\], which describes the rational map$\bar A \colon V_{m,n} \dashrightarrow V_{n(m-2),n}$ in codimension one. Namely, we study how $\bar A$ extends to the generic point of the discriminant divisor in the GIT compactification (see Corollary \[cor:n-1,n\]), and prove that for $n=2,3$ and $m\ge 4$, as well as for $n\geq 4$, $m\gg 0$, the morphism $\bar A$ contracts the discriminant divisor to a lower-dimensional subvariety in the target (see Corollary \[C:contraction\]). In the process, we prove that the image of $\bar A$ contains the orbit of the Fermat hypersurface in its closure and as a result obtain a new proof of the generic smoothness of associated forms (see Corollary \[C:generic-smoothness\]). Notation and conventions ------------------------ Let $S:=\operatorname{Sym}V\simeq \CC[x_1,\dots, x_n]$ be a symmetric algebra of an $n$-dimensional vector space $V$, with its standard grading. Let $\D:=\operatorname{Sym}V^{\vee} \simeq \CC[z_1,\dots, z_n]$ be the graded dual of $S$, with the structure of the $S$-module given by the *polar pairing* $S\times \D \to \D$, which is defined by $$\label{E:apolarity-action} g(x_1,\dots,x_n)\circ F(z_1,\dots,z_n):=g(\partial/\partial z_1, \dots, \partial/\partial z_1)F(z_1,\dots,z_n).$$ A homogeneous polynomial $f\in S_{m}$ is called a *direct sum* if, after a linear change of variables, it can be written as the sum of two non-zero polynomials in disjoint sets of variables: $$f=f_1(x_1,\dots, x_a)+f_2(x_{a+1},\dots, x_n).\label{E:direct-sum-def}$$ We will use the recognition criteria for direct sums established in [@fedorchuk-direct], and so we keep the pertinent terminology of that paper. We will say that $f\in S_{m}$ is a $k$-partial Fermat form for some $k\leq n$, if, after a linear change of variables, it can be written as follows: $$f=x_1^{m}+\cdots+x_k^{m}+g(x_{k+1},\dots, x_n).\label{E:partial-fermat}$$ Clearly, any $n$-partial Fermat form is linearly equivalent to the standard Fermat form. Furthermore, all $k$-partial Fermat forms are direct sums. We denote by $\operatorname{\mathfrak{DS}}_m$ the locus of direct sums in $S_{m}$. Associated form of a balanced complete intersection {#sectionmainres} =================================================== Fix $d\ge 2$. In what follows the trivial case $(n,d)=(2,2)$ will be excluded. A length $n$ regular sequence $g_1,\dots, g_n$ of elements of $S_d$ will be called a balanced complete intersection of type $(d)^n$. It defines a graded Gorenstein Artin $\CC$-algebra $$\cA(g_1,\dots,g_n):=S/(g_1,\dots, g_{n}),$$ whose socle lies in degree $n(d-1)$. In [@alper-isaev-assoc-binary] an element $\AA(g_1,\dots, g_n) \in \D_{n(d-1)}$, called *the associated form of $g_1,\dots,g_n$*, was introduced. The form $\AA(g_1,\dots, g_n)$ is a homogeneous Macaulay inverse system, or a dual socle generator, of the algebra $\cA(g_1,\dots,g_n)$. It follows that $[\AA(g_1,\dots, g_n)]\in \PP \D_{n(d-1)}$ depends only on the linear span $\langle g_1,\dots, g_n\rangle$, which we regard as a point in $\operatorname{Grass}(n, S_d)$. Recall that $g_1,\dots, g_n$ is a regular sequence in $S_d$ if and only if $\langle g_1,\dots, g_n\rangle$ does not in lie in the resultant divisor $\operatorname{\mathfrak{Res}}\subset \operatorname{Grass}(n, S_d)$. Setting $\operatorname{Grass}(n, S_d)_{\operatorname{Res}} := \operatorname{Grass}(n, S_d) \setminus \operatorname{\mathfrak{Res}}$, we obtain a morphism $$\AA \colon \operatorname{Grass}(n, S_d)_{\operatorname{Res}} \to \PP \D_{n(d-1)}.$$ Given $f\in S_{d+1}$, the partial derivatives $\partial f/\partial x_1,\dots, \partial f/\partial x_n$ form a regular sequence if and only if $f$ is non-degenerate. For a non-degenerate $f\in S_{d+1}$, in the *associated form of $f$* was defined to be $$A(f):=\AA(\partial f/\partial x_1,\dots, \partial f/\partial x_n) \in \D_{n(d-1)}.$$ Summarizing, we obtain a commutative diagram $$\begin{gathered} \xymatrix{ \PP(S_{d+1})_{\Delta} \ar[rd]_{\nabla} \ar[rr]^{A} & & \PP(\D_{n(d-1)}) \\ & \operatorname{Grass}(n, S_d)_{\operatorname{Res}} \ar[ur]_{\AA}, & } \end{gathered}\label{D:affine-triangle}$$ where $\PP(S_{d+1})_{\Delta}$ denotes the complement to the discriminant divisor in $\PP(S_{d+1})$ and $\nabla$ is the morphism sending a form into the linear span of its first partial derivatives. The above diagram is equivariant with respect to the standard $\operatorname{SL}(n)$-actions on $S$ and $\D$. By [@alper-isaev-assoc-binary], the morphism $\AA$ is a locally closed immersion, and it was proved in [@fedorchuk-isaev] that $\AA$ sends polystable orbits to polystable orbits. Passing to the GIT quotients, we thus obtain a commutative diagram $$\label{D:affine-triangle-GIT} \begin{gathered} \xymatrix{ \PP(S_{d+1})_{\Delta}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \ar[rd]_{\widetilde{\nabla}} \ar[rr]^{\bar{A}} & & \PP(\D_{n(d-1)})^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n) \\ & \operatorname{Grass}(n, S_d)_{\operatorname{Res}}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \ar[ur]_{\bar{\AA}}, & } \end{gathered}$$ where $\widetilde{\nabla}:=\nabla{/\hspace{-0.25pc}/}\operatorname{SL}(n)$ is a finite injective morphism (see [@fedorchuk-ss]) and $\bar{\AA}:=\AA{/\hspace{-0.25pc}/}\operatorname{SL}(n)$ is a locally closed immersion. The main focus of this paper is the geometry of diagram . Noting that by [@fedorchuk-ss] the map $\nabla$ extends to a morphism from $\PP(S_{d+1})^{ss}$ to$\operatorname{Grass}(n, S_d)^{ss}$ and thus induces a map $\overline{\nabla}$ of the corresponding GIT quotients, we will now state our two main results as follows: \[T:gradient\] The morphism $\overline{\nabla} \colon \PP(S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \to \operatorname{Grass}(n, S_d)^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n)$ is a closed immersion. \[T:barA1\] The rational map $$\bar{A}\colon \PP(S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \dashrightarrow \PP(\D_{n(d-1)})^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n)$$ extends to the generic point of the discriminant divisor $\Delta {/\hspace{-0.25pc}/}\operatorname{SL}(n)$ in the GIT compactification and contracts the discriminant divisor to a lower-dimensional variety for all sufficiently large $d$ as described in Corollaries [\[cor:n-1,n\]]{} and [\[C:contraction\]]{}. Preliminaries on dualities {#S:dualities} ========================== In this section we collect results on Macaulay inverse systems of graded Gorenstein Artin $\CC$-algebras. We also recall the duality between the Hilbert points of such algebras and the gradient points of their inverse systems. Recall that we regard $S=\CC[x_1,\dots,x_n]$ as a ring of polynomial differential operators on the graded dual ring $\D:=\CC[z_1,\dots,z_n]$ via polar pairing . For every positive $m$, the restricted pairing $$S_{m} \times \D_{m} \to \CC\label{E:pairing}$$ is perfect and so defines an isomorphism $$\label{E:duality} \D_{m} \simeq S_{m}^{\vee},$$ where, as usual, $V^{\vee}$ stands for the dual of a vector space $V$. Given $W\subset \D$, we define $$W^{\perp}:=\{f\in S \mid f\circ g=0, \text{ for all $g\in W$}\}\subset S.\label{E:perp-of-D}$$ Similarly given $U\subset S$, we define $$U^{\perp}:=\{g\in \D \mid f\circ g=0, \text{ for all $f\in U$}\}\subset \D.\label{E:perp-of-S}$$ \[C:explicit\] Isomorphism sends an element $\omega\in S_{m}^{\vee}$ to the element $$\mathfrak{D}_{\omega}:=\sum_{i_1+\cdots+i_n=m} \frac{\omega\bigl(x_1^{i_1}\cdots x_n^{i_n}\bigr)}{i_1!\cdots i_n!} z_1^{i_1} \cdots z_n^{i_n}\in \D_m.$$ Conversely, an element $g\in \D_{m}$ is mapped by isomorphism to the projection $$S_{m}\twoheadrightarrow S_m/(g^{\perp})_m \simeq \CC,$$ where the isomorphism with $\CC$ is chosen so that $1\in \CC$ pairs to $1$ with $g$. One observes that $f\circ \mathfrak{D}_{\omega}=\omega(f)$ for every $f\in S_{m}$, and the first part of the claim follows. The second part is immediate from definitions. \[C:evaluation\] Given $\omega\in S_{m}^{\vee}$, for every $(a_1,\dots,a_n)\in \CC^n$ we have $$\mathfrak{D}_{\omega}(a_1,\dots,a_n)=\omega\bigl((a_1x_1+\cdots+a_nx_n)^m/m!\bigr).\label{assocform}$$ $$\begin{aligned} \omega\bigl((a_1x_1+\cdots+a_nx_n)^m/m!\bigr) &=\frac{(a_1x_1+\cdots+a_nx_n)^m}{m!} \circ {\mathfrak D}_{\omega} \\ &=\frac{(a_1\partial /\partial z_1+\cdots+a_n\partial /\partial z_n)^m}{m!} {\mathfrak D}_{\omega} ={\mathfrak D}_{\omega}(a_1,\dots,a_n),\end{aligned}$$ where the last equality is easily checked, say on monomials. \[R:vanishing\] It follows from Corollary \[C:evaluation\] that all forms in a subset $W\subset \D_m$ vanish at a given point $(a_1,\dots,a_n)\in \CC^n$ if and only if $(a_1x_1+\cdots+a_nx_n)^m \in W^{\perp}$. Notice that the maps $$\bigl[\langle \mathfrak{D}_{\omega}\rangle \subset \D_{m}\bigr] \mapsto \\ \bigl[(\mathfrak{D}_{\omega}^{\perp})_m\subset S_m\bigr]= \bigl[\ker(\omega)\subset S_m\bigr]$$ define isomorphisms $$\operatorname{Grass}(1, \D_{m}) \simeq \operatorname{Grass}\left(\dim_{\CC} S_{m}-1, S_{m}\right).$$ More generally, for any $1\le m\le \binom{m+n-1}{n-1}-1$ the correspondence $$\bigl[W \subset \D_{m}\bigr] \mapsto \\ \bigl[ (W^{\perp})_m\subset S_m\bigr]$$ yields an isomorphism $$\label{E:grassmann} \operatorname{Grass}(k, \D_{m}) \simeq \operatorname{Grass}\left(\dim_{\CC} S_{m}-k, S_{m}\right).$$ Let $I\subset S$ be a Gorenstein ideal and $\nu$ the socle degree of the algebra $\cA=S/I$. Recall that a *(homogeneous) Macaulay inverse system* of $\cA$ is an element $f_{\A}\in \D_{\nu}$ such that $$f_\A^{\perp}=I$$ (see [@iarrobino-kanev Lemma 2.12] or [@eisenbud Exercise 21.7]). As $(f_\A^{\perp})_{\nu}=I_{\nu}$, we see that all Macaulay inverse systems are mutually proportional and $\langle f_{\A}\rangle=\bigl((I_{\nu})^\perp\bigr)_{\nu}$. Clearly, the line $\langle f_{\A}\rangle\in \operatorname{Grass}(1, \D_{\nu})$ maps to the $\nu^{th}$ Hilbert point $H_{\nu}\in \operatorname{Grass}(\dim_{\CC} S_{\nu}-1, S_{\nu})$ of $\cA$ under isomorphism with $k=1$. Papers [@eastwood-isaev1; @eastwood-isaev2], for any $\omega\in S_{\nu}^{\vee}$ with $\ker\omega=I_{\nu}$, introduced the [*associated form*]{} of $\cA$ as the element of $\D_{\nu}$ given by the right-hand side of formula (\[assocform\]) with $m=\nu$ (up to the factor $\nu!$). By Corollary \[C:evaluation\], under isomorphism with $k=1$ the span of every associated form in $\D_{\nu}$ also maps to the $\nu^{th}$ Hilbert point $H_{\nu}\in \operatorname{Grass}(\dim_{\CC} S_{\nu}-1, S_{\nu})$ of $\cA$. In particular, for the algebra $\cA$ any associated form is simply one of its Macaulay inverse systems, and equation with $m=\nu$ and $\ker\omega=I_{\nu}$ is an explicit formula for a Macaulay inverse system of $\cA$ (see [@isaev-criterion] for more details). Gradient points {#S:gradient} --------------- Given a polynomial $F\in \D_m$, we define the [*$p^{th}$ gradient point of $F$*]{} to be the linear span of all $p^{th}$ partial derivatives of $F$ in $\D_{m-p}$. We denote the $p^{th}$ gradient point by $\nabla^{p}(F)$. Note that $$\nabla^{p}(F)=\{g \circ F \mid g\in S_{p}\}$$ is simply the $(m-p)^{th}$ graded piece of the principal $S$-module $SF$. The $1^{st}$ gradient point $\nabla F:=\langle \partial F/\partial z_1, \dots, \partial F /\partial z_n\rangle$ will be called simply *the gradient point of $F$*. \[P:hilb-dual\] The $p^{th}$ gradient point of a Macaulay inverse system $f_{\A}\in \D_{\nu}$ maps to the $(\nu-p)^{th}$ Hilbert point $H_{\nu-p}$ of $\cA$ under isomorphism . Let $G$ be the $p^{th}$ gradient point of $f_{\A}$, that is $$G:=\left\langle \frac{\partial^p}{\partial z_1^{i_1} \cdots \partial z_n^{i_n}} f_{\A} \mid i_1+\cdots+i_n=p\right\rangle.$$ We need to verify that $I_{\nu-p}=(G^{\perp})_{\nu-p}$. We have $$\begin{aligned} (G^{\perp})_{\nu-p}& =\left\{f\in S_{\nu-p} \mid f\circ \frac{\partial^p}{\partial z_1^{i_1} \cdots \partial z_n^{i_n}} f_{\A} =0 \text{ for all $i_1+\cdots+i_n=p$}\right\}\\ &=\left\{f\in S_{\nu-p} \mid f x_1^{i_1}\cdots x_n^{i_n} \circ f_{\cA} =0 \text{ for all degree $p$ monomials}\right\} \\ &=\left\{f\in S_{\nu-p} \mid x_1^{i_1}\cdots x_n^{i_n} f \in f_\cA^{\perp} \text{ for all degree $p$ monomials}\right\} \\ &=\left\{f\in S_{\nu-p} \mid x_1^{i_1}\cdots x_n^{i_n}f \in I_{\nu} \text{ for all degree $p$ monomials}\right\} \\ &=I_{\nu-p},\end{aligned}$$ where the last equality comes from the fact that $I$ is Gorenstein. As a corollary of the above duality result, we recall in Proposition \[P:singular\] below a generalization of [@alper-isaev-assoc Lemma 4.4]. Although this statement is well-known (it appears, for example, in [@rota Proposition 4.1, p. 174]), we provide a short proof for completeness. We first recall that a non-zero homogeneous form $f$ in $n$ variables has multiplicity $\ell+1$ at a point $p\in \PP^{n-1}$ if and only if all partial derivatives of $f$ of order $\ell$ (hence of all orders $\leq \ell$) vanish at $p$, and some partial derivative of $f$ of order $\ell+1$ does not vanish at $p$. We define the *Veronese cone* $\mathcal{C}_m$ to be the variety of all degree $m$ powers of linear forms in $S_m$: $$\mathcal{C}_{m}:=\bigl\{L^{m} \mid L \in S_1\bigr\}\subset S_{m}.$$ \[P:singular\] Let $I\subset S$ be a Gorenstein ideal and $\nu$ the socle degree of the algebra $\cA=S/I$. Then a Macaulay inverse system $f_{\cA}$ of $\cA$ has a point of multiplicity $\ell+1$ if and only if there exists a non-zero $L\in S_1$ such that $L^{\nu-\ell}\in I_{\nu-\ell}$, and $L^{\nu-\ell-1} \not\in I_{\nu-\ell-1}$. In particular, $f_{\cA}$ has no points of multiplicity $\ell+1$ or higher if and only if $$I_{\nu-\ell} \cap \mathcal{C}_{\nu-\ell} = (0).$$ By Proposition \[P:hilb-dual\], the $\ell^{th}$ gradient point of $f_{\cA}$ is dual to the $(\nu-\ell)^{th}$ Hilbert point of $\cA$ $$H_{\nu-\ell}\colon S_{\nu-\ell} \twohead \cA_{\nu-\ell}.$$ We conclude by Remark \[R:vanishing\] that all partial derivatives of $f_{\cA}$ of order $\ell$ vanish at $(a_1,\dots,a_n)$ if and only if $$(a_1x_1+\cdots+a_nx_n)^{\nu-\ell}\in \ker H_{\nu-\ell}=I_{\nu-\ell}.$$ It follows that $L=a_1x_1+\cdots+a_nx_n$ satisfies $L^{\nu-\ell}\in I_{\nu-\ell}$ and $L^{\nu-\ell-1} \not\in I_{\nu-\ell-1}$ if and only if $f_{\cA}$ has multiplicity exactly $\ell+1$ at the point $(a_1,\dots, a_n)$. The gradient morphism $\nabla$ ============================== In this section, we prove Theorem \[T:gradient\]. Recall that we have the commutative diagram $$\begin{aligned} \xymatrix{ \PP\bigl(S_{d+1})^{ss} \ar[d]^{\pi_0} \ar[r]^{\nabla} & \operatorname{Grass}\bigl(n, S_d\bigr)^{ss}\ar[d]^{\pi_1} \\ \PP\bigl(S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \ar[r]^{\overline{\nabla}\qquad} &\operatorname{Grass}\bigl(n, S_d\bigr)^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n). } \end{aligned}\label{diagram}$$ Let $\operatorname{\mathfrak{DS}}^{ss}_{d+1}:=\PP(\operatorname{\mathfrak{DS}}_{d+1})^{ss}$ be the locus of semistable direct sums in $\PP\bigl(S_{d+1})^{ss}$. By [@fedorchuk-direct Section 3], the set $\operatorname{\mathfrak{DS}}^{ss}_{d+1}$ is precisely the closed locus in $\PP\bigl(S_{d+1})^{ss}$ where $\nabla$ has positive fiber dimension. Suppose $f\in S_{d+1}$ is a semistable form. Then, after a linear change of variables, we have a maximally fine direct sum decomposition $$\label{E:direct-sum} f=\sum_{i=1}^k f_i (\mathbf{x}^i),$$ where $V_i=\langle \mathbf{x}^i\rangle$ are such that $V=\oplus_{i=1}^k V_i$, and where each $f_i$ is not a direct sum in $\operatorname{Sym}V_i$. Set $n_i:=\dim_{\CC} V_i$. We define the canonical torus $\Theta(f)\subset \operatorname{SL}(n)$ associated to $f$ as the connected component of the identity of the subgroup $$\{g\in \operatorname{SL}(n) \mid \text{$V_i$ is an eigenspace of $g$, for every $i=1,\dots, k$}\}\subset\operatorname{SL}(n).\label{E:torus} $$ Clearly, $\Theta(f) \simeq (\CC^{*})^{k-1}$, and since $$\nabla([f])=\nabla([f_1])\oplus \cdots \oplus \nabla([f_k]), \ \text{where $\nabla([f_i]) \in \operatorname{Grass}(n_i, \operatorname{Sym}^{d} V_i)$},$$ we also have $\Theta(f)\subset \operatorname{Stab}(\nabla([f]))$, where $\operatorname{Stab}$ denotes the stabilizer under the $\operatorname{SL}(n)$-action. From the definition of $\Theta(f)$, it is clear that $\Theta(f)\cdot [f] \subset \nabla^{-1}(\nabla([f]))$, and in fact [@fedorchuk-direct Corollary 3.12] gives a set-theoretic equality $\nabla^{-1}(\nabla([f]))=\Theta(f)\cdot [f]$. We will now obtain a stronger result: \[L:gradient-fiber\] One has $\nabla^{-1}(\nabla([f]))=\Theta(f)\cdot [f]$ scheme-theoretically, or, equivalently, $$\ker(d \nabla_{[f]})={\mathbf T}_{[f]}(\Theta(f)\cdot [f]),$$ where ${\mathbf T}_{[f]}$ denotes the tangent space at $[f]$. Under the standard identification of ${\mathbf T}_{[f]} \PP(S_{d+1})$ with $S_{d+1}/\langle f\rangle$, the subspace ${\mathbf T}_{[f]}(\Theta(f)\cdot [f])$ is identified with $\langle f_1, \dots, f_{k}\rangle/\langle f\rangle$. It now suffices to show that every $g\in S_{d+1}$ that satisfies $\nabla[g] \subset \nabla[f]$ must lie in $\langle f_1, \dots, f_{k}\rangle$, where$\nabla[g]:=\langle \partial g/\partial x_1, \dots, \partial g/\partial x_n\rangle\subset S_d$. This is precisely the statement of [@fedorchuk-direct Corollary 3.12]. We note an immediate consequence: If $f\in S_{d+1}^{ss}$ is not a direct sum, then $\nabla$ is unramified at $[f]$. Further, since $\nabla$ is equivariant with respect to the $\operatorname{SL}(n)$-action, we have the inclusion $\operatorname{Stab}([f])\subset \operatorname{Stab}(\nabla([f]))$. As the following result shows, the difference between $\operatorname{Stab}([f])$ and $\operatorname{Stab}(\nabla([f]))$ is controlled by the torus $\Theta(f)$. \[C:gradient-stabilizer\] The subgroup $\operatorname{Stab}(\nabla([f]))$ is generated by $\Theta(f)$ and $\operatorname{Stab}([f])$. Suppose $\sigma\in \operatorname{Stab}(\nabla([f]))$. Then $\nabla(\sigma\cdot[f])=\nabla([f])$ implies by Lemma \[L:gradient-fiber\] that $\sigma\cdot[f]=\tau \cdot[f]$ for some $\tau \in \Theta(f)$. Consequently, $\tau^{-1}\circ \sigma \in \operatorname{Stab}([f])$ as desired. Next, we obtain the following generalization of [@alper-isaev-assoc-binary Proposition 6.3], whose proof we follow almost verbatim. The morphism $\nabla$ is a closed immersion along the open locus $\cU:=\PP(S_{d+1})^{ss} \setminus \operatorname{\mathfrak{DS}}^{ss}_{d+1}$ of all elements that are not direct sums. Since for every $[f]\in \cU$ we have that $\nabla$ is unramified at $[f]$ and $\nabla^{-1}(\nabla([f]))=[f]$, it suffices to show that $\nabla$ is a finite morphism when restricted to $\cU$. Since, by [@fedorchuk-ss], the induced morphism on the GIT quotients is finite, by [@luna-slices-etales p. 89, Lemme] it suffices to verify that $\nabla$ is quasi-finite and that $\nabla$ sends closed orbits to closed orbits. The former has already been established, and the latter is proved below in Proposition \[P:polystability\]. \[P:polystability\] Suppose $f\in S_{d+1}^{ss}$ is polystable and not a direct sum. Then the image $\nabla([f])\in \operatorname{Grass}(n,S_d)^{ss}$ is polystable. The above result is a generalization of [@fedorchuk-ss Theorem 1.1], whose method of proof we follow; we also keep the notation of *loc.cit.*, especially as it relates to monomial orderings. We begin with a preliminary observation. \[1-PS-DS\] Suppose $f\in S_{d+1}$ is such that there exists a non-trivial one-parameter subgroup $\lambda$ of $\operatorname{SL}(n)$ acting diagonally on $x_1,\dots, x_n$ with weights $\lambda_1,\dots, \lambda_n$ and satisfying $$w_{\lambda}(\operatorname{in}_{\lambda}(\partial f/\partial x_i))=d\lambda_i.$$ Then $f$ is a direct sum. We can assume that $$\lambda_1\leq \cdots \leq \lambda_{a}<\lambda_{a+1}=\cdots=\lambda_n$$ for some $1\le a<n$. Then the fact that $$w_{\lambda}(\operatorname{in}_{\lambda}(\partial f/\partial x_i))=d\lambda_i=d\lambda_n,$$ for all $i=a+1,\dots, n$, implies $$\partial f/\partial x_{a+1}, \dots, \partial f/\partial x_{n}\in \CC[x_{a+1},\dots, x_n].$$ Consequently, $f=g_1(x_1,\dots,x_a)+g_2(x_{a+1},\dots, x_n)$ is a direct sum. Since $f$ is polystable, by [@fedorchuk-ss Theorem 1.1] it follows that $\nabla([f])$ is semistable. Suppose $\nabla([f])$ is not polystable. Then there exists a one-parameter subgroup $\lambda$ acting on the coordinates $x_1,\dots, x_n$ with the weights $\lambda_1, \dots, \lambda_n$ such that the limit of $\nabla([f])$ under $\lambda$ exists and does not lie in the orbit of $\nabla([f])$. In particular, the limit of $[f]$ under $\lambda$ does not exist. Then by [@fedorchuk-ss Lemma 3.5], there is an upper triangular unipotent coordinate change $$\label{E:sub} \begin{aligned} x_1& \mapsto x_1+c_{12}x_2+\cdots+ c_{1n}x_n, \\ x_2&\mapsto\phantom{{}=1111} x_2+\cdots+ c_{2n}x_n, \\ \vdots \\ x_n&\mapsto \phantom{{}=1111111111111111}x_n \end{aligned}$$ such that for the transformed form $$h(x_1,\dots,x_n):=f(x_1+c_{12}x_2+\cdots+ c_{1n}x_n, x_2+\cdots+ c_{2n}x_n, \dots, x_n)$$ the initial monomials $$\operatorname{in}_{\lambda}(\partial h/\partial x_1), \dots, \operatorname{in}_{\lambda}(\partial h/\partial x_n)$$ are distinct. Now, setting $$\mu_i:=w_{\lambda}(\operatorname{in}_{\lambda}(\partial h/\partial x_i)),$$ by [@fedorchuk-ss Lemma 3.2] we have $$\mu_1+\cdots+\mu_n= 0.$$ It follows that with the respect to the one-parameter subgroup $\lambda'$ acting on $x_i$ with the weight $d\lambda_i-\mu_i$, all monomials of $h$ have non-negative weights (cf. [@fedorchuk-ss the proof of Lemma 3.6]). Write $h=h_0+h_1$, where all monomials of $h_0$ have zero $\lambda'$-weights and all monomials of $h_1$ have positive $\lambda'$-weights. Then $h_0\in \overline{\operatorname{SL}(n)\cdot h}=\operatorname{SL}(n)\cdot h$, by the polystability assumption on $f$. Furthermore, $h_0$ is stabilized by $\lambda'$. If $\lambda'$ is a trivial one-parameter subgroup, then $\mu_i=d\lambda_i$ for all $i=1,\dots, n$, and by Lemma \[1-PS-DS\] the form $h$ is a direct sum, which is a contradiction. Suppose now that $\lambda'$ is a non-trivial one-parameter subgroup. Clearly, we have $$w_{\lambda}(\operatorname{in}_{\lambda}(\partial h_0/\partial x_i)\geq w_{\lambda}(\operatorname{in}_{\lambda}(\partial h/\partial x_i),$$ since the state of $h_0$ is a subset of the state of $h$. If one of the inequalities above is strict, then $\nabla([h_0])$ is destabilized by $\lambda$, contradicting the semistability of $\nabla([h_0])$ established in [@fedorchuk-ss Theorem 1.1]. Thus $$w_{\lambda}(\operatorname{in}_{\lambda}(\partial h_0/\partial x_i))=w_{\lambda}(\operatorname{in}_{\lambda}(\partial h/\partial x_i))=\mu_i.$$ Moreover, since $h_0$ is $\lambda'$-invariant, we have that $\partial h_0/\partial x_i$ is homogeneous of degree $-w_{\lambda'}(x_i)=\mu_i-d\lambda_i$ with respect to $\lambda'$. Let $\mu$ be the one-parameter subgroup acting on $x_1,\dots, x_n$ with the weights $\mu_1,\dots, \mu_n$. It follows that $$w_{\mu}(\operatorname{in}_{\mu}(\partial h_0/\partial x_i))=dw_{\lambda}(\operatorname{in}_{\lambda}(\partial h_0/\partial x_i)+w_{\lambda'}(\operatorname{in}_{\lambda'}(\partial h_0/\partial x_i)=d\mu_i-\mu_i+d\lambda_i.$$ Then the one-parameter subgroup $\lambda+\mu$ acting on $x_1,\dots, x_n$ with the weights $\lambda_1+\mu_1,\dots, \lambda_n+\mu_n$ satisfies $$\begin{array}{l} w_{\lambda+\mu}(\operatorname{in}_{\lambda+\mu}(\partial h_0/\partial x_i))= w_{\lambda}(\operatorname{in}_{\lambda}(\partial h_0/\partial x_i))+w_{\mu}(\operatorname{in}_{\mu}(\partial h_0/\partial x_i)) =\\ \hspace{8cm}d\mu_i-\mu_i+d\lambda_i+\mu_i=d(\mu_i+\lambda_i). \end{array}$$ Applying Lemma \[1-PS-DS\], we conclude that either $h_0$ is a direct sum, or $$\lambda_i+\mu_i=0\quad \text{for all $i=1,\dots, n$}.$$ In the latter case, it follows that $\lambda$ is proportional to $\lambda'=d\lambda-\mu$. Since the limit of $h$ under $\lambda'$ exists and is equal to $h_0$, the limit under $\lambda$ of $h$ must exist and be equal to $h_0$ as well. Observing that the inverse of an upper-triangular matrix with 1’s on the diagonal has the same form, we see that the limit of $$f(x_1,\dots,x_n)=h(x_1+c_{12}'x_2+\cdots+ c_{1n}'x_n, x_2+\cdots+ c_{2n}'x_n, \dots, x_n)$$ under $\lambda$ also exists. This contradiction concludes the proof. \[C:nablaprespolystability\] The morphism $\nabla\colon \PP(S_{d+1})^{ss} \to \operatorname{Grass}(n,S_d)^{ss}$ preserves polystability. Suppose $f=f_1+\cdots+f_k$ is the maximally fine direct sum decomposition of a polystable form $f$, where $f_i\in \operatorname{Sym}^{d+1} V_i$, and where $V=\oplus_{i=1}^k V_i$. Then each $f_i$ is polystable and not a direct sum in $\operatorname{Sym}^{d+1} V_i$. Hence $\nabla([f_i])$ is polystable with respect to the $\operatorname{SL}(V_i)$-action. Since $\Theta(f)\subset \operatorname{Stab}(\nabla([f]))$ is a reductive subgroup, to prove that $\nabla([f])$ is polystable, it suffices to verify that $\nabla([f])$ is polystable with respect to the centralizer $C_{\operatorname{SL}(n)}(\Theta(f))$ of $\operatorname{Stab}(\Theta(f))$ in $\operatorname{SL}(n)$, see [@luna-adherences Corollaire 1 and Remarque 1]. We have $$C_{\operatorname{SL}(n)}(\Theta(f))=\left(\operatorname{GL}(V_1)\times \cdots \times \operatorname{GL}(V_k)\right)\cap \operatorname{SL}(n).$$ Arguing as on [@fedorchuk-ss p. 456], we see that every one-parameter subgroup $\lambda$ of $C_{\operatorname{SL}(n)}(\Theta(f))$ can be renormalized to a one-parameter subgroup of $\operatorname{SL}(V_1)\times \cdots \times \operatorname{SL}(V_k)$ without changing its action on $\nabla([f])$. Since $\nabla([f_i])$ is polystable with respect to $\operatorname{SL}(V_i)$, it follows that $$\nabla([f])=\nabla([f_1])\oplus \cdots \oplus \nabla([f_k])$$ is polystable with respect to the action of $\lambda$ thus proving the claim. Suppose that $f$ is polystable, consider its maximally fine direct sum decomposition and the canonical torus $\Theta(f)$ in $\operatorname{Stab}(\nabla([f]))$ as constructed above. In what follows, we will write $X$ to denote $\PP(S_{d+1})^{ss}$ and $Y$ to denote $\operatorname{Grass}(n, S_d)^{ss}$. Set $p:=\pi_0([f])\in X{/\hspace{-0.25pc}/}\operatorname{SL}(n)$. We will prove that $\overline{\nabla}$ is unramified at $p$. Let $N_{[f]}$ be the normal space to the $\operatorname{SL}(n)$-orbit of $[f]$ in $X$ at the point $[f]$, and $N_{\nabla([f])}$ the normal space to the $\operatorname{SL}(n)$-orbit of $\nabla([f])$ in $Y$ at the point $\nabla([f])$. We have a natural map $$\iota\colon N_{[f]} \to N_{\nabla([f])}$$ induced by the differential of $\nabla$. The map $\iota$ is injective by Lemma \[L:gradient-fiber\]. Since both $[f]$ and $\nabla([f])$ have closed orbits in $X$ and $Y$, respectively (see Corollary \[C:nablaprespolystability\]), to verify that $\overline{\nabla}$ is unramified at $p$, it suffices, by Luna’s étale slice theorem, to prove that the morphism $$\label{E:slice} s(f)\colon N_{[f]} {/\hspace{-0.25pc}/}\operatorname{Stab}([f]) \to N_{\nabla([f])}{/\hspace{-0.25pc}/}\operatorname{Stab}(\nabla([f]) $$ is unramified. As $\nabla$ is not necessarily stabilizer-preserving at $[f]$ (i.e., $\operatorname{Stab}([f])$ may not be equal to $\operatorname{Stab}(\nabla([f]))$), we cannot directly appeal to the injectivity of $\iota$. Instead, consider the $\Theta(f)$-orbit, say $F$, of $[f]$ in $X$. Let $\cN_{F/X}$ be the $\Theta(f)$-invariant normal bundle of $F$ in $X$. Since by Lemma \[L:gradient-fiber\] we have $\nabla^{-1}(\nabla([f]))=F$, there is a natural $\Theta(f)$-equivariant map $J:\cN_{F/X} \to N_{\nabla([f])}$. We now make a key observation that for the induced map $\tilde J:\cN_{F/X} {/\hspace{-0.25pc}/}\Theta(f)\to N_{\nabla([f])}$ one has $$\tilde J(\cN_{F/X} {/\hspace{-0.25pc}/}\Theta(f))=\iota\left(N_{[f]} \right).$$ Since $\overline{\nabla}$ is finite by [@fedorchuk-ss Proposition 2.1], the morphism $s(f)$ from Equation is quasi-finite. Applying Lemma \[L:GIT-lemma\] (proved below), with $\operatorname{Spec}A=N_{[f]}$, $\operatorname{Spec}B=N_{\nabla([f])}$, $T=\Theta(f)$, $H=\operatorname{Stab}([f])$, $G=\operatorname{Stab}(\nabla([f]))$, as well as Corollary \[C:gradient-stabilizer\], we obtain that $s(f)$ is in fact a closed immersion, and so is unramified. Note that here the group $G$ is reductive by Matsushima’s criterion. This proves that $\overline{\nabla}$ is unramified at $p$. We now note that $\overline{\nabla}$ is injective. Indeed, this follows as in the proof of [@fedorchuk-ss Part (2) of Proposition 2.1] from Corollary \[C:nablaprespolystability\] and the finiteness of $\overline{\nabla}$. We then conclude that $\overline{\nabla}$ is a closed immersion. \[L:GIT-lemma\] Suppose $G$ is a reductive group. Suppose $T\subset G$ is a connected reductive subgroup, and $H\subset G$ is a reductive subgroup such that $G$ is generated by $T$ and $H$. Suppose we have a $G$-equivariant closed immersion of normal affine schemes admitting an action of $G$ $$\operatorname{Spec}A \hookrightarrow \operatorname{Spec}B.$$ such that $\operatorname{Spec}A^H \to \operatorname{Spec}B^G$ is quasi-finite. Then $\operatorname{Spec}A^G\simeq \operatorname{Spec}A^H$ and, consequently, $\operatorname{Spec}A^H \to \operatorname{Spec}B^G$ is a closed immersion. We have the following commutative diagram $$\xymatrix{ \operatorname{Spec}A^H \ar[d] \ar[rd] \ar@{^(->}[r] & \operatorname{Spec}B^H \ar[d] \\ (\operatorname{Spec}A^H){/\hspace{-0.25pc}/}T\simeq \operatorname{Spec}A^G \ar@{^(->}[r] & (\operatorname{Spec}B^H){/\hspace{-0.25pc}/}T \simeq \operatorname{Spec}B^G. }$$ Since the diagonal arrow is quasi-finite by assumption, and the bottom arrow is a closed immersion, we conclude that the GIT quotient $\operatorname{Spec}A^H \to (\operatorname{Spec}A^H){/\hspace{-0.25pc}/}T$ is quasi-finite as well. Since this is a good quotient by a connected group, the morphism $\operatorname{Spec}A^H \to (\operatorname{Spec}A^H){/\hspace{-0.25pc}/}T\simeq \operatorname{Spec}A^G$ must be an isomorphism. \[corT:A\] The morphism $$\bar{A}\colon \PP(S_{d+1})_{\Delta}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \rightarrow \PP(\D_{n(d-1)})^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n)$$ is a locally closed immersion. The morphism $\AA_{\operatorname{Gr}}$ ====================================== In this section, we prove Theorem \[T:barA1\]. In fact, we study in detail the rational map $\bar A \colon (\PP S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \dashrightarrow \PP(\D_{n(d-1)})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n)$ in codimension one. As in Section \[sectionmainres\], fix $d\ge 2$. As always, we assume that $n\ge 2$ and disregard the trivial case $(n,d)=(2,2)$. Given $U\in \operatorname{Grass}(n, S_d)$, we take $I_U$ to be the ideal in $S$ generated by the elements in $U$. Consider the following locus in $\operatorname{Grass}(n, S_d)$: $$W_{n,d} =\{U\in \operatorname{Grass}(n, S_d) \mid \dim_{\CC} (S/I_U)_{n(d-1)-1} = n\}.\label{E:Wnd}$$ Since $\dim_{\CC} (S/I_U)_{n(d-1)-1}$ is an upper semi-continuous function on $\operatorname{Grass}(n, S_d)$ and for every $U\in \operatorname{Grass}(n, S_d)$ one has $\dim_{\CC} (S/I_U)_{n(d-1)-1} \geq n$, we conclude that $W_{n,d}$ is an open subset of $\operatorname{Grass}(n, S_d)$. Moreover, since for $U\in\operatorname{Grass}(n, S_d)_{\operatorname{Res}}$ the ideal $I_U$ is Gorenstein of socle degree $n(d-1)$, we have $\operatorname{Grass}(n, S_d)_{\operatorname{Res}} \subset W_{n,d}$. Applying polar pairing, we obtain a morphism $$\begin{aligned} \AA_{\operatorname{Gr}} \colon W_{n,d} \to \operatorname{Grass}(n, \D_{n(d-1)-1}), \\ \quad \AA_{\operatorname{Gr}}(U)=\left[(I_U)_{n(d-1)-1}^\perp \subset \D_{n(d-1)-1}\right]. \end{aligned}\label{E:A-Gr}$$ From the duality between Hilbert and gradient points it follows that $$\nabla(\AA(U))=\AA_{\operatorname{Gr}}(U) \text{ for every $U\in \operatorname{Grass}(n, S_d)_{\operatorname{Res}}$.}$$ We conclude that we have the commutative diagram: $$\makebox[250pt]{$\begin{gathered} \xymatrix{ \PP(S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n)&&\PP(\D_{n(d-1)})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n)\\ \PP(S_{d+1})^{ss} \ \ar[u]^{\pi_0} \ar[d]_{\nabla} & \PP(S_{d+1})_{\Delta} \ar@{_{(}->}[l] \ar[r]^{A} \ar[d]^{\nabla} & \PP(\D_{n(d-1)})^{ss} \ar[d]^{\nabla} \ar[u]_{\pi_2} \\ \operatorname{Grass}(n, S_d)^{ss} \ \ar[d]_{\pi_1} & \operatorname{Grass}(n, S_d)_{\operatorname{Res}} \ar@{_{(}->}[l] \ar@{_{(}->}[d] \ar[ur]^{\AA} \ar[r] & \operatorname{Grass}(n, \D_{n(d-1)-1})^{ss} \ \ar[d]^{\pi_3} \\ \operatorname{Grass}(n, S_d)^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) & W_{n,d} \ar[ru]^{\AA_{\operatorname{Gr}}} & \operatorname{Grass}(n, \D_{n(d-1)-1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n). } \end{gathered}$}$$ \[P:A-Gr-defined\] Suppose $U\in \operatorname{Grass}(n, S_{d})$ is such that $$\VV(I_U)=\{p_1,\dots,p_k\}$$ is scheme-theoretically a set of $k$ distinct points in general linear position in $\PP^{n-1}$. Then $U\in W_{n,d}$. A set $\{p_1,\dots,p_k\}$ points in $\PP^{n-1}$ is in general linear position if and only if $k\leq n$, and, up to the $\operatorname{PGL}(n)$-action, $$\begin{aligned} p_i=\{x_1=\cdots=\widehat{x_i}=\cdots=x_n=0\},\quad i=1,\dots, k,\end{aligned}$$ in the homogeneous coordinates $[x_1:\dots:x_n]$ on $\PP^{n-1}$. Since $\operatorname{depth}(I_U)=n-1$, we can choose degree $d$ generators $g_1,\dots, g_n$ of $I_U$ such that $g_1,\dots, g_{n-1}$ form a regular sequence. Then $\Gamma:=\VV(g_1,\dots,g_{n-1})$ is a finite-dimensional subscheme of $\PP^{n-1}$. By Bézout’s theorem, $\Gamma$ is a set of $d^{n-1}$ points, counted with multiplicities. Set $R:=S/(g_1,\dots,g_{n-1})$. Consider the Koszul complex $K_{\bullet}:=K_{\bullet}(g_1,\dots, g_n)$. We have $$\HH_0(K_{\bullet})=S/(g_1,\dots,g_n)=S/I_U. $$ Since $g_1,\dots,g_{n-1}$ is a regular sequence, we also have $$\begin{aligned} \HH_i(K_{\bullet})= 0 \quad \text{for all $i>0$}\end{aligned}$$ and $$\begin{aligned} \HH_{1}(K_{\bullet})=\bigl(((g_1,\dots,g_{n-1}):_{\, S}\hspace{-0.1cm}(g_1,\dots, g_n))/(g_1,\dots,g_{n-1})\bigr)(-d)\simeq \operatorname{Ann}_{R}(g_n)(-d).\end{aligned}$$ To establish the identity $$\operatorname{codim}\bigl((I_U)_{n(d-1)-1}, S_{n(d-1)-1}\bigr)=n$$ it suffices to prove $$\HH_{1}(K_{\bullet})_{n(d-1)-1}=0.$$ Indeed, in this case the graded degree $n(d-1)-1$ part of the Koszul complex will be an exact complex of vector spaces and so the dimension of $\left(S/I_U\right)_{n(d-1)-1}$ will coincide with that in the situation when $g_1,\dots,g_n$ is a regular sequence, that is, with $n$. As we have already observed, we have $$\HH_{1}(K_{\bullet})_{n(d-1)-1}=\operatorname{Ann}_R(g_n)_{n(d-1)-1}(-d)=\operatorname{Ann}_R(g_n)_{n(d-1)-1-d}.$$ Hence it suffices to prove that $\operatorname{Ann}_R(g_n)_{n(d-1)-1-d}=0$. Write $\Gamma=\Gamma'\cup \Gamma''$, where $\Gamma'\neq\varnothing$ and $\Gamma'':=\{p_1,\dots,p_k\}$. Since $g_n$ vanishes on all of $\Gamma''$ but does not vanish at any point of $\Gamma'$, every element of $\operatorname{Ann}_{R}(g_n)_{n(d-1)-1-d}$ comes from a degree $n(d-1)-1-d$ form that vanishes on all of $\Gamma'$. We apply the Cayley-Bacharach Theorem [@cayley-bacharach Theorem CB6], which implies the following statement: \[C:CB\] Set $s:=d(n-1)-(n-1)-1=n(d-1)-d$. If $r \leq s$ is a non-negative integer, then the dimension of the family of projective hypersurfaces of degree $r$ containing $\Gamma'$ modulo those containing all of $\Gamma$ is equal to the failure of $\Gamma''$ to impose independent conditions on projective hypersurfaces of complementary degree $s-r$. In our situation $r=s-1$, and $\Gamma''$ imposes independent conditions on hyperplanes by the general linear position assumption. Hence we conclude by Claim \[C:CB\] that every form of degree $n(d-1)-1-d$ that vanishes on all of $\Gamma'$ also vanishes on all of $\Gamma''$ and therefore, as the ideal $(g_1,\dots,g_{n-1})$ is saturated, maps to $0$ in $R$. We thus see that $\operatorname{Ann}_{R}(g_n)_{n(d-1)-1-d}=0$. This finishes the proof. Motivated by the result above, we consider the following partial stratification of the resultant divisor $\operatorname{\mathfrak{Res}}\subset \operatorname{Grass}(n, S_d)$. For $1\leq k \leq n$, define $Z_{k}$ to be the locally closed subset of $\operatorname{Grass}(n, S_d)$ consisting of all subspaces $U$ such that $\VV(I_U)$ is scheme-theoretically a set of $k$ distinct points in general linear position in $\PP^{n-1}$. Clearly, $Z_1$ is dense in $\operatorname{\mathfrak{Res}}$, and $$\overline{Z}_{k} \supset Z_{k+1}\cup \cdots \cup Z_n.$$ We will also set $\Sigma_{k}:=\nabla^{-1}(Z_k) \subset \PP(S_{d+1})$. By the Jacobian criterion, $\Sigma_k$ is the locus of hypersurfaces with only $k$ ordinary double points in general linear position and no other singularities. \[L:non-empty\] For every $1\leq k\leq n$, one has that $Z_k$ is a non-empty and irreducible subset of $\operatorname{Grass}(n, S_d)$, and $\Sigma_k$ is a non-empty and irreducible subset of $\PP(S_{d+1})^{ss}$. It follows from the Hilbert-Mumford numerical criterion that any hypersurface in $\PP^{n-1}$ of degree $d+1$ with at worst ordinary double point singularities is semistable. Having $k$ singularities at $k$ fixed points $p_1,\dots, p_k$ (resp., having $k$ fixed base points $p_1,\dots, p_k$) in general linear position is a linear condition on the elements of $\PP(S_{d+1})$ (resp., the elements of the Stiefel variety over $\operatorname{Grass}(n, S_d)$) and so defines an irreducible closed subvariety $\Sigma(p_1,\dots,p_k)$ in $\PP(S_{d+1})$ (resp., $Z(p_1,\dots,p_k)$ in $\operatorname{Grass}(n, S_d)$). The property of having exactly ordinary double points at $p_1,\dots, p_k$ (resp., having the base locus being equal to $\{p_1,\dots, p_k\}$ scheme-theoretically) is an open condition in $\Sigma(p_1,\dots,p_k)$ in $\PP(S_{d+1})$ (resp., $Z(p_1,\dots,p_k)$ in $\operatorname{Grass}(n, S_d)$) and so defines an irreducible subvariety $\Sigma^{0}(p_1,\dots,p_k)$ (resp., $Z^0(p_1,\dots,p_k)$). We conclude the proof of irreducibility by noting that $\Sigma_k=\operatorname{PGL}(n)\cdot \Sigma^{0}(p_1,\dots,p_k)$ (resp., $Z_k=\operatorname{PGL}(n)\cdot Z^{0}(p_1,\dots,p_k)$). Since $\Sigma_{k}=\nabla^{-1}(Z_k)$, it suffices to check the non-emptiness of $\Sigma_{k}$. If $F\in \Sigma_n$ has ordinary double points at $p_1,\dots, p_n$, then by the deformation theory of hypersurfaces, there exists a deformation of $F$ with ordinary double points at $p_1,\dots, p_k$ and no other singularities. Indeed, if $G\in S_{d+1}$ is a general form vanishing at $p_1,\dots, p_k$ and non-vanishing at $p_{k+1},\dots, p_n$, then $F+tG \in \Sigma^{0}(p_1,\dots,p_k)$ will have ordinary double points at $p_1,\dots, p_k$ and no other singularities for $0<t\ll 1$. It remains to prove that $\Sigma_n$ is non-empty. Indeed, the following is an element of $\Sigma_n$: $$(d-1)(x_1+\cdots+x_n)^{d+1}-(d+1)(x_1+\cdots+x_{n})^{d-1}(x_1^2+\cdots+x_n^2) +2(x_1^{d+1}+\cdots+x_n^{d+1}).$$ In fact, a generic linear combination of all degree $(d+1)$ monomials with the exception of $x_i^{d+1}$, for $i=1,\dots, n$, and $x_i^{d}x_j$, for $i, j=1,\dots, n$, $i<j$, is a form with precisely $n$ ordinary double point singularities in general linear position. By Proposition \[P:A-Gr-defined\], we know that $\AA_{\operatorname{Gr}}$ is defined at all points of $Z_{1}\cup \cdots \cup Z_n$. In fact, we can explicitly compute $\AA_{\operatorname{Gr}}(U)$ for all $U\in Z_n$, as well as the orbit closure of $\AA_{\operatorname{Gr}}(U)$ for all $U\in Z_{n-1}$. We need a preliminary fact. \[P:power\] Suppose $U\in \operatorname{Grass}(n, S_d)$ and $p\in \VV(I_U)\subset \PP V^{\vee}$. Let $L \in V^{\vee}$ be a non-zero linear form corresponding to $p$. Then $L^{n(d-1)-1} \in (I_U)_{n(d-1)-1}^\perp$. Since $p\in \VV(I_U)$, all elements of $(I_U)_{n(d-1)-1}$ vanish at $p$, and it follows that $F\circ L^{n(d-1)-1}=0$ for all $F\in (I_U)_{n(d-1)-1}$ (cf. Remark \[R:vanishing\]). \[C:Fermat\] Suppose $U\in Z_k$ is such that $$\VV(I_U)=\{p_1:=[1:0:\cdots:0], p_2:=[0:1:\cdots:0], \dots, p_k:=[0:\cdots:1:\cdots :0]\}.$$ Then $$\AA_{\operatorname{Gr}}(U) = \langle z_1^{n(d-1)-1}, \dots, z_k^{n(d-1)-1}, g_{k+1}(z_1,\dots,z_n), \dots, g_n(z_1,\dots,z_n) \rangle,$$ for some $g_{k+1},\dots, g_n\in \D_{n(d-1)-1}$. In particular, for $U\in Z_n$ one has $$\AA_{\operatorname{Gr}}(U) = \langle z_1^{n(d-1)-1}, \dots, z_n^{n(d-1)}\rangle=\nabla\bigl(\bigl[z_1^{n(d-1)}+\cdots+z_n^{n(d-1)}\bigr]\bigr).$$ Moreover, for a generic $U\in Z_k$, we have $\AA_{\operatorname{Gr}}(U) \in \operatorname{Grass}(n, \D_{n(d-1)})_{\operatorname{Res}}$. Since the point $p_i=\VV(x_1,\dots, \widehat{x_i}, \dots, x_n) \in \PP V^{\vee}$ corresponds to the linear form $z_i \in V^{\vee}$, Proposition \[P:power\] implies that $z_i^{n(d-1)-1}\in \AA_{\operatorname{Gr}}(U)$ for every $i=1,\dots, k$. As $Z_n \subset \overline{Z}_{k}$ and $\AA_{\operatorname{Gr}}(U)\in \operatorname{Grass}(n, \D_{n(d-1)})_{\operatorname{Res}}$ for every $U\in Z_n$, it follows that $\AA_{\operatorname{Gr}}(U)$ is also spanned by a regular sequence for a generic $U\in Z_{k}$. The claim follows. Consider the rational maps $$\begin{gathered} \xymatrix{ \PP(S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \ar[rd]_{\overline{\nabla}} \ar@{-->}[rr]^{\bar{A}} & & \PP(\D_{n(d-1)})^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n) \\ & \operatorname{Grass}(n, S_d)^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n) \ar@{-->}[ur]_{\bar{\AA}} & }\label{D:rational-triangle-GIT} \end{gathered}$$ of projective GIT quotients. \[T:Fermat\] There is a dense open subset $Y_k$ of $Z_k$ such that $$\bar \AA \colon \operatorname{Grass}(n,S_d)^{ss} {/\hspace{-0.25pc}/}\operatorname{SL}(n) \dashrightarrow \PP(\D_{n(d-1)})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n)$$ is defined on $\pi_1(Y_k)$, $k=1,\dots,n$. Moreover, for $U\in Y_k$ the value $\bar \AA(\pi_1(U))$ is the image under $\pi_2$ of a polystable $k$-partial Fermat form. In particular, for every $U\in Z_{n}$ and for a generic $U\in Z_{n-1}$ $$\bar \AA(\pi_1(U))=\pi_2\left(z_1^{n(d-1)}+\cdots+z_n^{n(d-1)}\right).$$ Recall that $Z_k$ is non-empty by Lemma \[L:non-empty\]. Suppose $U\in Z_k$ is generic, then by Corollary \[C:Fermat\] in suitable coordinates we have $$\AA_{\operatorname{Gr}}(U) = \langle z_1^{n(d-1)-1}, \dots, z_k^{n(d-1)-1}, g_{k+1}(z_1,\dots,z_n), \dots, g_n(z_1,\dots,z_n) \rangle,$$ and $\AA_{\operatorname{Gr}}(U)\notin \operatorname{\mathfrak{Res}}$. It follows (as in the proof of [@fedorchuk-isaev Proposition 2.7]) that the closure of the $\operatorname{SL}(n)$-orbit of $\AA_{\operatorname{Gr}}(U)$ contains $$\label{E:decomposable-A} \langle z_1^{n(d-1)-1}, \dots, z_k^{n(d-1)-1}, \bar g_{k+1}(z_{k+1},\dots,z_n), \dots, \bar g_n(z_{k+1},\dots,z_n) \rangle,$$ where $\bar g_i:=g_i(0,\dots,0, z_{k+1},\dots, z_n)$ for $i=k+1,\dots, n$. Then the claim follows for for $k=n-1$ and $k=n$ as in these cases we necessarily have $\bar g_n=z_n^{n(d-1)-1}$. For $k$ arbitrary, since $\overline{\nabla}$ is a closed immersion by Theorem \[T:gradient\], we conclude that $\bar \AA$ is defined at $\pi_1(U)$. Let $F\in \PP(\D_{n(d-1)})^{ss}$ be a polystable element with $\pi_2(F)=\bar\AA(\pi_1(U))$. Then we must have $\nabla(F) \in \overline{\operatorname{SL}(n)\cdot \AA_{\operatorname{Gr}}(U)}$, and so $\nabla(F)$ is linearly equivalent to an element of the form . It follows at once that $$\bar \AA(\pi_1(U))=\pi_2\left(z_1^{n(d-1)}+\cdots+z_k^{n(d-1)-1}+G(z_{k+1},\dots,z_n)\right)$$ is the image under $\pi_2$ of a polystable $k$-partial Fermat form. We will now establish Theorem \[T:barA1\] as detailed in the next two corollaries. \[cor:n-1,n\] The rational map $$\bar A \colon (\PP S_{d+1})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n) \dashrightarrow \PP(\D_{n(d-1)})^{ss}{/\hspace{-0.25pc}/}\operatorname{SL}(n)$$ is defined at a generic point of $\pi_0(\Sigma_{n-1})$ and at every point of $\pi_0(\Sigma_n)$. For a generic $f\in \Sigma_{n-1}$ and for every $f\in \Sigma_n$, we have $$\bar A(\pi_0(f))=\pi_2(z_1^{n(d-1)}+\cdots+z_n^{n(d-1)}).$$ \[C:contraction\] When $n=2$, the rational map $\bar A$ contracts the discriminant divisor to a point (corresponding to the orbit of the Fermat form in $\D_{2d-4}$) for all $d\ge 3$. When $n=3$, the rational map $\bar A$ contracts the discriminant divisor to a lower-dimensional subvariety if $d\ge 3$. More generally, for every $n\ge 4$ there exists $d_0$ such that for all $d\ge d_0$ the map $\bar A$ contracts the discriminant divisor to a lower-dimensional subvariety. Notice that $\Sigma_1$ is dense in the discriminant divisor $\Delta$. Hence, for $n=2$ the statement follows from Corollary \[cor:n-1,n\]. When $n=3$, Theorem \[T:Fermat\] implies that $\bar A(\pi_0(\Sigma_1))$ lies in the locus of a $1$-partial Fermat form in $\D_{3(d-1)}$. The linear equivalence classes of $1$-partial ternary Fermat forms are in bijection with the linear equivalence classes of binary degree $3(d-1)$ forms. The dimension of this locus is $3d-6$, which for $d\ge 3$ is strictly less than the dimension $\binom{d+3}{2}-10$ of the discriminant divisor. If $n\ge 4$, by Theorem \[T:Fermat\] the set $\bar A(\pi_0(\Sigma_1))$ lies in the locus of a $1$-partial Fermat form in $\D_{n(d-1)}$. The linear equivalence classes of $1$-partial Fermat forms in $n$ variables are in bijection with the linear equivalence classes of degree $n(d-1)$ forms in $n-1$ variables. The dimension of this locus is ${n(d-1)+(n-2)}\choose{n-2}$, which for sufficiently large $d$ is strictly less than the dimension of the discriminant divisor ${{(d+1)+(n-1)}\choose{n-1}}-(n^2+1)$. We conclude the paper with an alternative proof of the main fact of [@alper-isaev-assoc] (see Proposition 4.3 therein). \[C:generic-smoothness\] The closure of $\operatorname{Im}A$ in $\PP(\D_{n(d-1)})^{ss}$ contains the orbit $$\operatorname{SL}(n)\cdot \left\{z_1^{n(d-1)}+\cdots+z_n^{n(d-1)}\right\}$$ of the Fermat hypersurface. Consequently, $A(f)$ is a smooth form for a generic smooth $f \in S_{d+1}$. By Corollary \[cor:n-1,n\], we have $$\pi_2(z_1^{n(d-1)}+\cdots+z_n^{n(d-1)}) \in \operatorname{Im}(\bar A).$$ Since the orbit of the Fermat hypersurface is closed in $\PP(\D_{n(d-1)})^{ss}$, it lies in the closure of $\operatorname{Im}A$. [^1]: The first author was partially supported by the NSA Young Investigator grant H98230-16-1-0061 and Alfred P. Sloan Research Fellowship.

ArXiv

--- author: - 'T. Sperling' - 'W. B[ü]{}hrer' - 'C.M. Aegerter' - 'G. Maret' title: Direct determination of the transition to localization of light in three dimensions --- [**In the diffusive transport of waves in three dimensional media, there should be a phase transition with increasing disorder to a state where no transport occurs. This transition was first discussed by Anderson in 1958 [@anderson58] in the context of the metal insulator transition, but as was realized later it is generic for all waves [@anderson85; @john84]. However, the quest for the experimental demonstration of “Anderson” or “strong” localization of waves in 3D has been a challenging task. For electrons [@bergmann] and cold atoms [@kondov11], the challenge lies in the possibility of bound states in a disordered potential well. Therefore, electromagnetic and acoustic waves have been the prime candidates for the observation of Anderson localization [@kuga; @albada; @wolf; @drakegenack; @wiersma97; @scheffold99; @fiebig08; @stoerzer06; @stoerzer06n2; @acousticexp; @hu08]. The main challenge using light lies in the distinction between effects of absorption and localization [@wiersma97; @scheffold99]. Here we present measurements of the time-dependence of the transverse width of the intensity distribution of the transmitted waves, which provides a direct measure of the localization length and is independent of absorption. From this we find direct evidence for a localization transition in three dimensions and determine the corresponding localization lengths.**]{} In the diffusive regime ($kl^* \gg 1$) the mean square width $\sigma^2$ of the transmitted pulse, i.e. the spread of the photon cloud, is described by a linear increase in time $\sigma^2 = 4 D t$ [@lenke00]. Here, $D$ is the diffusion coefficient for light, $k$ the wave-vector and $l^*$ the transport mean free path. When considering interference effects of the diffusive light, Anderson et. al [@abrahams] predicted a transition to localization in three dimensional systems at high enough turbidity $(kl^*)^{-1}$. The criterion for where this transition should occur is known as the Ioffe-Regel criterion, namely $kl^* \lesssim 1$ [@ioffe60]. At such high turbidities, light will be localized to regions of a certain length scale, namely the localization length $\xi$, which diverges at the transition to localization. This implies that $\sigma^2$ initially increases linearly with time, but saturates at a later time $t_{\text{loc}}$ (localization time) towards a constant value given by $\sigma^2 = \xi^2$, where $\xi$ is the localization length. In this work we present measurements of light propagation in 3D open, highly scattering TiO$_2$ powders. Given the high turbidity of the samples studied and the large slab thickness ($L$ varying from 0.6 mm to 1.5 mm) the transmitted light undergoes typically a few million scattering events in any of the three spatial directions before leaving the sample. Thus our samples present a true bulk 3D medium for light transport. The great advantage of determining the time dependence of the width of the transmission profile lies in the fact that since the width is obtained at a specified time, absorption effects are present on all paths equally. This means that the width of the profile at a given time is [*independent of absorption*]{}. This can be seen from the general definition of the width in terms of the spatial dependence of the photon density $T(\rho,t)$, where $\rho$ is a vector in the 2D transmission plane with the origin at the center of the beam: $\sigma^2(t) = \int\rho^2 T(\boldsymbol{\rho},t)\text{d}^2\boldsymbol{\rho}/\int T(\boldsymbol{\rho},t)\text{d}^2\boldsymbol{\rho}.$ In this definition, an exponential decrease due to absorption enters $T(\rho,t)$ both in the nominator and in the denominator and thus cancels out. In the diffusive regime, the profile will be given by a Gaussian: $T(\rho) \propto \exp(-\frac{\rho^2}{8Dt})$, i.e. with a width $\sigma^2 = 4Dt$. Hence we fit a 2D Gaussian to the intensity profile at a given time (see Fig. \[fig:raw\], which shows the gated intensity profile at three different time points demonstrating the increase in width with time). This fit yields the width of the Gaussian in both the x- and y-direction. In localizing samples, the intensity distribution is expected to be exponential, with a characteristic length scale $\xi$. This can be seen in our samples, however at small distances $\rho$, the profile can be well approximated by a Gaussian (see supplementary material). Hence, we fit a Gaussian to all our samples, which gives qualitatively similar fits as an exponential function in the localized case (see supplementary material). The fitted widths are then plotted as a function of time to yield the results shown in Fig. \[fig:tmax\]. In the case of a diffusive sample, Aldrich anatase, with $kl^*_{AA} = 6.4$, the square of the width increases linearly over the whole timespan (see Fig. \[fig:tmax\] a) as expected. The small deviation from linearity around the diffusion time $\tau_{max}$ is a result of the gating of the high rate intensifier (HRI) [@HRI] (see supplementary material). The slope of the increase is in very good accord with the diffusion coefficient determined from time dependent transmission experiments [@stoerzer06]. Note also that the time dependent width can exceed the thickness of the sample, which is a consequence of the fact that we are studying the transmission profile at specific times. The width $\sigma^2$ of the transmitted pulse gives a direct measure of the localization length $\xi$ in the localizing regime. This is because the 2D transmission profile of the photon cloud is confined to within a localization length. When considering an effective diffusion coefficient corresponding to the slope of the temporal increase in width, one thus obtains an effective decrease of the diffusion coefficient with time as $D(t) \propto 1/t$ after a time scale corresponding to the localization length [@berkovits]. In this picture, for large $L$, one expects a time dependence of the width, which is linear up to the localization length and then remains constant as time goes on. Numerical calculations of self-consistent theory [@skipetrov06; @cherroret10] give a different increase at short times as $\sigma^2 \propto t^{1/2}$ and a plateau value of $\sigma^2 = 2L\xi$ for $L \gg \xi$. These predictions can be directly tested from data of samples with high turbidity, which show non-classical diffusion in time dependent transmission measurements. This is shown in Fig. \[fig:tmax\] b) and c). Taking a closer look at the short time behavior one can see that $\sigma^2$ increases linearly in time contrary to the self-consistent theory calculation. This is similar to the behavior found in acoustic waves [@hu08]. However in contrast to the diffusive sample, a plateau of the width can be clearly seen. This is in good accord with the theoretical prediction and a direct sign of Anderson localization. This plateau can also be seen directly from the transmission profiles shown in Fig. \[fig:raw\], where the normalized intensity profile is shown for three different time points. At late times, the width does no longer increases indicating a localization of light. The data shown in Fig. \[fig:tmax\] also show results for samples of different thickness. These samples of different thickness are made from the same particles but may vary slightly in terms of filling fraction. However as checked by coherent backscattering, samples made up from the same particles have very comparable turbidity (see supplementary material). If the thickness $L$ of the sample becomes comparable to the localization length, a decrease of the width of the photon distribution with time can be observed. This surprising fact can be understood in a statistical picture of localization, where a range of localization lengths exists in the sample corresponding to different sizes of closed loops of photon transport. In finite slabs, larger localized loops will be cut off by the surfaces leading to a lower population of such localized states at longer times. Thus on average, the observed width will correspond to increasingly shorter localization lengths and thus a decrease of $\sigma^2$ with time can be observed. This is schematically illustrated in Fig. \[fig:tmax\] d). Such a peak in the width of the intensity distribution has also been seen in calculations of self-consistent theory, albeit in thicker samples [@cherroret10]. When the thickness decreases even more, such that it is shorter than the localization length, the plateau in the width is lost altogether and $\sigma^2$ increases over the whole time window. In fact, the behavior then corresponds to that predicted for the mobility edge [@berkovits], where a sub-linear increase of $\sigma^2 \propto t^{2/3}$ is predicted. At the transition one observes a kink in $\sigma^2$ and the ratio of the initial slope to that at the kink corresponds to the sub-diffusive exponent $a$. In addition, this thickness dependence can be used as an alternative determination of the localization length. The evaluation of the plateaus of the localizing samples for different thicknesses, yields a localization length independent of $L$. In case the time dependence showed a maximum rather than a plateau, the maximum value was used. Thus we identify $\sigma_\infty^2 = \xi^2$ and obtain $\xi_\text{R104} = 717(6) \mu\text{m}$ for R104, $\xi_\text{R902} = 717(9)\mu\text{m}$ for R902 and $\xi_\text{R700} = 670(9)\mu\text{m}$ for R700. These are mean values for all thicknesses investigated. As expected, sample R700 has the smallest localization length $\xi$, as has already been concluded by time of flight experiments [@stoerzer06] and corresponds to the lowest value of $kl^*_{R700} = 2.7$ in this sample. In terms of localization, R104 and R902 are very similar, which again is in good accord with the fact that their turbidities are very similar, $kl^*_{R104} = 3.7$ and $kl^*_{R902} = 3.4$ respectively, even though their other sample properties are rather different. As stated above, this determination of $\xi$ is in good accord with that from the thickness dependence of the occurrence of a plateau. As seen in Fig. \[fig:tmax\], R104 with a thickness of $L = 0.71\text{mm}$ behaves sub-diffusively, but the sample with $L = 0.75\text{mm}$ shows a plateau, indicating a localization length of $\xi = 0.73(2)\text{mm}$. The same transitional behavior can be seen for R902 between $0.7\text{mm}<L<0.8\text{mm}$ as well. So far, we have shown that for different samples showing a range of $kl^*$ close to unity a qualitative change in the transport properties occurs which is consistent with the transition to Anderson localization. In order to show that these are not sample intrinsic properties, we now study one and the same sample at different incoming wavelengths. The turbidity depends quite strongly on the wavelength $\lambda$ of light, which we tuned from [$550 \, \text{nm}$]{} to [$650 \, \text{nm}$]{}. For these wavelengths, we have determined that the turbidity changes from $kl^*_\text{550nm} \approx 2.1$ up to $kl^*_\text{650nm} \approx 3.45$, thus spanning a range similar to that of the different samples above. At the highest and lowest wavelengths, the values of $kl^*$ were interpolated from the accessible values, which is a good approximation, since for the investigated region $kl^*$ are found to scale linearly with $\lambda$ (see supplementary material). The result of such a spectral measurement of a R700 sample ($L = 0.98\, \text{mm}$ and $m = 377\, \text{mg}$) is shown in Fig. \[fig:spectral\]. For the wavelengths of [$640 \, \text{nm}$]{} and [$650 \, \text{nm}$]{}, corresponding to the largest values of $kl^*$, $\sigma^2$ does not saturate, which shows that the mobility edge is approached. This allows a direct characterization of the localization transition with a continuous change of the order parameter. We have determined the same spectral information also from a R104 sample, which is closer to the mobility edge at a wavelength of 590 nm and for a rutile sample from Aldrich, which shows classical diffusion at 590 nm. For all of these samples, we have determined the value of $kl^*$ [@gross07]. With the value of $\xi$, and the scattering strength $kl^*$ we are able to determine the approach to the mobility edge at $kl^*_\text{crit}$, as shown in Fig. \[fig:kl\*-sigma\]. At the mobility edge, we can determine the qualitative change in behavior from the ratio of the slopes of $\sigma^2$ as a function of time in the localized or sub-linear regime and the initial diffusive regime (see supplementary material). This gives a direct estimate of the exponent $a$ with which the width increases with time, $\sigma^2 \propto t^{a}$ shown in Fig. \[fig:transition\]. There is a clear transition in the behavior with $kl^*$, showing a critical value of $kl^*_\text{crit}=4.5(4)$, above which $a = 1$ and below which $a = 0$. This is in good accord with the determination from time of flight measurements on similar samples yielding $kl^*_\text{crit,ToF}=4.2(2)$ [@stoerzer06n2]. Note that with an effective refractive index of the samples of $n_\text{eff} \simeq 1.75$, a critical value of $kl^*_\text{crit} = 4.2$ corresponds to an onset of localization at the point of $l^*/\lambda_\text{eff} = 1$, which is a reasonable expectation for the onset of localization. The dependence of the inverse width on the turbidity, as shown in Fig. \[fig:kl\*-sigma\], also indicates the critical behavior around the transition. Below the critical turbidity, $\sigma^2$ increases at all times and the corresponding inverse localization length is zero. At the mobility edge, the localization length is limited by the sample thickness, which in the case shown here was approximately 1 mm and a more detailed determination of the intrinsic localization length is not possible. For highly turbid samples, well below the transition, the inverse localization length seems to increase linearly with decreasing $kl^*$ indicating an exponent of unity. However, there is insufficient dynamic range close to the transition for a full determination of a critical exponent. In conclusion, we have shown direct evidence for localization of light in three dimensions and the corresponding transition at the mobility edge. This has been achieved using the time dependence of the mean square width $\sigma^2$ of the transmission profile, which is an excellent measure for the onset of localization of light. In contrast to other measures, it is completely independent of absorption and allows a [*direct*]{} determination of the localization length for samples close to the mobility edge. We find that for highly turbid samples, $\sigma^2$ shows a plateau, which changes to a sub-linear increase for critical turbidities and becomes linear for purely diffusive samples. This allows a detailed characterization of the behavior of transport close to the transition, which is not possible with other techniques. By evaluating the plateau $\sigma^2_\infty$ of localizing samples one can directly access the localization length $\xi$. For sample thicknesses close to the localization length, we moreover observe a decrease in the width of the photon cloud, which we associate with a statistical distribution of microscopic localization lengths. A description of these data will stimulate further theoretical work and comparison between such quantitative theories, such as self-consistent theory [@cherroret10] or direct numerical simulation [@gentilini10] and the data can then yield valuable information about the statistical distribution of localization lengths close to the transition. In addition, we have shown that the transition to localization can be observed in one and the same sample using spectral measurements, thus continuously varying the control parameter of turbidity through the transition. For highly turbid samples, the width of the transmission profile saturates at a value, which increases with decreasing turbidity until the localization length is comparable to the sample thickness. At this point the width increases at all times, albeit with a sub-linear increase at long times. This behavior is expected from the diffusion coefficient at the mobility edge [@berkovits]. Such measurements close to the transition between Anderson localization and diffusion allow a determination of the critical turbidity $kl^*_\text{crit} = 4.5(4)$, which is in good agreement with an indirect determination using time of flight measurements. In addition, our determination of the localization length during the approach to the localization transition allows an estimate of the critical exponent of the transition. Well away from the critical regime, we find a value close to unity, which is not incompatible with theoretical determinations [@abrahams; @john84; @numeric]. A complete description of the transition in open media taking finite size effects into account will be a great challenge for future theoretical descriptions of Anderson localization. [**Methods**]{} The samples are slabs made up of nano particles of sizes ranging from 170 to 540 nm in diameter with polydispersities ranging between 25 and 45 $\%$. Powders were provided by DuPont and Sigma Aldrich. These samples are slightly compressed and have been used previously [@stoerzer06] to demonstrate non-classical transport behaviour in time dependent transmission. TiO$_2$ has a relatively high refractive index in the visible of $n=2.7$ in the rutile phase and 2.5 in the anatase phase. The extremely high turbidity of the samples implies the use of a high power laser system to be able to measure this transmitted light. We use a frequency doubled Nd:YAG laser (Verdi V18), operated at [$18 \, \text{W}$]{} output power, to pump a titanium sapphire laser (HP Mira). The HP Mira runs mode locked with a repetition rate of [$75 \, \text{MHz}$]{} at a maximum of about [$4 \, \text{W}$]{}. To convert the laser light from about [$790 \, \text{nm}$]{} to orange laser light ([$590 \, \text{nm}$]{}) a frequency doubled OPO is used. The laser wavelength emitted by the OPO can be tuned from approx. [$550 \, \text{nm}$]{} to [$650 \, \text{nm}$]{}. To approximate a point-like source the laser beam was focused onto the flat front surface of the sample with a waist of 100 $\mu$m. The transmitted light was imaged from the flat backside by a magnifying lens ($f =$[$25 \, \text{mm}$]{}, mounted in reverse position) onto a high rate intensifier (HRI, LaVision PicoStar). The HRI can be gated on a time scale of about [$1 \, \text{ns}$]{} and the gate can be shifted in time steps of [$0.25 \, \text{ns}$]{}. The HRI is made of gallium arsenide phosphide which has a high quantum efficiency of maximum 40.6 % at about [$590 \, \text{nm}$]{}. A fluorescent screen images the signal onto a [$16 \, \text{bit}$]{} CCD Camera with a resolution of 512 $\times$ 512 pixel. With this system we were able to record the transmitted profile with a time resolution below a nanosecond. To measure the turbidity of a sample we used a backscattering set-up described elsewhere [@fiebig08; @gross07]. With this setup covering the full angular range, it is possible to determine $kl^*$ from the inverse width of the backscattering cone. Since this system used different laser sources, the spectral range of the set-up is more limited in wavelength ([$568 \, \text{nm}$]{} to [$619 \, \text{nm}$]{} and [$660 \, \text{nm}$]{}). This work was funded by DFG, SNSF, as well as the Land Baden-Württemberg, via the Center for Applied Photonics. Furthermore we like to thank Nicolas Cherroret for his support and fruitful discussions. [99]{} Anderson, P.W., ’Absence of diffusion in certain random lattices.’, Phys. Rev. [**109**]{}, 5 (1958). Anderson, P.W., ’The question of classical localization: a theory of white paint?’, Philosophical Magazine. Lett [**52**]{}, 3 (1985). John, S., ’Electromagnetic Absorption in a Disordered Medium near a Photon Mobility Edge.’, Phys. Rev. Lett. [**53**]{}, 2169 (1984). Altshuler, B.L. et al., Mesoscopic Phenomena in Solids (North-Holland, Amsterdam, 1991). Kondov, S.S., [*et al.*]{}, ’Three-Dimensional Anderson Localization of Ultracold Matter.’, Science [**334**]{}, 66 (2011). Kuga, Y. and Ishimaru, A., ’Retroreflectance from a dense distribution of spherical particles.’, J. Opt. Soc. Am. A [**1**]{}, 831 (1984). van Albada, M.P. and Lagendijk, A., ’Observation of weak localization of light in a random medium.’, Phys. Rev. Lett. [**55**]{}, 2696 (1985). Wolf, P.E. and Maret, G., ’Weak localization and coherent backscattering of photons in disordered media.’, Phys. Rev. Lett. [**55**]{}, 2696 (1985). Drake, J.M. and Genack, A.Z., ’Observation of nonclassical optical diffusion.’, Phys. Rev. Lett. [**63**]{}, 259 (1989). Wiersma, D.S., Bartolini, P., Lagendijk, A., and Righini, R., ’Localization of light in a disordered medium.’, Nature [**390**]{}, 671 (1997). Scheffold, F., Lenke, R., Tweer, R., and Maret, G., ’Localization or classical diffusion of light?’, Nature [**398**]{}, 206 (1999). Fiebig, S., Aegerter, C.M., Bührer, W., Störzer, M., Akkermans, E., Montambuax, G. and Maret, G., ’Conservation of energy in coherent backscattering at large angles.’, EPL [**81**]{}, 64004 (2008). Störzer, M., Gross, P., Aegerter, C.M. and Maret, G., ’Observation of the critical regime in the approach to Anderson localization of light.’, Phys. Rev. Lett. [**96**]{}, 063904 (2006). Aegerter, C.M., Störzer, M. and Maret, G., ’Experimental determination of critical exponents in Anderson localization of light.’, Europhys. Lett. [**75**]{}, 562 (2006). Bayer, G. and Niederdränk, T., ’Weak localization of acoustic waves in strongly scattering media.’, Phys. Rev. Lett. [**70**]{}, 3884 (1993). Hu, H., Strybulevych, A., Page, J.H., Skipetrov, S.E. and van Tiggelen, B.A., ’Localization of ultrasound in a three-dimensional elastic network.’, Nature Phys.[**4**]{}, 945 (2008). Lenke, R. and Maret, G., [*Multiple Scattering of Light: Coherent Backscattering and Transmission*]{}, Gordon and Breach Science Publishers (2000). Abrahams E., Anderson P.W., Licciardello D.C., and Ramakrishnan T.V., ’Scaling theory of localization: absence of quantum diffusion in two dimensions.’, Phys. Rev. Lett. [**42**]{}, 673 (1979). Ioffe, A.F. and Regel, A.R., ’Non-crystalline, amorphous and liquid electronic semiconductors.’, Progress in Semiconductors [**4**]{}, 237 (1960). Berkovits, R. and Kaveh, M., ’Propagation of waves through a slab near the Anderson transition: a local scaling approach.’, J. Phys. C: Cond. Mat. [**2**]{}, 307 (1990). Skipetrov, S.E. and van Tiggelen, B.A., ’Dynamics of Anderson localization in open 3D media.’, Phys. Rev. Lett. [**96**]{}, 043902 (2006). Cherroret, N., Skipetrov, S.E. and van Tiggelen, B.A., ’Transverse confinement of waves in random media.’, Phys. Rev. E [**82**]{}, 056603 (2010). Gross, P., Störzer, M., Fiebig, S., Clausen, M., Maret, G. and Aegerter, C.M., ’A precise method to determine the angular distribution of backscattered light to high angles.’, Rev. Sci. Instrum. [**78**]{}, 033105 (2007). The finite gating window leads to an averaging of the time-dependent width weighted by the time-dependent transmitted intensity. Due to the non-linear time-dependence of this intensity, the averaged width can be different around the maximum intensity for longer gating times. Gentilini, S., Fratalocchi, A., and Conti, C., ’Signatures of Anderson localization excited by an optical frequency comb.’, Physical Review B [**81**]{}, 014209 (2010). MacKinnon, A. and Kramer, B., ’One-parameter scaling of localization length and conductance in disordered systems.’, Phys. Rev. Lett. [**47**]{}, 1546 (1981). ![\[fig:tmax\] The mean square width scaled with the sample size $\frac{\sigma^2}{L^2}$ is shown for different samples. The time axis is scaled with the diffusion time $\tau_\text{max}$ (see supplementary material). In a) Aldrich anatase is shown, which behaves diffusively. Samples showing localizing effects are b) R104 and c) R700. The legends show the slab thickness $L$ in mm. d) Schematic illustration of the expectation for the time dependence of the width in the presence of statistically distributed localization lengths as discussed in the text. The decreasing population of the different grey lines at late times for larger localization lengths, leads to an overall decrease of the width in particular for sample thicknesses close to the average localization length, because big loops are leaking out of the sample. The different coloured lines correspond to the time dependence of the width with increasing microscopic localization length from small (green) to large (red). ](all_tmax4.pdf){width="1\linewidth"} ![\[fig:spectral\] The spectral measurement of a R700 sample ranging from [$550 \, \text{nm}$]{} – [$650 \, \text{nm}$]{}, corresponding to $kl*$ values between 2.1 and 3.6, is shown. With decreasing wavelength $\lambda$ the turbidity $kl^*$ increases, as well as localizing effects are getting stronger. This can be seen via the lower mean square width $\sigma^2_\infty$ of the plateaus. For the wavelengths above [$640 \, \text{nm}$]{} one can observe a breakdown of localization to a sub-diffusive behavior. The legend shows the wavelength of light in nm.](spectral3.pdf){width="\linewidth"} ![\[fig:kl\*-sigma\] The inverse of the mean square width $\sigma^2_\infty$ of the plateau against $kl^*$ for different samples. As can be seen, the width, corresponding to the localization length diverges at a value of $kl^* \simeq 4.5$, indicating the transition from a localized to a non-localized state. The increase of the localization length approaching the critical turbidity can also be used to estimate the critical exponent. ](spectral_kl_sigma.pdf){width="\linewidth"} ![\[fig:transition\] The value of the exponent $a$ describing the temporal increase of the mean square width. In the diffusive regime, the exponent should be unity, whereas in the fully localized regime a value of zero is expected. At the mobility edge the sub-diffusive increase corresponds to intermediate values. This allows a determination of the critical turbidity. ](transition.pdf){width="\linewidth"}

ArXiv

--- abstract: 'We predicted residual fluid intelligence scores from T1-weighted MRI data available as part of the ABCD NP Challenge 2019, using morphological similarity of grey-matter regions across the cortex. Individual structural covariance networks (SCN) were abstracted into graph-theory metrics averaged over nodes across the brain and in data-driven communities/modules. Metrics included degree, path length, clustering coefficient, centrality, rich club coefficient, and small-worldness. These features derived from the training set were used to build various regression models for predicting residual fluid intelligence scores, with performance evaluated both using cross-validation within the training set and using the held-out validation set. Our predictions on the test set were generated with a support vector regression model trained on the training set. We found minimal improvement over predicting a zero residual fluid intelligence score across the sample population, implying that structural covariance networks calculated from T1-weighted MR imaging data provide little information about residual fluid intelligence.' author: - 'Neil P. Oxtoby[[](https://orcid.org/0000-0003-0203-3909)]{}' - 'Fabio S. Ferreira[[](https://orcid.org/0000-0002-0977-2539)]{}' - 'Agoston Mihalik[[](https://orcid.org/0000-0002-4510-4933)]{}' - 'Tong Wu[[](https://orcid.org/0000-0002-7468-2249)]{}' - 'Mikael Brudfors[[](https://orcid.org/0000-0002-2884-2336)]{}' - 'Hongxiang Lin[[](https://orcid.org/0000-0001-6643-327X)]{}' - 'Anita Rau[[](https://orcid.org/0000-0002-4759-2846)]{}' - 'Stefano B. Blumberg[[](https://orcid.org/0000-0002-7150-9918)]{}' - 'Maria Robu[[](https://orcid.org/0000-0003-0106-0542)]{}' - 'Cemre Zor[[](https://orcid.org/0000-0002-6141-2610)]{}' - 'Maira Tariq[[](https://orcid.org/0000-0002-2826-4046)]{}' - Maria Del Mar Estarellas Garcia - 'Baris Kanber[[](https://orcid.org/0000-0003-2443-8800)]{}' - 'Daniil I. Nikitichev[[](https://orcid.org/0000-0001-5877-9174)]{}' - 'Janaina Mourao-Miranda[[](https://orcid.org/0000-0002-3309-8441)]{}' bibliography: - 'ABCD\_Refs.bib' title: 'ABCD Neurocognitive Prediction Challenge 2019: Predicting individual residual fluid intelligence scores from cortical grey matter morphology' --- Introduction {#INTRO} ============ Establishing the neurobiological mechanisms underlying intelligence is a key area of research in Neuroscience [@Goriounova2019]. A strong correlation has been observed between cognitive ability measured at a very young age with the socioeconomic status [@Foverskov2017], as well as longevity and health [@Lam2017], at an older age. Moreover, intelligence has been shown to be very stable from young to old age in the same individuals [@Deary2007][@Deary2013]. Thus understanding the mechanisms of cognitive abilities has implications for health of the general population and can be used to enhance such abilities, for example through education or environment [@Gottfredson1997]. Neuroimaging plays a key role in advancing our knowledge of the neurological mechanisms of intelligence. Several brain-imaging studies have shown the link between brain features and intelligence, including a positive correlation with cortical volume and thickness, specifically in the frontal and temporal regions [@HulshoffPol2006; @Narr2007; @Choi2008; @Karama2011; @Jung2007]. A link has also been observed between intelligence and the structural integrity of white matter [@Penke2012] and the function integrity of the temporal, frontal and parietal cortices [@Wang2009]. Studies have also involved both adult and children [@Muetzel2015; @Yu2008]. The ABCD NP Challenge asks the question “How predictable is fluid intelligence from brain imaging data?” To answer this, we took a data-driven, exploratory approach of trying many models and image-based features — starting with a hackathon led by the UCL Centre for Medical Image Computing (CMIC). CMIC aims to make an impact on key medical challenges facing 21st century society through performing world-leading research on problems in medical imaging and image-analysis. Our expertise extends from feature extraction/generation through to image-based modelling [@Oxtoby2017; @Young2014], machine learning [@Schrouff2018; @Blumberg2018], and beyond. The hackathon took place one afternoon in February 2019 and involved researchers across research groups in UCL CMIC, in addition to colleagues from the affiliated UCL Wellcome Centre for Human Neuroimaging, UCL Department of Clinical and Experimental Epilepsy, and Max Planck UCL Centre for Computational Psychiatry and Ageing Research. Regular followup progress meetings followed the hackathon. The brain is a complex organ widely touted as operating as a cliquish small-world network [@Bassett2006], although this may not be the whole story [@Bassett2017]. The ABCD NP Challenge lacks the diffusion MRI data necessary to estimate anatomical connectivity via tractography. However, it is possible to quantify morphological similarity of an individual’s cortex using a graph called a “structural covariance network” (SCN), which can be used to distinguish between clinical groups [@Tijms2013]. We calculate SCNs for each individual in the ABCD NP Challenge data set and input them as features to train predictive models of residual fluid intelligence (rFIQ). The paper is structured as follows. The next section describes the challenge data and our methods. Section \[RESULTS\] presents our results which we discuss in section \[DISCUSSION\] then conclude. Methods {#METHODS} ======= Data ---- The ABCD NP Challenge data consists of a cross-section of imaging data and intelligence scores for children aged 9–10 years. The T1-weighted MRI data was acquired using the protocol detailed on the challenge website [@ABCDWEB] and in [@Casey2018a], and split into training ($N=3739$), validation ($N=415$), and test ($N=4515$) sets. The training and validation sets also include scores of fluid intelligence, which the ABCD Study measures using the NIH Toolbox Neurocognition battery [@Akshoomoff2013]. For the challenge, fluid intelligence was residualized to remove dependence upon brain volume, data collection site, age at baseline, sex at birth, race/ethnicity, highest parental education, parental income, and parental marital status. While we understand the motivation — the challenge is to predict intelligence from imaging — this pre-residualization choice in the challenge design is somewhat limiting because it completely removes any ability to include covariance of these factors with image-based features. The MRI data provided was already in pre-processed form. Pre-processing included skull-stripping, removing noise, correcting for field inhomogeneities [@Hagler2018; @Pfefferbaum2018] and affine alignment of all images to the SRI24 adult brain atlas [@Rohlfing2010]. The SRI24 segmentations and corresponding volumes were also provided. Unsurprisingly, the regional volumes were not predictive of a target that had been adjusted for total brain volume. Structural Covariance Network Features -------------------------------------- It has been shown that cortical morphology is predictive of cognitive deficits in individuals with Alzheimer’s disease [@Tijms2013]. We wanted to explore whether the same could be said for predicting intelligence, so we generated a structural covariance network (SCN) following [@Lawrie2012] (code available on GitHub) for each individual in the ABCD NP Challenge data set. The SCN is a graph where the nodes are small cortical regions (3 voxels cubed) and the edges quantify structural similarity (morphology) between nodes. From each SCN we generated nodal graph-theory features using the Brain Connectivity Toolbox [@Rubinov2010], which were then averaged across the brain and also within each of the largest three modules (communities) of the graph. We also considered measures of variation in these features (standard deviation and median-absolute deviation). Our 26 features include small-worldness, rich club coefficient, path length, node degree, clustering coefficient, and betweenness centrality (Table \[Table-1\]). See Figure \[Fig-0\] for a graphical representation of the pipeline. ![Structural Covariance Network feature-generation pipeline.[]{data-label="Fig-0"}](SCN-figure.pdf){width="0.9\columnwidth"} Generating approximately ten thousand SCNs and corresponding graph-theory features is an intensive computational task. When the pipeline failed for a given individual, or time was not permitting (such as the late addition of 868 additional test subjects), this resulted in missing data. For these few individuals ($\leq8\%$: Table \[Table-1\]) we inserted a prediction of zero (nominally the mean). Predictive Models {#predmod} ----------------- We trained two models to predict rFIQ from features based on morphological similarity. The first was the event-based model (EBM) of progression [@Young2014; @Fonteijn2012]. The second was support-vector regression (SVR) [@Drucker:1996:SVR:2998981.2999003]. We trained each model on data from the training set, and assessed performance using MSE on the validation set (Table \[Table-1\]). The best-performing model (SVR) was used to generate our submission to the challenge: predictions for the test set. The EBM learns a discrete sequence of progression events from normal/low state to abnormal/high. It was designed for neurodegenerative diseases but can be applied to any monotonic phenomenon. Here we define low rFIQ as more than one standard deviation (std) below the mean and high rFIQ as more than one std above the mean. If rFIQ is a monotonic function of structural covariance, then the EBM should be able to find a probabilistic sequence of events that represent this function. “Events” are structural covariance graph-theory features, and they must differ statistically between low-rFIQ and high-rFIQ for them to be included in the model — otherwise they contain no “signal” for this trajectory. We excluded features that “did not pass” ($p>0.10$) the Mann-Whitney U test of the null hypothesis that the distributions (low/high rFIQ) are equal. EBM stage and rFIQ score was input into a Kernel Ridge Regression model (default parameters, scikit-learn: [@Pedregosa2011]) to make the predictions. The SVR was run in PRoNTo version 3 (Pattern Recognition for Neuroimaging Toolbox) [@PRONTO; @Schrouff2013] — a software toolbox of pattern recognition techniques for the analysis of neuroimaging data. Model performance on the training set was assessed using 5-fold nested cross-validation (i.e. the internal and external loops had 5 folds) to optimise the penalty parameter C (we use 6 different logarithmically-spaced values: 0.01, 0.1, 1, 10, 100 and 1000) and compute the MSE per fold, which were averaged across folds to compute the final prediction error (Table \[Table-2\]). [|r|c|c|c|]{} & **Training** & **Validation** & **Test**\ & (N=3579 of 3739) & (N=390 of 415) & (N=4156 of 4515)\ Small-world & 1.68 (0.03) & 1.68 (0.02) & 1.68 (0.02)\ **Rich Club** – median & 0.29 (0.01) & 0.29 (0.01) & 0.29 (0.01)\ – mad & 0.11 (0.03) & 0.11 (0.01) & 0.11 (0.01)\ **Path Length** – median & 2.48 (0.03) & 2.48 (0.01) & 2.48 (0.01)\ – std & 1.15 (0.03) & 1.15 (0.02) & 1.16 (0.02)\ **Degree** – median & 1050 (45) & 1052 (45) & 1053 (40)\ – mad & 295 (19) & 294 (15) & 294 (15)\ **Centrality** – median & 6584 (171) & 6590 (150) & 6578 (152)\ – mad & 5153 (157) & 5157 (118) & 5158 (117)\ **Clustering** – median & 0.53 (0.01) & 0.53 (0.01) & 0.53 (0.01)\ – mad & 0.063 (0.005) & 0.063 (0.005) & 0.063 (0.005)\ \ **Avg. Degree** – 1 & 995 (233) & 1004 (232) & 995 (239)\ – 2 & 996 (240) & 997 (235) & 999 (239)\ – 3 & 1019 (242) & 1004 (243) & 1014 (238)\ Avg. degree z-score (all) & 0.2 (0.1) & 0.2 (0.1) & 0.2 (0.1)\ Avg. path length (all) & 1.5 (0.1) & 1.5 (0.1) & 1.5 (0.1)\ **Centrality** – 1 & 6020 (3750) & 6180 (3800) & 6100 (3780)\ – 2 & 6290 (3790) & 6030 (3740) & 6290 (3770)\ – 3 & 6660 (3830) & 6550 (3760) & 6520 (3810)\ **Clustering** – 1 & 0.53 (0.06) & 0.53 (0.06) & 0.52 (0.06)\ – 2 & 0.53 (0.06) & 0.53 (0.06) & 0.53 (0.06)\ – 3 & 0.53 (0.06) & 0.53 (0.06) & 0.52 (0.06)\ Results {#RESULTS} ======= We included 26 SCN graph-theory features that represent morphological similarity across the cortex. Table \[Table-1\] summarises the features we derived from the T1 images, and the level of completeness in each challenge data set (see Section \[predmod\]). For the EBM, only three features passed through our Mann-Whitney U test filter (see Methods): small-worldness, betweenness centrality (median), and degree. Even for these features, there was very little difference between the low- and high-rFIQ groups (see Table \[Table-1\]), with Cohen’s d effect sizes of $-0.11/0.06$ (small-world), $0.07/-0.10$ (degree), and $-0.09/0.009$ (centrality) in the training/validation sets. In light of the opposing effect direction (signs), the model’s poor generalisation performance is unsurprising (see Table \[Table-2\]). For the SVR model, two features were most important: small-worldness (weight $w = 11.42$); and clustering coefficient in community 2 ($w = 6.04$). Among the next most important were average path length and other clustering coefficients. Table \[Table-2\] shows our prediction results for both models: mean-squared errors and Pearson’s squared correlation coefficient for training and validation. It is clear that both the approaches did not generalise well under validation. Our submission to the challenge (SVR) was positioned near the middle of the testing leaderboard with $\mathrm{MSE}=93.8335$. ![Training and validation errors for the EBM approach.[]{data-label="Fig-1"}](Fig-1.pdf){width="0.95\columnwidth"} --------- ------- ------------- ---------- ------------- ------------- Test set MSE Correlation MSE Correlation MSE SVR 85.82 0.02 71.19 0.01 **93.8335** EBM+KRR 85.46 0.001 71.58 0.003 N/A --------- ------- ------------- ---------- ------------- ------------- : Mean-squared error (MSE) and correlation for the predictive models. For reference, the variance of the training set was $85.85$ and the validation set was $71.53$.[]{data-label="Table-2"} ![SVR prediction errors: (left) training set using 5-fold CV; (right) validation set.[]{data-label="Fig-2"}](Fig-2_1.pdf "fig:"){width="0.9\columnwidth"}\ ![SVR prediction errors: (left) training set using 5-fold CV; (right) validation set.[]{data-label="Fig-2"}](Fig-2_2.pdf "fig:"){width="0.9\columnwidth"} Discussion {#DISCUSSION} ========== The ABCD NP Challenge was certainly challenging. Our MSE for predicting residual fluid intelligence was only nominally better than simply predicting zero, i.e., the mean. This implies that the residual fluid intelligence is not explainable by graph theory features derived from structural covariance networks. We found similar results for all combinations of models and features attempted during and after our hackathon — from basic regression to deep learning. Moreover, the validation leader board (see challenge website) demonstrated that other entries into the challenge had similarly meagre performance improvement on simply predicting the mean. While the residualization process precluded the use of models that include covariance of the residualization factors [@Schrouff2018] with image-based features, it is difficult to say whether or not this would have improved the results dramatically. Including variables in the residualization procedure that are correlated with the predicted variable is likely to remove important variability in the data leading to predictive models with low performance [@Rao2017]. Conclusion ========== Based on our results, and those on the validation leaderboard for the challenge, we are inclined to conclude that structural imaging is probably incapable of predicting more than a couple of points worth of residual fluid intelligence.

ArXiv

--- abstract: 'Note: This short article was submitted to Nature Physics as a Correspondence. The intention was to provide a brief albeit significant criticism of the work of J. Dunkel and S. Hilbert, *Consistent Thermostatistics Forbids Negative Absolute Temperatures*, Nature Physics **10**, (2014). The respected editor decided not to publish the Correspondence. We have therefore decided to submit the paper to arXiv. Comments/criticisms are welcomed, particularly from the authors of the mentioned paper.' --- **To the editor —** Recently Dunkel and Hilbert [@dunkel2014] have argued that the consistency relation $T(\pdv*{S}{A_\mu})_E=-(\pdv*{E}{A_\mu})_S=-\expval{\pdv*{H}{A_\mu}}$ favors the Gibbs entropy ($S_G$) over the Boltzmann entropy ($S_B$) since this relation is satisfied if and only if $S$ is an adiabatic invariant. The first equality follows simply from invertibility of entropy $S(E,A_\mu)\rightarrow E(S,A_\mu)$; however, the second equality is rooted in assuming that a mechanical adiabatic process is equivalent to a slow thermodynamic adiabatic process. For a mechanical adiabatic process one can write $\expval{\dd H/\dd t}=\dd E/\dd t=\sum_\mu \expval{\pdv*{H}{A_\mu}}\dd A_\mu/\dd t$. On the other hand, for a thermodynamic adiabatic process ($\dd S/\dd t=0$), one can write $\dd E/\dd t=\sum_\mu (\pdv*{E}{A_\mu})_S \dd A_\mu/\dd t$. The equivalence of these two leads to the second relation in the “consistency relation", i.e. $(\pdv*{E}{A_\mu})_S=\expval{\pdv*{H}{A_\mu}}$. As is well-known $N$ is an important thermodynamic variable, which contributes to the total energy through chemical work, $\; \dbar W=\mu \dd N$, through the intensive parameter of chemical potential, $\mu \equiv (\pdv*{E}{N})_S$. In fact in the Supplementary Information, Dunkel and Hilbert explicitly mention that $A_\mu$ can be an external parameter, *“such as volume, particle number, magnetic field strength, etc"* [@dunkel2014]. Here we note that such a relation fails to hold for the important parameter $A_\mu=N$ even when $S$ is chosen to be the Gibbs entropy. In other words we show that thermodynamic chemical potential $\mu_{th}=(\pdv*{E}{N})_S=-T(\pdv*{S}{N})_E$ is not the same as “mechanical" chemical potential $\mu_m=\expval{\pdv*{H}{N}}$. This lack of consistency remains true regardless of choice of entropy $S_G$ or $S_B$. We therefore argue that the consistency relation is invalid and should not be used to argue in favor of one entropy ($S_G$) over the other ($S_B$). To see that the consistency relation is invalid for $S=S_G=\ln \lbrace \Tr[\Theta(E-H)]\rbrace$, for $ A_\mu=N$, one must simply realize that the trace operation (or integration in phase space for classical systems) is not interchangeable with differentiation with respect to $N$, as the degrees of freedom of a quantum or classical system (which determines the dimension of phase space in classical system), crucially depends on $N$. This interchangeability is assumed to always hold true (for any allowed $A_\mu$) by Dunkel and Hilbert in order to prove the consistency relation for $S_G$[@dunkel2014]. To see the above argument more explicitly note that in the microcanonical ensemble, $\expval{\pdv*{H}{N}}\equiv \Tr [\pdv*{H}{N}\delta (E-H)]/\omega=-\Tr[\pdv*{\Theta(E-H)}{N}]/\omega$, where $\omega=\Tr[\delta(E-H)]$. On the other hand, $T_G(\pdv*{S_G}{N})_E=(\pdv*{S_G}{N})_E/(\pdv*{S_G}{E})_N=\pderivative*{N} (\Tr [\Theta (E-H)])/\omega$. Therefore, one can see that the consistency relation holds only if one can move in the differentiation into the integration on state space, i. e. trace operation. This is clearly not allowed and has in fact been proved to be false in specific and explicit cases[@tavassoli2015]. If one replaced $\Tr$ with $\int ...\int \dd ^{3N}p \dd ^{3N}q/h^{3N}$ for a classical system, the lack of interchangeability will become more evident. As an example of why $\mu_{th} \neq \mu_m$, consider a simple classical ideal gas. It is well-known that the chemical potential of such a system is a large negative number $\mu_{th}=(\pdv*{E}{N})_S<<0$ [@swendsen2012]. On the other hand, the “mechanical" chemical potential $\mu_m=\expval{\pdv*{H}{N}}$ is clearly a nonnegative number as one can imagine by measuring the change of total energy by adding (a zero energy) particle. We note that lack of validity of consistency relation is rooted in the assumption that a slow thermodynamical adiabatic process is equivalent to a mechanically adiabatic process, which is in turn rooted in a mechanical interpretation of thermodynamics. A mechanical adiabatic process is simply a slow process where $ \dd A_\mu / \dd t$ is assumed to be very small. A thermodynamics adiabatic (and quasi-static) process is one which entropy remains constant. The main idea of Dunkel and Hilbert is that if entropy is a mechanical adiabatic invariant then the two processes are equivalent, i.e. $S=$const. However, we have argued that Gibbs entropy cannot be an adiabatic invariant with respect to $N$. Therefore, since $N$ is a fundamental thermodynamic parameter, arguments based on mechanical properties of Gibbs entropy such as consistency and/or adiabatic invariance should be viewed with extra care. It is important to note that since consistency relation is in effect an equivalence relation between various energy terms, such as mechanical work or chemical work, it is important for all such terms to obey this equivalence. Therefore, if such equivalence breaks down for a given term (i.e. with respect to $N$), then the fundamental thermodynamic relation $\dd E = T\dd S -p\dd V + \mu\dd N + ...$ cannot have a mechanical analog even if $S$ is considered to be an adiabatic invariant with respect to all but one energy term. In other word, since all systems of interest possess chemical potential, the lack of equivalence in this quantity casts doubt on the reliability of such consistency relation as a criterion to judge the validity of various entropy definitions. Next, we address some potential objections. The entire formulation of the consistency relation is based on the microcanonical (entropy) representation where one assumes that $S=S(E,V,N, ...)$. Although the system is isolated or closed (i.e. fixed $E,V,N,...$), one can clearly differentiate the entropy function with respect to its independent variables, i.e. $(\pdv*{S}{N})_{E,V} = f(E,V,N)$. Otherwise, intensive parameters such as temperature $T(E,V,N)= \lbrace(\pdv*{S}{E})_{V,N}\rbrace^{-1}$ cannot be defined. Furthermore, differentiation with respect to discrete variables is a standard practice in thermodynamics [@swendsen2012]. Quantum mechanics teaches us that energy is *not* a continuous variable. However, for large macroscopic systems (and even otherwise), we perform partial differentiation with respect to energy since $(\Delta E/E) \rightarrow 0$. In much the same way, particle number is considered a continuous variable since $(\Delta N / N=1/N) \rightarrow 0$. Therefore, one cannot exempt $N$ from the requirement of consistency relation, $A_\mu = N$. We end by making a general comment about the “mechanical foundation of thermodynamics". Thermodynamics is a formal theory based on lack of mechanical information about constituents of macroscopic systems, which naturally leads to the existence (and definition) of entropy. Mechanics, on the other hand, assumes availability of all relevant mechanical variables. Clearly, these two important theories are based on opposing set of assumptions. The insistence to base thermodynamics on a mechanical foundation has historically led to controversial issues, the most famous of which are thermodynamic irreversibility vs. mechanical reversibility as well as the problem of Maxwell’s demon[@swendsen2012]. The work of Dunkel and Hilbert yet provides another example of how such an insistence leads to controversial and inconsistent results. We finally emphasize that our results should not be interpreted as favoring one entropy definition over another, but a critique of the so called “consistency" relation which equates mechanical and thermodynamical adiabatic processes. [40]{} Dunkel, J. & Hilbert, S. *Nature Phys.* **10**, 67–72(2014). Tavassoli, A. & Montakhab, A. *arXiv preprint* arXiv:1512.02855(2015). Swendsen, R. H. *An Introduction to Statistical Mechanics and Thermodynamics*(Oxford University Press, 2012). ------------------------------------------------------------------------ **Afshin Montakhab\* and Arash Tavassoli** Department of Physics, College of Sciences, Shiraz University, Shiraz 71946-84795, Iran \*email: [email protected]

ArXiv

--- abstract: 'If $f$ is a polynomial of degree $d$ in ${{\mathbb{F}}}_{q}[x]$, let $c_{k}(f)$ be the number of cycles of length $k$ in the directed graph on ${\mathbb F}_{q}$ with edges $\lbrace (v,f(v))\rbrace_{v\in {\mathbb F}_{q}}.$ For random polynomials, the numbers $c_{k}, 1\leq k\leq b,$ have asymptotic behavior resembling that for the cycle lengths of random functions $f:[q]\rightarrow [q].$ However random polynomials differ from random functions in important ways. For example, given the set of cyclic (periodic) points, it is not necessarily true that all permutations of those cyclic points are equally likely to occur as the restriction of $f$. This, and the limitations of Lagrange interpolation, together complicate research on ${\mathbf{T}}(f)=$ the ultimate period of $f$ under compositional iteration. We prove a lower bound for the average value of $\log {\mathbf{T}}$: if $d=d(q)\rightarrow\infty$, but $d=o(\sqrt{q})$, then the expected value of $\log {\mathbf{T}}$ is $${\mathbb E}(\log {\mathbf{T}}):=\frac{1}{q^{d}(q-1)}\sum\limits_{f} \log {\mathbf{T}}(f) > \frac{d}{2}(1+o(1)),$$ where the sum is over all $q^{d}(q-1)$ polynomials of degree $d$ in ${{\mathbb{F}}}_{q}[x]$. Similar results are proved for rational functions.' author: - | Charles Burnette\ Department of Mathematics\ Drexel University\ Philadelphia, PA 19104-2875\ [email protected] - | Eric Schmutz\ Department of Mathematics\ Drexel University\ Philadelphia, PA 19104-2875\ [email protected] title: 'Periods of Iterated Rational Functions over a Finite Field\' --- Introduction ============ The classical theory of random mappingsconcerns functions chosen randomly from among the $q^{q}$ functions whose domain and codomain are a given $q$-element set $V.$ (See, for example, [@FO; @Harris60; @K].) Basic facts about random mappings have motivated conjectures and theorems in arithmetic dynamics[@Bach; @BGHK; @RodPan; @Pol; @Silverman; @Silvermanbook]. This paper includes ${\mathbb F}_{q}[x]$-analogs of known facts about random mappings cycle lengths and the lengths of their (ultimate) periods under function composition. If $f$ is any function from ${\mathbb F}_{q}$ to ${\mathbb F}_{q}$, let [cyclic]{}$(f)$ be the set of periodic points, i.e the set of vertices that lie on cycles in the functional digraph $G_{f}$ that is formed by putting an edge from $v$ to $f(v)$ for each $v$. Let $\mathbf{Z}(f) = |\textsc{cyclic}(f)|$ and let $\sigma_{f}$ be the restriction of $f$ to [cyclic]{}$(f)$. Finally, let ${{\mathbf{T}}}(f)$ be the least common multiple of the cycle lengths, i.e. the order of $\sigma_{f}$ as an element of the group Sym([cyclic]{}$(f)$). Equivalently, $ {{\mathbf{T}}}(f)$ is the ultimate period of $f$, defined as the smallest positive integer $t$ such that, for every $n\geq q, f^{(n+t)}=f^{(n)}. $ A theorem of Harris[@Harris] tells us how large ${\mathbf{T}}$ normally is for random mappings: if $\epsilon >0$, then for all but $o(q^{q})$ functions $f$, $$\label{bounds} e^{(\frac{1}{8}-\epsilon)\log^{2} q} <{{\mathbf{T}}}(f) < e^{(\frac{1}{8}+\epsilon)\log^{2} q}.$$ The average value of ${\mathbf{T}}$ is much larger [@EJS]. Igor Shparlinski and Alina Ostafe proposed the challenging problem of estimating the average value of ${\mathbf{T}}$ in the much harder case of degree $d$ polynomials and rational functions. We require that $d$ be less than $q$, because otherwise all $q^{q}$ mappings can be realized by Lagrange interpolation as degree $d$ polynomials, and we are back to random mappings. Let $\Omega(q,d)$ be the set of all $q^{d}(q-1)$ polynomials of degree $d$ with coefficients in the finite field ${\mathbb F}_{q}$. Let ${\mathbb{M}}={\mathbb{M}}_{q,d}$ be the uniform probability measure on $\Omega(q,d)$; for all $A\subseteq \Omega(q,d)$, ${\mathbb{M}}(A)=\frac{|A|}{q^{d}(q-1)}.$ Then ${{\mathbf{T}}}$, $\log {\mathbf{T}},$ ${\mathbf{Z}}$, and any other quantities of interest may be regarded as a random variables on $\Omega(q,d)$, and the theorems of combinatorial probability are applicable. In particular, by grouping together elements with the same period, we see that the average value of ${\mathbf{T}}$ is just the expected value ${\mathbb{E}_{\mathbb M}}({\mathbf{T}})$: $$\frac{1}{q^{d}(q-1)}\sum\limits_{f\in \Omega(q,d)}{\mathbf{T}}(f) =\sum\limits_{j\geq 0}j{\mathbb{M}}(\lbrace f: {\mathbf{T}}(f)=j\rbrace )= {\mathbb{E}_{\mathbb M}}({\mathbf{T}}).$$ Here we prove, using elementary methods, that $$\frac{1}{q^{d}(q-1)}\sum\limits_{f\in \Omega(q,d)}\log({\mathbf{T}}(f)) \geq \sum^{*}\limits_{p\leq \frac{d}{2}} \frac{\log p}{p}\prod\limits_{j=0}^{p-1}(1-\frac{j}{q}) - \sum\limits^{*}_{p\leq \frac{d}{2}} \frac{\log p}{2p^{2}} \prod\limits_{j=0}^{2p-1}(1-\frac{j}{q}) ,$$ where the $*$ indicates that the sum is restricted to prime numbers. As one consequence, we deduce the following crude lower bound for the average value of ${\mathbf{T}}$ as an incomplete answer to Shparlinski and Ostafe’s question. If $d=d(q)\rightarrow \infty$ but $d=o(\sqrt{q})$, then $$\frac{1}{q^{d}(q-1)}\sum\limits_{f\in \Omega(q,d)}{\mathbf{T}}(f) \geq (1+o(1)) \frac{d}{2}.$$ In some ways, it is easier to work with random mappings than with random polynomials. In both Harris’ proof [@Harris], and in subsequent work on the average period [@EJS], essential use was made of the following observation about classical random mappings and conditional probabilities: given the set of cyclic vertices, all permutations of those vertices are equally likely to occur as the restriction of $f.$ A more formal statement in terms of conditional probabilities is (Folklore) \[uniform\] Suppose ${\mathbb M}^{*}$ is the uniform probability measure on the $|V|^{|V|}$ functions $f:V\rightarrow V$. Then, for all non-empty $A\subseteq V$ and all permutations $\sigma$ of $A$, $${\mathbb M}^{*}\left(\sigma_{f}=\sigma\, |\, \text{{\sc cyclic}}(f)=A\right) : = \frac{{\mathbb M}^{*}\left(\sigma_{f}=\sigma\right)}{ {\mathbb{M}}^{*}(\text{\sc cyclic} (f)=A ) }=\frac{1}{|A|!}.$$ However, a simple counting argument shows that the analogue of Observation \[uniform\] cannot be true in general for $\Omega(d,q)$. Consider, for example, the case where $A$ is all of ${\mathbb{F}}_{q}$. If it were true that ${\mathbb{M}}\left( \sigma_{f}=\sigma\, |\, \text{{\sc cyclic}}(f)=A\right) =\frac{1}{|A|!}>0$ for all $\sigma$ in $\text{Sym}(A)$, then, for each of the $|A|! $ permutations $\sigma$, $\Omega(q,d)$ would contain at least one element $f$ for which $\sigma_{f}=\sigma$. This implies that $|\Omega(q,d)|\geq |A|! .$ But if $A$ is all of ${{\mathbb{F}}}_{q}$, and if $d=o(q)$, then $|\Omega(q,d)|< |A|! $ for all sufficiently large prime powers $q$. Nevertheless, $\sigma_{f}$ is always a permutation of [cyclic]{}$(f).$ By an old theorem of Landau, [@Landau], [@N1], the maximum order that a permutation of an $m$-element set can have is $e^{\sqrt{m\log m}(1+\epsilon(m))}$, where $\epsilon(m)\rightarrow 0$. However we do not know enough about the distribution of ${\mathbf Z}$ to draw any conclusions about $ {\mathbb{E}_{\mathbb M}}({\mathbf{T}})$. To date, we have only the trivial upper bound $ {\mathbb{E}_{\mathbb M}}({\mathbf{T}})\leq e^{\sqrt{q\log q}(1+o(1))}$ that follows from ${\mathbf Z}\leq q.$ Small Cycle Lengths. ==================== Flynn and Garton[@FG] used Lagrange interpolation to calculate the expected number of cycles of length $k$. In this section, Flynn and Garton’s methods are combined with the method of factorial moments to determine the asymptotic behaviour of the cycle length multiplicities $c_{1},c_{2},\dots ,c_{b}$ for fixed $b$. As we shall see, the cycle lengths are not independent, but they are asymptotically independent as $q\rightarrow \infty$. They behave very much like the cycle lengths of random permutations and random mappings. \[decomp\] Let $A\subset {\mathbb{F}}_{q},$ and let $\sigma$ be a permutation of $A$. If $d\geq |A|,$ then there are exactly $q^{d-|A|}(q-1)$ polynomials in ${\mathbb{F}}_{q}[x]$ of degree $d$ that extend $\sigma$ to ${\mathbb{F}}_{q}$. Let ${\cal E}_{d}$ consist of those polynomials $f$ of degree $d$ that extend $\sigma$, i.e. such that $f(x)=\sigma(x)$ for all $x\in A$. Also let ${\cal K}_{d}$ be the set of polynomials of degree $d$ such that $f(x)=0$ for all $x\in A$. By the Lagrange interpolation theorem, $f\in {\cal E}_{d}$ if and only if $$f(x)=L(x)+k(x),$$ where $L$ is the minimum degree interpolating polynomial (whose degree is less than $d$), and $k(x)\in {\cal K}_{d}.$ Since ${\mathbb F}_{q}[x]$ is a Euclidean domain, a polynomial of degree $d$ is in ${\cal K}_{d}$ if and only if it is divisible by $g(x)=\prod\limits_{a\in A}(x-a).$ Thus $k\in {\cal K}_{d}$ if and only if there is a polynomial $h$ of degree $d- |A| $ such that $k(x)=g(x)h(x)$. The number of ways to choose the polynomial $h$ is $q^{d-|A|} (q-1)$. Therefore $|{\cal E}_{d}|=|{\cal K}_{d}|=q^{d-|A|}(q-1)$. Before proving the next theorem, we review some basic facts from probability. Many readers will be familiar with the Method of Moments. A glib summary is that, to prove that a sequence of random variables $X_{n}$ converges to $X$, it suffices (under mild conditions) to prove that, for each $r$, the expected value of $X_{n}^{r}$ converges to the expected value of $X^{r}.$ For multivariable extensions, it is necessary to consider cross moments: to prove that $(X_{n},Y_{n})$ converges to $(X,Y)$, one proves that ${\mathbb E}(X_{n}^{r_{1}}X_{n}^{r_{2}})$ converges to ${\mathbb E}(X^{r_{1}}Y^{r_{2}})$ for non-negative integers $r_{1},r_{2}$. Section 30 of Billingsley[@PB1] contains a more precise and detailed presentation of the Method of Moments, together with a number-theoretic application. Note that each of the powers $X^{r}$ can be written as a linear combination of falling the falling factorials $$(X)_{s}:=\prod\limits_{j=0}^{s-1}(X-j), s=1,2,\dots ,r$$ and vice versa. Hence there is a corresponding Method of Factorial Moments; convergence of the factorial moments implies convergence of the moments, and vice versa. This method is especially convenient for arithmetic and enumerative applications because $(X)_{r} $ occurs naturally as the number of injective functions from an $r$-element set to an $X$-element set. It is what George Andrews ([@Andrews], page 32) refers to as the number of $r$-permutations of an $X$ element set. If $X_{1}$, $X_{2}$ have a Poisson distributions with parameters $\lambda_{1}, \lambda_{2}$, i.e. if $\Pr(X_{i}=m) = e^{-\lambda_{i}}\frac{\lambda_{i}^{m}}{m!}, i=1,2; m=0,1,2,\dots,$ then for any $r$ the $r$’th factorial moment of $X_{i}$ is $\lambda_{i}^{r}$, i.e. ${\mathbb E}((X)_{r})=\lambda_{i}^{r}.$ With the additional hypothesis that $X_{1}$ and $X_{2}$ are independent, the falling factorials are also independent and the expectation of their product is the product of their expectations: ${\mathbb E}((X)_{r_{1}}(X)_{r_{2}})= \lambda_{1}^{r_{1}}\lambda_{2}^{r_{2}}$. Section 6.1 of [@JLR] is directly relevant to our application of the Method of Factorial Moments. It includes (bottom of page 144) a generalization of Lemma \[RGlemma\] below. Before stating the lemma, we fix some notation. If $\sigma$ is a permutation of a subset of ${\mathbb F}_{q}$, define an indicator(i.e. characteristic function) by $I_{\sigma}(f)=1$ if $\sigma$ is the restriction of $f$ to the set of elements that $\sigma$ permutes, and $I_{\sigma}(f)=0$ otherwise. Let ${\cal Z}_{k}$ be the set of all $k$-cycles that can be formed using elements of ${\mathbb F}_{q}$. Then $$c_{k}(f)=\sum\limits_{C\in {\cal Z}_{k}}I_{C}(f)$$ is the number of cycles of length $k$ that $f$ has. Also note that a product of indicators is itself an indicator with a concrete interpretation: $ I_{C_{1}}I_{C_{2}}\cdots I_{C_{r}}(f)$ is one if and only if the $r$ cycles $C_{1},C_{2}, \dots , C_{r}$ are all cycles of $f$, i.e $ I_{C_{1}}I_{C_{2}}\cdots I_{C_{r}}(f)=I_{\sigma}(f),$ where $\sigma$ is the permutation having cycles $C_{1}, C_{2}, \dots ,{C_{r}}.$ \[RGlemma\] With the notation $I_{C}(f)=1$ if $C$ is a $k$-cycle for $f$, and zero otherwise, we have $$c_{k}(c_{k}-1)\cdots (c_{k}-r+1)=\sum I_{C_{1}}I_{C_{2}}\dots I_{C_{r}},$$ where the sum is over all sequences of $r$ disjoint $k$-cycles. In other words, there are exactly $(c_{k}(f))_{r} $ ways to pick a sequence of $r$ disjoint $k$ cycles of $f$. \[identity\] Recall that ${\mathbb M}(A)=\frac{|A|}{q^{d}(q-1)}$ for all sets $A$ of degree $d$ polynomials. \[smallcyclesthm\] If $d=d(q)\rightarrow\infty$, then for any fixed $b$ the random variables $c_{k}$, $ k=1,2,\dots , b,$ are asymptotically independent Poisson($1/k$). In other words, for for any non negative integers $m_{1},m_{2},\dots , m_{b}$, $$\lim\limits_{q\rightarrow\infty} {\mathbb M}(c_{k}=m_{k},1\leq k\leq b)=\prod\limits_{k=1}^{b}e^{-\frac{1}{k}}\frac{1}{m_{k}!k^{m_{k}}}.$$ .2cm[**Comment:**]{} The conclusion might well be true when $d$ is fixed. However the hypothesis $d(q)\rightarrow\infty$ is needed for our proof because of the way Lagrange interpolation is used. .2cm By \[Theorem 6.10,13\] (or \[Theorem 21,6\]), it suffices to show that the joint factorial moments ${ {\mathbb{E}_{\mathbb M}}}\left((c_{1})_{r_{1}}(c_{2})_{r_{2}}\cdots (c_{b})_{r_{b}}\right)$ converge to those of the corresponding independent Poisson distributions. In other words, it suffices to check that, for any choice of $r_{1},r_{2},\dots ,r_{b}$, $$\begin{aligned} \lim\limits_{q\rightarrow\infty} {\mathbb {\mathbb{E}_{\mathbb M}}}\left( \prod\limits_{k=1}^{b} (c_{k})_{r_{k}}\right) &=\prod\limits_{k=1}^{b}(r_{k}\text{'th factorial moment of Poisson(1/k) random var.}) \\ &= \prod\limits_{k=1}^{b} \frac{1}{k^{r_{k}}}.\end{aligned}$$ Let $\Lambda_{\vec{r}}$ denote a given choice of an integer partition with $r_{k}$ parts of size $k, $ and let $m=\sum_{k}kr_{k}$ be the number that $\Lambda_{\vec{r}}$ partitions. In the product $ \prod\limits_{k=1}^{b} (c_{k})_{r_{k}}$, we can apply Lemma \[identity\] to each factor $ (c_{k})_{r_{k}}$: $$\label{b4avg} \prod\limits_{k=1}^{b} (c_{k})_{r_{k}}= \prod\limits_{k=1}^{b}( \sum I_{C_{k,1}}I_{C_{k,2}}\cdots I_{C_{k,r_{k}}}).$$ If we now expand the right hand side, there are an unpleasantly large number of terms. On the other hand, each term is nothing more than a product of indicators. We have $I_{C}I_{C^{'}}=0$ whenever two cycles $C$,$C'$ are not disjoint. Therefore, when (\[b4avg\]) is expanded, each non-zero term in the sum corresponds to a permutation $\sigma$ of of $m$ field elements having type $\Lambda_{\vec{r}}$ (i.e. having $r_{k}$ cycles of length $k$ for $1\leq k\leq b$). Because there are $r_{k}!$ possible ways to order the cycles of length $k$, the indicator $ I_{\sigma}=\prod\limits_{k=1}^{b}\prod\limits_{j=1}^{r_{k}}I_{C_{k,j} } $ occurs $\prod\limits_{k}r_{k}!$ times. Combining the terms with same permutation we get $$\label{bypartition} \prod\limits_{k=1}^{b}( \sum I_{C_{k,1}}I_{C_{k,2}}\cdots I_{C_{k,r_{k}}})=(\prod\limits_{k}r_{k}!) \sum\limits_{\sigma}I_{\sigma},$$ where the sum is over all permutations of type $\Lambda_{\vec{r}}$ of $m$ elements of ${\mathbb F}_{q}$. Averaging over all polynomials in $\Omega(q,d)$, and using the fact that expectation ${\mathbb E}(-)$ is linear (with no assumption of independence), we get $$\label{postavg} {\mathbb{E}_{\mathbb M}}( \prod\limits_{k=1}^{b} (c_{k})_{r_{k}})=(\prod\limits_{k}r_{k}!) \sum\limits_{\sigma}{\mathbb{E}_{\mathbb M}}(I_{\sigma}).$$ Because $d(q)\rightarrow\infty$, whereas $\sigma$ is fixed permutation, we can invoke Lemma \[decomp\] to get $${\mathbb{E}_{\mathbb M}}(I_{\sigma})= {\mathbb{M}}(\lbrace f \text{ extends }\ \sigma\rbrace )=\frac{q^{d-m}(q-1)}{q^{d}(q-1)}=q^{-m}$$ for all sufficiently large $q$. Thus, in (\[postavg\]), all the terms in the sum are equal. There are ${q\choose m}$ ways to choose the $m$ elements that $\sigma$ permutes. By a well known theorem of Cauchy (see, for example, page 18 proposition 1.3.2 of Stanley [@Stanleyvol1]), an $m$-element set has exactly $\frac{m!}{\prod\limits_{\ell} \ell^{r_{\ell}}r_{\ell}!}$ permutations with the given partition $\Lambda_{\vec{r}}$ Therefore, for all sufficiently large $q$, $${\mathbb{E}_{\mathbb M}}( \prod\limits_{k=1}^{b} (c_{k})_{r_{k}})= \prod\limits_{k}r_{k}!\cdot {q\choose m}\frac{m!}{\prod\limits_{\ell} \ell^{r_{\ell}}r_{\ell}!}q^{-m}= \frac{ (q)_{m}}{q^{m}} \prod\limits_{k}\frac{1}{ k^{r_{k}}}.$$ Because $m$ is fixed, $$\lim\limits_{q\rightarrow \infty} \frac{ (q)_{m}}{q^{m}}=\lim\limits_{q\rightarrow \infty} \prod\limits_{j=0}^{m-1}(1-\frac{j}{q})=1.$$ Therefore $$\lim\limits_{q\rightarrow \infty} {\mathbb{E}_{\mathbb M}}( \prod\limits_{k=1}^{b} (c_{k})_{r_{k}})= \prod\limits_{k=1}^{b}\frac{1}{ k^{r_{k}}}$$ as was to be shown. Bounds for $ {\mathbb{E}_{\mathbb M}}(\log {\mathbf{T}}) $, ${\mathbb{E}_{\mathbb M}}({\mathbf{T}})$ , and the Typical Period. ============================================================================================================================== Recall that, for $f\in \Omega(q,d)$, ${\mathbf{T}}(f)$ is the least common multiple of the cycle lengths, and $c_{k}$ is the number of cycles of length $k$. The goal in this section is to prove lower bounds for the expected values of ${\mathbf{T}}$ and $ \log {\mathbf{T}}$ as well as a lower bound for ${\mathbf{T}}$ that holds with asymptotic probability one. For any any choice of $\xi$, the period ${\mathbf{T}}$ is at least as large as the product (without multiplicity) of prime cycle lengths in the interval $[2,\xi]$. Define $\delta_{k}(f)=1$ if $c_{k}>0$, and $\delta_{k}(f)=0$ otherwise. Then $$\label{primeproduct} {\mathbf{T}}(f)\geq \prod\limits_{p\leq \xi }^{*}p^{\delta_{p}},$$ where the $*$ indicates that the product is restricted to primes. Since $e^{x}>x$ for all $x\geq 0$, it is clear that ${\mathbf{T}}(f) =e^{\log {\mathbf{T}}(f)} \geq \log {\mathbf{T}}(f).$ Combining this with (\[primeproduct\]), and averaging over degree $d$ polynomials, we get $$\label{jensen} {\mathbb{E}_{\mathbb M}}({\mathbf{T}})\geq {\mathbb{E}_{\mathbb M}}(\log {\mathbf{T}})\geq \sum\limits_{p\leq \xi}^{*}{\mathbb{E}_{\mathbb M}}(\delta_{p})\log p.$$ We postpone the choice of $\xi$, but note that the bound can only improve if $\xi$ is increased. However our ability estimate $ {\mathbb{E}_{\mathbb M}}(\delta_{p})$ will depend on $d$ being larger than $2\xi$, so the price for a better estimate is that the polynomials must have larger degree. In order to calculate ${\mathbb{E}_{\mathbb M}}(\delta_{p}),$ we need to calculate the cardinality of ${\cal Z}_{p}$, the set of all possible $p$-cycles. There are ${q\choose p} $ ways to choose $p$ elements of ${\mathbb F}_{q}$ to form a $p$-cycle. As a special case of the aforementioned theorem of Cauchy, there $(p-1)!$ ways to form a $p$-cycle from $p$ elements (There are $p!$ ways to write down the $p$ elements in a cycle. This overcounts by a factor $p$ because, calculating subscripts$\mod p$, we have $(x_{0},x_{1},\dots ,x_{p-1})= (x_{i},x_{i+1},\dots , x_{i+p})$ for $0\leq i<p$ ). Therefore $$\label{Zpcount} |{\cal Z}_{p}|= (p-1)! {q\choose p}.$$ To estimate the quantity ${\mathbb{E}_{\mathbb M}}(\delta_{p})$ in equation (\[jensen\]), note that $|\Omega(q,d)|{\mathbb{E}_{\mathbb M}}(\delta_{p})$ is the number of polynomials in $\Omega(q,d)$ that have at least one cycle of length $p$. This number can be estimated using inclusion-exclusion and Bonferroni inequalities (See, for example, equation (7) page 66 of [@Stanleyvol1]. For each possible $p$-cycle $C$, having $C$ as a cycle is a property that a polynomial may have.) One option is to use cardinalities of sets in the formulae, and divide by $|\Omega(q,d)|$ at the end to get ${\mathbb{E}_{\mathbb M}}(\delta_{p})$. Another equivalent option is to work directly with the probabilities as weights.) If $C$ is a cycle, let $A_{C}$ be the event that $f$ has $C$ as a cycle; $A_{C}=\lbrace f\in \Omega(q,d): I_{C}(f)=1\rbrace.$ Also define $$S_{r}=S_{r}(p)=\sum\limits_{\lbrace C_{1},C_{2},\dots ,C_{r}\rbrace }{\mathbb{M}}\!\left(\bigcap_{i=1}^{r}A_{C_{i}} \right)= \sum\limits_{\lbrace C_{1},C_{2},\dots ,C_{r}\rbrace}{\mathbb{E}_{\mathbb M}}(\prod\limits_{i=1}^{r}I_{C_{i}}),$$ where the sums are over all [*unordered*]{} $r$ element subsets of $ {\cal Z}_{p}.$ If $p$ is prime, then by inclusion-exclusion and the alternating inequalities(see, for example, page 91 of [@Stanleyvol1]), we have $ {\mathbb{E}_{\mathbb M}}(\delta_{p})=\sum\limits_{j\geq 1}(-1)^{j+1}S_{j},$ and for any $m$, $$\sum\limits_{j=1}^{2m}(-1)^{j+1}S_{j}\leq {\mathbb{E}_{\mathbb M}}(\delta_{p})\leq \sum\limits_{j=1}^{2m-1}(-1)^{j+1}S_{j} .$$ In particular, with $m=1$ we get a convenient lower bound: $$\label{EdeltaBounds} S_{1}-S_{2}\leq {\mathbb{E}_{\mathbb M}}(\delta_{p}) \leq S_{1}.$$ So long as $d\geq p$, we can apply Lemma \[decomp\] to each cycle $C$ in $ {\cal Z}_{p}:$ $${\mathbb{M}}\!\left(A_{C}\right)=\frac{q^{d-p}(q-1)}{q^{d}(q-1)} =q^{-p}.$$ Combining this with our formula (\[Zpcount\]) for the cardinality of ${\cal Z}_{p}$, we get an exact formula for $S_{1}:$ $$\label{S1id} S_{1}=\sum\limits_{C\in {\cal Z}_{p}} q^{-p}= {q\choose p}\frac{(p-1)!}{q^{p}}= \frac{1}{p} \frac{(q)_{p}}{q^{p}}= \frac{1}{p}\prod\limits_{j=0}^{p-1}(1-\frac{j}{q}).$$ Similarly, if $d\geq 2p$, we can apply Lemma \[decomp\] to any permutation that consists of two disjoint $p$-cycles. (As before, intersecting cycles contribute zero). Therefore $S_{2}=\sum\limits_{\lbrace C_{1},C_{2}\rbrace}q^{-2p}$, where the sum is restricted to disjoint pairs of cycles. For the number of ways to choose a set of two disjoint $p$-cycles, we again have ${q\choose 2p}$ choices of field elements to permute. Again, by Cauchy’s theorem, there are $\frac{(2p)!}{2p^{2}}$ permutations of of those elements that have two $p-$cycles. Thus $$\label{S2id} S_{2}= \frac{(2p)!}{2p^{2}}{q\choose 2p}q^{-2p}=\frac{1}{2p^{2}}{\prod\limits_{j=0}^{2p-1}}(1-\frac{j}{q}).$$ Putting (\[S1id\]) and (\[S2id\]) into (\[EdeltaBounds\]) and then (\[jensen\]), we get \[ETtheorem\] If $d\geq 2\xi ,$ then $${\mathbb{E}_{\mathbb M}}(\log {\mathbf{T}}) \geq \sum^{*}\limits_{p\leq \xi} \frac{\log p}{p}{\prod\limits_{j=0}^{p-1}}(1-\frac{j}{q}) - \sum\limits^{*}_{p\leq \xi} \frac{\log p}{2p^{2}} \prod\limits_{j=0}^{2p-1}(1-\frac{j}{q}) ,$$ where the $*$ indicates that the sum is restricted to prime numbers. Clearly the second sum in Theorem \[ETtheorem\] is bounded by an absolute constant, regardless of how large $d$ and $\xi$ are, since the unrestricted series $\sum\limits_{n=1}^{\infty}\frac{\log n}{n^{2}}$ is convergent. For the main term, one can make various asymptotic approximations, depending on the choice of $d$ and $\xi$. For example, if $d=d(q)\rightarrow \infty$ but $d =o(\sqrt{q})$, then we can choose $\xi=\frac{d}{2}$ and use the approximation $\log(1-u)=O(u)$ to simplify the product: $$\label{logapprox} \exp(\sum\limits_{j=1}^{p-1}\log(1-\frac{j}{q})) =\exp( \sum\limits_{j=1}^{p-1}O(\frac{j}{q})) =e^{o(1)}=1+o(1).$$ It is well known fact ([@Hua], page 89), reportedly due to Mertens, that $\sum\limits^{*}_{p\leq \xi}\frac{\log p}{p}=\xi+O(1)$ as $\xi\rightarrow\infty.$ Therefore we have \[logcor\] If $d=d(q)\rightarrow \infty$, but $d =o(\sqrt{q})$,then $${\mathbb{E}_{\mathbb M}}(\log {\mathbf{T}})\geq (1+o(1))\frac{d}{2}.$$ .5cm Corollary \[logcor\] does not necessarily mean that most polynomials have period greater $e^{(\frac{1}{2}-\epsilon)d}$. Without more information, one cannot even rule out the existence of a constant upper bound $\kappa$ such that ${\mathbf{T}}\leq \kappa$ for all but $o(q^{d}(q-1))$ polynomials in $\Omega(q,d)$. However the next proposition shows that, in fact, most polynomials have order larger than $\sqrt{\frac{d}{2}}.$ .2cm If $d=d(q)\rightarrow \infty$ but $d =o(\sqrt{q})$, then all but $o(q^{d}(q-1)$) polynomials in $\Omega(q,d)$ have at least one cycle whose length is in the interval $J=[\beta ,\beta^{2}]$, where $ \beta(q)=\sqrt{\frac{d}{2}}.$ .2cm For $f\in \Omega(q,d)$, let ${\mathbf{N}}(f)=\sum\limits_{k\in J}c_{k}$ be the number of cycles of $f$ that have length in $J$. The goal is to show ${\mathbb{M}}({\mathbf{N}}=0)=o(1)$. If ${\mathbf{N}}=0$, then obviously $|{\mathbf{N}}-\mu|\geq \mu$ for any real number $\mu$. Therefore $$\label{setupChebyshev} {\mathbb{M}}({\mathbf{N}}=0)\leq {\mathbb{M}}( |{\mathbf{N}}-\mu|\geq \mu).$$ A standard approach is to set $\mu:={\mathbb{E}_{\mathbb M}}({\mathbf{N}}),$ and $\sigma^{2}:={\mathbb{E}_{\mathbb M}}( ({\mathbf{N}}-\mu)^{2}),$ and use Chebyshev’s inequality(see, for example, page 75 of \[5\]) to show that the right side of (\[setupChebyshev\]) approaches zero. By Chebyshev’s inequality, $$\label{chebsyshev} {\mathbb{M}}(|{\mathbf{N}}-\mu|\geq \mu) \leq \frac{\sigma^{2}}{\mu^{2}}.$$ It therefore suffices to show that $ \sigma^{2}=o(\mu^{2}).$ By elementary algebra, $$\label{algebra} ({\mathbf{N}}-\mu)^{2}={\mathbf{N}}({\mathbf{N}}-1)+(1-2\mu){\mathbf{N}}+ \mu^{2}.$$ Now average over polynomials in $\Omega(q,d)$; take the expected value of both sides of (\[algebra\]), to get $$\label{variance} \sigma^{2}={\mathbb{E}_{\mathbb M}}(({\mathbf{N}})_{2})+(1-2\mu)\mu + \mu^{2} ={\mathbb{E}_{\mathbb M}}(({\mathbf{N}})_{2})+\mu -\mu^{2}.$$ To calculate $\mu$, use Lemma \[decomp\] just like before in (\[S1id\]). (See also [@FG]). We chose $\beta$ so that $d=2\beta^{2}\geq k$, therefore Lemma \[decomp\] applies. If ${\cal Z}_{k}$ denotes the set of all possible $k$-cycles, then ${\mathbf{N}}=\sum\limits_{k\in J}\sum\limits_{C\in {\cal Z}_{k}}I_{C}, $ and $$\mu= \sum\limits_{k\in J}\sum\limits_{C\in {\cal Z}_{k}}{\mathbb{E}_{\mathbb M}}(I_{C}) = \sum\limits_{k\in J}|{\cal Z}_{k}| q^{-k} = \sum\limits_{k\in J}\frac{(q)_{k}}{q^{k}}\frac{1}{k}.$$ Since $k\leq d=o(\sqrt{q})$, we get $\frac{(q)_{k}}{q^{k}}= \prod\limits_{j=0}^{k-1}(1-\frac{j}{q})=1+o(1)$, just as in (\[logapprox\]). Thus $$\mu =(1+o(1))\sum\limits_{k=\beta}^{\beta^{2}}\frac{1}{k}=(1+o(1))\log \beta.$$ As in Lemma \[RGlemma\], we have $({\mathbf{N}})_{2}=\sum\limits_{C_{1},C_{2}}I_{C_{1}}I_{C_{2}}$ where the sum is over distinct pairs of disjoint cycles $C_{1},C_{2}$ whose lengths lie in $J$. If $C_{1}$ and $C_{2}$ are disjoint cycles of lengths $k_{1},k_{2}$ respectively, then by Lemma \[decomp\], ${\mathbb{E}_{\mathbb M}}(I_{C_{1}}I_{C_{2}})= q^{-k_{1}-k_{2}}$. (We chose $\beta$ so that $d=2\beta^{2}\geq k_{1}+k_{2}$.) The number of ways to choose an ordered pair of disjoint cycles of length $k_{1}$ and $k_{2}$ is $\frac{q!}{k_{1}!k_{2}!(q-k_{1}-k_{2})!}(k_{1}-1)!(k_{2}-1)!= \frac{(q)_{k_{1}+k_{2}}}{k_{1}k_{2}}$ just as in the proof of Theorem \[smallcyclesthm\]. Hence $${\mathbb{E}_{\mathbb M}}(({\mathbf{N}})_{2})= \sum\limits_{k_{1},k_{2}\in J} \frac{(q)_{k_{1}+k_{2}}}{ q^{k_{1}+k_{2}}} \frac{1}{k_{1}k_{2} }.$$ Because $\beta^{2}=o(\sqrt{q})$, we have $ \frac{(q)_{k_{1}+k_{2}}}{ q^{k_{1}+k_{2}}} =1+o(1)$ for all $k_{1},k_{2}\in J$. Therefore $${\mathbb{E}_{\mathbb M}}(({\mathbf{N}})_{2})=(1+o(1))\sum\limits_{k_{1},k_{2}\in J}\frac{1}{k_{1}k_{2}} =(1+o(1))(\log\beta)^{2} = \mu^{2}(1+o(1)).$$ Putting this back into (\[variance\]), we get $ \sigma^{2}=o(\mu^{2}).$ By (\[chebsyshev\]), this completes the proof. Rational Functions ================== If $f$ and $g$ are polynomials in ${\mathbb F}_{q}[x]$, let $\rho(g)=$ the degree of $g$ and let $mgcd(f,g)$ be the greatest monic common divisor of $f$ and $g$. Let $$U(q,d)=\lbrace (f,g): \rho(f)=\rho(g)=d \text{ and } mgcd(f,g)=1 \text{ and } g \text{ is monic}\rbrace.$$ It is known [@BenBen], [@Wilf] that $$\label{wilf}|U(q,d)|=q^{2d+1}(1-\frac{1}{q})^{2}.$$ For each pair $(f,g)\in U(q,d)$, the rational function $R(x)=\frac{f(x)}{g(x)}$ can be regarded as a function from ${\mathbb{F}}_{q} \cup\lbrace\infty\rbrace$ to ${\mathbb{F}}_{q}\cup\lbrace\infty\rbrace$ with the convention that $R(x)=\infty$ whenever $g(x)=0$ and $R(\infty)=$ the leading coefficient of $f$. Note that, because of the conditions on $f$ and $g$, the polynomials $f$ and $g$ are uniquely determined by the rational function $R$. Let ${\mathbb{P}}_{d}$ be the uniform probability measure on $U(q,d)$, and let ${\mathbb E}_{d}$ denote the corresponding expectation. Since $F_{q}\cup\lbrace\infty\rbrace$ is finite, the ultimate period is well-defined. Rather than introduce new notation, we reuse the same notation ${\mathbf{T}}$ for the ultimate period; in this section ${\mathbf{T}}$ defined on $U(q,d)$ instead of $\Omega(q,d)$. Our goal in this section is to prove a lower bound for ${\mathbb{E}}_{d}(\log {\mathbf{T}})$. To avoid dealing with the exceptional point $\infty$, define $\hat{c}_{k}$ to be the number of cycles of length $k$ that do not include $\infty$. Also let $\hat{\delta}_{k}=1$ if $\hat{c}_{k}>0$, $\hat{\delta}_{k}=0$ otherwise. With this notation, ${\mathbf{T}}$ may be strictly larger than the least common multiple of the $k$’s for which $\hat{c}_{k}>0$. (It is possible that the only cycle of length $k$ happens to contain $\infty$.) That poses no difficulties because we are proving lower bounds. As before, we have $$\label{primeproduct2} {\mathbf{T}}(f)\geq \prod\limits_{p\leq \zeta }^{*}p^{\hat{\delta}_{p}},$$ where the $*$ indicates that the product is restricted to primes, and $\zeta$ is a parameter to be specified later. Averaging over rational functions in $U(q,d)$, we get $$\label{Rjensen} {\mathbb{E}}_{d}({\mathbf{T}})\geq {\mathbb{E}}_{d}(\log {\mathbf{T}})\geq \sum\limits_{p\leq \zeta }^{*}{\mathbb{E}}_{d}(\hat{\delta}_{p})\log p,$$ where the choice of the parameter $\zeta$ is postponed. Next we prove a rational function analogue of lemma \[decomp\]. \[decomp2\] Let $A\subset {\mathbb{F}}_{q} $, and let $\sigma$ be a permutation of $A$. If $d\geq 2|A|,$ then $${\mathbb{P}}_{d}(\lbrace (f,g)\in U(q,d): \frac{f}{g} \text{ extends } \sigma\rbrace )= q^{-|A|}\!\left(1+O\!\left(\frac{|A|}{q}\right)\right).$$ From the definitions we have $${\mathbb{P}}_{d}(\lbrace (f,g) \in U(q,d): \frac{f}{g} \text{ extends } \sigma\rbrace )=$$ $$\label{unrelaxed} \frac{1}{|U(q,d)| }\sum\limits_{g}|\lbrace f: \frac{f}{g}\in U(q,d) \text{ and } \frac{f(x)}{g(x)}=\sigma(x) \text{ for all } x\in A \rbrace |,$$ where the sum is over monic degree $d$ polynomials $g$ that have no roots in $A$. As an auxiliary device, consider a superset of $U(q,d)$ in which the coprimality restriction is removed: $$U^{*}(q,d)= \lbrace (f,g): \rho(f)=\rho(g)=d \text{ and } g \text{ is monic}\rbrace.$$ If $(f,g)\in U^{*}(q,d)$, and $mgcd(f,g)=h$ has degree $k$, then $f^{*}:=\frac{f}{h}$ and $g^{*}:=\frac{g}{h}$ are coprime polynomials of degree $d-k$. Define $$\Phi_{1}=\frac{1}{|U(q,d)| }\sum\limits_{g}|\lbrace f: \frac{f}{g}\in U^{*}(d,q) \text{ and } {\frac{f}{g}}=\sigma(x) \text{ for all } x\in A \rbrace |,$$ where, as in (\[unrelaxed\]), the sum is over monic degree $d$ polynomials $g$ that have no roots in $A$. Comparing $\Phi_{1}$ with (\[unrelaxed\]), we see that $\Phi_{1}$ is an overestimate because $f$ is chosen from a larger set. We can therefore write $$\label{sieve} {\mathbb{P}}_{d}(\lbrace R\in U(q,d): R \text{ extends } \sigma\rbrace )= \Phi_{1}-\Phi_{2}-\Phi_{3},$$ where $\Phi_{2}$ is the excess contribution from those pairs in ${U}^{*}(q,d)-U(q,d)$ for which the degree of the mgcd is is larger than one, but not larger than $|A|$, and $\Phi_{3}$ is the remaining excess contribution from pairs with for which the $mgcd$ has degree larger than $|A|$. To fit our equations on one line, abbreviate $U=|U(q,d)|$. Then we can be more explicit: $$\Phi_{2}= \frac{1}{U } \sum\limits_{\lbrace h: 2\leq \rho(h)\leq |A| \rbrace}|\lbrace (f,g): mgcd(f,g)=h,g \text{ monic, w.o. roots in } A, {\frac{f}{g}}\ \text{extends } \sigma\rbrace|,$$ and $$\Phi_{3}= \frac{1}{|U| } \sum\limits_{\lbrace h: \rho(h)> |A| \rbrace}|\lbrace (f,g): mgcd(f,g)=h,g \text{ monic, w.o. roots in } A, {\frac{f}{g}}\ \text{extends } \sigma\rbrace|.$$ First we estimate the main term $\Phi_{1}$. In $\Phi_{1}$, the number of terms in the sum is the number of degree $d$ monic polynomials with no root in $A$. We use inclusion-exclusion, with property $i$ being that $g$ has the $i$’th element of $A$ as a root. For any particular set $\alpha_{1},\alpha_{2}, \dots \alpha_{j}$ of roots, the number of monic degree $d$ polynomials having those roots is the number that can written as $(x-\alpha_{1})(x-\alpha_{2}) \cdots (x-\alpha_{j}) \tilde{g}(x)$ for some monic degree $d-j$ poly $\tilde{g}$, namely $q^{d-j}.$ By inclusion-exclusion, the number of terms in the sum $\Phi_{1}, $ i.e. the number of degree $d$ monic polynomials with no root in $A$, is $$\sum\limits_{j=0}^{|A|}{|A| \choose j}(-1)^{j}q^{d-j} = q^{d}(1-\frac{1}{q})^{|A|}= q^{d}(1+O(\frac{|A|}{q})).$$ Because $d> |A|$, we can use Lemma \[decomp\] again to calculate the terms in the sum $\Phi_{1}$ for a given $g$: $$|\lbrace f: \rho(f)=d \text{ and } f(x)=\sigma(x)g(x) \text{ for all } x\in A \rbrace | = q^{d-|A|}(q-1).$$ Combining our estimates for the size of terms, the number of terms, and the cardinality of $U(q,d)$, we get $\Phi_{1}= \left( \frac{1}{ q^{2d+1}(1-\frac{1}{q})^{2}} \right) \left( q^{d}(1-\frac{1}{q})^{|A|} \right)\left( q^{d-|A|}(q-1) \right).$ Thus $$\label{T1} \Phi_{1} =q^{-|A|}(1+O(\frac{|A|}{q})).$$ Next we show that $\Phi_{2}$ is negligible compared to $\Phi_{1}$ . If the degree of $h= mgcd(f,g)$ is $k$, then $f^{*}:=f/h$ and $g^{*}:=g/h$ have degree $d-k$. Given $k$, there are $q^{k}$ choices of a monic polynomial $h=mgcd(f,g)$, and at most $q^{d-k}$ choices of a monic polynomial $g^{*}$ of degree $d-k$. We have $d\geq 2|A|$ as a hypothesis, and we have $k\leq |A|$ built into the definition of $\Phi_{2}$. Therefore $d-k\geq |A|$ and Lemma \[decomp\] is applicable: there are, for a given $g^{*}$, at most $q^{d-k-|A|}$ choices for $f^{*}$ such that $f^{*}(x)=\sigma(x)g^{*}(x)$ for all $x\in A$. Therefore $$\begin{aligned} \Phi_{2} &\leq \frac{1}{|U(q,d)| } \sum\limits_{k=2}^{|A|}q^{k} q^{d-k} q^{d-k-|A|} \\ &= \frac{q^{2d-|A|}}{ q^{2d+1}(1-\frac{1}{q})^{2}} \sum\limits_{k=2}^{|A|} q^{-k} \\ &= q^{-|A|}O(\frac{1}{q^{3}})=O(\frac{\Phi_{1}}{q^{3}}).\end{aligned}$$ Thus $$\label{T2} \Phi_{2}=O(\frac{\Phi_{1}}{q^{3}}).$$ We can be very crude about estimating $\Phi_{3}$. We get a satisfactory bound by ignoring completely the requirement that $\frac{f^{*}}{g^{*}}$ extends $\sigma$, and $g$ has no roots in $A$. We retain only the requirements that $g$ and $h$ are monic, and that $\rho(f^{*})=\rho(g^{*})=d-\rho(h).$ Thus $$\Phi_{3} \leq \frac{1}{|U(q,d)| } \sum\limits_{\lbrace \text{monic } h: \rho(h)> |A| \rbrace}|\lbrace (f^{*},g^{*}): \rho(f^{*})=\rho(g^{*})=d-\rho(h) , g^{*} \text{\ monic} \rbrace|.$$ Given $k>|A|,$ there are $q^{k}$ choices of a monic polynomial $h$ of degree $k$, $q^{d-k}$ choices of a monic polynomial $q^{*}$ of degree $d-k$, and $q^{d-k}(q-1)=q^{d-k+1}(1-\frac{1}{q})$ choices of a polynomial $f^{*}$ of degree $d-k$. Therefore $$\begin{aligned} \Phi_{3} &\leq \frac{1}{|U(q,d)| } \sum\limits_{k=|A|+1}^{d}q^{k}q^{d-k} q^{d-k+1}(1-\frac{1}{q})\\ &= \frac{1}{q^{2d+1}(1-\frac{1}{q})^{2}}\sum\limits_{k=|A|+1}^{d}q^{2d+1-k}(1-\frac{1}{q}) &= \frac{1}{(1-\frac{1}{q})}\sum\limits_{k=|A|+1}^{d}q^{-k}= O\!\left(\frac{1}{q^{|A|+1}}\right).\end{aligned}$$ Thus $$\label{T3} \Phi_{3}=O\left(\frac{\Phi_{1}}{q}\right).$$ Putting (\[T1\]), (\[T2\]), and (\[T3\]) into (\[sieve\]), we get lemma (\[decomp2\]). With Lemma \[decomp2\] in hand, the proof of a lower bound for ${\mathbb{E}}_{d}(\log {\mathbf{T}})$ is essentially the same as in the preceding section. We need a new notation for the rational function analogues of $S_{1}$ and $S_{2}$. Let $$\hat{S}_{r}=\hat{S}_{r}(p)=\sum\limits_{\lbrace C_{1},C_{2},\dots ,C_{r}\rbrace }{\mathbb P}_{d}\!\left(\bigcap_{i=1}^{r}A_{C_{i}} \right)= \sum\limits_{\lbrace C_{1},C_{2},\dots ,C_{r}\rbrace}{\mathbb E}_{d}(\prod\limits_{i=1}^{r}I_{C_{i}}),$$ (Cycles that contain $\infty$ are not included in this sum. As, before ${\cal Z}_{p}$ is the set of $p$-cycles that can be formed from the elements of ${\mathbb F}_{q}$.) Taking $\sigma$ in Lemma \[decomp2\] to be a $p$-cycle, we can estimate the sum $\hat{S}_{1}.$ If $d\geq 2p$, $$\begin{aligned} \hat{S}_{1}&=\sum\limits_{C\in {\cal Z}_{p}}{\mathbb{P}}_{d}(C\ \text{is a cycle of } R) \\ &= {q\choose p}(p-1)!q^{-p}\left(1+O\!\left(\frac{p}{q}\right)\right).\end{aligned}$$ Thus $$\label{S1Re} \hat{S}_{1} = \frac{1}{p}\prod\limits_{j=0}^{p-1}(1-\frac{j}{q})\left(1+O\!\left(\frac{p}{q}\right)\right)=\frac{1}{p}\prod\limits_{j=0}^{p-1}(1-\frac{j}{q})+O(\frac{1}{q}). $$ Similarly, if $d\geq 4p$, we can apply Lemma \[decomp2\] to any permutation that consists of two $p$-cycles. Therefore $$\begin{aligned} \hat{S}_{2}&=\sum\limits_{\lbrace C_{1},C_{2}\rbrace}{\mathbb P}_{d}(C_{1},C_{2}\text{ are cycles})\\ &= \frac{q(q-1)\cdots (q-2p+1)}{2p^{2}} q^{-2p}\left(1+O\!\left(\frac{p}{q}\right)\right)\end{aligned}$$ After simplification, this is $$\label{S2Re} \hat{S}_{2}=\frac{1}{2p^{2}}{\prod\limits_{j=0}^{2p-1}}(1-\frac{j}{q})+O(\frac{1}{pq}) $$ As in (\[EdeltaBounds\]), we have $$\label{REdeltaBounds} \hat{S}_{1}-\hat{S}_{2}\leq {\mathbb E}_{d}(\hat{\delta}_{p}) \leq \hat{S}_{1}.$$ Putting (\[S1Re\]) and (\[S2Re\]) into (\[REdeltaBounds\]) and then (\[jensen\]), we get $${\mathbb E}_{d}(\log {\mathbf{T}}) \geq \sum^{*}\limits_{p\leq \zeta} \frac{\log p}{p}{\prod\limits_{j=0}^{p-1}}(1-\frac{j}{q}) - \sum\limits^{*}_{p\leq \zeta} \frac{\log p}{2p^{2}}{\prod\limits_{j=0}^{2p-1}}(1-\frac{j}{q}) + O(\sum\limits_{p\leq \zeta}\frac{\log p}{q}).$$ If we take $\zeta=\frac{d}{4}$, and require $d=o(\sqrt{q})$, then $\sum\limits_{p\leq \zeta}\frac{\log p}{q} =o(1)$ and, as in (\[logapprox\]), the products are $1+o(1)$. We therefore have a rational function analogue of Corollary \[logcor\]: If $d=d(q)\rightarrow\infty$ is such a way that $d=o(\sqrt{q})$, then \[ETtheorem2\] $${\mathbb E}_{d}(\log {\mathbf{T}}) \geq \frac{d}{4}(1+o(1)).$$ .5cm[**Acknowledgement:**]{} We thank Igor Shparlinski and BIRS, for providing reference [@FG] and stimulating our interest in this subject. We also thank Derek Garton for helpful comments. A referee’s suggestions helped us improve both the proofs and the exposition. [99]{} , [Number theory]{}, (Corrected reprint of the 1971 original) [Dover Publications, Inc., New York]{}, [1994]{}, ISBN [0-486-68252-8]{} , [Toward a theory of [P]{}ollard’s rho method]{}, [Inform. and Comput.]{}, [*Information and Computation*]{}, [**90**]{}, (1991), [no.2]{}, [139–155]{}. , [Periods of rational maps modulo primes]{}, [*Math. Ann.*]{}, [**355**]{}, [(2013)]{}, [no.2]{},[637–660]{}. , [The probability of relatively prime polynomials]{}, [*Math. Mag.*]{}, [**80**]{}, [2007]{}, [no.3]{}, [196–202]{}. , [Probability and measure]{},second edition, [Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics]{}, [John Wiley & Sons, Inc., New York]{}, [1986]{}, ISBN [0-471-80478-9]{}. , [Random graphs]{}, [Academic Press, Inc. \[Harcourt Brace Jovanovich, Publishers\], London]{}, [1985]{}, ISBN = [0-12-111755-3; 0-12-111756-1]{}. , [A pentagonal number sieve]{}, [*J. Combin. Theory Ser. A*]{}, [**82**]{} ([1998]{}), [no. 2]{}, [186–192]{}. , [Random mapping statistics]{}, in [Advances in cryptology—[EUROCRYPT]{} ’89 ([H]{}outhalen, 1989)]{}, [Lecture Notes in Comput. Sci.]{}, [**434**]{}, [329–354]{}, [Springer, Berlin]{}, [1990]{}. , [Graph components and dynamics over finite fields]{}, [*Int. J. Number Theory*]{}, [**10**]{}, [2014]{}, [no. 3]{}, [779–792]{}. , [Probability distributions related to random mappings]{}, [*Ann. Math. Statist.*]{}, [1960]{}, [1045–1062]{}. , [The asymptotic distribution of the order of elements in symmetric semigroups]{}, [*J. Combinatorial Theory Ser. A*]{}, [**15**]{}, [(1973)]{}, [66–74]{}. , [Introduction to number theory]{}, [Translated from the Chinese by Peter Shiu]{}, [Springer-Verlag, Berlin-New York]{}, [1982]{}. , [Random graphs]{}, [Wiley-Interscience Series in Discrete Mathematics and Optimization]{}, [Wiley-Interscience, New York]{}, [2000]{}, ISBN [0-471-17541-2]{}. , [Random mappings]{}, [Translation Series in Mathematics and Engineering]{} [Optimization Software, Inc., Publications Division, New York]{}, [1986]{},ISBN [0-911575-16-2]{}, , [Handbuch der [L]{}ehre von der [V]{}erteilung der [P]{}rimzahlen. 2 [B]{}ände]{},[2d ed]{}, [Chelsea Publishing Co., New York]{}, [1953]{}. Martins, Rodrigo S. V. and Panario, Daniel, On the Heuristic of Approximating Polynomials over Finite Fields by Random Mappings, arXiv:1505.02983. , [Évaluation asymptotique de l’ordre maximum d’un élément du groupe symétrique]{}, [*Acta Arith.*]{}, [**50**]{}, [1988]{}, no. [3]{}, [221–242]{}. , [A [M]{}onte [C]{}arlo method for factorization]{}, [*Nordisk Tidskr. Informationsbehandling (BIT)*]{}, [**15**]{}, (1975), [no.3]{}, [331–334]{}. , [Period lengths for iterated functions]{}, [*Combin. Probab. Comput.*]{}, [**20**]{}, [2011]{}, [no. 2]{}, [289–298]{}. , [Variation of periods modulo [$p$]{} in arithmetic dynamics]{}, [*New York J. Math.*]{}, [**14**]{}, (2008), [601–616]{}. , [The arithmetic of dynamical systems]{}, [Graduate Texts in Mathematics]{}, [**241**]{}, [Springer, New York]{}, (2007), ISBN [978-0-387-69903-5]{}. , [Enumerative combinatorics. [V]{}ol. [I]{}]{}, [The Wadsworth & Brooks/Cole Mathematics Series]{}, [1986]{} [0-534-06546-5]{}.

ArXiv

--- abstract: | In recent synchrotron radiation facilities, the use of short (picosecond) electron bunches is a powerful method for producing giant pulses of Terahertz Coherent Synchrotron Radiation (THz CSR). Here we report on the first direct observation of these pulse shapes with a few picoseconds resolution, and of their dynamics over a long time. We thus confirm in a very direct way the theories predicting an interplay between two physical processes. Below a critical bunch charge, we observe a train of identical THz pulses (a broadband Terahertz comb) stemming from the shortness of the electron bunches. Above this threshold, a large part of the emission is dominated by drifting structures, which appear through spontaneous self-organization. These challenging single-shot THz recordings are made possible by using a recently developed photonic time stretch detector with a high sensitivity. The experiment has been realized at the SOLEIL storage ring. PACS numbers : 41.60.Ap, 29.27.Bd, 42.65.Re, 07.57.Hm author: - 'C. Evain' - 'E. Roussel' - 'M. Le Parquier' - 'C. Szwaj' - 'M.-A. Tordeux' - 'L. Manceron' - 'J.-B. Brubach' - 'P. Roy' - 'S. Bielawski' title: 'Direct observation of spatio-temporal dynamics of short electron bunches in storage rings ' --- In last generation storage ring facilities, it has become possible to produce pulses of Coherent Synchrotron Radiation (CSR), with powers that exceed “classical” (incoherent) synchrotron radiation by factors greater than 10000 [@Carr2001; @Abo-Bakr2002; @Byrd2002; @Mochihashi2006; @Karantzoulis2010; @Barros2013; @Billinghurst2015], and may be organized in Terahertz combs [@Steinmann2015; @Tammaro2015]. A priviledged way for obtaining this type of emission with the stability required for users, consists in operating the storage ring facilities with short (picosecond) electron bunches [@Abo-Bakr2002; @Abo-Bakr2003; @Mathis2003; @Feikes2011; @Martin2011; @Barros2013; @Billinghurst2015; @Barros2015]. As the main idea, if the longitudinal shape of an electron bunch presents Fourier components at terahertz frequencies, an intense and stable emission of coherent terahertz radiation is generally expected in bending magnets of accelerators, as illustrated in Figure \[fig:setup\]. This simple conceptual picture is nevertheless deeply complicated by a physical phenomenon of very fundamental nature: the interaction of electron bunches with their own emitted radiation leads to a spatiotemporal instability, which is characterized by the spontaneous appearance of microstructures at the millimeter scale [@Stupakov2002; @Venturini2002; @Roussel2014a]. In turn, these structures also emit CSR. Hence the interplay between these two modes of emission (i.e., from the “shortness” of the electron bunches and from the spontaneously formed microstructures) plays a central role in the knowledge and improvement of the present and future CSR sources. However, it is only recently that direct experimental observation of the emitted THz electric field pulse shapes became possible. As an important milestone, electro-optic sampling has become possible in storage rings either for recording CSR pulses [@Katayama2012; @Mueller2012], and electron bunch near-field shapes [@Hiller2014]. Then the so-called photonic time-stretch strategy [@Coppinger1999; @Goda2013] opened the way to direct recordings of CSR electric field pulse shapes, over successive shots at ultra-high repetition rates [@Roussel2015]. Nevertheless, the limited sensitivity of these new high-repetition-rate picosecond detectors limited experimental investigation of THz signals to special conditions of high bunch charge and long electron bunches. This configuration corresponds to highest possible peak powers, but also to strongly unstable operation [@Roussel2014a; @Roussel2015]. Thus, characterization of CSR pulse shape dynamics remained an open question for short (picosecond-scale) and low-charge electron bunches, i.e., the specific conditions where exploiting this fundamental phenomenon, as a THz source, are foreseen. {width="88.00000%"} In this Letter, we present the actual shapes (the electric field evolution, including envelope and carrier) of the Terahertz CSR pulses emitted in short bunch operation in a storage ring. This analysis has been possible by combining two recent advances in Terahertz electric field detection: high-sensitivity electro-optic sampling [@Ahmed2014; @Savolainen2013] for obtaining the required signal-to-noise ratio with picosecond resolution, and the so-called time-stretch strategy [@Coppinger1999; @Goda2013; @Roussel2015], which enables single-shot recordings at ultra-high acquisition rates. As a main result, we present a direct characterization of the transition from stable Terahertz emission to dynamically evolving pulses: the microbunching instablity threshold, which plays a central role in theories of CSR [@Venturini2002; @Stupakov2002]. We also show that numerical integration of the Vlasov-Fokker-Planck equation reproduces the main characteristics of the experimental signals.\ The experiment has been performed at the SOLEIL storage ring in a so-called [*low-alpha*]{} user shift. This mode is characterized by the operation of short bunches (of about 5 ps RMS [@Tordeux2012]), which emit CSR with a stability which is compatible with THz user experiments. Moreover the 209 bunches which circulate in the storage ring have typically different charges \[Fig. \[fig:setup\](b)\]. This allows investigations as a function of bunch charge by simply studying the emission of each individual electron bunches, without change of the storage ring operating conditions. \]. For clarity, the signals are shifted vertically by 0.1 V.[]{data-label="fig:shots"}](fig2.pdf){width="1\linewidth"} {width="\textwidth"} The THz CSR pulses are recorded on the AILES beamline, using a photonic time-stretch electro-optic sampling setup, which is specially designed for high sensitivity \[Fig. \[fig:setup\](c)\]. This detection system is composed of two parts, which are partly intertwined (see also Ref. [@Szwaj2016] for details). The first part is a single-shot EOS [@Jiang1998; @Wilke2002] system based on the design developped at SLS [@Mueller2012], FLASH [@Steffen2009] and ANKA [@Hiller2014], which imprints the THz electric evolution onto the longitudinal profile of a chirped laser pulse at 1030 nm. As a key point for reaching the required high sensitivity, we follow the special strategy recently published by Ahmed, Savolainen and Hamm [@Ahmed2014; @Savolainen2013], which consists in using balanced detection, and introducting a set of Brewster plates as shown in Figure \[fig:setup\](c). In order to obtain the necessary high acquisition rates, we associated this EOS setup with the so-called photonic time stretch technique developed by B. Jalali and coworkers [@Coppinger1999; @Wong2011]. The chirped laser pulses containing the THz pulse shape information are stretched within a 2 km-long fiber [@Goda2013], up to the nanosecond scale. At the balanced photodetector’s output, we thus obtain a “replica” of CSR pulse, that is temporally magnified, by a factor of about 200. The optical performances of the setup are described in detail in Ref. [@Szwaj2016]. The acquisition window is of the order of 15 ps, and the sensitivity is of 2.5 $\mu$V/cm/$\sqrt{\textnormal{Hz}}$ at 300 GHz in the electro-optic crystal (i.e., 1.25 V/cm). Since the frequency range of the CSR pulses is typically in the 150-600 GHz range [@Barros2013; @Barros2015], the “stretched signal” to be recorded by the oscilloscope is in the 0.75-3 GHz range. Here we use a Teledyne LeCroy LabMaster 10-65Zi oscilloscope (with an – overdimensioned – 30 GHz bandwidth), and the acquired data are numerically low-pass filtered at 6 GHz. We recorded 3 ms long time series, which corresponds to approximately 2500 turns in the storage ring. Typical single-shot recordings of the THz CSR pulses (electric field versus time) are represented in Figure \[fig:shots\]. In order to obtain a synthetic view of the CSR pulse dynamics over a large number of turns in the storage ring, we also represented the pulse series in a colorscale diagram, versus time and number of turns, as displayed in Figure \[fig:exp\_2d\](a)-(c), for three values of bunch charge. These diagrams clearly reveal two different emission processes. First, the data evidence the coherent emission which was predicted to occur from the “shortness” of the electron bunch [@Sannibale2004]. This process is directly visible below the microbunching instability threshold \[Fig. \[fig:exp\_2d\](a)\]. The emission occurs as a few-period coherent THz pulses, without dynamical evolution from turn-to-turn. This corresponds to a quasi-perfect Terahertz frequency comb [@Tammaro2015]. A second process occurs when the charge exceeds the so-called microbunching instability threshold \[Figs. \[fig:exp\_2d\](b,c)\]. An important part of the THz radiation is then emitted in the form of an oscillation, that drifts at each turn in the storage ring. This carrier-envelope phase drift stems from the appearance of microstructures which are continuously moving inside the electron bunch, as described in the numerical simulations presented later. In addition, the first process, i.e. the coherent emission due to the bunch shortness, is expected to exist above the microbunching instability threshold. The signature of this contribution from the data presented in Figs. \[fig:exp\_2d\](b,c) can be evidenced by computing the electric field pulse average over a large number of turns: $$V_{avg}^{EOS}(t)=\sum_{n=n0}^{n0+N-1}V_{n}^{EOS}(t),$$ where $V_{n}^{EOS}(t)$ is the electro-optic sampling signal of the THz pulses. This processing filters out the drifting component and may thus be viewed as a measure of the component due to the bunch shortness. This contribution $V_{avg}^{EOS}(t)$ (with $N\simeq 850$) is visible in Figures \[fig:exp\_2d\](g,h,i). Note that this component is linked to the short-bunch mode operation, i.e., is not observed when long electron bunches are used [@Roussel2014a; @Roussel2015]. From these data, it is also possible to reconstruct the turn-by-turn evolution of the THz pulse energy $E_{THz}^n$, with $$E_{THz}^n=\int_{t_1}^{t_2} |V_{n}^{EOS}(t)|^2 dt,$$ where $t_1$ and $t_2$ are taken near the boundaries of the THz pulse. This signal corresponds to the signal that would be recorded with a standart “slow” detector \[see Figs. \[fig:exp\_2d\](d-f)\]. It permits to confirm in a very direct way that the modulation in the THz signal, which can be clearly seen in Fig. \[fig:exp\_2d\](e) and which had been observed in numerous studies [@Kuske2009; @Feikes2011; @Evain2012; @Roussel2014; @Judin2012; @Bartolini2011; @Billinghurst2016], is directly related to the period of apparition of the micro-structures in the bunch \[Fig. \[fig:exp\_2d\](b)\]. We have then performed numerical integrations of analytic models, to make a comparison with these experimental data (obtained for the first time in a short-bunch mode), and also to access the 2D longitudinal phase-space. We present here the results obtained with one of the most elementary models: the one-dimensional Vlasov-Fokker-Planck model with shielded CSR wakefield [@Venturini2002]. The evolution of the electron density is described by the phase space distribution $f(q, p, \theta)$, where $\theta$ is the time associated to the number of round-trips in the ring, and $q$ and $p$ are longitudinal coordinate, and electron energy variables respectively. The dynamical evolution of $f (q, p,\theta)$ is determined by the following Vlasov-Fokker-Planck equation [@Venturini2002]: $$\begin{aligned} \textstyle \frac{\partial f(q,p,\theta)}{\partial \theta} &-& \textstyle p \frac{\partial f}{\partial q} + \frac{\partial f}{\partial p} \left( q - I_c E_{wf} \right) = 2\epsilon \frac{\partial}{\partial p}\left( p f + \frac{\partial f }{\partial p}\right) \label{eq:VFP1}\end{aligned}$$ The time $\theta$ associated to round-trips is a dimensionless variable: $\theta = 2\pi f_s t$, where $t$ is the time (in seconds) and $f_s$ is the synchrotron frequency. The longitudinal position $q$ and relative momentum $p$ are the deviation from the so-called synchronous electron (with position $z_0$ and energy $E_0$). $q$ and $p$ are expressed in units of the equilibrium bunch length $\sigma_z$ and energy spread $\sigma_E$ at zero current: $q=(z-z_0)/\sigma_z$, and $p = (E - E_0)/\sigma_E$. $\epsilon = 1/(2π f_s \tau_s )$, where $\tau_s$ is the synchrotron damping time. $I_c=I \frac{e2\pi R_c}{2\pi f_s \sigma_E T_0}$, with $I$ is the average beam current ($I=\frac{Q}{T_0}$), $R_c$ the dipole radius of curvature, and $T_0$ the revolution period. All parameters are in MKS units. ![Numerical simulations of the bunch dynamics, with a charge Q of 82.6 pC, 94.4 pC, and 141.6 pC, left, middle and right column respectively. (a)-(c): electron distribution in the phase-space (at the turn number $n\simeq 846$). (d)-(f): evolution of the electric field pulses (inside the chamber) $E_{wf}(t,n)$ versus turn number, with $t=-q\sigma_z/c$, $n=\theta/(2\pi f_s T_0)$, $c$ the light velocity. (g)-(i): average over the turns of the emitted pulses. Simulation parameters: $E_0=2.75~$GeV, $f_s=928$ Hz, $\sigma_E=E_0 \times 1.017\times 10^{-3}$, $\sigma_z=0.92~$mm, $\tau_s=3.27~$ms, $R_c=5.36$ m, $T_0=1.181~\mu$s, $h=12.5~$mm.[]{data-label="fig:num"}](fig4.pdf){width="1\linewidth"} The interaction between electrons is described by the term $E_{wf}(q,\theta)$ which represents the electron bunch wakefield (expressed in V/m, per Ampere of average current in the storage ring). We use here the CSR wakefield created by an electron bunch in a circular trajectory, between two parallel plates with distance 2h. The exact expression of $E_{wf}(q,\theta)$ is well known [@Murphy1997], and detailed in [@Roussel2014]. Integration of Eq. (1) is performed using the semi-Lagrangian scheme, described in Ref. [@Warnock2000], and parameters are listed in the figure \[fig:num\] caption. Figure \[fig:num\] shows the results of the numerical integration of equation (1) for three bunch charges $Q=82.6~$pC, 94.4 pC and 141.6 pC. We retreive here qualitatively important features observed experimentally. Below the instability threshold, no micro-structures are observable on the phase-space \[Fig. \[fig:num\](a)\] and the shape of the electron bunch is constant from turn-to-turn \[Fig. \[fig:num\](d)\]. Just above the instability threshold, “branches” continuously appears on the bottom part of the phase-space, and display a rotational motion at the synchrotron frequency \[see Fig. \[fig:num\](b) and supplemental movie\]. As a main consequence, the wakefield presents an oscillation which is drifting toward the bunch head \[Fig. \[fig:num\](e)\]. For higher currents, the micro-structures are more prominent \[see Fig. \[fig:num\](c) and supplemental movie\] and can reach the upper part of phase space. In this case, the wakefield \[Fig. \[fig:num\](f)\] also presents an oscillation which is drifting toward the tail of the bunch (in addition to the oscillation drifting to the head). We believe that the availability of this new type of data (CSR pulse shapes versus round-trips) is a precious tool for testing models of bunch dynamics (including wakefields), in a more clear-cut way than with classical means (i.e., average spectra, THz pulse energies, microbunching instability threshold, etc.). In the example presented here (where one of the most elementary model of wakefields is used), agreement as well as differences can clearly be seen. For instance below threshold, the wakefield displays a longer period in the numerical simulation \[Fig. \[fig:num\](g)\] than experimentally \[Fig. \[fig:exp\_2d\](g)\]. This may be attributed to differences in the wakefield modeling and/or in filter effects in the beamline wich are not taken into account in the modeling. Differences at higher current are also perceptible, as experimental data \[Fig. \[fig:exp\_2d\](c)\] are more irregular than numerical ones \[Fig. \[fig:num\](f)\]. We think that such comparisons of numerical and experimental diagrams can now be used as an important guide for the modeling process, including the crucial work of accelerator impedance modeling.\ In conclusion, it is now possible to analyze in a complete way the coherent synchrotron radiation emission due to short electron bunches in storage rings. As a main result, we could directly visualize the predicted transition between a stable THz coherent emission (due to the “bunch shortness”) and the emission due to the so-called microbunching instability. In the future, we believe that this viewing of CSR dynamics enables new and stringent tests for modeling the present and future storage ring facilities. In another perspective, there results show that THz spectra can be obtained at a MHz repetition rate, paving the way for ultra high repetition rate time resolved THz spectroscopy. We would like to thanks Menlo Systems for important advices. The work has been supported by the BQR of Lille University (2015), by the Ministry of Higher Education and Research, Nord-Pas de Calais Regional Council and European Regional Development Fund (ERDF) through the Contrat de Projets État-Région (CPER) 20072013, and by the LABEX CEMPI project (ANR-11-LABX-0007). [40]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevLett.88.254801) [****,  ()](\doibase 10.1103/PhysRevLett.89.224801) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****, ()]{} [****,  ()](\doibase 10.1103/PhysRevLett.114.204801) @noop [ ()]{} @noop [**** ()]{} [****,  ()](\doibase 10.1103/PhysRevLett.90.094801) [****,  ()](\doibase 10.1023/A:1024429801191) [****,  ()](\doibase 10.1103/PhysRevSTAB.14.030705) [****, ()](\doibase 10.1103/PhysRevSTAB.14.040705) [****,  ()](\doibase http://dx.doi.org/10.1016/j.jms.2015.03.012) [****,  ()](\doibase 10.1103/PhysRevSTAB.5.054402) [****,  ()](\doibase 10.1103/PhysRevLett.89.224802) [****,  ()](\doibase 10.1103/PhysRevLett.113.094801) [****,  ()](http://scitation.aip.org/content/aip/journal/apl/100/11/10.1063/1.3694049) [****,  ()](\doibase 10.1103/PhysRevSTAB.15.070701) @noop [ ()]{} [****,  ()](\doibase 10.1109/22.775471) @noop [****,  ()]{} [****, ()](http://dx.doi.org/10.1038/srep10330) [****,  ()](http://scitation.aip.org/content/aip/journal/rsi/85/1/10.1063/1.4862657) [****,  ()](\doibase 10.1073/pnas.1317459110) @noop [ ()]{} @noop [ ()]{} [****,  ()](\doibase http://dx.doi.org/10.1063/1.121231) [****,  ()](\doibase 10.1103/PhysRevLett.88.124801) [****,  ()](\doibase 10.1103/PhysRevSTAB.12.032802) [****,  ()](http://jlt.osa.org/abstract.cfm?URI=jlt-29-20-3099) [****,  ()](\doibase 10.1103/PhysRevLett.93.094801) @noop [ ()]{} [****,  ()](http://iopscience.iop.org/article/10.1209/0295-5075/98/40006/meta#citations) @noop [**** ()]{} @noop [ ()]{} @noop [ ()]{} [****,  ()](\doibase 10.1103/PhysRevAccelBeams.19.020704) @noop [****,  ()]{} @noop [ ()]{}

ArXiv

--- abstract: 'We study the intermittency and noise of dislocation systems undergoing shear deformation. Simulations of a simple two-dimensional discrete dislocation dynamics model indicate that the deformation rate exhibits a power spectrum scaling of the type $1/f^{\alpha}$. The noise exponent is far away from a Lorentzian, with $\alpha \approx 1.5$. This result is directly related to the way the durations of avalanches of plastic deformation activity scale with their size.' author: - 'L. Laurson and M. J. Alava' title: '$1/f$ noise and avalanche scaling in plastic deformation' --- Introduction ============ Traditionally, the plastic deformation of solids has been thought of as a smooth process, so that the fluctuations can be neglected on large enough scales as the behavior becomes homogeneous. Contrary to this old paradigm, very recent evidence from experiments points out clearly that plastic deformation proceeds via bursts of dislocation activity [@miguel; @dimiduk; @weiss; @marsan; @zaiser; @weiss_jcp; @zaiserreview]. This activity, due to the long-range interactions, should have extended spatial and temporal correlations. Supporting this new picture, a number of simulation studies of simplified models [@miguel; @miguel2; @koslowski; @morettizaiser_jstat] have demonstrated that such bursts or avalanches can be characterized by, apparently, scale-free size distributions. These advances bring plasticity close to a multitude of systems exhibiting “crackling noise” [@sethna]. This is a generic property of many non-equilibrium systems, such as earthquakes [@gr], Barkhausen noise in ferromagnets [@durin], fracture [@ciliberto; @salminen], vortex avalanches in superconductors [@field] and martensitic shape memory alloys [@vives] to name a few. The common feature among these are metastability and avalanches of activity that intercept quiescent periods. The realization that plastic deformation has rich internal dynamics is of interest both for physics and materials science. Dislocations provide an interesting example of the physics of systems with long-range interactions competing with pinning from obstacles such as impurities, grain boundaries, and other dislocations [@zaiserreview; @morettizaiser] (“forest hardening” being a well-known example of the ensuing phenomena). Such avalanching systems are most simply studied by the one-dimensional activity time series $V(t)$. In plasticity, $V$ could be the global strain rate, or the acoustic emission activity during deformation, or in a constant strain rate experiment the shear stress. In this work we consider the temporal characterization of dislocation activity and show that its power spectrum (PS) exhibits what is known as $1/f$-noise, $P(f)\sim 1/f^{\alpha}$. This “flicker” noise is a well-known but relatively little understood phenomenon in various fields of science [@1/f]. Then, we relate the scaling exponent $\alpha$ to the characteristic scaling of the avalanches. This has to be done with care due to the low-level background activity included in $V$. At the end, we are able to demonstrate that the noise of irreversible deformation may follow generic, experimentally verifiable scaling laws. We use a simple two-dimensional (2d) discrete dislocation dynamics (DDD) model [@miguel2], and study its behavior in the steady-state regime with a constant shear stress and a small (on the average constant) deformation rate. While the statistics of avalanche sizes in this kind of models has been established to be of power law type [@miguel; @miguel2], no temporal scaling analysis has been presented so far. The PS $P(f)$ of a time series $V(t)$ is the absolute square of the Fourier transform of $V(t)$, $$P(f) \sim \left|\int e^{i2\pi ft}V(t)dt\right|^2.$$ Since the time-time correlation function and the PS are related, the latter is a measure of temporal correlations in the system. $\alpha$-values less than two indicate the presence of complex time correlations. Recently it has been realized that under certain fairly general conditions the exponent $\alpha$ can be derived from the scaling exponents characterizing the avalanche distributions, the examples ranging from models of Barkhausen noise [@KS], to self-organized criticality [@oma] to fluid invasion into disordered media [@rost]. The intermittency is easier to understand in toy models as sandpile models of self-organized criticality [@btw], where $V(t)$ drops to zero between avalanches. In most experiments and more realistic model systems - such as here - there is a background (from noise, or processes that co-exist with the avalanches). Also, finite drive rates can lead to merging of avalanches, thus to problems in characterizing the underlying activity. The structure of the paper is as follows: In Sec. II we introduce the 2d DDD model, and in Sec. III the scaling relation relating the power spectrum exponent to the avalanche statistics is presented. Results of the numerical simulations are presented in Sec. IV. Section V finishes the paper with conclusions. DDD model ========= Despite its simplicity, the DDD model we study here has been shown to capture many features of realistic plasticity, such as avalanches with scale-free size distributions [@miguel; @miguel2] as well as the Andrade creep law [@miguelprl]. It is a two-dimensional model representing a cross section of a crystal with single-slip geometry. The dislocations are assumed to be straight parallel edge dislocations, with the dislocation lines oriented along the $z$ axis. They glide along directions parallel to their Burgers vectors ${\bf b}=\pm b {\bf u}_x$, with ${\bf u}_x$ the unit vector in the $x$-direction. It is thus sufficient to consider a two-dimensional system (i.e. the $xy$ plane) with $N$ pointlike edge dislocations gliding in the $x$-direction. Equal numbers of dislocations with positive and negative Burgers vectors are assumed. For simplicity other processes contributing to dislocation motion such as dislocation climb are not considered. An essential feature of the model is that the dislocations interact with each other through their anisotropic long-range stress fields $$\label{stress} \sigma_s({\bf r}) = Db\frac{x(x^2-y^2)}{(x^2+y^2)^2},$$ where $D=\mu/2\pi(1-\nu)$, with $\mu$ the shear modulus and $\nu$ the Poisson ratio of the material. We assume overdamped dynamics with a linear force-velocity relationship, giving rise to the equations of motion of the form $$\label{eom} \frac{\chi_d^{-1}v_n}{b} = s_n b \left[ \sum_{m \neq n} s_m \sigma_s({\bf r}_{nm}) + \sigma \right],$$ where $v_n$ is the velocity of the $n$’th dislocation, $\chi_d$ is the dislocation mobility, $s_n$ refers to the sign of the Burgers vector of the $n$’th dislocation and $\sigma$ is the external shear stress experienced by the dislocation. These equations are put into a dimensionless form by measuring lengths in units of $b$, times in units of $1/(\chi_d D b)$ and stresses in units of $D$. Due to the long range nature of the dislocation-dislocation interaction, periodic boundary conditions are implemented with an infinite sum of images [@hirth]. The positions ${\bf r}_n$, $n=1,\dots,N$, of the $N$ dislocations as a function of time are computed by integrating the equations of motion (\[eom\]) numerically by using an adaptive step size fifth-order Runge-Kutta algorithm. For small distances, Eq. (\[stress\]) ceases to be valid. Thus for distances smaller than $2b$ from the dislocation, we set the dislocation stress field to zero. This procedure removes the unphysical singularity in Eq. (\[stress\]). Furthermore, when the distance $r_{nm}$ between two dislocations with Burgers vectors of opposite sign gets small, i.e. $r_{nm}<2b$, we employ a phenomenological annihilation reaction by simply removing them from the system. To compensate, and to include dislocation multiplication (as in real plasticity through Frank-Read sources), we introduce a mechanism to create new dislocations. The system is split into smaller cells and then one monitors both the local stress and the number of pinned dislocations (i.e. those moving slower than a threshold velocity) in each cell. Given pinned dislocations and that the local stress exceeds a threshold value, a dislocation pair with opposite Burgers vectors is generated with a probability proportional to the absolute local stress value. This means that on the average it takes a finite time for the source to create a new dislocation pair, as is the case with real dislocation sources. The two new dislocations of opposite sign are inserted into random locations inside the neighboring cells of the one containing the activated source. This is done with the constraint that their combined stress field must decrease the magnitude of the local stress at the source location, while elsewhere in the system further dislocation activity may be triggered, resulting in a correlated sequence of dislocation activity. Scaling of the power spectrum ============================= Consider a bursty time series $V(t)$ consisting of temporally separated avalanches. The usual definition of an avalanche is a connected sequence of values of $V(t)$ exceeding some threshold value $V_{th}$, to e.g. subtract uncorrelated background noise. If an avalanche starts at $t=0$ and ends at $t=T$, the size $s$ of an avalanche of duration $T$ is defined as $s(T)=\int_0^T[V(t)-V_{th}]dt$. Assume that the average avalanche size $\langle s(T) \rangle$ of avalanches of a given duration $T$ scales as a power law of the duration, $$\label{sT} \langle s(T) \rangle \sim T^{\gamma_{st}},$$ and that the avalanches are self-similar so that the averaged avalanche shapes $V(T,t)$ of avalanches of different durations $T$ can be collapsed onto a single curve by using the ansatz $$\label{shape_ansatz} V(T,t) = T^{\gamma_{st}-1} f_{shape}(t/T).$$ Here, $f_{shape}(x)$ is a scaling function. The total energy is obtained as the $\theta=0$ component of the stationary time-time correlation function, defined by $C(\theta) = \int V(t)V(t+\theta)dt$. By cosine transforming the time-time correlation function $C(\theta|s)$ of avalanches of a given size $s$, one obtains the scaling form $$\label{energy} E(f|s) = s^2g_E(f^{\gamma_{st}}s)$$ for the corresponding energy spectrum. The scaling of the total power spectrum then follows by averaging $E(f|s)$ over the avalanche size probability distribution $D_s(s)$, e.g. a power law $D_s(s) \sim s^{-\tau_s}$ cut off at $s=s^*$, so that $$P(f) = f^{-\gamma_{st}(3-\tau_s)} \int^{s^* f^{\gamma_{st}}} dx x^{2-\tau_s} g_E(x).$$ The result will depend on the value of $\tau_s$ as well as on the form of the scaling function $g_E(x)$ [@KS; @oma]. In the case $g_E(x) \sim 1/x$ [@KS; @oma] and $\tau_s < 2$ (for the DDD models at hand, the latter condition seems to be fulfilled, with $\tau_s \approx 1.6$ [@miguel2]) the power spectrum scales as $$\label{tulos} P(f) \sim f^{-\gamma_{st}}.$$ This indicates a scaling relation $\alpha = \gamma_{st}$, which possibility we next check with simulations. Results ======= A natural choice for the time series is the “collective velocity” of the dislocations, $V(t)=\sum_i |v_i|$. It is proportional to the energy dissipated per unit time by the dislocation system. Thus the statistical properties of $V(t)$ could be related to the acoustic emission (AE) statistics, as suggested by a number of authors, see Ref. [@zaiserreview]. Another possibility is $V_s(t)=\sum_i b_i v_i$, i.e. the global strain rate. We mainly present results here for the first choice. The numerical simulations are started from a random initial configuration of $N_0$ dislocations with equal numbers of positive and negative Burgers vectors (here $L=200b$ and $L=300b$, with $N_0=600$ and $N_0=1000$, respectively). The system is first let to relax in the absence of external stress, until it reaches a metastable arrangement with $N<N_0$. Then a small constant external stress is applied and the evolution of the system is monitored. In the absence of dislocation multiplication, the system undergoes a “jamming transition” at a critical value $\sigma_c$ of the external stress [@miguelprl]. For $\sigma=\sigma_c$, the strain rate decays as a power law in time, with the Andrade’s power law creep exponent close to $-2/3$. For $\sigma>\sigma_c \approx 0.01$, the system crosses over to a constant strain rate regime (linear creep) which we study next. With the dislocation multiplication turned on, the system displays intermittent avalanche-like bursts of dislocation activity. The number of dislocations fluctuates around $N \approx 300$ for $L=200b$ and $N \approx 600$ for $L=300b$. Fig. \[signal\] shows that $V(t)$ displays intermittent behavior, consisting of avalanches. These are defined as a connected sequence of $V(t)$-values exceeding a threshold value $V_{th}$ and have a wide range of sizes. Here, we monitor $V(t)$ for external stress values close to but above the critical value $\sigma_c$. Fig. \[shape\] exhibits the average avalanche shape (one needs a minimum duration, of the order of 10, for scaling), for varying $V_{th}$. They appear to agree with Eq. (\[shape\_ansatz\]). The data has been obtained by scaling all individual avalanches in a given duration range according to Eq. (\[shape\_ansatz\]). The threshold value $V_{th}$ does not have any effect on the $s(T)$ statistics (given that $V_{th}$ is high enough such that a “noise level” corresponding to [*incoherent motion*]{} of dislocations is not taken into account), but the avalanche shapes appear to depend on $V_{th}$. For small $V_{th}$-values, avalanches are clearly asymmetric to the left (in agreement with experiments [@weiss3]), but become more symmetric upon increasing $V_{th}$. This asymmetry is also manifest in the creation rate of dislocations during an avalanche (inset of Fig. \[shape\]). In Fig. \[1\_per\_x\] we show the energy spectra computed using the Lomb periodogram, scaled according to Eq. (\[energy\]). The behavior of the scaling function $g_E(x)$ indicates that $g_E(x) \sim 1/x$. The main result is shown in Fig. \[svst\], where we compare the scaling of the total power spectrum, from $V(t)$, with that of $\langle s(T) \rangle$. We observe a PS of the form $P(f) \sim f^{-\gamma_{st}}$ spanning almost two decades, with $\gamma_{st} \approx 1.5$. Thus the PS of dislocation activity is related here to the intrinsic scaling of the avalanches. The extension of the scaling regime increases with $L$, the system size (Fig. \[ps\_L\]), with a cut-off frequency $f_{cut-off}$ roughly equal to the system size dependent inverse duration of the longest avalanche, $f_{cut-off}\sim T_{max}^{-1}(L)$. The absence of scaling for the very highest frequencies ($f>0.1$) is due to a finite crossover time, before the avalanches have a self-similar structure. Similarly to the exponent of the avalanche size distribution [@miguel], the exponent of the power spectrum is interestingly insensitive to the value of the external stress. In the inset of Fig. \[svst\] we consider the PS of the total strain rate. While the scaling region appears somewhat narrower in this case, the exponent $\alpha$ is observed to be unchanged from $1.5$. These results are largely independent of the details of the model, such as the threshold value for the local stress to create new dislocations. We have checked that the same results can be recovered even in the case with no dislocation multiplication, where the avalanches are due to different threshold mechanisms, such as unpinning of dislocation dipoles. As this approach suffers from the fact that simulations are computationally more demanding due to the longer time scales one must reach, we have here restricted ourselves to considering only the case where dislocation multiplication does occur. Knowing that the the avalanches are self-affine one can also compute $\gamma_{st}$ from the other known exponents. Consider the probability $D_V(V|s)$ that a value $V$ will occur at some point during an avalanche of size $s$. Assuming scaling and requiring normalization, $\int_0^{\infty} D_V(V|s) dV = 1$, as well as $\langle V \rangle = s/T \sim s^{1-1/\gamma_{st}}$, one obtains (see also [@KS]) $$D_V(V|s) = V^{-1}f_V(Vs^{1/\gamma_{st}-1}),$$ where $f_V(x)$ is a scaling function. The total probability $D_V(V)$ is then obtained by integrating $D_V(V|s)$ over the avalanche size distribution $D_s(s) \sim s^{-\tau_s}$, giving rise to $D_V(V) \sim V^{-(\frac{\gamma_{st}\tau_s-1}{\gamma_{st}-1})} \equiv V^{-\tau_V}$. Miguel et al. [@miguel] studied the statistics of the quantity $E = (\sum_i |v_i|)^2$, and found that its distribution behaves like $D_E(E) \sim E^{-\tau_E}$, with $\tau_E \approx 1.8$. From $\tau_E = \frac{1}{2}(\tau_V + 1)$ one may solve for $\gamma_{st}$, resulting in (with $\tau_s \approx 1.6$ [@miguel2]) $$\label{gammaon} \gamma_{st} = \frac{2\tau_{E}-2}{2\tau_{E}-\tau_s-1} \approx 1.6,$$ in reasonable agreement with our results above. Conclusions =========== In this work, we have demonstrated that a simple two-dimensional dislocation system exhibits $1/f^\alpha$-noise, so that the relation $\alpha=\gamma_{st}$ can be explained by the scaling properties of the avalanche dynamics. We have also considered the PS of the global strain rate, with similar conclusions. These findings lend themselves to experimental tests, e.g. as in the creep experiments of ice single crystals [@weiss], in experiments on deforming metallic single (micro)crystals [@weiss2; @dimiduk], and could also be considered in experiments on colloidal crystals [@pertsinidis]. In AE experiments performed with three-dimensional systems, values consistent with mean field exponents have been reported [@weiss; @weiss2], suggesting $\alpha=\gamma_{st}=2$ (see Eq. (\[gammaon\])). It is thus possible that to get quantitative agreement with experiments, the full three-dimensional problem should be studied, e.g. by means of simulations of a three-dimensional dislocation dynamics model. Other issues not considered in the present study include the possible relevance of quenched disorder e.g. in the form of forest dislocations, which would provide strong pinning centers to resist dislocation motion. One should note, however, that also the present model contains a pinning force landscape generated by the dislocations themselves [@miguel]. In polycrystals, avalanches interact with the grain boundaries, which will probably lead to a size-dependent avalanche shape [@richeton]. This means that Eq. (\[shape\_ansatz\]) is no longer directly applicable, presenting an interesting theoretical question. Our theory assumes negligible correlations between avalanches. In experiments, dislocation avalanches have been found to exhibit a tendency to cluster in time, such that a “mainshock” is typically followed by a sequence of few “aftershocks” [@weiss; @dimiduk]. While this may modify the low frequency part of the PS, the high frequency part, corresponding to correlations within individual avalanches, is still expected to scale according to Eq. (\[tulos\]). Finally we note that in addition to materials with crystalline structure, one field where the use of spectral tools and the study of avalanches might be used to characterize the spatio-temporal behavior is the plasticity of non-crystalline media, e.g. amorphous glasses, where the localization and intermittency of plastic events has been recently demonstrated in simulations [@lemaitre; @tanguy]. [**Acknowledgments**]{} The Center of Excellence program of the Academy of Finland is thanked for financial support. [99]{} M. Zaiser, Advances in Physics [**55**]{}, 185 (2006). M.-C. Miguel, A. Vespignani, S. Zapperi, J. Weiss, and J.-R. Grasso, Nature [**410**]{}, 667 (2001). D. M. Dimiduk, C. Woodward, R. LeSar, M. D. Uchic, Science [**312**]{} 1188 (2006). J. Weiss, J.-R. Grasso, M.-C. Miguel, A. Vespignani, and S. Zapperi, Mat. Sci. Eng. A [**309-310**]{}, 360 (2001). J. Weiss and D. Marsan, Science [**299**]{}, 89 (2003). M. Zaiser, F. Madani, V. Koutsos, and E. C. Aifantis. Phys. Rev. Lett. [**93**]{}, 195507 (2004). J. Weiss and J.-R. Grasso, J. Phys. Chem. B [**101**]{}, 6113 (1997). For a typical 2d DDD model, see: M.-C. Miguel, A. Vespignani, S. Zapperi, J. Weiss, and J.-R. Grasso, Mat. Sci. Eng. A [**309-310**]{}, 324 (2001). M. Koslowski, R. LeSar, and R. Thomson, Phys. Rev. Lett. [**93**]{}, 125502 (2004). M. Zaiser and P. Moretti, J. Stat. Mech. (2005) P08004. J. P. Sethna, K. A. Dahmen, and C. Myers, Nature [**410**]{}, 242 (2001). B. Gutenberg and C. F. Richter, Bull. Seismol. Soc. Am. [**34**]{}, 185 (1944). G. Durin and S. Zapperi, J. Magn. Mat. [**242-245**]{}, 1085 (2002). A. Garcimartin [*et al.*]{}, Phys. Rev. Lett [**79**]{}, 3202 (1997). L.I. Salminen, A.I. Tolvanen, and M.J. Alava, Phys. Rev. Lett. [**89**]{}, 185503 (2002). S. Field, J. Witt, F. Nori, and X. Ling, Phys. Rev. Lett. [**74**]{}, 1206 (1995). E. Vives, J. Ortín, L. Mañosa, I. Ràfols, R. Pérez-Magrané, and A. Planes, Phys. Rev. Lett. [**72**]{}, 1694 (1994). P. Moretti, M.-C. Miguel, M. Zaiser, and S. Zapperi, Phys. Rev. B [**69**]{}, 214103 (2004). P. Dutta and P. M. Horn, Rev. Mod. Phys. [**53**]{}, 497 (1981); M. B. Weissmann, Rev. Mod. Phys. [**60**]{}, 537 (1988); E. Milotti, physics/0204033. M. C. Kuntz and J. P. Sethna, Phys. Rev. B [**62**]{}, 11699 (2000). L. Laurson, M. J. Alava, and S. Zapperi, J. Stat. Mech. (2005) L11001. M. Rost, L. Laurson, M. Dubé, and M. Alava, submitted. P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. Lett. [**59**]{}, 381 (1987). M.-C. Miguel, A. Vespignani, M. Zaiser, and S. Zapperi, Phys. Rev. Lett. [**89**]{}, 165501 (2002). J. P. Hirth and J. Lothe, [*Theory of Dislocations*]{} (Krieger, Malabar, Florida, 1992). T. Richeton, J. Weiss, and F. Louchet, Acta Mater. [**53**]{}, 4463 (2005). T. Richeton, P. Dobron, F. Chmelik, J. Weiss, and F. Louchet, Mat. Sci. Eng. A [**424**]{}, 190 (2006). A. Pertsinidis and X. S. Ling, New J. Phys. [**7**]{}, 33 (2005). T. Richeton, J. Weiss, and F. Louchet, Nature Materials [**4**]{}, 465 (2005). C. Maloney and A. Lemaître, Phys. Rev. Lett. [**93**]{}, 016001 (2004). A. Tanguy, F. Leonforte, and J.-L. Barrat, Eur. Phys. J. E [**20**]{}, 355 (2006).

ArXiv

--- abstract: 'We compute, within the Schrödinger functional scheme, a renormalization group invariant renormalization constant for the first moment of the non-singlet parton distribution function. The matching of the results of our non-perturbative calculation with the ones from hadronic matrix elements allows us to obtain eventually a renormalization group invariant average momentum of non-singlet parton densities, which can be translated into a preferred scheme at a specific scale.' author: - | Marco Guagnelli$^1$, Karl Jansen$^{2,}$[^1] $\;$ and Roberto Petronzio$^{1}$\ [$^1$ Dipartimento di Fisica, Università di Roma [*Tor Vergata*]{} ]{}\ [and INFN, Sezione di Roma II]{}\ [Via della Ricerca Scientifica 1, 00133 Rome, Italy ]{}\ [ $^2$ CERN, 1211 Geneva 23, Switzerland ]{} title: | [ ]{}\ Renormalization group invariant average momentum of non-singlet parton densities --- Physical quantities that need renormalization, such as the coupling constant, the quark mass or the matrix elements of operators appearing in the Wilson operator product expansion with a non-zero anomalous dimension, are “running” with the renormalization scale. The choice of the scale is in general motivated by the kinematics of the Green functions involving such renormalized quantities, but the final physical predictions of the theory without perturbative approximations are independent of such a choice. This leads to the well-known renormalization group equations that put the independence on a formal basis. The redundancy in a parametrisation of the theory in terms of renormalized quantities and the relative renormalization scale can be avoided by considering renormalization group invariant quantities, such as the $\Lambda$ parameter of QCD or the renormalization group invariant quark masses (RGIM). The advantage of the latter choice in non-perturbative lattice determinations of the quark mass has recently been stressed by the authors of ref. [@qmass] where an essential part of the RGIM programme was carried out. In particular, the definition of the RGIM, which corresponds, roughly speaking, to a running mass at infinite renormalization scale, is free of the renormalization scheme dependence that usually affects quantities renormalized (in a given scheme) at a fixed scale. It can hence be evolved back to an arbitrary finite scale in a preferred scheme. The purpose of this letter is to present a similar calculation for the operator that corresponds to the average momentum of non-singlet parton densities. A lattice — perturbative and non-perturbative — study of the evolution of such an operator has been discussed in refs. [@roberto_pert] and  [@paper_1] to which we address the reader for more details about the calculation that we here only shortly summarize as follows. We calculate the renormalization constant of the twist-two non-singlet operator for the first moment of the quark parton distribution defined by: $${\cal{O}}_{\mu \nu}^{qNS} = \bigl( \frac{i}{2}\bigl)^{n-1} {\bar{\psi}}(x) \gamma_{\{\mu} {\stackrel{\leftrightarrow}{D}}_{\nu\}} \frac{\lambda^f}{2} \psi(x)\ -\ \mbox{trace terms}\; , \label{eq:twist_two_continuum}$$ where $\left\{\dots\right\}$ means symmetrization of the indices. We remark that the technique discussed here can be extended to higher moments in an anloguous way. The basic ingredient for the reconstruction of the non-perturbative scale dependence of the renormalization constants of the above operator is the finite-size step scaling function $\sigma_Z$ defined by: $$\label{Zs} Z(sL) = \sigma_Z(s, \bar{g}^2(L))Z(L)\; ,$$ where $L$ is the physical length that plays the role of the renormalization scale, $s$ parametrizes the step size of the change in the scale, and $Z$ is the renormalization constant of the operator, which is defined by: $${\cal O}^{R}(\mu) = Z(1/\mu)^{-1}{\cal O}^{\rm bare}(1/L)\; .$$ $Z$ is obtained from the Schrödinger Functional (SF) matrix element, $\langle\dots\rangle_{\rm SF}$, of the operator on a finite volume $L^3T$, normalized to its tree level[^2]: $$\langle{\cal O}^{\rm bare}(1/L)\rangle_{\rm SF} = Z(L) \langle{\cal O}^{\rm tree}\rangle_{\rm SF}\; .$$ The renormalized operator then satisfies $\langle {\cal O}^R(\mu=1/L)\rangle_{\rm SF}=\langle {\cal O}^{\rm tree}\rangle_{\rm SF}$. The framework of the Schrödinger Functional [@schrfunc; @sint], which describes the quantum time evolution between two fixed classical gauge and fermion configurations, defined at times $t=0$ and $t=T$, has been used extensively in the recent literature [@letter; @paper4; @qmass] to calculate non-perturbative renormalization constants of local operators. Among the advantages of the method, we only quote the possibility of performing the computations at zero physical quark mass and of using non-local gauge-invariant sources for the fermions without need of a gauge-fixing procedure. In our particular case, we exploit both features. Our observable is defined by [@roberto_pert]: $$Z = \frac{f_2(x_0=L/4)}{\sqrt{f_1}}\left/\left(\frac{f_2(x_0=L/4)}{\sqrt{f_1}}\right)_{\rm tree}\right.\; , \label{eq:obs}$$ with $f_2$ given by $$f_2(x_0) = -a^6\sum_{\bf{y},\bf{z}} \rm{e}^ {i\bf{p}(\bf{y}-\bf{z})} \langle \frac{1}{4} \bar\psi(x) \gamma_{\{1} {\stackrel{\leftrightarrow}{D}}_{2\}}\frac{1}{2} \tau^3 \psi(x) \bar\zeta({\bf{y}}) \Gamma \frac{1}{2} \tau^3 \zeta({\bf{z}})\rangle \label{eq:f2}$$ and $f_1$ by $$f_1 = -a^{12}\sum_{\bf{y},\bf{z},\bf{v},\bf{w}}\langle \bar\zeta'(\bf{v})\frac{\tau^3}{2}\zeta'(\bf{w}) \bar\zeta(\bf{x})\frac{\tau^3}{2}\zeta(\bf{y})\rangle\; , \label{eq:f1}$$ where $\zeta=\delta /\delta\bar{\psi}_c$ and $\bar\zeta=-\delta/\delta\psi_c$ are the derivatives with respect to the two-component classical fermion fields ($\bar{\psi}_c$ and $\psi_c$, respectively) at the boundary $x_0=0$, while $\zeta'$ and $\bar\zeta'$ are the corresponding derivatives at the boundary $x_0=T$. The projection on the classical components is achieved by the projector $P_{\pm}$ defined by $\frac{1}{2}(1\pm\gamma_0)$. On the boundaries, the theory possesses only a [*global*]{} gauge invariance that is preserved by the quantities defined above. The values of $x_0$ (set to $T/4$) and of the non-zero component of the momentum $p_x$ (set to $2\pi$/$L$) are both scaled in units of $L$, which therefore remains the only scale besides the lattice spacing $a$. The quantity $f_1$ serves as a normalization factor that removes the wave function renormalization constant of the $\zeta$ fields in order to isolate the running associated with the operator in eq. (\[eq:twist\_two\_continuum\]) only. The determination of the step scaling function in the continuum has been shown to be universal with respect to the lattice action used in ref. [@paper_2]. From a fit to its dependence upon the running coupling constant $\bar{g}^2$, renormalized in the SF scheme, we can extract the following “running” step scaling function: $$\sigma ( \mu/\mu_0,\bar{g}^2(\mu_0) ) = Z(1/\mu)/ Z(1/\mu_0) \label{eq:step_running}$$ i.e. the renormalization constant normalized to the one at a reference scale $\mu_0$. The running operator matrix element at the scale $\mu$, which we denote generically by the symbol $O$, can be defined in terms of the one at scale $\mu_0$ simply by: $$O^{ren}(\mu) = O^{ren}(\mu_0)\sigma ( \mu/\mu_0,\bar{g}^2(\mu_0) )\; . \label{eq:operator_running}$$ The scale dependence of the renormalized operator just reflects the one of its renormalization constant governed by the equation: $$\frac{{\rm d} Z(1/\mu)}{{\rm d}\log(\mu)} = Z(1/\mu)\cdot\gamma_O(g^2(\mu))\; , \label{eq:ren_g_Z}$$ from which follows: $$\frac{{\rm d}O^{ren}(\mu)}{{\rm d}\log(\mu)} = O^{ren}(\mu)\cdot\gamma_O(g^2(\mu))\; . \label{eq:ren_g_O}$$ Following ref. [@qmass] but using a slightly different normalization in taking out the factor of $2b_0$, we define, for operators entering the Wilson operator product expansion, a renormalization group invariant matrix element: $$O^{ren}_{\rm INV} = O^{ren}(\mu)\cdot (\bar{g}^2)^{-\gamma_0/2b_0} \exp\left\{ -\int_0^{\bar{g}} dg\left[\frac{\gamma(g)}{\beta(g)} - \frac{\gamma_0}{b_0g}\right]\right\}\; , \label{eq:O_inv}$$ where for the anomalous dimension function $\gamma(g)$ and the $\beta$-function the expressions up to three loops may be inserted for values of $g$ small enough that perturbation theory can be trusted: $$\label{gamma_3loop} \gamma(g^2(\mu)) = \gamma_0 g^2(\mu) + \gamma_1 g^4(\mu) + \gamma_2 g^6(\mu),$$ $$\label{beta_3loop} \beta(g^2(\mu)) = \beta_0 g^4(\mu) + \beta_1 g^6(\mu) + \beta_2 g^8(\mu).$$ We note that for $\gamma(g)$ we know the effective three-loop term from our non-perturbative computation of $\gamma_2$ [@paper_1], while $\gamma_0$ and $\gamma_1$ are given from perturbation theory. From eq. (\[eq:step\_running\]) once the $O^{ren}(\mu_0)$ is known in some scheme, for example the SF scheme we have described, we can obtain the renormalization scheme invariant matrix element by introducing an “ultraviolet invariant” running step scaling function [^3] defined by: $$\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0) =\sigma ( \mu/\mu_0,\bar{g}^2(\mu_0) )\cdot (\bar{g}^2(\mu))^{-\gamma_0/2b_0} \exp\left\{ -\int_0^{\bar{g}(\mu)} dg\left[\frac{\gamma(g)}{\beta(g)} - \frac{\gamma_0}{b_0g}\right]\right\} \label{eq:step_running_inv}$$ as follows: $$O^{ren}_{\rm INV} = O^{ren}(\mu_0)\cdot\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)\; . \label{eq:O_inv_2}$$ The scale $\mu_0$ is in general a low-energy scale, where the hadronic matrix element can be calculated without severe finite volume effects. In our case, it can be identified with a low-energy scale at which the evolution of the renormalization constant can be started. In particular we shall fix this scale to be $2L_{\rm max}$ as in ref. [@qmass]. Recently, $L_{\rm max}$ has been computed in terms of the low energy reference quantity $r_0$ [@rainer] in [@rainerlmax]. In order to “step down” from this scale, we will need the step scaling function with $s=2$, i.e. starting from $\bar{g}^2(L_{\rm max})=3.48$, our largest value of $\bar{g}^2$, we evolve with a step size of 2 until contact with perturbation theory can be made. In this paper, we calculate, as a first step towards the computation of the renormalization group invariant matrix element, the quantity $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0 = \frac{1}{2L_{\rm max}})$. Note that $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ still depends on the reference scale $\mu_0$. The dependence on $\mu_0$ will only disappear later, when it will be matched with the proper hadronic matrix element, making $O_{\rm INV}^{ren}$ renomalization scheme independent. In order to rely on the perturbative expansion for the $\beta$- and $\gamma$-functions appearing in eq. (\[eq:step\_running\_inv\]), we had to extend the calculation of our non-perturbative running to higher scales. We added four more values of $\bar{g}$ for the step scaling function that now covers, in total, values of $\bar{g}^2(L)$ ranging from $\bar{g}^2(L) = 3.48$ to $\bar{g}^2(L) = 0.8873$. For the results at the four lowest values of $\bar{g}$ we used the non-perturbatively improved clover action [@paper3]; in figs. 1 and 2 we report the continuum extrapolation, for the values of $\bar{g}^2$ not presented already in ref. [@paper_2] of the step scaling function of the quantities $f_1$ and $f_2$ of eq. (\[eq:obs\]) ($\sigma_{\bar{Z}}$ and $\sigma_{f_1}$, respectively, see [@paper_1]): at smaller length scales the effects of lattice artefacts for $\sigma_{\bar{Z}}$ are progressively reduced and the extrapolations become flatter. From the results for $\sigma_{Z}$ at the, in total, nine values of $\bar{g}$, we can make a fit to the step scaling function as a function of $\bar{g}^2(L)$. The results for $\sigma_{Z}$ at the five largest values of $\bar{g}^2(L)$ are taken from the combined data presented in [@paper_2]. In ref. [@paper_1] we have shown that, in the scheme we adopted, the coefficient of the two-loop anomalous dimension is very large, when compared for example to the one in the $\overline{\rm MS}$ scheme. We have also shown that this coefficient reduces by changing the expansion parameter, i.e. by using $\bar{g}^2(L/4)$ instead of $\bar{g}^2(L)$. The step scaling function as a function of $\bar{g}^2(L/4)$ is well fitted numerically by a polynomial in $\bar{g}^2(L/4)$ of the form: $$\sigma(\bar{g}^2(L/4)) = 1 - \gamma_0\log(2)\bar{g}^2 + c_4\cdot \bar{g}^4 + c_6\cdot \bar{g}^6 + c_8\cdot \bar{g}^8\; , \label{eq:step_fit}$$ where $\gamma_0 = 4/(9\pi^2)$. The final results stay unchanged when we switch to a two-parameter fit that also gives a very good $\chi^2$. We show our data for $\sigma_{Z}$ as a function of $\bar{g}^2(L/4)$ together with the fit of eq. (\[eq:step\_fit\]) in fig. 3. ------------- ---------------------------------------- ---------------------------------------- ---------------------------------------- ---------------------------------------- $\mu/\mu_0$ $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ $\;$ $\bar{g}(L/4)$ $\bar{g}(L)$ $\bar{g}(L/4)$ $\bar{g}(L)$ $\;$ $\gamma_O$ 2-loop $\gamma_O$ 2-loop $\gamma_O$ 3-loop $\gamma_O$ 3-loop $2^1$ $1.09(1)$ $ 1.33(1)$ $1.16(1)$ $1.18(1)$ $2^2$ $1.10(2)$ $ 1.24(2)$ $1.15(2)$ $1.16(2)$ $2^3$ $1.11(2)$ $ 1.20(2)$ $1.14(2)$ $1.15(2)$ $2^4$ $1.11(3)$ $ 1.18(2)$ $1.14(3)$ $1.14(2)$ $2^5$ $1.11(3)$ $ 1.16(3)$ $1.13(3)$ $1.14(3)$ $2^6$ $1.11(4)$ $ 1.15(3)$ $1.13(4)$ $1.13(3)$ $2^7$ $1.11(4)$ $ 1.14(3)$ $1.12(4)$ $1.13(3)$ $2^8$ $1.10(5)$ $ 1.14(3)$ $1.12(5)$ $1.13(3)$ $2^9$ $1.10(5)$ $ 1.13(3)$ $1.11(5)$ $1.12(3)$ ------------- ---------------------------------------- ---------------------------------------- ---------------------------------------- ---------------------------------------- : The values for $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ when different scales $\mu$ are taken for matching with perturbation theory.[]{data-label="tab:sigmamu0"} From this fit we can construct the running step scaling function of eq. (\[eq:step\_running\_inv\]) with $\mu_0=(2L_{\rm max})^{-1}$. The result is shown in fig. 4, where we have used the two-loop expression for $\gamma(g)$ and the 3-loop expression for $\beta(g)$. By using eq. (\[eq:step\_running\_inv\]) we can finally estimate the value of $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$: in the second column of table 1 we report the values of $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0)$ as a function of the scale $\mu$: for large values of $\mu$ the function, within the errors, approaches a plateau. We make a fit to a constant for the results ranging from $\mu/\mu_0 = 2^5$ to $\mu/\mu_0 = 2^9$, and we finally quote: $$\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0=(2L_{\rm max})^{-1}) = 1.11(2)\; .$$ The invariant step scaling function is still scheme-dependent, because of the presence of the reference scale $\mu_0$. This will be cancelled only in the combination that defines the invariant matrix element. However, at fixed $\mu_0$, it should be independent of the choice of $\bar{g}^2(L/4)$ or of $\bar{g}^2(L)$ in the fit to the step scaling function. We therefore repeated the whole procedure described above by fitting the step scaling function as a function of $\bar{g}^2(L)$ and by using the correspondingly modified gamma function to two loops. The results are given in the third column of table 1. They are fully compatible with those obtained from the case $L/4$, although the plateau starts at higher energies, as expected. We report the comparison of both cases also in fig. \[fig:rgi\]. In the fourth and fifth column of table 1 we report the result for the case “$L/4$” and “$L$” respectively, after including our estimate of the three-loop anomalous dimensions for the two cases, determined in [@paper_1; @paper_2]. Not surprisingly, the two cases get close to each other more precociously. An estimate of the renormalization group invariant yields $\EuFrak{S}_{\rm INV}^{\rm UV}(\mu_0=\frac{1}{2L_{\rm max}})=1.14(2)$, again consistent with our earlier results using $\bar{g}(L/4)$ as expansion parameter. Matching the results of this paper with a non-perturbative calculation of the hadronic matrix element, in the continuum, in the same scheme and at the same reference energy scale, leads to the definition of a renormalization group invariant matrix element that can be confronted with experiment at any scale and in a preferred scheme. Such a calculation is in progress. Acknowledgement {#acknowledgement .unnumbered} =============== We thank Martin Lüscher for discussions and helpful comments. [99]{} S. Capitani, M. Lüscher, R. Sommer and H. Wittig, DESY preprint, DESY-98-154, hep-lat/9810063. A. Bucarelli, F. Palombi, R. Petronzio and A. Shindler, hep–lat/9808005. M. Guagnelli, K. Jansen and R. Petronzio, CERN preprint, CERN-TH/98-265, hep-lat/9809009, to be published in Nucl. Phys. B. M. Lüscher, R. Narayanan, P. Weisz and U. Wolff, Nucl. Phys. B384 (1992) 168. S. Sint, Nucl. Phys. B421 (1994) 135. K. Jansen, C. Liu, M. Lüscher, H. Simma, S. Sint, R. Sommer, P. Weisz and U. Wolff, Phys. Lett. B372 (1996) 275. M. Lüscher, S. Sint, R. Sommer and H. Wittig, Nucl. Phys. B491 (1997) 344. M. Guagnelli, K. Jansen and R. Petronzio, CERN preprint, CERN-TH/99-10, hep-lat/9801016. R. Sommer, Nucl. Phys. B 411 (1994) 839. M. Guagnelli, R. Sommer and H. Wittig, Nucl. Phys. B535 (1998) 389. M. Lüscher, S. Sint, R. Sommer, P. Weisz and U. Wolff, Nucl. Phys. B491 (1997) 232. [^1]: Heisenberg Foundation Fellow [^2]: In the following we choose $T = L$. [^3]: We remark that the invariance holds with respect to a change of the “ultraviolet” scale $\mu$ and not of the “infrared” scale $\mu_0$.

ArXiv

--- abstract: 'We provide a simple, self-contained proof of a trapped surface formation theorem that sharpens the previous results both of Christodoulou and An-Luk. Our argument is based on a systematic extension of the scale-critical arguments in An-Luk, to connect Christodoulou’s short-pulse method and Klainerman-Rodnianski’s signature counting argument to the peeling properties previously used in small-data results such as Klainerman-Nicolo. This in particular allows us to avoid elliptic estimates and geometric renormalizations, and gives us our new technical simplifications.' address: 'Department of Mathematics, National University of Singapore, Singapore 119076' author: - Xinliang An title: | A Scale-Critical Trapped Surface Formation Criterion:\ A New Proof via Signature for Decay Rates --- i ł[(]{} §[S\_[u,]{}]{} ø Ø Introduction ============ Background ---------- In this paper, we study the evolution of Einstein vacuum equations $$\label{1.1} \mbox{Ric}_{\mu\nu}=0$$ for a (3+1) dimensional Lorentzian manifold $(\mathcal{M}, g)$. (0,2)–(1,1)–(2,2)–(0,4)–(0,2); (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_u$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{u}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); We will introduce coordinates $u$ and ${\underline{u}}$ in $(\mathcal{M}, g)$ through a *double null foliation*[^1]. With coordinates $u, {\underline{u}}$, characteristic initial data will be prescribed along incoming null hypersurface $\Hb_0$, where ${\underline{u}}=0$, and outgoing null hypersurface $H_{u_{\infty}}$, where $u=u_{\infty}$. If the characteristic initial data are small enough, by Christodoulou-Klainerman’s monumental work [@Chr-Kl] we have [completeness of all forward geodesics]{}, which [implies]{} that no singularity would form in the light gray region above. On the other hand, if the initial data are large, in their domain of influence (gray region above) a geometric object, *trapped surface*[^2], may form dynamically. In 1965, Penrose proved the celebrated incompleteness theorem: (Penrose [@Penrose]) For spacetime $(\mathcal{M}, g)$ containing a non-compact Cauchy hypersurface and $g$ satisfying (\[1.1\]), if $M$ contains a compact trapped surface, then it is future causally geodesically incomplete. Therefore, in this setting, proving singularity formation in general relativity is transferred into deriving trapped surface formation. And it is crucial to design initial data prescribed along $\Hb_0$ and $H_{u_{\infty}}$. In order to form a trapped surface, according to stability of Minkowski, the initial data picked cannot be small. Moreover we cannot prescribe spherically symmetric data along both $\Hb_0$ and $H_{u_{\infty}}$ either. This is due to a classic theorem of Birkhoff: spherically symmetric Einstein vacuum spacetimes must be either (flat) Minkowskian[^3][^4] or (static) Schwarzschild[^5] metrics. Hence, *large and non-spherically symmetric* initial data are required. At the same time, solving (\[1.1\]) with large initial data is really hard. For general large data problem, for the evolution of Einstein vacuum equations, we only have local existence result. However, forming a trapped surface at a later time requires [ a mathematical result *beyond* local existence]{}. These render the problem of trapped surface formation to be a really hard one. And it was open for a long time.\ In 2008, Christodoulou solved this long-standing open problem with a 587-page monumental work [@Chr:book]. He designed an open set of large initial data, which have a special structure, called *short pulse* ansatz. In particular, this ansatz allows one to consider a hierarchy of large and small quantities, parametrized by a small parameter $\d$. For initial data these quantities behave differently, being of various sizes in term of $\d$. And their sizes form a hierarchy. But for each quantity, surprisingly, its size is almost preserved by the nonlinear evolution. Therefore, once this hierarchy is designed for initial data, it remains for later time. With this philosophy, despite being a large data problem, a long time global existence theorem can be established. Moreover, these initial conditions indeed lead to trapped-surface formation in the future of the characteristic initial data prescribed along $\Hb_0$ and $H_{u_\infty}$.\ Einstein vacuum equations are a nonlinear hyperbolic system, containing many unknowns. Christodoulou controlled all of them term by term. Later, two systematical approaches by Klainerman-Rodnianski [@KR:Trapped] and An [@An:Trapped] were provided to simplify Christodoulou’s main result in [@Chr:book]. In [@KR:Trapped], an index $s_1$ called *signature for short pulse* is introduced. With this index, Klainerman and Rodnianski systematically tracked the $\d$-weights used in the estimates. And they gave a simplified and shorter proof of $\d$-hierarchy’s almost preserving in a finite region. In [@Chr:book], besides $\d$-weights, Christodoulou also employed weights related to decay and prove his main theorem that *a trapped surface could form dynamically with initial data prescribed arbitrary dispersed at past null infinity*. In [@An:Trapped], An introduced a new index $s_2$ called *signature for decay rates*. With the help of this new index, An extended Klainerman and Rodnianski’s result [@KR:Trapped] from a finite region to an infinite region and re-proved Christodoulou’s main theorem in [@Chr:book] with around 120 pages. The proof in [@An:Trapped] [ is still quite long because of]{}: - : Even with the systematical approach as in [@KR:Trapped], there are still a few anomalous terms in $\d$-weights. To deal with $\d$-anomaly, it takes some pages. Moreover, the $\d$-anomaly would be more severe when using more angular derivatives. Hence in [@KR:Trapped] Klainerman and Rodnianski tried to use the least amount of angular derivatives and they didn’t use angular derives with order higher than 2. - : In [@An:Trapped], An wanted to re-prove the main result in [@Chr:book] with the same amount of angular derivatives used in Christodoulou’s proof. In both [@Chr:book] and [@An:Trapped], two angular derivatives of curvature components are employed. For energy estimates with such limited angular derives, to avoid losing of derivatives we have to go through an additional technical-and-difficult section *elliptic estimates for the third derivatives of Ricci coefficients*. This also prolongs the proof. In this paper, we find new ways to avoid both obstructions: New Ingredients --------------- 1. In Einstein vacuum spacetimes, peeling ties conformal compactification and plays an important role in small data problems (see [@KN:peeling] and reference therein). In this paper, via systematically capturing peeling properties with signature for decay rates $s_2$, we find that peeling would also be vital for problems in large data regime.\ 2. In all preceding works, a colored region on the left in the below is considered, where $\d$ is a small positive parameter and all a priori estimates are established with $\d$ and $|u|$ weights. For example, for geometric quantities ${\hat{\chi}}, \rho$ (to be defined in Section \[define geometric quantities\]) we have $$\|{\hat{\chi}}\|_{L^{\infty}(\S)}\leq \f{\d^{-\f12}}{|u|} \,\, \mbox{in \cite{Chr:book, KR:Trapped}, and} \,\, \|\rho\|_{L^{\infty}(\S)}\leq \f{\d a}{|u|^3} \,\, \mbox{in \cite{A-L}.}$$ In this paper, with a large positive universal number $a$, we consider a different spacetime region (the colored region on the right). And for the characteristic initial data, we construct a new hierarchy based on geometric peeling properties: we design new weighted norms, and the weights are only depending on index $s_2$ (*signature for decay rates*), which was introduced by An in [@An:Trapped]. Since we don’t use parameter $\d$, we don’t need the index $s_1$ (signature for short-pulse) any more. *With these new norms and new approach, we can avoid all the $\d$-anomaly.* (This overcomes .) (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-1}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{\delta}({\underline{u}}=\delta)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); 3. In this paper, we also employ and generalize a direct method [introduced in [@Hol] and used in [@An:Trapped] and other papers]{} for deriving energy estimates. The direct method is based on pairing second Bianchi equations $D_{[\iota}R_{\nu\tau]\varphi\lambda}=0$, putting suitable weight for each equation, and doing integration by parts to cancel the borderline terms.[^6] This approach works well even for higher order energy estimates for Einstein vacuum equations, since potential borderline terms are cancelled and there is no new type of borderline term popping up. In this paper, we adopt and generalize this method and give a *systematical* approach for deriving energy estimates with angular [derivatives]{} of any (high) order. This enables us to use Sobolev’s inequality directly. *We avoid all technical and long calculations for elliptic estimates.* (This overcomes .) In [@A-A] we are extending the method and result of this paper to the Einstein-Maxwell system. And there we notice that, even for Einstein-Maxwell system, the additional elliptic-estimate part cannot be avoided. The simplification of avoiding elliptic estimates in this paper is because of the special Ricci-flat structures of Einstein vacuum equations. 4. In both [@KR:Trapped] and [@An:Trapped], the following Hölder’s inequality in scale invariant norms is crucial. $$\|\phi_1\cdot\phi_2\|_{L_{sc}^{2}(\S)}\leq\frac{\d^{\f12}}{|u|}\|\phi_1\|_{L_{sc}^{\infty}(S)}\|\phi_2\|_{L_{sc}^{2}(\S)}.$$ This inequality tells us if all terms are normal (their scale invariant norms are of size $1$), the nonlinear terms are lower order compared with linear terms. Hence, in the proof we only need to track the linear terms and few anomalous terms, which reduces the workload significantly. For $|u|\geq 1$, the smallness gained in above inequality is coming from $\d$ being sufficient small.\ While for here, when rewriting Hölder’s inequality in the new scale invariant norms, we have (see \[Holder’s\]) $$\label{new holder} \|\phi_1\cdot\phi_2\|_{L_{sc}^{2}(\S)}\leq\frac{1}{|u|}\|\phi_1\|_{L_{sc}^{\infty}(S)}\|\phi_2\|_{L_{sc}^{2}(\S)}.$$ For $|u|\geq a/4$ and $a$ being sufficiently large, the smallness gained in (\[new holder\]) is coming from $|u|$ weight. In another word, in the new spacetime region, *the peeling property (encoded in scale invariant norms through the signature for decay rates $s_2$) provides the crucial gain of smallness. And signature $s_2$ captures the information of geometric peeling properties in a systematical way.*\ Putting all the ingredients together, in Section \[sec setting\]-Section \[TSF\] we obtain a very direct and short (self-contained) proof, showing that a trapped surface ($S_{-a/4, 1}$) could form in evolution.*The new ansatz and hierarchies designed in this paper are interesting extensions of the established short-pulse method.*\ 5. The results above are also related to a scale-critical theorem near the center. In Section \[sec rescale\] we observe a new coordinate transformation (rescaling). Under this rescaling, we establish a correspondence between the following two regions: (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=\d u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{\d a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=\d)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); We can translate mathematical results in one picture (left) into the other (right). Since all the estimates derived in Section \[sec setting\]-Section \[TSF\] are uniform for $u_{\infty}$, in the right picture we could keep $\d$ and let $u_{\infty}\rightarrow -\infty$, which gives *a scale-critical trapped surface formation criterion from past null infinity*. Here $S_{-\d a/4, \, \d}$ is a tiny trapped surface formed with radius $\d a$. *This is the (sharp) scale critical extension of Christodoulou’s monumental work [@Chr:book]*. If we let $a=\d^{-1}$, we then recover Christodoulou’s main theorem in [@Chr:book].\ 6. In [@A-L], there are three parameters $\{a, b, \d\}$ satisfying $1\ll b\leq \at \leq \d^{-\f12}$. With renormalization techniques, in [@A-L] An and Luk derived results scale-critical for $\d$ and also sharp for $a$. If we let $b=\at$, in Section \[rescale bounds\] we will see that *by using signature $s_2$ (peeling property), the new approach in this paper would not only systematically sharpen a priori estimates obtained by An-Luk in [@A-L], but also it avoids the technical geometric renormalizations in [@A-L] completely.* This paper serves as a more intrinsic and more concise reproof and extension of [@A-L] (assuming $b=\at$).\ For a note on the development of this direction, [by designing and employing a different hierarchy]{}, in [@A-L] An and Luk [improved [@Chr:book] and]{} proved the first scale-critical result for Einstein vacuum equations. With the same small parameter $\d$, with relatively larger initial data Christodoulou formed a trapped surface of radius 1; while with much smaller initial data An and Luk formed a trapped surface of radius $\d a$, where $a$ is a universal large constant like $1000$.[^7] [An and Luk want to form a tiny trapped surface with radius $\d a$]{}, hence they have to deal with the region very close to [the center]{}. In this region all the geometric quantities have growth rates. To bounded these growth rates, they employed weighted estimates as well as several crucial geometric renormalizations. Since [@A-L] is scale critical, one can keep $a$ as a universal constant and let $\d\rightarrow 0$. Hence a series of trapped surfaces (with radius shrinking to $0$) are obtained. In [@An:; @AH], An further explored this idea. Together with an elliptic approach to identify the boundary, An showed that a whole black hole region could emerge dynamically from just a “point" $O$ in the spacetime. For an open set of initial data, this boundary (apparent horizon) is proved to be smooth and spacelike except at $O$. The second law of black hole mechanics is further verified and a conjecture of Ashtekar was proved in [@An:; @AH].\ 7. Since An-Luk only dealed with finite spacetime regions in [@A-L], the main result (Theorem \[thm3\]) in this paper could also be viewed as an extension of [@A-L]. (There is another way to extend An-Luk [@A-L] by doing a rescaling. In Section \[rescale and comparison\] we outline that approach and make the comparison.)\ In summary, the approach in this paper synthesizes new ideas outlined above and captures the geometric structures of Einstein vacuum equations in a systematical way via signature for decay rates $s_2$. *The intrinsic[^8] peeling property plays a crucial role!* Now the new proof of trapped surface formation is self-contained and is less than 50 pages. It simplifies and extends Christodoulou’s monumental work [@Chr:book] to a (sharp) scale-critical result. It gives another proof, a systematical improvement and an extension of one of the main conclusions in [@A-L] by An-Luk. It also has a few very interesting applications [@An:TSA]. Other Related Results --------------------- Besides the results described above, many other improvement or extensions of [@Chr:book] have also been achieved. In [@Chr:book], Christodoulou required both a homogenous upper and a homogenous lower bound for his short-pulse initial data. The upper bound ensures the semi-global existence of Einstein vacuum equations up to the region, where a trapped surface is about to emerge. The homogenous lower bound is used to confirm trapped-surface formation. With the same initial data upper bound as in [@Chr:book], in [@K-L-R] Klainerman, Luk and Rodnianski relaxed the lower bound requirement vastly. They replaced *inf* by *sup* and obtained a remarkable anisotropic result. For Einstein vacuum equations, interested readers are also referred to [@Dafermos; @Le; @L-Y; @L-R; @R-S] and references therein. For Einstein equations coupled with matter, Yu [@Yu1; @Yu2] extended the result of [@KR:Trapped] and obtained similar results for Einstein-Maxwell system with signature for short-pulse. In a recent paper [@L-L] by Li and Liu, they studied Einstein-scalar field system and an almost scale-critical trapped surface formation criterion is achieved.\ Next, we start to explain the physical intuition behind trapped surface formation. Heuristic Argument {#heuristic} ------------------ We consider a spacetime region foliated by incoming and outgoing null hypersurfaces, i.e. $\Hb_{{\underline{u}}}$ and $H_u$, respectively. Here $\Hb_{{\underline{u}}}$ and $H_u$ are level sets of two optical functions, which satisfy $$g^{\mu\nu}\partial_{\mu}u\partial_{\nu}u=0, \quad \mbox{and} \quad g^{\mu\nu}\partial_{\mu}{\underline{u}}\partial_{\nu}{\underline{u}}=0.$$ For the colored region, we have $u_{\infty}\leq u \leq -a/4 <0$ and $0\leq {\underline{u}}\leq 1$. Here each point ($S_{u,{\underline{u}}}=H_u \cap \Hb_{{\underline{u}}}$) in the Penrose’s diagram is corresponding to a 2-sphere.\ (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_u$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{\delta}({\underline{u}}=\delta)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); Let $e_3$ and $e_4$ be null vectors and be tangent to each $\Hb_{{\underline{u}}}$ and $H_u$, respectively. Moreover, we require $g(e_3, e_4)=-2$. These $\{e_3,e_4\}$ are called a null pair.\ On each $\S$, we also define [$\{e_A, e_B\}_{A, B=1,2}$ an arbitrary tangent frame on it.]{}\ We then define null second fundamental forms $\chi_{AB}$, ${\underline{\chi}}_{AB}$ associated with $\S$: $$\chi_{AB}:=g(D_{e_A}e_4, e_B), \quad {\underline{\chi}}_{AB}:=g(D_{e_A}e_3, e_B).$$ We further decompose $\chi_{AB}$ and ${\underline{\chi}}_{AB}$ into trace part ${\mbox{tr}}\chi$, ${\mbox{tr}}{\underline{\chi}}$ and traceless part ${\hat{\chi}}_{AB}, \,\, {\hat{{\underline{\chi}}}}_{AB}$: $$\chi_{AB}=\f12{\mbox{tr}}\chi \cdot \gamma_{AB}+{\hat{\chi}}_{AB}, \quad \quad {\underline{\chi}}_{AB}=\f12{\mbox{tr}}{\underline{\chi}}\cdot \gamma_{AB}+{\hat{{\underline{\chi}}}}_{AB},$$ where $\gamma_{AB}$ is the induced metric on $\S$. A *trapped surface* is a $2$-sphere, of which both null expansions are negative, i.e. $${\mbox{tr}}\chi<0 \, \, \mbox{\textit{and}} \, \, {\mbox{tr}}{\underline{\chi}}<0 \, \, \mbox{\textit{hold pointwisely on}} \, \, \S.$$ We will prescribe Minkowkian data along $\Hb_0$, i.e. each $S_{u,0}$ is a standard $2$-sphere embedded in Minkowski spacetime with radius $|u|$. For Minkowskian data, we have $${\mbox{tr}}\chi(u,0)=\frac{2}{|u|}, \quad {\mbox{tr}}{\underline{\chi}}(u,0)=-\frac{2}{|u|}.$$ It is easy to show that ${\mbox{tr}}{\underline{\chi}}$ is always negative: for initial data along $H_{u_{\infty}}$, we have ${\mbox{tr}}{\underline{\chi}}(u_{\infty},{\underline{u}})=-{2}/{|u_{\infty}|}+l.o.t.\footnote{In this article, we use $l.o.t.$ to mean lower order terms.}<0$. Moreover ${\mbox{tr}}{\underline{\chi}}$ is decreasing along $e_3$ direction $$\nab_3{\mbox{tr}}{\underline{\chi}}=-\f12({\mbox{tr}}{\underline{\chi}})^2-|{\hat{{\underline{\chi}}}}|^2+l.o.t.,$$ this implies ${\mbox{tr}}{\underline{\chi}}<0$ in the whole colored region. For $\chi$, from $\mbox{Ric}_{44}=0$, we have two transport equations: $$\label{trapped surface 1} \nabla_4{\mbox{tr}}\chi+\frac{1}{2}({\mbox{tr}}\chi)^2=-|\hat{\chi}|^{2}+l.o.t.,$$ and $$\label{trapped surface 2} \nabla_3\hat{\chi}+\frac{1}{2}{\mbox{tr}}{\underline{\chi}}\hat{{\chi}}=l.o.t..$$ Using $\nab_4 {\mbox{tr}}\chi\leq -|{\hat{\chi}}|^2,$ we have $${\mbox{tr}}\chi(u,{\underline{u}})\leq {\mbox{tr}}\chi(u,0)-\int_0^{{\underline{u}}}|{\hat{\chi}}|^2(u,{{\underline{u}}}')d{{\underline{u}}}'=\frac{2}{|u|}-\int_0^{{\underline{u}}}|{\hat{\chi}}|^2(u,{{\underline{u}}}')d{{\underline{u}}}'.$$ With the derived fact ${\mbox{tr}}{\underline{\chi}}=-{2}/{|u|}+l.o.t.$, would [ imply]{} $$|u|^2|{\hat{\chi}}|^2(u,{\underline{u}})= |u_{\infty}|^2|{\hat{\chi}}|^2(u_{\infty},{\underline{u}})+l.o.t.$$ These imply that along $H_{-a/4}$ $$\label{TP1} \begin{split} {\mbox{tr}}\chi(-\f14a,{\underline{u}})\leq& {\mbox{tr}}\chi(-\f14a,0)-\int_0^{{\underline{u}}}|{\hat{\chi}}|^2(-\f14a,{{\underline{u}}}')d{{\underline{u}}}'+l.o.t.\\ =&\frac{2}{|\f14 a|}-\frac{|u_{\infty}|^2}{|\f14 a|^2}\int_0^{{\underline{u}}}|{\hat{\chi}}|^2(u_{\infty},{{\underline{u}}}')d{{\underline{u}}}'+l.o.t. \end{split}$$ If we choose ${\hat{\chi}}(u_{\infty}, {\underline{u}})$ along $H_{u_{\infty}}$ such that $$\label{chih infinity} \|{\hat{\chi}}\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})} \approx \f{\at}{|u_{\infty}|}, \quad \mbox{and} \quad |u_{\infty}|^2\int_0^{1}|{\hat{\chi}}|^2(u_{\infty}, {\underline{u}}')d{\underline{u}}'\geq a.$$ Then from (\[TP1\]) we arrive at $${\mbox{tr}}\chi(-\f14a,1)\leq\frac{2}{|\f14 a|}-\frac{|u_{\infty}|^2}{|\f14 a|^2}\int_0^{1}|{\hat{\chi}}|^2(u_{\infty},{{\underline{u}}}')d{{\underline{u}}}'+l.o.t.<\f{8}{a}-\f{16}{a}+l.o.t.< 0.$$ Hence, $S_{-a/4, 1}$ is a trapped surface. In the argument above, choosing ${\hat{\chi}}$ which satisfies (\[chih infinity\]) is crucial. We make the following choice $$\label{chih initial} |{\hat{\chi}}|(u_{\infty}, {\underline{u}})\approx \at/|u_{\infty}|,$$ which will provide a new hierarchy (in terms of $a$ and $u$) for all geometric components. At the same time, to rigorously verify this heuristic argument, we need to overcome two main difficulties: 1. We need to make sure that all lower order terms listed above are truly negligible compared with main terms. Since Einstein vacuum equations are a coupled system of many geometric quantities, this requires detailed understandings of nonlinear interaction of all geometric quantities. 2. We need to prove a global-existence theorem in the large data regime. The physical intuition behind is that *focusing of gravitational waves leads to trapped-surface formation*. With arbitrary dispersed data at past null infinity, we need to ensure that the gravitational radiation can go sufficiently far inside from past null infinity. From PDE point of view, this means to establish a global existence result for Einstein vacuum equations (\[1.1\]) with no symmetric assumption. This will be a large data problem for an energy super-critical hyperbolic system.\ In the below we outline an approach, which overcomes these two difficulties. Geometric Quantities and Signature for Decay Rates {#define geometric quantities} -------------------------------------------------- In our dynamical spacetime, different curvature components and Ricci coefficients would behave distinguishingly. We hence decompose them with respect to a null frame $e_3, e_4$ and a frame $e_1, e_2$ tangent to the 2-sphere $S_{u,{\underline{u}}}$. Denote the indices $A,B$ to be $1,2$. With frames $\{e_3, e_4, e_A, e_B\}$, we define the (null) curvature components: $$\begin{split} \a_{AB}&=R(e_A, e_4, e_B, e_4),\quad \, \,\, \ab_{AB}=R(e_A, e_3, e_B, e_3),\\ \b_A&= \frac 1 2 R(e_A, e_4, e_3, e_4) ,\quad \bb_A =\frac 1 2 R(e_A, e_3, e_3, e_4),\\ \rho&=\frac 1 4 R(e_4,e_3, e_4, e_3),\quad \sigma=\frac 1 4 \,^*R(e_4,e_3, e_4, e_3). \end{split}$$ Here $\, ^*R$ is the Hodge dual of $R$. Denote $D_A:=D_{e_{A}}$. We define Ricci coefficients: $$\begin{split} &\chi_{AB}=g(D_A e_4,e_B),\, \,\, \quad {\underline{\chi}}_{AB}=g(D_A e_3,e_B),\\ &\eta_A=-\frac 12 g(D_3 e_A,e_4),\quad \etab_A=-\frac 12 g(D_4 e_A,e_3),\\ &\omega=-\frac 14 g(D_4 e_3,e_4),\quad\,\,\, \omegab=-\frac 14 g(D_3 e_4,e_3),\\ &\zeta_A=\frac 1 2 g(D_A e_4,e_3), \end{split}$$ We further decompose $\chi$ and ${\underline{\chi}}$ into trace and traceless part. Denote ${\hat{\chi}}_{AB}$ and ${\hat{{\underline{\chi}}}}_{AB}$ are the traceless part of $\chi_{AB}$ and ${\underline{\chi}}_{AB}$ respectively. To capture the information of their behaviours, to each $\phi\in \{\a,\beta,\rho,\sigma, {\underline{\beta}},\ab, \chi,{\underline{\chi}},\zeta,\eta,{\underline{\eta}},\omega,{\underline{\omega}}\}$ we assign a number called *signatures* $s_2(\phi)$ to it. The rule is the following: $$\label{signature 21} s_2(\phi):=0\cdot{N_4}(\phi)+0.5\cdot{N_A}(\phi)+1\cdot{N_3}(\phi)-1.$$ Here $N_4(\phi)$ is the number of times $e_4$ appears in the definition of $\phi$. Similarly we define $N_3(\phi)$ and $N_A(\phi)$ where $A=1,2$. For example, in the definition for $\etab_A=-\f12 g(D_4 e_A, e_3)$, we have one $e_4$, one $e_A$ and one $e_3$. Hence $$N_4(\etab_A)=1,\quad N_A(\etab_A)=1, \quad N_3(\etab_A)=1.$$ According to (\[signature 21\]), $\etab_A$ has signature $$s_2(\etab_A)=0\cdot 1+0.5\cdot 1+1\cdot 1-1=0.5.$$ Similarly, for $\chi_{AB}=g(D_A e_4, e_B)$ we have $$N_4(\chi_{AB})=1, \quad N_A(\chi_{AB})=2, \quad N_3(\chi_{AB})=0.$$ Hence (\[signature 21\]) implies $$s_2(\chi_{AB})=0\cdot 1+0.5\cdot 2+1\cdot 0-1=0.$$ Gather these signatures, we have the *signature table*:\ ------- ---------- --------- -------- ---------- --------------------- ---------------------- -------- ---------- --------- -------- -------------------- ----------------------------- -------------------------- ---------------------- -- -- $\alpha$ $\beta$ $\rho$ $\sigma$ $\underline{\beta}$ $\underline{\alpha}$ $\chi$ $\omega$ $\zeta$ $\eta$ $\underline{\eta}$ $\mbox{tr}\underline{\chi}$ $\hat{\underline{\chi}}$ $\underline{\omega}$ $s_2$ 0 0.5 1 1 1.5 2 0 0 0.5 0.5 0.5 1 1 1 ------- ---------- --------- -------- ---------- --------------------- ---------------------- -------- ---------- --------- -------- -------------------- ----------------------------- -------------------------- ---------------------- -- -- \ Based on signature $s_2(\phi)$, we then define *scale invariant norms*: $$\label{si norms} \begin{split} \|\phi\|_{L^{\infty}_{sc}(\S)}:=&a^{-s_2(\phi)}|u|^{2s_2(\phi)+1}\|\phi\|_{L^{\infty}(\S)},\\ \|\phi\|_{L^{2}_{sc}(\S)}:=&a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^{2}(\S)}. \end{split}$$ A main reason for using scale invariant norms is that for most geometric quantities $\phi$, we will show that $\|\phi\|_{L^{\infty}_{sc}(\S)}$ and $\|\phi\|_{L^{2}_{sc}(\S)}$ are of size $1$. Later we call these $\phi$ normal terms. Through the definitions in (\[si norms\]), the $a$-weights and $u$-weights are naturally built in the norms. Furthermore, one important identity holds for nonlinear interactions. From the definition of signature[^9], we have $$s_2(\phi_1\cdot\phi_2)=s_2(\phi_1)+s_2(\phi_2).$$ With it, we could rewrite Hölder’s inequality in scale invariant norms and obtain $$\|\phi_1\cdot\phi_2\|_{L^2_{sc}(\S)}\leq \f{1}{|u|}\|\phi_1\|_{L^{\infty}_{sc}(\S)}\|\phi_2\|_{L^{2}_{sc}(\S)}.$$ In the spacetime region studied, we have $1/|u|\leq 1/\at \ll 1.$ Since $a$ is a large universal number, the above inequality tells us, if all the terms are normal, then the nonlinear interactions can be treated as lower order terms. Therefore, only rare anomalous terms are left for further analysis. Using signature $s_2$ will also help a lot in deriving (higher order) energy estimates, which are the core of the global existence result. Main Results ------------ In Sections \[secbasic\]-\[energy estimate\] we will first derive (An Existence Result near Past Null Infinity)\[main.thm1\] (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); Given $\M I^{(0)}$, there exists a sufficiently large $a_0=a_0(\M I^{(0)})$. For $0<a_0<a$, with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|\nab^{k}_4(|u_{\infty}|\nab)^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}$ along $u=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}=0$, Einstein vacuum equations (\[1.1\]) admit a unique smooth solution in the colored region: $$u_{\infty}\leq u \leq -a/4, \quad 0\leq {\underline{u}}\leq 1.$$ We then verify the heuristic argument with estimates derived in Sections \[secbasic\]-\[energy estimate\]. In Section \[TSF\], we prove (Formation of Trapped Surfaces)\[main.thm2\] (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); Given $\M I^{(0)}$, there exists a sufficiently large $a_0=a_0(\M I^{(0)})$. For $0<a_0<a$, solving Einstein vacuum equations (\[1.1\]) with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|\nab^{k}_4(|u_{\infty}|\nab)^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}$ along $u=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}=0$, - $\int_0^1|u_{\infty}|^2|{\hat{\chi}}_0|^2(u_{\infty}, {\underline{u}}')d{\underline{u}}'\geq a$ for every direction along $u=u_{\infty}$, we have that $S_{-a/4,1}$ is a trapped surface. In Section \[sec rescale\], we will describe a new coordinate transformation. With it we convert above results into our main conclusion (A Scale-Critical Trapped Surface Formation Criterion from Past Null Infinity)\[thm3\] (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{\d a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{\d}({\underline{u}}=\d)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); Given $\M I^{(0)}$, for fixed $\d$ there exists a sufficiently large $a_0=a_0(\M I^{(0)},\d)$. For $0<a_0<a$, solving Einstein vacuum equations (\[1.1\]) with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|(\d\nab_4)^k(|u_{\infty}|\nab)^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}$ along $u=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}=0$, - $\int_0^{\delta}|u_{\infty}|^2|{\hat{\chi}}_0|^2(u_{\infty}, {\underline{u}}')d{\underline{u}}'\geq \d a$ for every direction along $u=u_{\infty}$, we have that $S_{-\d a/4,\d}$ is a trapped surface. Theorem \[thm3\] is an extension of An-Luk [@A-L] to allow characteristic initial data prescribed at very far away (at $u=u_{\infty}$). At the same time, Theorem \[thm3\] could also be viewed as a scale-critical extension of Christodoulou [@Chr:book]. In Theorem \[thm3\], if we further choose $a=4c\cdot\d^{-1}$, where $0< c\leq 1$ being of size $1$, we then have \[Corollary1.5\] (Recovery of Christodoulou’s monumental work [@Chr:book]) (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-c}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{\d}({\underline{u}}=\d)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); Given $\M I^{(0)}$ and constant $c$ (where $0<c \leq 1$ being of size $1$), there exists a sufficiently small $\d_0=\d_0(\M I^{(0)},c)$. For $0<\d<\d_0\ll c$, solving Einstein vacuum equations (\[1.1\]) with initial data: - $\sum_{i\leq 10, k\leq 3}\d^{\frac12}\|(\d\nab_4)^k(|u_{\infty}|\nab)^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}$ along $u=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}=0$, - $\int_0^{\delta}|u_{\infty}|^2|{\hat{\chi}}_0|^2(u_{\infty}, {\underline{u}}')d{\underline{u}}'\geq 4c$ for every direction along $u=u_{\infty}$, we have that $S_{-c, \d}$ is a trapped surface. Note: we could also obtain a priori bounds that are in line with [@Chr:book] by Christodoulou. Setting, Equations and Notations {#sec setting} ================================ Double Null Foliation {#secdnf} --------------------- We construct a double null foliation in a neighborhood of $S_{u_{\infty},0}$ as follows: (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (2,2)–node \[midway,sloped,above,black\] [$\Hb_1({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(2,-2); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)– node\[midway,sloped,above,black\][$H_{u}$]{}(2,2); (3.3,-0.6)– node\[midway, sloped, above,black\][$L'$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$L'$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][${\underline{L}}'$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][${\underline{L}}'$]{}(1.7,0.4); Starting from a point $p$ on $2$-sphere $S_{u_{\infty}, 0}$, in $2$-dimensional $T_p^{\perp}S_{u_{\infty}, 0}$, we could find two future-directed vectors $L'_p, \underline{L}'_p$ such that $$g(L'_p, L'_p)=0, \, \, g(\underline{L}'_p, \underline{L}'_p)=0, \, \, g(L'_p, {\underline{L}}'_p)=-2.\footnote{ $\{L'_p, {\underline{L}}'_p\}$ are uniquely determined up to a scaling factor $\lambda>0$: \quad $\{L'_p, {\underline{L}}'_p\}\rightarrow \{\lambda L'_p, \lambda^{-1}{\underline{L}}'_p\}$. }$$ Based on $p$ and along $L'_p$ direction, a unique geodesic $l_p$ is sent out. We extend $L'$ along $l_p$ such that $D_{L'}L'=0$. We then have $l_p$ is null. This is because $g(L'_p, L'_p)=0$ and $$L'(g(L', L'))=2g(D_{L'}L', L')=0.$$ We hence have $g(L',L')=0$ along $l_p$. Gathering all the $\{l_p\}$ together, we then have an outgoing null hypersurface called $H_{u_{\infty}}$. Similarly, we obtain the incoming null hypersurface $\Hb_0$ emitting from $S_{u_{\infty},0}$.\ Note that, by above construction, for each point $q$ on $H_{u_{\infty}}$ or $\Hb_0$, in $T_q H_{u_{\infty}}$ or $T_q \Hb_0$, there is the preferred null vector $L'_q$ or ${\underline{L}}'_q$ associated with $q$.\ We define function $\O$ to be 1 on $S_{u_{\infty},0}$ and extend $\O$ as a continuous function along $H_{u_{\infty}}$ and $\Hb_0$. [^10] We consider vector fields $$L=\O^2 L' \, \,\, \mbox{along}\, \, \, H_{u_{\infty}}, \,\, \mbox{and} \, \, {\underline{L}}=\O^2 {\underline{L}}' \, \, \,\mbox{along}\, \, \, \Hb_0,$$ and define functions $${\underline{u}}\,\, \mbox{on} \,\, H_{u_{\infty}}\,\,\, \mbox{satisfying}\, \, \, L{\underline{u}}=1 \,\, \mbox{and} \,\, {\underline{u}}=0\, \,\mbox{on} \,\, S_{u_{\infty}, 0},$$ $$u \, \, \mbox{on} \, \, \Hb_0 \, \, \, \mbox{satisfying} \, \, \,{\underline{L}}u=1 \, \, \mbox{and} \, \, u=u_{\infty}\, \,\mbox{on} \, \, S_{u_{\infty}, 0}.$$ Let $S_{u_{\infty}, {\underline{u}}'}$ be the embedded $2$-surface on $H_{u_{\infty}}$, such that ${\underline{u}}={\underline{u}}'$. Similarly, define $S_{u', 0}$ to be the embedded $2$-surface on $\Hb_0$, such that $u=u'$. At each point $q$ on $2$-surface $S_{u_{\infty}, {\underline{u}}'}$, we already have the preferred outgoing null vector $L'_q$ tangent to $H_{u_{\infty}}$. Hence, at $q$ we can also fix a unique incoming null vector ${\underline{L}}_q'$ via requiring $$g({\underline{L}}'_q, {\underline{L}}'_q)=0 \quad \mbox{and} \quad g({\underline{L}}'_q, L'_q)=-2\O^{-2}|_q.$$ There exists a unique geodesic $\underline{l}_q$ emitting from $q$ with direction ${\underline{L}}'$. We then extend ${\underline{L}}'$ along $\underline{l}_q$ through $D_{{\underline{L}}'}{\underline{L}}'=0$. Gathering all the $\{\underline{l}_q\}$ for $q\in S_{u_{\infty}, {\underline{u}}'}$, we have constructed the incoming null hypersurface $\Hb_{{\underline{u}}'}$ emitting from $S_{u_{\infty}, {\underline{u}}'}$. Similarly, from $S_{u',0}$ we also construct the outgoing null hypersurface $H_{u'}$. We further define $2$-sphere $S_{u', {\underline{u}}'}:=H_{u'}\cap \Hb_{{\underline{u}}'}$. At each point $p$ of $S_{u',{\underline{u}}'}$, we define positive-valued function $\O$ via $$\label{define Omega} g(L'_p, {\underline{L}}'_p)=:-2\O^{-2}| _p.$$ Note $L'_p$ is well-defined on $H_{u'}$, along an outgoing null geodesic $l$ passing through $p$; ${\underline{L}}'_p$ is also well-defined on $\Hb_{{\underline{u}}'}$, along an incoming null geodesic $\underline{l}$ crossing $p$.\ These $3$-dimensional incoming null hypersurfaces $\{\Hb_{{\underline{u}}'}\}_{0\leq {\underline{u}}' \leq 1}$ and outgoing null hypersurfaces $\{H_{u'}\}_{u_{\infty}\leq u' \leq -a/4}$ together with their intersections $S_{u', {\underline{u}}'}=H_{u'}\cap \Hb_{{\underline{u}}'}$ give us the so called *double null foliation*.\ On $\S$, by (\[define Omega\]) we have $g(L', {\underline{L}}')=-2\O^{-2}$. Thus, $g(\O L', \O {\underline{L}}')=-2$. Throughout this paper we will work with the normalized null pair $(e_3, e_4)$: $$e_3:=\Omega{\underline{L}}', \quad e_4:=\Omega L', \,\, \mbox{and} \quad g(e_3, e_4)=-2.$$ Moreover, for the characteristic initial data, we choose the following gauge: $$\Omega\equiv 1 \quad\mbox{on $H_{u_{\infty}}$ and $\Hb_0$}.$$ Functions $u$ and ${\underline{u}}$ defined above also satisfy the eikonal equations $$g^{\mu\nu}\partial_\mu u\partial_\nu u=0,\quad g^{\mu\nu}\partial_\mu{\underline{u}}\partial_\nu {\underline{u}}=0.$$ And it is straight forward to check $$L'^\mu=-2g^{\mu\nu}\partial_\nu u,\quad {\underline{L}}'^\mu=-2g^{\mu\nu}\partial_\nu {\underline{u}}, \quad L{\underline{u}}=1, \quad {\underline{L}}u=1.$$ Here ${\underline{L}}:=\Omega^2{\underline{L}}', \quad L:=\Omega^2 L'$ are also called equivariant vector fields. The Coordinate System {#coordinates} --------------------- We will use a coordinate system $(u,{\underline{u}}, \theta^1, \theta^2)$. Here $u$ and ${\underline{u}}$ are defined above. To get $(\theta^1, \theta^2)$ for each point on $\S$, we follow the approach in Chapter 1 of [@Chr:book]: we first define a coordinate system $(\theta^1, \theta^2)$ on $S_{u_{\infty},0}$. Since $S_{u_{\infty},0}$ is the standard $2$-sphere in Minkowskian spacetime, here we use the coordinates of stereographic projection. Then we extend this coordinate system to $\Hb_0$ by requiring $${{\mathcal L} \mkern-10mu /\,}_{{\underline{L}}} \theta^A=0\mbox{ on $\Hb_0$}. \footnote{On $\Hb_0$, we have $\O=1$ and ${{\mathcal L} \mkern-10mu /\,}_{{\underline{L}}} \theta^A=\f{\partial}{\partial u}\theta^A$.}$$ Here ${{\mathcal L} \mkern-10mu /\,}_L$ is the restriction of the Lie derivative to $TS_{u,{\underline{u}}}$. In another word, given a point $p$ on $S_{u_{\infty}, 0}$, assuming $l_p$ is the incoming null geodesics on $\Hb_0$ emitting from $p$, then all the points along $l_p$ have the same angular coordinate $(\theta^1, \theta^2)$. We further extend this coordinate system from $\Hb_0$ to the whole spacetime under requirement $${{\mathcal L} \mkern-10mu /\,}_L \th^A=0,$$ i.e. all the points along the same outgoing null geodesics (along $L$) on $H_u$ have the same angular coordinate. We hence have established a coordinate system in a neighborhood of $S_{u_{\infty}, 0}$. With this coordinate system, we can rewrite $e_3$ and $e_4$ as $$e_3=\Omega^{-1}\left(\frac{\partial}{\partial u}+d^A\frac{\partial}{\partial \th^A}\right), e_4=\Omega^{-1}\frac{\partial}{\partial {\underline{u}}}.$$ And the Lorentzian metric $g$ takes the form $$\label{equation g} g=-2\O^2(du\otimes d{\underline{u}}+d{\underline{u}}\otimes du)+\gamma_{AB}(d\theta^A-d^A du)\otimes(d\theta^B-d^B du).$$ We require [$d^A$ to satisfy $d^A=0$]{} on $\Hb_0$. Equations {#seceqn} --------- We then decompose curvature components and Ricci coefficients with respect to null frames $e_3, e_4$ and frames $e_1, e_2$ tangent to the 2-sphere $S_{u,{\underline{u}}}$. Denote the indices $A,B$ to be $1,2$. With frame $\{e_3, e_4, e_A, e_B\}$, we define null curvature components: $$\label{def curvatures} \begin{split} \a_{AB}&=R(e_A, e_4, e_B, e_4),\quad \, \,\, \ab_{AB}=R(e_A, e_3, e_B, e_3),\\ \b_A&= \frac 1 2 R(e_A, e_4, e_3, e_4) ,\quad \bb_A =\frac 1 2 R(e_A, e_3, e_3, e_4),\\ \rho&=\frac 1 4 R(e_4,e_3, e_4, e_3),\quad \sigma=\frac 1 4 \,^*R(e_4,e_3, e_4, e_3). \end{split}$$ Here $\, ^*R$ stands for the Hodge dual of $R$. Denote $D_A:=D_{e_{A}}$. We define Ricci coefficients: $$\label{def Ricci coefficients} \begin{split} &\chi_{AB}=g(D_A e_4,e_B),\, \,\, \quad {\underline{\chi}}_{AB}=g(D_A e_3,e_B),\\ &\eta_A=-\frac 12 g(D_3 e_A,e_4),\quad \etab_A=-\frac 12 g(D_4 e_A,e_3),\\ &\omega=-\frac 14 g(D_4 e_3,e_4),\quad\,\,\, \omegab=-\frac 14 g(D_3 e_4,e_3),\\ &\zeta_A=\frac 1 2 g(D_A e_4,e_3). \end{split}$$ Let $\gamma_{AB}$ be the induced metric on $\S$, we further decompose $\chi, {\underline{\chi}}$ into $$\chi_{AB}=\f12{\mbox{tr}}\chi\cdot \gamma_{AB}+{\hat{\chi}}_{AB}, \quad {\underline{\chi}}_{AB}=\f12{\mbox{tr}}{\underline{\chi}}\cdot \gamma_{AB}+{\hat{{\underline{\chi}}}}_{AB}.$$ Here ${\hat{\chi}}_{AB}$ and ${\hat{{\underline{\chi}}}}_{AB}$ are the corresponding traceless parts.\ Denote $\nab$ to be the induced covariant derivative operator on $S_{u,{\underline{u}}}$. And let $\nab_3$ and $\nab_4$ to be the projections of covariant derivatives $D_3$ and $D_4$ to $S_{u,{\underline{u}}}$. By the definitions of Ricci coefficients, one can verify: $$\begin{split} &\omega=-\frac 12 \nab_4 (\log\Omega),\qquad \omegab=-\frac 12 \nab_3 (\log\Omega),\\ &\eta_A=\zeta_A +\nab_A (\log\Omega),\quad \etab_A=-\zeta_A+\nab_A (\log\Omega). \end{split}$$\ We then define several different contractions between tensors. Let $$(\phi^{(1)}\hot\phi^{(2)})_{AB}:=\phi^{(1)}_A\phi^{(2)}_B+\phi^{(1)}_B\phi^{(2)}_A-\gamma_{AB}(\phi^{(1)}\cdot\phi^{(2)}) \quad\mbox{for one forms $\phi^{(1)}_A$, $\phi^{(2)}_A$,}$$ $$(\phi^{(1)}\wedge\phi^{(2)})_{AB}:={{\epsilon} \mkern-8mu /\,}^{AB}(\gamma^{-1})^{CD}\phi^{(1)}_{AC}\phi^{(2)}_{BD}\quad\mbox{for symmetric $2$-tensors $\phi^{(1)}_{AB}$, $\phi^{(2)}_{AB}$}.$$ Here ${{\epsilon} \mkern-8mu /\,}$ is the volume form associated to the metric $\gamma$. For simplicity, we employ $\phi^{(1)}\cdot\phi^{(2)}$ to represent an arbitrary contraction of the tensor product of $\phi^{(1)}$ and $\phi^{(2)}$ with respect to the metric $\gamma$. We also use ${\mbox{div }}$, ${\mbox{curl }}$ and ${\mbox{tr}}$ operators. For totally symmetric tensors, define these operators by $$({\mbox{div }}\phi)_{A_1...A_r}:=\nabla^B\phi_{BA_1...A_r},$$ $$({\mbox{curl }}\phi)_{A_1...A_r}:={{\epsilon} \mkern-8mu /\,}^{BC}\nabla_B\phi_{CA_1...A_r},$$ $$(\mbox{tr}\phi)_{A_1...A_{r-1}}:=(\gamma^{-1})^{BC}\phi_{BCA_1...A_{r-1}}.$$ [ We also define by $^*$ for $1$-forms and symmetric $2$-tensors respectively as follows (note that on $1$-forms this is the Hodge dual on $\S$): $${^*}\phi_A:=\gamma_{AC}{{\epsilon} \mkern-8mu /\,}^{CB}\phi_B,$$ $${^*}\phi_{AB}:=\gamma_{BD}{{\epsilon} \mkern-8mu /\,}^{DC}\phi_{AC}.$$ And define the operator $\nab\hot$ on a $1$-form $\phi_A$ by $$(\nab\hot\phi)_{AB}:=\nab_A\phi_B+\nab_B\phi_A-\gamma_{AB}{\mbox{div }}\phi.$$ ]{}\ We are ready to state the transport equations for curvature components and Ricci coefficients. Rewrite the second Bianchi equations $D_{[\iota}R_{\nu\tau]\varphi\lambda}=0$ with null frames, we arrive at $$\label{eq:null.Bianchi} \begin{split} &\nab_3\alpha+\frac 12 {\mbox{tr} {\underline{\chi}}}\alpha=\nabla\hot \beta+ 4\omegab\alpha-3({\hat{\chi}}\rho+^*{\hat{\chi}}\sigma)+ (\zeta+4\eta)\hot\beta, \\ &\nab_4\beta+2\trch\beta = {\mbox{div }}\alpha - 2\omega\beta + \eta \alpha,\\ &\nab_3\beta+{\mbox{tr} {\underline{\chi}}}\beta=\nabla\rho + 2\omegab \beta +^*\nabla\sigma +2{\hat{\chi}}\cdot\betab+3(\eta\rho+^*\eta\sigma),\\ &\nab_4\sigma+\frac 32\trch\sigma=-{\mbox{div }}^*\beta+\frac 12{\hat{{\underline{\chi}}}}\cdot ^*\alpha-\zeta\cdot^*\beta-2\etab\cdot ^*\beta,\\ &\nab_3\sigma+\frac 32{\mbox{tr} {\underline{\chi}}}\sigma=-{\mbox{div }}^*\betab+\frac 12{\hat{\chi}}\cdot ^*\alphab-\zeta\cdot ^*\betab-2\eta\cdot ^*\betab,\\ &\nab_4\rho+\frac 32\trch\rho={\mbox{div }}\beta-\frac 12{\hat{{\underline{\chi}}}}\cdot\alpha+\zeta\cdot\beta+2\etab\cdot\beta,\\ &\nab_3\rho+\frac 32{\mbox{tr} {\underline{\chi}}}\rho=-{\mbox{div }}\betab- \frac 12{\hat{\chi}}\cdot\alphab+\zeta\cdot\betab-2\eta\cdot\betab,\\ &\nab_4\betab+\trch\betab=-\nabla\rho +^*\nabla\sigma+ 2\omega\betab +2{\hat{{\underline{\chi}}}}\cdot\beta-3(\etab\rho-^*\etab\sigma),\\ &\nab_3\betab+2{\mbox{tr} {\underline{\chi}}}\, \betab=-{\mbox{div }}\alphab-2\omegab\betab+\etab \cdot\alphab,\\ &\nab_4\alphab+\frac 12 \trch\alphab=-\nabla\hot \betab+ 4\omega\alphab-3({\hat{{\underline{\chi}}}}\rho-^*{\hat{{\underline{\chi}}}}\sigma)+ (\zeta-4\etab)\hot \betab. \end{split}$$ Here $^*$ denotes the Hodge dual on $S_{u,{\underline{u}}}$. The above transport equations for curvature are called *null Bianchi equations*. We then rewrite $\mbox{Ric}_{\mu\nu}=0$ with null frames. For $\chi$ and ${\underline{\chi}}$ we have $$\label{null.str1} \begin{split} \nab_4 \trch+\frac 12 (\trch)^2&=-|{\hat{\chi}}|^2-2\omega \trch,\\ \nab_4{\hat{\chi}}+\trch {\hat{\chi}}&=-2 \omega {\hat{\chi}}-\alpha,\\ \nab_3 {\mbox{tr} {\underline{\chi}}}+\frac 12 ({\mbox{tr} {\underline{\chi}}})^2&=-2\omegab {\mbox{tr} {\underline{\chi}}}-|{\hat{{\underline{\chi}}}}|^2,\\ \nab_3{\hat{{\underline{\chi}}}}+ {\mbox{tr} {\underline{\chi}}}\, {\hat{{\underline{\chi}}}}&= -2\omegab {\hat{{\underline{\chi}}}}-\alphab,\\ \nab_4 {\mbox{tr} {\underline{\chi}}}+\frac1 2 \trch {\mbox{tr} {\underline{\chi}}}&=2\omega {\mbox{tr} {\underline{\chi}}}+2\rho- {\hat{\chi}}\cdot{\hat{{\underline{\chi}}}}+2{\mbox{div }}\etab +2|\etab|^2,\\ \nab_4{\hat{{\underline{\chi}}}}+\frac 1 2 \trch {\hat{{\underline{\chi}}}}&=\nab\widehat{\otimes} \etab+2\omega {\hat{{\underline{\chi}}}}-\frac 12 {\mbox{tr} {\underline{\chi}}}{\hat{\chi}}+\etab\widehat{\otimes} \etab,\\ \nab_3 \trch+\frac1 2 {\mbox{tr} {\underline{\chi}}}\trch &=2\omegab \trch+2\rho- {\hat{\chi}}\cdot{\hat{{\underline{\chi}}}}+2{\mbox{div }}\eta+2|\eta|^2,\\ \nab_3{\hat{\chi}}+\frac 1 2 {\mbox{tr} {\underline{\chi}}}{\hat{\chi}}&=\nab\widehat{\otimes} \eta+2\omegab {\hat{\chi}}-\frac 12 \trch {\hat{{\underline{\chi}}}}+\eta\widehat{\otimes} \eta. \end{split}$$ For the remaining Ricci coefficients, we have $$\label{null.str2} \begin{split} \nabla_4\eta&=-\chi\cdot(\eta-\etab)-\b,\\ \nabla_3\etab &=-{\underline{\chi}}\cdot (\etab-\eta)+\bb,\\ \nabla_4\omegab&=2\omega\omegab+\frac 34 |\eta-\etab|^2-\frac 14 (\eta-\etab)\cdot (\eta+\etab)- \frac 18 |\eta+\etab|^2+\frac 12 \rho,\\ \nabla_3\omega&=2\omega\omegab+\frac 34 |\eta-\etab|^2+\frac 14 (\eta-\etab)\cdot (\eta+\etab)- \frac 18 |\eta+\etab|^2+\frac 12 \rho.\\ \end{split}$$ These above transport equations for Ricci coefficients are call *null structure equations*. In this article, we will also need another form of equation $\nab_4{\mbox{tr}}{\underline{\chi}}$ $$\label{eqn 4 trchib} \begin{split} &\nab_4({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})+\f12{\mbox{tr}}\chi({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})\\ =&\f{1}{|u|}{\mbox{tr}}\chi-\f{4}{|u|}\o+2\o({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})+2\rho-{\hat{\chi}}\cdot{\hat{{\underline{\chi}}}}+2{\mbox{div }}{\underline{\eta}}+2|{\underline{\eta}}|^2\\ =&\f12({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|}){\mbox{tr}}\chi-\f12{\mbox{tr}}{\underline{\chi}}{\mbox{tr}}\chi+2{\mbox{tr}}{\underline{\chi}}\o+2\rho-{\hat{\chi}}\cdot{\hat{{\underline{\chi}}}}+2{\mbox{div }}{\underline{\eta}}+2|{\underline{\eta}}|^2, \end{split}$$ and another form of $\nab_3{\mbox{tr}}{\underline{\chi}}$ $$\label{eqn 2 trchib} \begin{split} \nab_3({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})+{\mbox{tr}}{\underline{\chi}}({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})=\f{2}{|u|^2}(\O^{-1}-1)+\f12({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})-2{\underline{\omega}}{\mbox{tr}}{\underline{\chi}}-|{\hat{{\underline{\chi}}}}|^2. \end{split}$$ When embedding $\S$ into $4$-dimensional spacetime, we have Gauss-Codazzi equations and in null frames we have $$\label{null.str3} \begin{split} {\mbox{div }}{\hat{\chi}}&=\frac 12 \nabla \trch - \frac 12 (\eta-\etab)\cdot ({\hat{\chi}}-\frac 1 2 \trch\cdot\gamma) -\beta,\\ {\mbox{div }}{\hat{{\underline{\chi}}}}&=\frac 12 \nabla {\mbox{tr} {\underline{\chi}}}+ \frac 12 (\eta-\etab)\cdot ({\hat{{\underline{\chi}}}}-\frac 1 2 {\mbox{tr} {\underline{\chi}}}\cdot\gamma) +\betab,\\ {\mbox{curl }}\eta &=-{\mbox{curl }}\etab=\sigma +\frac 1 2{\hat{{\underline{\chi}}}}\wedge{\hat{\chi}},\\ K&=-\rho+\frac 1 2 {\hat{\chi}}\cdot{\hat{{\underline{\chi}}}}-\frac 1 4 \trch {\mbox{tr} {\underline{\chi}}}. \end{split}$$ Here $K$ is Gaussian curvature of spheres $S_{u,{\underline{u}}}$. Integration ----------- Let $U$ be a coordinate patch on $\S$. Denote $p_U$ to be the corresponding partition of unity. For a function $\phi$, we define its integration on $S_{u,{\underline{u}}}$, $H_u$ and $\Hb_{{\underline{u}}}$ via $$\label{int S} \int_{S_{u,{\underline{u}}}} \phi :=\sum_U \int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\phi\cdot p_U\cdot \sqrt{\det\gamma}\,d\th^1 d\th^2,$$ $$\int_{H_{u}^{(0,{\underline{u}})}} \phi :=\sum_U \int_0^{{\underline{u}}}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\phi\cdot 2p_U\cdot\Omega\cdot\sqrt{\det\gamma}\,d\th^1 d\th^2d{\underline{u}}',$$ $$\int_{H_{{\underline{u}}}^{({u_{\infty}},u)}} \phi :=\sum_U \int_{u_{\infty}}^{u}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\phi\cdot 2p_U\cdot\Omega\cdot\sqrt{\det\gamma}\,d\th^1 d\th^2du'.$$ Let $D_{u ,{\underline{u}}}$ be the region $u_{\infty}\leq u'\leq u$, $0\leq {\underline{u}}'\leq {\underline{u}}$. We define the integration of $\phi$ in $D_{u,{\underline{u}}}$ as $$\begin{split} \int_{D_{u,{\underline{u}}}} \phi :=&\sum_U \int_{u_{\infty}}^u\int_0^{{\underline{u}}}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\phi\cdot p_U\cdot \O^2\cdot\sqrt{-\det g}\, d\th^1 d\th^2 du' d{\underline{u}}'.\\ \end{split}$$ We further define the $L^p$ ($1\leq p < \infty$) norms for an arbitrary tensorfield $\phi$: $$||\phi||_{L^p(S_{u,{\underline{u}}})}^p:=\int_{S_{u,{\underline{u}}}} <\phi,\phi>_\gamma^{p/2},$$ $$||\phi||_{L^p(H_u)}^p:=\int_{H_{u}} <\phi,\phi>_\gamma^{p/2},$$ $$||\phi||_{L^p(\Hb_{{\underline{u}}})}^p:=\int_{\Hb_{{\underline{u}}}} <\phi,\phi>_\gamma^{p/2}.$$ When $p=\infty$, we define the $L^\infty$ norm by $$||\phi||_{L^\infty(S_{u,{\underline{u}}})}:=\sup_{\th\in S_{u,{\underline{u}}}} <\phi,\phi>_\gamma^{1/2}(\th).$$ We also employ mixed-type norms in this paper: $$||\phi||_{L^2_{{\underline{u}}}L^\infty_u L^p(\S)}:=\left(\int_0^{1}(\sup_{u_{\infty}\leq u \leq -\f{a}{4}}||\phi||_{L^p(S_{u,{\underline{u}}'})})^2d{\underline{u}}'\right)^{\frac{1}{2}},$$ $$||\phi||_{L^2_{u}L^\infty_{{\underline{u}}} L^p(\S)}:=\left(\int_{u_{\infty}}^{-\f{a}{4}}(\sup_{0\leq {\underline{u}}\leq 1}||\phi||_{L^p(S_{u',{\underline{u}}})})^2du'\right)^{\frac{1}{2}}.$$ In this paper the following Minkowski’s inequality will be used frequently: $$||\phi||_{L^\infty_{{\underline{u}}}L^2_{u} L^p(\S)}\leq||\phi||_{L^2_{u}L^\infty_{{\underline{u}}} L^p(\S)}.$$ Definition of Signatures {#Signatures} ------------------------ As explained in heuristics, we want to prescribe ${\hat{\chi}}$ such that $$|{\hat{\chi}}|(u_{\infty}, {\underline{u}})\approx \f{\at}{|u_{\infty}|} \quad \mbox{ along } \quad H_{u_{\infty}}.$$ Following the same procedures explained in details in Chapter 2 of [@Chr:book], we obtain the following estimates on $H_{u_{\infty}}$: $$\label{initial hierarchy} \begin{split} &|\a|\lesssim \f{\at}{|u_{\infty}|}, \quad |\beta|\lesssim \f{\at}{|u_{\infty}|^2}, \quad |\rho|\lesssim \f{a}{|u_{\infty}|^3},\\ &|\sigma|\lesssim \f{a}{|u_{\infty}|^3}, \quad |{\underline{\beta}}|\lesssim \f{a}{|u_{\infty}|^4}, \quad |\ab|\lesssim \f{a^{\f32}}{|u_{\infty}|^5},\\ &|\o|\lesssim \f{1}{|u_{\infty}|}, \quad |{\mbox{tr}}\chi|\lesssim \f{1}{|u_{\infty}|}, \quad |\eta|\lesssim \f{\at}{|u_{\infty}|^2}, \quad |{\underline{\eta}}|\lesssim \f{\at}{|u_{\infty}|^2},\\ &|{\mbox{tr}}{\underline{\chi}}-\f{2}{u_{\infty}}|\lesssim \f{a}{|u_{\infty}|^3}\lesssim \f{1}{|u_{\infty}|^2}, \quad |{\underline{\omega}}|\lesssim \f{a}{|u_{\infty}|^3}, \quad |{\hat{{\underline{\chi}}}}|\lesssim \f{\at}{|u_{\infty}|^2}. \end{split}$$ Note that near $S_{u_{\infty}, 0}$ all geometric quantities have decay rates and *they obey peeling property* (see [@KN:peeling]). In PDE estimates, it will be hard to track these $|u|$ and $a$ weights term by term. We hope to design a “*scale invariant norm* -$L^{\infty}_{sc}(\S)$” with $|u|$ and $a$ weights built in, such that for most geometric quantities $\phi$, we have $$\|\phi\|_{L^{\infty}_{sc}(\S)}\lesssim 1.$$ To achieve this, first we need to find some connections between the definitions of various geometric quantities in (\[def curvatures\]), (\[def Ricci coefficients\]) and the $|u_{\infty}|$-weights, $a$-weights listed above. By relaxing the above estimate for ${\underline{\beta}}\mbox{ and } \ab$ $$\label{relax beb ab} \begin{split} &\mbox{from }\quad |{\underline{\beta}}|\lesssim \f{a}{|u_{\infty}|^4} \quad \mbox{to } \quad |{\underline{\beta}}|\lesssim \f{a^{\f32}}{|u_{\infty}|^4},\\ &\mbox{from }\quad |\ab|\lesssim \f{a^{\f32}}{|u_{\infty}|^5} \quad \mbox{to } \quad |\ab|\lesssim \f{a^2}{|u_{\infty}|^5}, \end{split}$$ and keeping the other estimates for now, we find a systematical way to define $L^{\infty}_{sc}(\S)$. We first introduce *signature for decay rates*: to $\phi\in \{\a,\beta,\rho,\sigma, K, {\underline{\beta}},\ab, \chi,{\underline{\chi}},\zeta,\eta,{\underline{\eta}},\omega,{\underline{\omega}}, \gamma\}$, we assign signatures $s_2(\phi)$ according to the rule: $$\label{signature 2} s_2(\phi):=0\cdot{N_4}(\phi)+\frac{1}{2}\cdot{N_a}(\phi)+1\cdot{N_3}(\phi)-1.$$ $N_4(\phi)$ is the number of times $e_4$ appears in the definition of $\phi$. Similarly we define $N_3(\phi)$ and $N_a(\phi)$ where $a=1,2$. Following the definition, we then have the *signature table*\ ------- ---------- --------- -------- ---------- ----- --------------------- ---------------------- -------- ---------- --------- -------- -------------------- ----------------------------- -------------------------- ---------------------- ---------- $\alpha$ $\beta$ $\rho$ $\sigma$ $K$ $\underline{\beta}$ $\underline{\alpha}$ $\chi$ $\omega$ $\zeta$ $\eta$ $\underline{\eta}$ $\mbox{tr}\underline{\chi}$ $\hat{\underline{\chi}}$ $\underline{\omega}$ $\gamma$ $s_2$ 0 0.5 1 1 1 1.5 2 0 0 0.5 0.5 0.5 1 1 1 0 ------- ---------- --------- -------- ---------- ----- --------------------- ---------------------- -------- ---------- --------- -------- -------------------- ----------------------------- -------------------------- ---------------------- ---------- \ With above definition, we also have $$\label{signature of derivative} s_2(\nabla_4{\phi})=s_2(\phi), \quad s_2(\nabla{\phi})=s_2(\phi)+\frac{1}{2}, \quad s_2(\nabla_3{\phi})=s_2(\phi)+1.$$ Scale Invariant Norms {#Scale invariant norms} --------------------- For any horizontal tensor-field $\phi$ with signature $s_2(\phi)$, we further define *scale invariant norms on $S_{u,{\underline{u}}}$*: $$\label{scale invariant norms} \begin{split} \|\phi\|_{L_{sc}^{\infty}(\S)}:=&a^{-s_2(\phi)}|u|^{2s_2(\phi)+1}\|\phi\|_{L^{\infty}(\S)},\\ \|\phi\|_{L_{sc}^{2}(\S)}:=&a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^{2}(\S)},\\ \|\phi\|_{L^1_{sc}(\S)}:=&a^{-s_2(\phi)}|u|^{2s_2(\phi)-1}\|\phi\|_{L^1(\S)}. \end{split}$$ For convenience, along $H_u^{(0,{\underline{u}})}$ and $\Hb_{{\underline{u}}}^{(u_{\infty},u)}$ we also define *scale invariant norms along null hypersurfaces* $$\label{scale invariant norms 2} \begin{split} \|\phi\|^2_{L^2_{sc}(H_u^{(0,\underline{u})})}:=& \int_0^{\underline{u}}\|\phi\|^2_{L^2_{sc}(S_{u,\underline{u}'})}d\underline{u}',\\ \|\phi\|^2_{L^2_{sc}(\underline{H}_{\underline{u}}^{(u_{\infty},u)})}:=& \int_{u_{\infty}}^{u}{\frac{a}{|u'|^2}}\|\phi\|^2_{L^2_{sc}(S_{u',\underline{u}})}du'. \end{split}$$ Let $\phi\in \{\beta,\rho,\sigma, {\underline{\beta}},\ab, {\mbox{tr}}\chi, {\mbox{tr}}{\underline{\chi}}+\f{2}{|u|}, \eta,{\underline{\eta}},\omega,{\underline{\omega}}\}$. After relaxing estimates for ${\underline{\beta}}, \ab$ as in (\[relax beb ab\]), along $H_{u_{\infty}}$, (\[initial hierarchy\]) could be rewritten in scale invariant norms $$\label{initial scale invariant} \begin{split} \f{1}{\at}\|{\hat{\chi}}\|_{L^{\infty}_{sc}(S_{u_{\infty},{\underline{u}}})}&+\f{\at}{|u_{\infty}|}\|{\hat{{\underline{\chi}}}}\|_{L^{\infty}_{sc}(S_{u_{\infty},{\underline{u}}})}+\f{a}{|u_{\infty}|^2}\|{\mbox{tr}}{\underline{\chi}}\|_{L^{\infty}_{sc}(S_{u_{\infty},{\underline{u}}})}\\ &+\f{1}{\at}\|\a\|_{L^{\infty}_{sc}(S_{u_{\infty},{\underline{u}}})}+\|\phi\|_{L^{\infty}_{sc}(S_{u_{\infty},{\underline{u}}})}\leq 1. \end{split}$$ For most geometric terms, their scale invariant norms are of size 1. But for ${\hat{\chi}}$ and $\a$ it requires an additional smallness $1/\at$, for ${\hat{{\underline{\chi}}}}$ it requires an additional smallness $\at/|u_{\infty}|$ and for ${\mbox{tr}}{\underline{\chi}}$ it requires an additional smallness $a/|u_{\infty}|^2$. We hence call ${\hat{\chi}}, {\hat{{\underline{\chi}}}}, {\mbox{tr}}{\underline{\chi}}, \a$ *anomalous terms*. Conservation of Signatures -------------------------- A key property of signature $s_2$ is that the induced metric $\gamma_{ab}$ on $\S$ satisfies $s_2(\gamma_{ab})=0$. This ensures *signature conservation*: $$\label{signature conservation 2} s_2(\phi_1\cdot{\phi_2})=s_2(\phi_1)+s_2(\phi_2).$$ For example, we have one of the null structure equations $$\label{nab 3 o} \nabla_3{\omega}=2\omega\underline{\omega}+\frac{3}{4}|\eta-\underline{\eta}|^2+\frac{1}{4}(\eta-\underline{\eta})\cdot(\eta+\underline{\eta})-\frac{1}{8}|\eta+\underline{\eta}|^2+\frac{1}{2}\rho.$$ From signature table and (\[signature of derivative\]), it can be read that all the nonlinear terms and linear terms have the same signature $s_2$, that is 1: $$s_2(\o{\underline{\omega}})=s_2(\omega)+s_2({\underline{\omega}})=0+1=1, \quad s_2(\eta\cdot\eta)=s_2(\eta)+s_2(\eta)=\f12+\f12=1,$$ $$s_2(\eta\cdot{\underline{\eta}})=s_2(\eta)+s_2({\underline{\eta}})=\f12+\f12=1, \quad s_2({\underline{\eta}}\cdot{\underline{\eta}})=s_2({\underline{\eta}})+s_2({\underline{\eta}})=\f12+\f12=1,$$ $$s_2(\nab_3\o)=s_2(\o)+1=0+1=1, \quad s_2(\rho)=1.$$ This delightful fact is true not only for the equation of $\nab_3 \o$, but also true for all null structure equations, null Bianchi equations and constrain equations [^11]. When using scale invariant norms, this key feature enables us to treat all the terms on the right hand side of (\[nab 3 o\]) as one term, since they share the same signature $s_2$. Moreover, when using scale invariant norms to rewrite Hölder’s inequalities, we get Hölder’s Inequality in Scale Invariant Norms -------------------------------------------- $$\label{Holder's} \begin{split} \|\phi_1\cdot\phi_2\|_{L_{sc}^{2}(\S)}\leq&\frac{1}{|u|}\|\phi_1\|_{L_{sc}^{\infty}(S)}\|\phi_2\|_{L_{sc}^{2}(\S)},\\ \|\phi_1\cdot\phi_2\|_{L^1_{sc}(\S)}\leq&\f{1}{|u|}\|\phi_1\|_{L^{\infty}_{sc}(\S)}\|\phi_2\|_{L^1_{sc}(\S)},\\ \|\phi_1\cdot\phi_2\|_{L^1_{sc}(\S)}\leq&\f{1}{|u|}\|\phi_1\|_{L^{2}_{sc}(\S)}\|\phi_2\|_{L^2_{sc}(\S)}. \end{split}$$ Note in the region of study we have $1/|u|\leq 4/a<<1$. This means if all terms are normal, the nonlinear terms in (\[nab 3 o\]) or in other equations could be treated as lower order terms. This will simply the proof a lot. Norms ----- Here we define norms, which will be used throughout the paper. Let $$\label{psi Psi} \t \psi\in\{\o, {\mbox{tr}}\chi, \eta, {\underline{\eta}}, {\underline{\omega}}\}, \t \Psi\in \{\b,\rho, \sigma,{\underline{\beta}},\ab\}, \t \Psi'\in \{\rho, \sigma,{\underline{\beta}},\ab\}.$$ We also denote ${\widetilde{{\mbox{tr}}{\underline{\chi}}}}:={\mbox{tr}}{\underline{\chi}}+\f{2}{|u|}$. For $0\leq i \leq 6$, we define $$\label{O i infty} \begin{split} \mathcal{O}_{i,\infty}(u,\underline{u}):= &\f{1}{\at}\|(\at\nab)^i{\hat{\chi}}\|_{L^{\infty}_{sc}(\S)}+\|(\at\nab)^i\t \psi\|_{L^{\infty}_{sc}(\S)}+\f{\at}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^{\infty}_{sc}(\S)}\\ &+\f{a}{|u|^2}\|(\at\nab)^i{\mbox{tr}}{\underline{\chi}}\|_{L^{\infty}_{sc}(\S)}+\f{a}{|u|}\|(\at\nab)^i {\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^{\infty}_{sc}(\S)}, \end{split}$$ $$\label{R i infty} \begin{split} \mathcal{R}_{i,\infty}(u,\underline{u}):= \f{1}{\at}\|(\at\nab)^i\a\|_{L^{\infty}_{sc}(\S)}+\|(\at\nab)^i\t \Psi\|_{L^{\infty}_{sc}(\S)}. \end{split}$$ For $0\leq i \leq {\color{black}10}$, we define $$\label{O i 2} \begin{split} \mathcal{O}_{i,2}(u,\underline{u}):=& \f{1}{\at}\|(\at\nab)^i{\hat{\chi}}\|_{L^{2}_{sc}(\S)}+\|(\at\nab)^i\t \psi\|_{L^{2}_{sc}(\S)}+\f{\at}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^{2}_{sc}(\S)}\\ &+\f{a}{|u|^2}\|(\at\nab)^i{\mbox{tr}}{\underline{\chi}}\|_{L^{2}_{sc}(\S)}+\f{a}{|u|}\|(\at\nab)^i {\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^{2}_{sc}(\S)}. \end{split}$$ For $0\leq i \leq 9$, we define $$\label{R i 2} \begin{split} \mathcal{R}_{i,2}(u,\underline{u}):= \f{1}{\at}\|(\at\nab)^i\a\|_{L^{2}_{sc}(\S)}+\|(\at\nab)^i\t \Psi\|_{L^{2}_{sc}(\S)}. \end{split}$$ For $0\leq i \leq 10$, we define $$\label{R i H} \mathcal{R}_i(u,\underline{u}):=\f{1}{\at}\|(\at\nab)^i\alpha\|_{L^{2}_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i\t\Psi\|_{L^{2}_{sc}(H_u^{(0,{\underline{u}})})},$$ $$\label{R i Hb} \mathcal{\underline{R}}_i(u,\underline{u}):=\f{1}{\at}\|(\at\nab)^i \beta\|_{L^{2}_{sc}(\underline{H}_{{\underline{u}}}^{(u_{\infty}, u)})}+\|(\at\nab)^i\t\Psi'\|_{L^{2}_{sc}(\underline{H}_{{\underline{u}}}^{(u_{\infty}, u)})}.$$ We then set $\mathcal{O}_{i,\i},\, \mathcal{O}_{i,2},\, \mathcal{R}_{i,\infty},\, \mathcal{R}_{i, 2},\, \mathcal{R}_i,\, \underline{\mathcal{R}}_i$ to be the supremum over $u,{\underline{u}}$ in our spacetime region of $\mathcal{O}_{i,\i}(u,{\underline{u}}), \mathcal{O}_{i,2}(u,{\underline{u}}), \mathcal{R}_{i,\infty}(u,{\underline{u}}), \mathcal{R}_{i, 2}(u,{\underline{u}}), \mathcal{R}_i(u,{\underline{u}}), \underline{\mathcal{R}}_i(u,{\underline{u}})$, respectively. Finally, we define $\mathcal{O}, \mathcal{R}$: $$\mathcal{O}:=\sum_{i\leq 6}(\mathcal{O}_{i,\infty}+\mathcal{R}_{i,\infty})+\sum_{i\leq 9}(\M O_{i,2}+ \M R_{i,2}),$$ $$\mathcal{R}:=\sum_{i\leq 10}\mathcal{R}_i+\mathcal{\underline{R}}_i.$$ And let $\mathcal{O}^{(0)}, \mathcal{R}^{(0)}, \underline{\M R}^{(0)}$ be the corresponding norms of the initial hypersurfaces $H_{u_{\infty}}$ and $\Hb_0$. Lastly, we define the initial data quantity $$\mathcal{I}^{(0)}:=\sup_{0\leq {\underline{u}}\leq 1}\mathcal{I}^{(0)}({\underline{u}}),$$ where $$\begin{split} \mathcal{I}^{(0)}({\underline{u}}):=&\f{|u_{\infty}| }{\at}\|{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}+\sum_{0\leq k \leq 10,}\sum_{0\leq m \leq 20}\f{1}{\at}\|(|u_{\infty}|\nab)^{m}(\nab_4)^k{\hat{\chi}}_0\|_{L^2(S_{u_{\infty},{\underline{u}}})}. \end{split}$$ Here ${\hat{\chi}}_0$ denotes ${\hat{\chi}}$ along $H_{u_{\infty}}^{(0,{\underline{u}})}$. Notation -------- We collect the notations that are employed for convenience throughout the article: - We denote $\sup_{u,{\underline{u}}}$ to be the supremum over all values of $u,{\underline{u}}$, where $u_{\infty}\leq u\leq -\f{a}{4}$ and $0\leq {\underline{u}}\leq 1$. - If $A$ and $B$ are two quantities, we often use $A\lesssim B$ meaning that there exists a constant $C>0$, which is independent of $a$, such that $A\leq CB$. Whenever there is no danger of confusion, we substitute $\leq$ for $\lesssim$. - For equations involving many terms, the coefficients on the left are kept precise. Whenever there is no danger of confusion, the coefficients on the right are allowed to vary up to a nonzero constant. - We will employ $(\,,\,)$ to denote sum of all terms, which have one of the components in the bracket. For instance, the notation $\phi_1(\phi_2, \phi_3)$ means the sum of all terms in the form of $\phi_1\phi_2$ or $\phi_1\phi_3$. - Denote $D$ to be the spacetime region $\{(u,{\underline{u}})\,\,| \,\, u_{\infty}\leq u \leq -a/4, \quad 0\leq{\underline{u}}\leq 1\}.$ - [ For integers $i_1\geq 0$ and $i_2\geq 1$, sometimes we use $\nab^{i_1}\p^{i_2}$ to express a product of $i_2$ terms: $$\nab^{i_1}\p^{i_2}=\nab^{j_1}\p\cdot\nab^{j_2}\p\cdot\cdot\cdot\nab^{j_{i_2}}\p, \mbox{ where } j_1, j_2,..., j_{i_2}\in \mathbb{N} \mbox{ and } i_1=j_1+j_2+...+j_{i_2}.$$ Here we assume that $j_{i_2}$ is the largest number. ]{} The Preliminary Estimates {#secbasic} ========================= An Approach of Bootstrap ------------------------ In this article, we will employ a bootstrap argument to derive uniform upper bounds of $\M O, \M R, \underline{\M R}$ for nonlinear Einstein vacuum equations. Along $H_{u_{\infty}}$ and $\Hb_{0}$, by analysing characteristic initial data we have $$\label{EBA0} \M O^{(0)}+\M R^{(0)}+\underline{\M R}^{(0)}\ls \M I^{(0)}.$$ Here we have $\M I^{(0)}\lesssim 1$. Our goal is to show that in $D=\{(u,{\underline{u}})\, | \, u_{\infty}\leq u \leq -a/4, \,\, 0\leq {\underline{u}}\leq 1\}$ we have $$\label{EBA1} \M O(u,{\underline{u}})+\M R(u,{\underline{u}})+\underline{\M R}(u,{\underline{u}})\ls \M I^{(0)}+(\M I^{(0)})^2+1.$$ Once these uniform bounds are obtained, by [ characteristic-initial-data local existence result]{}[^12], the solutions can always be extended a bit towards the future direction of $u$. Hence, uniform estimate (\[EBA1\]) for $u_{\infty}\leq u \leq -a/4$ implies global existence of Einstein vacuum equations in $D=\{(u,{\underline{u}})\, | \, u_{\infty}\leq u \leq -a/4, \,\, 0\leq {\underline{u}}\leq 1\}$. To derive the uniform bound (\[EBA1\]), we make bootstrap assumptions $$\label{BA.0} \M O(u, {\underline{u}})\leq O, \quad \M R(u,{\underline{u}}) +\underline{\M R}(u,{\underline{u}}) \leq R.$$ Here $O$ and $R$ are large numbers, such that $$\M I^{(0)}+(\M I^{(0)})^2+1\ll O, \quad \M I^{(0)}+(\M I^{(0)})^2+1\ll R, \quad \mbox{but} \quad (O+R)^{20}\leq {a^{\f{1}{16}}}.$$ We also define $\Upsilon=\{u\,|\,\, u_{\infty}\leq u \leq-a/4 \mbox{ and } (\ref{BA.0}) \mbox{ hold for every } 0\leq {\underline{u}}\leq 1\}$. First, we hope to prove $\Upsilon=[u_{\infty}, -a/4].$ At $u=u_{\infty}$, we have (\[EBA0\]). By continuity of solutions (via local existence), for small $\epsilon>0$ it holds for $u_{\infty}\leq u \leq u_{\infty}+\epsilon$ $$\M O^{(0)}\ls \M I^{(0)}\ll O, \quad \M R^{(0)} +\underline{\M R}^{(0)} \ls \M I^{(0)}\ll R,$$ $$\M O(u, {\underline{u}})\ls 2\,\M I^{(0)}\ll O, \quad \M R(u,{\underline{u}}) +\underline{\M R}(u,{\underline{u}}) \ls 2\,\M I^{(0)}\ll R.$$ This implies $[u_{\infty}, u_{\infty}+\epsilon]\subseteq \Upsilon$ and $\Upsilon$ is not empty. Since $\Upsilon\subseteq [u_{\infty}, -a/4]$, if we are able to prove that $\Upsilon$ is a set being both open and closed, then we prove $\Upsilon= [u_{\infty}, -a/4]$. Closeness follows from uniform estimates and continuity of solutions in $u$ variable, which doesn’t rise a challenge. Efforts are dedicated to verifying that $\Upsilon$ is open. The main parts of this paper are to show that for any $u\in\Upsilon$ we have [^13] $$\M O(u,{\underline{u}})\ls \M I^{(0)}+\M R(u,{\underline{u}})+{\underline{\mathcal{R}}}(u,{\underline{u}})+1+\f{C_1}{a^{\f18}}\cdot (O+R)^{20},\quad \M R(u,{\underline{u}})+\underline{\M R}(u,{\underline{u}})\ls \M I^{(0)}+(\M I^{(0)})^2+1+\f{C_2}{a^{\f18}}\cdot (O+R)^{20}.$$ Here $C_1$ and $C_2$ are integers independent of $a$ and basically count how many terms popping up in the estimates. By employing $(O+R)^{20}\leq {a^{\f{1}{16}}}$ and further requiring $a$ to be sufficiently large, we obtain $$\label{EBA3} \begin{split} &\M R(u,{\underline{u}})+\underline{\M R}(u,{\underline{u}})\ls \M I^{(0)}+(\M I^{(0)})^2+1+\f{C_2}{a^{\f18}}\cdot a^{\f{1}{16}}\ls \M I^{(0)}+(\M I^{(0)})^2+1,\\ &\M O(u,{\underline{u}})\ls \M I^{(0)}+\M R(u,{\underline{u}})+{\underline{\mathcal{R}}}(u,{\underline{u}})+1+\f{C_1}{a^{\f18}}\cdot a^{\f{1}{16}}\ls \M I^{(0)}+(\M I^{(0)})^2+1. \end{split}$$ These are improvements of the upper bounds in bootstrap assumptions (\[BA.0\]): $$\M O(u,{\underline{u}})\leq O, \,\, \M R(u,{\underline{u}})+\underline{\M R}(u,{\underline{u}})\leq R,$$ where $\M I^{(0)}+(\M I^{(0)})^2+1\ll O, \, \, \M I^{(0)}+(\M I^{(0)})^2+1\ll R.$ By continuity of solutions via local existence, $\Upsilon$ could be extended a bit towards larger $u$. This implies $\Upsilon$ being open. Together with $\Upsilon$ being closed and non-empty, we have $\Upsilon=[u_{\infty}, -a/4]$. Thus, for the whole region $D=\{(u,{\underline{u}})\, | \, u_{\infty}\leq u \leq -a/4, \,0\leq {\underline{u}}\leq 1\}$, estimates in (\[BA.0\]) hold. They imply (\[EBA3\]) and bounds in (\[EBA1\]): $$\M O(u,{\underline{u}})+\M R(u,{\underline{u}})+\underline{\M R}(u,{\underline{u}})\ls \M I^{(0)}+(\M I^{(0)})^2+1\, \mbox{ in } \,D.$$ Estimates for Metric Components {#metric} ------------------------------- We derive bound for $\Omega$ first: \[Omega\] Under the assumptions of Theorem \[main.thm1\] and bootstrap assumption , we have $$\|\Omega-1\|_{L^\i(\S)}\ls \f{O}{|u|}.$$ Consider the equation $$\label{Omegatransport} \omega=-\frac{1}{2}\nabla_4\log\Omega=\frac{1}{2}\Omega\nabla_4\Omega^{-1}=\frac{1}{2}\frac{\partial}{\partial {\underline{u}}}\Omega^{-1}.$$ We integrate respect to $d{\underline{u}}$. On $\Hb_0$ we have $\Omega^{-1}=1$ and this leads to $$\label{Omega -1} ||\Omega^{-1}-1||_{L^\infty(S_{u,{\underline{u}}})}\ls \int_0^{{\underline{u}}}||\omega||_{L^\infty(S_{u,{\underline{u}}'})}d{\underline{u}}'\ls \f{O}{|u|}.$$ Here we have used the bootstrap assumption : $$\|\o\|_{L^{\infty}_{sc}(\S)}\leq O \Leftrightarrow \|\o\|_{L^{\infty}(\S)}\leq \f{O}{|u|}.$$ Finally, notice that $$\|\Omega-1\|_{L^\i(\S)}\leq\|\Omega\|_{L^\i(\S)}\|\Omega^{-1}-1\|_{L^\i(\S)}\ls(1+\f{O}{|u|})^{-1}\cdot\f{O}{|u|}\ls \f{O}{|u|}.$$ We then move to control induced metric $\gamma$ on $\S$: \[gamma\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , for metric $\gamma$ on $\S$ we have $$c'\leq \det\gamma\leq C'.$$ Here $C'$ and $c'$ are constants depending only on initial data. Moreover, in $D$ $$|\gamma_{AB}|,|(\gamma^{-1})^{AB}|\leq C'.$$ We employ the first variation formula ${{\mathcal L} \mkern-10mu /\,}_L\gamma=2\Omega\chi.$ In coordinates, it states $$\label{ub gamma} \frac{\partial}{\partial {\underline{u}}}\gamma_{AB}=2\Omega\chi_{AB}.$$ This implies $$\frac{\partial}{\partial {\underline{u}}}\log(\det\gamma)=2\Omega\trch.$$ Let $\gamma_0(u,{\underline{u}},\th^1,\th^2)=\gamma(u,0,\th^1,\th^2)$. Then with $|2\O{\mbox{tr}}\chi|\leq O/|u|$ it follows $$\f{\det \gamma}{\det \gamma_0}=e^{\int_0^{{\underline{u}}}2\O{\mbox{tr}}\chi d{\underline{u}}'}\leq e^{\f{O}{a}}.$$ Via Taylor expansion, this implies $$\label{detgaper} |\det\gamma-\det(\gamma_0)|\leq \det(\gamma_0)|1-e^{\f{O}{a}}|\ls \f{O}{a},$$ which gives lower and upper bound for $\det \gamma$. For $\gamma$, denote $\Lambda$ to be the greater eigenvalue. We have $$\label{La} \Lambda\leq\sup_{A,B=1,2}\gamma_{AB},$$ $$\sum_{A,B=1,2}|\chi_{AB}|\leq\Lambda ||\chi||_{L^\infty(S_{u,{\underline{u}}})},$$ $$|\gamma_{AB}-(\gamma_0)_{AB}|\leq \int_0^{{\underline{u}}}|\chi_{AB}|d{\underline{u}}'\leq\Lambda\f{a^{\f12}}{|u|}O\ls\f{O}{\at}.$$ We hence bound $|\gamma_{AB}|$ from above. We further bound $|(\gamma^{-1})^{AB}|$ from above by using the upper bound for $|\gamma_{AB}|$ and the lower bound for $\det\gamma$. For metric $\gamma$, we will also need the following \[gamma2\] We continue to work under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions . Fix a point $(u,\theta^1, \theta^2)$ on the initial hypersurface $\Hb_0$. Along the outgoing null geodesics emitting from $(u, \theta^1, \theta^2)$, denote $\Lambda({\underline{u}})$ and $\lambda({\underline{u}})$ to be the larger and smaller eigenvalue of $\gamma^{-1}(u, {\underline{u}}=0, \theta^1, \theta^2)\gamma(u, {\underline{u}}, \theta^1, \theta^2)$. Then we have $$|\Lambda({\underline{u}})-1|+|\lambda({\underline{u}})-1|\leq \f{1}{\at}$$. Define $\nu({\underline{u}}):=\sqrt{\f{\Lambda({\underline{u}})}{\lambda({\underline{u}})}}.$ Following the derivation of (5.93) in [@Chr:book], by (\[ub gamma\]) we can derive $$\nu({\underline{u}})\leq 1+\int_0^{{\underline{u}}}|\O{\hat{\chi}}({\underline{u}}')|_{\gamma} \nu({\underline{u}}')d{\underline{u}}'.$$ Via Grönwall’s inequality, this [implies]{} $$\label{nu} |\nu({\underline{u}})|\ls 1 \quad \mbox{ and } \quad |\nu({\underline{u}})-1|\leq \f{\at\cdot O}{|u|^2}\leq \f{O}{a^{\f32}}\leq \f{1}{a}.$$ The desired estimate follows from (\[detgaper\]) and (\[nu\]). The above two propositions also imply \[area\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , in $D$ we have $$\sup_{{\underline{u}}}|\mbox{Area}(S_{u,{\underline{u}}})-\mbox{Area}(S_{u,0})|\leq\f{O^{\f12}}{a^{\f12}}|u|^2.$$ This follows from definition in (\[int S\]) and estimate in . Estimates for Transport Equations {#transportsec} --------------------------------- In latter sections, we will employ following propositions for transport equations: Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , for an $S_{u,{\underline{u}}}$ tangent tensor $\phi$ of arbitrary rank, we have $$\label{transport1} ||\phi||_{L^2(S_{u,{\underline{u}}})}\ls ||\phi||_{L^2(S_{u,{\underline{u}}'})}+\int_{{\underline{u}}'}^{{\underline{u}}} ||\nabla_4\phi||_{L^2(S_{u,{\underline{u}}''})}d{{\underline{u}}''},$$ $$\label{transport3} ||\phi||_{L^2(S_{u,{\underline{u}}})}\ls ||\phi||_{L^2(S_{u',{\underline{u}}})}+\int_{u'}^{u} ||\nabla_3\phi||_{L^2(S_{u'',{\underline{u}}})}d{u''}.$$ Here we first prove (\[transport1\]). For a scalar function $f$, by variation of area formula, we have $$\frac{d}{d{\underline{u}}}\int_{\S} f=\int_{\S} \left(\frac{df}{d{\underline{u}}}+\Omega \trch f\right)=\int_{\S} \Omega\left(e_4(f)+ \trch f\right).$$ Taking $f=|\phi|_{\gamma}^2$, using Cauchy-Schwarz inequality on the sphere and $L^\infty$ bounds for $\Omega$ and $\trch$, we obtain $$2\|\phi\|_{L^2(\S)}\cdot \f{d}{d{\underline{u}}}\|\phi\|_{L^2(\S)}\ls \|\phi\|_{L^2(\S)}\cdot \|\nab_4\phi\|_{L^2(\S)}+\f{O}{|u|}\|\phi\|^2_{L^2(\S)}.$$ This implies $$\f{d}{d{\underline{u}}}\|\phi\|_{L^2(\S)}\ls \|\nab_4\phi\|_{L^2(\S)}+\f{O}{|u|}\|\phi\|_{L^2(\S)}.$$ And (\[transport1\]) can be concluded by applying Grönwall’s inequality for ${\underline{u}}$ variable. Inequality (\[transport3\]) could be proved in [a similar fashion. For a scalar function $f$, we arrive at $${\underline{L}}\int_{\S} f=\int_{\S} \left({\underline{L}}f+\Omega {\mbox{tr}}{\underline{\chi}}f\right)=\int_{\S} \Omega\left(e_3(f)+ {\mbox{tr}}{\underline{\chi}}f\right).$$ Taking $f=|\phi|_{\gamma}^2$, using Cauchy-Schwarz inequality on the sphere and the fact $\O>0, {\mbox{tr}}{\underline{\chi}}<0$, we obtain $$2\|\phi\|_{L^2(\S)}\cdot {\underline{L}}\|\phi\|_{L^2(\S)}\ls \|\phi\|_{L^2(\S)}\cdot \|\nab_3\phi\|_{L^2(\S)}.$$ This implies $ {\underline{L}}\|\phi\|_{L^2(\S)}\ls \|\nab_3\phi\|_{L^2(\S)}$ and (\[transport3\]) follows. ]{} We then rewrite the above inequalities in scale invariant norms: \[transport\] For an $\S$ tangent tensor $\phi$ of arbitrary rank, we have $$\|\phi\|_{L^2_{sc}(S_{u,\underline{u}})}\ls \|\phi\|_{L^2_{sc}(S_{u,0})}+\int_0^{\underline{u}} \|\nabla_4\phi\|_{L^2_{sc}(S_{u,\underline{u}'})}d\underline{u}',$$ $$\|\phi\|_{L^2_{sc}(S_{u,\underline{u}})}\ls \|\phi\|_{L^2_{sc}(S_{u_{\infty},\underline{u}})}+\int_{u_{\infty}}^{u}\frac{a}{|u'|^2}\|\nabla_3\phi\|_{L^2_{sc}(S_{u',\underline{u}})}du'.$$ For $\nab_3$ equations, sometimes we need more precise estimates to deal with borderline terms. Typically, a borderline term contains ${\mbox{tr}}{\underline{\chi}}$. Thus, the coefficients in front of ${\mbox{tr} {\underline{\chi}}}$ play an important role. \[el\] We work under the assumptions of Theorem \[main.thm1\] and bootstrap assumptions . Let $\phi$ and $F$ be $\S$-tangent tensor fields of rank $k$ satisfying the following transport equation: $$\nab_3 \phi_{A_1...A_k}+\lambda_0{{\mbox{tr}}\underline{\chi}}\phi_{A_1...A_k}=F_{A_1...A_k}.$$ Denoting $\lambda_1=2(\lambda_0-\frac{1}{2})$, for $\phi$ we have $$|u|^{\lambda_1}\|\phi\|_{L^{2}(\S)}\lesssim |u_{\infty}|^{\lambda_1}\|\phi\|_{L^{2}(S_{u_{\infty},\underline{u}})}+\int_{u_{\infty}}^u|u'|^{\lambda_1}\|F\|_{L^{2}(S_{u',\underline{u}})}du'.$$ We use variation of area formula for equivariant vector ${\underline{L}}$ [^14] and a scalar function $f$: $${\underline{L}}\int_{\S} f=\int_{\S} \left({\underline{L}}f+\Omega {\mbox{tr} {\underline{\chi}}}f\right)=\int_{\S} \Omega\left(e_3(f)+ {\mbox{tr} {\underline{\chi}}}f\right).$$ With this identity, we obtain $$\label{evolution.id} \begin{split} &{\underline{L}}(\int_{\S}|u|^{2\lambda_1 }|\phi|^{2})\\ =&\int_{\S}\Omega\l -2\lambda_1 |u|^{2\lambda_1-1}(e_3 u)|\phi|^{2}+2|u|^{2\lambda_1}<\phi,\nab_3\phi>+ {\mbox{tr}}\underline{\chi}|u|^{2\lambda_1}|\phi|^{2}\r\\ =&\int_{\S}\Omega\l 2|u|^{2\lambda_1}<\phi, \nab_3\phi+{\lambda_0}{\mbox{tr} {\underline{\chi}}}\phi>\r\\ &+\int_{\S}\Omega |u|^{2\lambda_1}\l -\f{2\lambda_1 (e_3u)}{|u|}+(1-2\lambda_0){\mbox{tr} {\underline{\chi}}}\r|\phi|^2. \end{split}$$ Observe that we have $$\label{trchib additional} \begin{split} &-\f{2\lambda_1 (e_3u)}{|u|}+(1-2\lambda_0){\mbox{tr} {\underline{\chi}}}\\ = &-\f{2\lambda_1 \Omega^{-1}}{|u|}+(1-2\lambda_0){\mbox{tr} {\underline{\chi}}}\\ = &-\f{2\lambda_1 (\Omega^{-1}-1)}{|u|}+(1-2\lambda_0)({\mbox{tr} {\underline{\chi}}}+\f{2}{|u|})-\f{2\lambda_1+2-4\lambda_0}{|u|}\\ \ls &\f{O}{|u|^2}. \end{split}$$ For the last inequality, we employ (\[Omega -1\]), bootstrap assumption $\|{\mbox{tr}}{\underline{\chi}}+\f{2}{|u|}\|_{L^{\infty}(\S)}\leq \f{O}{|u|^2}$ and $\lambda_1=2(\lambda_0-1/2)$. Using Cauchy-Schwarz for the first term and applying Grönwall’s inequality for the second term, we obtain $$\begin{split} &|u|^{\lambda_1}\|\phi\|_{L^2(\S)}\\ \ls &e^{O\|u^{-2}\|_{L^1_u}}\l|u_{\infty}|^{\lambda_1}\|\phi\|_{L^2(S_{u_{\infty},\underline{u}})}+\int_{u_{\infty}}^u |u'|^{\lambda_1}\|F\|_{L^2(S_{u',\underline{u}})}du'\r\\ \ls &|u_{\infty}|^{\lambda_1}\|\phi\|_{L^2(S_{u_{\infty},\underline{u}})}+\int_{u_{\infty}}^u |u'|^{\lambda_1}\|F\|_{L^2(S_{u',\underline{u}})}du'. \end{split}$$ In the last step, we use $O\|u^{-2}\|_{L^1_u}\ls O/a\leq 1$. Sobolev Embedding {#Embedding} ----------------- With the derived estimates for metric $\gamma$, we follow [@Chr:book] to obtain a bound on the isoperimetric constant for a $2$-sphere $S$ $$I(S)=\sup_{\substack{U\subset S\\\partial U \in C^1}} \f{\min\{\mbox{Area}(U),\mbox{Area}(U^c)\}}{(\mbox{Perimeter}(\partial U))^2}.$$ \[isoperimetric\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumption , the isoperimetric constant obeys an upper bound $$I(\S)\leq \f1{\pi},$$ where $u_{\infty}\leq u\leq -a/4$ and $0\leq {\underline{u}}\leq 1$. Fix $u$. For $U_{{\underline{u}}}\subset S_{u,{\underline{u}}}$, we denote $U_0\subset S_{u,0}$ to be the backward image of $U_{{\underline{u}}}$ under the diffeomorphism generated by the equivariant vector $L$. Using Proposition \[gamma\], Proposition \[gamma2\] and their proof, we obtain $$\f{\mbox{Perimeter}(\partial U_{{\underline{u}}})}{\mbox{Perimeter}(\partial U_0)}\geq \sqrt{\inf_{S_{u,0}} \lambda({\underline{u}})}$$ and $$\f{\mbox{Area}(U_{{\underline{u}}})}{\mbox{Area}(U_0)}\leq \sup_{S_{u,0}} \f{\det(\gamma_{{\underline{u}}})}{\det(\gamma_0)},\quad\f{\mbox{Area}(U^c_{{\underline{u}}})}{\mbox{Area}(U^c_0)}\leq \sup_{S_{u,0}} \f{\det(\gamma_{{\underline{u}}})}{\det(\gamma_0)}.$$ The conclusion then follows from the fact that $I(S_{u,0})=\f{1}{2\pi}$ and the bounds in Proposition \[gamma\] and Proposition \[gamma2\]. We will use an $L^2-L^\i$ Sobolev embedding inequality in this article. In order to derive it, we will use two propositions quoted directly from [@Chr:book]: \[Lp\] For any Riemannian $2$-manifold $(S,\gamma)$, it holds $$(\mbox{Area}(S))^{-\f1p}\|\phi\|_{L^p(S)}\leq C_p\sqrt{\max\{I(S),1\}}\bigg(\|\nab\phi\|_{L^2(S)}+\big(\mbox{Area}(S)\big)^{-\f12}\|\phi\|_{L^2(S)}\bigg)$$ for $2<p<\infty$ and for any tensor field $\phi$. \[Linfty\] For any Riemannian $2$-manifold $(S,\gamma)$, we have $$\|\phi\|_{L^\infty(S)}\leq C_p\sqrt{\max\{I(S),1\}}(\mbox{Area}(S))^{\f12-\f1p}\bigg(\|\nab\phi\|_{L^p(S)}+\big(\mbox{Area}(S)\big)^{-\f12}\|\phi\|_{L^p(S)}\bigg)$$ for $p>2$ and for any tensor field $\phi$. Note by Proposition \[area\], we have $\mbox{Area}(\S)\sim |u|^2$. Combining Propositions \[isoperimetric\], \[Lp\] and \[Linfty\], we have \[Sobolev\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumption , it holds $$\label{Sobolev 1} \begin{split} \|\phi\|_{L^\infty(S_{u,{\underline{u}}})} \ls & \sum_{i\leq 2}\|u^{i-1}\nab^i\phi\|_{L^2(\S)}+\|\phi\|_{L^2(\S)}. \end{split}$$ Written in scale invariant norms: $$\label{Sobolev 2} \|\phi\|_{L_{sc}^{\infty}(S_{u,{\underline{u}}})}\ls \sum_{i\leq 2}\|(\at\nab)^i\phi\|_{L_{sc}^{2}(S_{u,{\underline{u}}})}+\|\phi\|_{L_{sc}^{2}(S_{u,{\underline{u}}})}.$$ Commutation Formula {#commutation} ------------------- We move to derive general commutation formulae. We first list the following formula from [@KNI:book]: \[commute0\] For a scalar function $f$, it holds $$[\nab_4,\nab]f=\frac 12 (\eta+{\underline{\eta}})\nab_4f-\chi\cdot\nab f,$$ $$[\nab_3,\nab]f=\frac 12 (\eta+{\underline{\eta}})\nab_3f-{\underline{\chi}}\cdot\nab f.$$ \[commute1\] For a 1-form $U_b$ tangent to $S_{u,{\underline{u}}}$, we have $$[\nab_4,\nab_a]U_b=-\chi_{ac}\nab_cU_b+\epsilon_{ac}{^*\b_b} U_c+\frac 12(\eta_a+{\underline{\eta}}_a)\nab_4U_b-\chi_{ac}{\underline{\eta}}_bU_c+\chi_{ab}{\underline{\eta}}\cdot U,$$ $$[\nab_3,\nab_a]U_b=-{\underline{\chi}}_{ac}\nab_cU_b+\epsilon_{ac}{^*{\underline{\beta}}_b} U_c+\frac 12(\eta_a+{\underline{\eta}}_a)\nab_3U_b-{\underline{\chi}}_{ac}\eta_bU_c+{\underline{\chi}}_{ab}\eta\cdot U.$$ \[commute2\] For a 2-form $V_{bc}$ tangent to $\S$, we have $$\begin{split} [\nab_4,\nab_a]V_{bc}=&\frac 12(\eta_a+{\underline{\eta}}_a)\nab_4V_{bc}-{\underline{\eta}}_bV_{dc}\chi_{ad}-{\underline{\eta}}_cV_{bd}\chi_{ad}-\epsilon_{bd}{^*\b_a}V_{dc}-\epsilon_{cd}{^*\b_c}V_{bd}\\ &+\chi_{ac}V_{bd}{\underline{\eta}}_d+\chi_{ab}V_{dc}{\underline{\eta}}_d-\chi_{ad}\nab_dV_{bc}, \end{split}$$ $$\begin{split} [\nab_3,\nab_a]V_{bc}=&\frac 12(\eta_a+{\underline{\eta}}_a)\nab_3V_{bc}-\eta_bV_{dc}{\underline{\chi}}_{ad}-\eta_cV_{bd}{\underline{\chi}}_{ad}+\epsilon_{bd}{^*{\underline{\beta}}_a}V_{dc}+\epsilon_{cd}{^*{\underline{\beta}}_c}V_{bd}\\ &+{\underline{\chi}}_{ac}V_{bd}\eta_d+{\underline{\chi}}_{ab}V_{dc}\eta_d-{\underline{\chi}}_{ad}\nab_dV_{bc}. \end{split}$$ In this article, we require $i_1, i_2,..., i_n\geq 0$. [Applying Proposition \[commute1\] and Proposition \[commute2\] through a mathematical induction, we then give the below general formulas (see also [@A-L; @L-R:Propagation]):]{} \[commute3\] Assume $\nabla_4\phi=F_0$. Let $\nabla_4\nabla^i\phi=F_i$. Then we have $$\begin{split} F_i= &\sum_{i_1+i_2+i_3=i}\nabla^{i_1}(\eta+\underline{\eta})^{i_2}\nabla^{i_3} F_0+\sum_{i_1+i_2+i_3+i_4=i-1} \nabla^{i_1}(\eta+\underline{\eta})^{i_2}\nabla^{i_3}\beta\nabla^{i_4} \phi\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}(\eta+\underline{\eta})^{i_2}\nabla^{i_3}\chi\nabla^{i_4} \phi. \end{split}$$ Similarly, assume $\nabla_3\phi=G_{0}$. Let $\nabla_3\nabla^i\phi=G_{i}$. We get $$\begin{split} G_{i}{\color{black}+}\frac{i}{2}{\mbox{tr}}{\underline{\chi}}\nab^i \phi= &\sum_{i_1+i_2+i_3=i}\nabla^{i_1}(\eta+\underline{\eta})^{i_2}\nabla^{i_3} G_{0}\\ &+\sum_{i_1+i_2+i_3+i_4=i-1} \nabla^{i_1}(\eta+\underline{\eta})^{i_2}\nabla^{i_3}\underline{\beta}\nabla^{i_4} \phi\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}(\eta+\underline{\eta})^{i_2}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \phi\\ &+\sum_{i_1+i_2+i_3+i_4=i-1}\nabla^{i_1}(\eta+\underline{\eta})^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} \phi. \end{split}$$ $$$$ Finally, by replacing $\b, {\underline{\beta}}$ via Codazzi equations: $$\beta=-{\mbox{div }}{\hat{\chi}}+\f12\nab {\mbox{tr}}\chi-\f12(\eta-\etab)\cdot({\hat{\chi}}-\f12{\mbox{tr}}\chi),$$ $${\underline{\beta}}={\mbox{div }}{\hat{{\underline{\chi}}}}-\f12\nab{\mbox{tr}}{\underline{\chi}}-\f12(\eta-\etab)\cdot({\hat{{\underline{\chi}}}}-\f12{\mbox{tr}}{\underline{\chi}}),$$ and substituting $\eta, {\underline{\eta}}, {\mbox{tr}}\chi, {\mbox{tr}}{\underline{\chi}}+\f{2}{|u|}$ with $\p$, we arrive at \[commute\] Suppose $\nabla_4\phi=F_0$. Let $\nabla_4\nabla^i\phi=F_i$. Then $$\begin{split} F_i= &\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3} F_0+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} \phi.\\ \end{split}$$ Similarly, suppose $\nabla_3\phi=G_{0}$. Let $\nabla_3\nabla^i\phi=G_{i}$. Then $$\begin{split} G_{i}{\color{black}+}\f {i}{2} {\mbox{tr}}{\underline{\chi}}\nab^i\phi =&\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3} G_{0}\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \phi\\ &+\sum_{i_1+i_2+i_3+i_4=i-1} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nab^{i_4}\phi.\\ \end{split}$$ $L^2(\S)$ estimates for Ricci coefficients {#secRicci} ========================================== We start from several useful estimates. [Denote]{} $$\label{psi Psi 2} \p\in \{\f{{\hat{\chi}}}{\at}, {\mbox{tr}}\chi, \o, \eta, {\underline{\eta}}, \zeta, {\underline{\omega}}, \f{a}{|u|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}, \f{\at}{|u|}{\hat{{\underline{\chi}}}}, \f{a}{|u|^2}{\mbox{tr}}{\underline{\chi}}\}, \quad\mbox{ and } \quad \Psi\in \{\f{\a}{\at}, \beta,\rho,\sigma, {\underline{\beta}},\ab\}.$$ Under the assumption of Theorem \[main.thm1\] and bootstrap assumption (\[BA.0\]), we have $$\label{4.1} \sum_{\substack{i_1+i_2\leq 9 \\}}\|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\leq |u|,$$ $$\label{4.2} \sum_{\substack{i_1+i_2\leq 9\\}}\|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2+1}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\leq O,$$ $$\label{4.3} \sum_{\substack{i_1+i_2\leq 9\\}} \|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2+2}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\leq \f{O^2}{|u|},$$ $$\label{4.4} \sum_{\substack{i_1+i_2\leq 9\\}} \|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2+3}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\leq \f{O^3}{|u|^2},$$ $$\label{4.5} \sum_{\substack{i_1+i_2+i_3\leq 9\\ }} \|(\at)^{i_1+i_2+i_3}\nabla^{i_1}\p^{i_2}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\leq O,$$ $$\label{4.6} \sum_{i_1+i_2+i_3\leq 9}\|(\at)^{i_1+i_2+i_3+1}\nabla^{i_1}\p^{i_2+1}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\leq \f{\at}{|u|}\cdot O^2,$$ $$\label{4.7} \sum_{i_1+i_2+i_3\leq 9}\|(\at)^{i_1+i_2+i_3+2}\nabla^{i_1}\p^{i_2+2}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\leq \f{a}{|u|^2}\cdot O^3.$$ We will prove (\[4.1\]) first. For $i_2=0$, (\[4.1\]) is true because naturally we could let $s_2(1)=0$ and $$\|1\|_{L^2_{sc}(\S)}=|u|.$$ For $i_2\geq 1$, we could rewrite $\nab^{i_1}\p^{i_2}$ [as]{} a product of $i_2$-terms $$\nab^{i_1}\p^{i_2}=\nab^{j_1}\p\cdot\nab^{j_2}\p\cdot\cdot\cdot\nab^{j_{i_2}}\p, \mbox{ where } i_1=j_1+j_2+...+j_{i_2},$$ and assume that $j_{i_2}$ is the largest number. We then rewrite $$(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2}=a^{i_2}\cdot(\at\nab)^{j_{i_2}}\p \cdot \Pi_{k=1}^{k=i_2-1}(\at\nab)^{j_k}\p.$$ We bound $(\at\nab)^{j_{i_2}}\p$ in $L^2_{sc}(\S)$ norm and bound other terms in $L^{\infty}_{sc}(\S)$ norms. By employing (\[Holder’s\]) for $i_2-1$ times, we obtain $$\begin{split} &\f{1}{|u|}\cdot\sum_{i_1+i_2\leq 9}\|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\\ \leq&\f{1}{|u|}\cdot \sum_{i_1+i_2\leq 9} \f{(\at)^{i_2}}{|u|^{i_2-1}} \|(\at\nab)^{j_{i_2}}\p\|_{L^{2}_{sc}(S_{u,{\underline{u}}})} \Pi_{k=1}^{k={i_2-1}}\|(\at\nab)^{j_k}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})},\\ \leq& \f{(\at)^{i_2}\cdot O^{i_2}}{|u|^{i_2}}\leq 1. \end{split}$$ We prove (\[4.2\]) in the same fashion. If $i_2=0$, (\[4.2\]) is true according to the definition of $O$. For $i_2\geq 1$, assume $i_1=j_1+j_2+...+j_{i_2+1}$. And assume $j_{i_2+1}$ is the largest. It follows $$\begin{split} &\sum_{i_1+i_2\leq 9} \|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2+1}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\\ \leq& \sum_{i_1+i_2\leq 9} \f{(\at)^{i_2}}{|u|^{i_2}} \|(\at\nab)^{j_{i_2+1}}\p\|_{L^{2}_{sc}(S_{u,{\underline{u}}})} \Pi_{k=1}^{k={i_2}}\|(\at\nab)^{j_k}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})},\\ \leq& \f{(\at)^{i_2}\cdot O^{i_2+1}}{|u|^{i_2}}\leq O. \end{split}$$ For (\[4.3\]), we use (\[Holder’s\]), (\[4.2\]) and derive $$\begin{split} &|u|\cdot\sum_{i_1+i_2\leq 9}\|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2+2}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\\ \leq&|u|\cdot\f{1}{|u|}\cdot \sum_{i_1+i_2\leq 9} \|(\at\nab)^{i_3}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\|(\at)^{i_2+i_4}\nab^{i_4}\p^{i_2+1}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})},\, \, \mbox{where} \, \, i_3+i_4=i_1\\ \leq&O\cdot O=O^2. \end{split}$$ With the same approach, for (\[4.4\]), we use (\[Holder’s\]), (\[4.3\]) and obtain $$\begin{split} &|u|^2\cdot\sum_{i_1+i_2\leq 9}\|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2+3}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\\ \leq&|u|^2\cdot\f{1}{|u|}\cdot \sum_{i_1+i_2\leq 9} \|(\at\nab)^{i_3}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\|(\at)^{i_2+i_4}\nab^{i_4}\p^{i_2+2}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})},\, \, \mbox{where} \, \, i_3+i_4=i_1\\ \leq&|u|\cdot O\cdot \f{O^2}{|u|}\leq O^3. \end{split}$$ We then move to (\[4.5\]). If $i_3\leq 7$, we bound $(\at\nab)^{i_3}\Psi$ with $L^{\infty}_{sc}(\S)$ norm; otherwise we bound $(\at\nab)^{i_3}\Psi$ with $L^{2}_{sc}(\S)$ norm. As before assume $i_1= j_1+j_2+...+j_{i_2}$. With (\[Holder’s\]) and (\[4.1\]) we obtain $$\begin{split} &\sum_{i_1+i_2+i_3\leq 9} \|(\at)^{i_1+i_2+i_3}\nabla^{i_1}\p^{i_2}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\\ \leq&\f{1}{|u|}\cdot \sum_{i_1+i_2+i_3\leq 9} \|(\at)^{i_1+i_2}\nab^{i_1}\p^{i_2}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\|(\at\nab)^{i_3}\Psi\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\\ &+\sum_{i_1+i_2+i_3\leq 9}\f{(\at)^{i_2}}{|u|^{i_2}} \Pi_{k=1}^{k=i_2}\|(\at\nab)^{j_k}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\cdot \|(\at\nab)^{i_3}\Psi\|_{L^{2}_{sc}(S_{u,{\underline{u}}})},\\ \leq& O. \end{split}$$ Similarly, for (\[4.6\]) we decompose $i_1=j_1+j_2+...+j_{i_2}$ and derive $$\begin{split} &\sum_{i_1+i_2+i_3\leq 9} \|(\at)^{i_1+i_2+i_3+1}\nabla^{i_1}\p^{i_2+1}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\\ \leq&\f{1}{|u|}\cdot\sum_{i_1+i_2+i_3\leq 9}\|(\at)^{i_1+i_2+1}\nab^{i_1}\p^{i_2+1}\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}\|(\at\nab)^{i_3}\Psi\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\\ &+\f{(\at)^{i_2+1}}{|u|^{i_2+1}}\sum_{i_1+i_2+i_3\leq 9} \Pi_{k=1}^{k=i_2+1}\|(\at\nab)^{j_k}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\cdot \|(\at\nab)^{i_3}\Psi\|_{L^{2}_{sc}(S_{u,{\underline{u}}})}, \\ \leq&\f{\at}{|u|}\cdot O^2. \end{split}$$ Finally, we prove (\[4.7\]). We use (\[Holder’s\]) once and reduce it to (\[4.6\]). $$\begin{split} &\sum_{i_1+i_2+i_3\leq 9} \|(\at)^{i_1+i_2+i_3+2}\nabla^{i_1}\p^{i_2+2}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\\ \leq&\sum_{\substack{i_1+i_2+i_3\leq 9\\i_4+i_5=i_1}}\f{\at}{|u|}\|(\at\nab)^{i_4}\p\|_{L^{\infty}_{sc}(S_{u,{\underline{u}}})}\cdot\|(\at)^{i_2+i_3+i_5+1}\nabla^{i_5}\p^{i_2+1}\nab^{i_3}\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})}\\ \leq&\f{\at}{|u|}\cdot O\cdot \f{\at}{|u|}\cdot O^2=\f{a}{|u|^2}\cdot O^3.\\ \end{split}$$ We are now ready to estimate Ricci coefficients and we start from $\o$ \[o.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10}\|(\at\nab)^i\o\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls \f{\at}{|u|^{\f12}}\big(\underline{\M R}[\rho]+1\big).$$ We use the following schematic null structure equation for $\o$: $$\nab_3\o=\f12\rho+\p\p.$$ Commuting it with angular derivative for $i$ times, we have $$\begin{split} &\nab_3 \nab^i\o +\frac i2 {\mbox{tr} {\underline{\chi}}}\nab^i\o\\ = &\nab^i\rho+\sum_{i_1+i_2+i_3+1=i}\nab^{i_1}\p^{i_2+1}\nab^{i_3}\rho+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \p\\ &+\sum_{i_1+i_2+i_3+i_4=i-1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} \p. \end{split}$$ Denote the above equality as $$\nab_3 \nab^i\o +\frac i2 {\mbox{tr} {\underline{\chi}}}\nab^i\o=G.$$ Applying Proposition \[el\], it holds $$\begin{split} |u|^{i-1}\|\nab^i\o\|_{L^2(\S)}\leq&|u_{\infty}|^{i-1}\|\nab^i\o\|_{L^2(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u|u'|^{i-1}\|G\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ Times $|u|$ on both sides and using $|u|\leq |u'|, \, |u|\leq |u_{\infty}|$ we have $$\label{o G} \begin{split} |u|^{i}\|\nab^i\o\|_{L^2(\S)}\leq&|u_{\infty}|^{i}\|\nab^i\o\|_{L^2(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u|u'|^{i}\|G\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ From signature table and property (\[signature of derivative\]), we have $$s_2(\nab^i\o)=s_2(\o)+i\cdot\f{1}{2}=0+\f{i}{2}=\f{i}{2}.$$ By conversation of signatures in each equation and property (\[signature of derivative\]), it holds $$s_2(G)=s_2(\nab_3 \nab^i\o)=s_2(\nab^i \o)+1=\f{i}{2}+1.$$ Using the definition of $L^2_{sc}(\S)$ norms $$\|\phi\|_{L_{sc}^{2}(\S)}:=a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^{2}(\S)},$$ we have $$\|\nab^i\o\|_{L^2_{sc}(\S)}=a^{-\f{i}{2}}|u|^i\|\nab^i\o\|_{L^2(\S)}, \quad \|G\|_{L^2_{sc}(\S)}=a^{-\f{i}{2}-1}|u|^{i+2}\|G\|_{L^2(\S)}.$$ That is equivalent to $$|u|^i\|\nab^i\o\|_{L^2(\S)}=\|(\at\nab)^i\o\|_{L^2_{sc}(\S)}, \quad |u|^{i}\|G\|_{L^2(\S)}= \f{a}{|u|^2}\|(\at)^iG\|_{L^2_{sc}(\S)}.$$ We then rewrite (\[o G\]) in $L^2_{sc}(\S)$ norms $$\begin{split} \|(\at\nab)^i\o\|_{L^2_{sc}(\S)}\leq&\|(\at\nab)^i\o\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|(\at\nab)^{i}\rho\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|\sum_{i_1+i_2+i_3+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nab^{i_3}\rho\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \p\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i-1}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} \p\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'. \end{split}$$ For the first term, since we prescribe $\O|_{u=u_{\infty}}=1$, note by $$\o=-\f12\nab_4(\log\O), \mbox{ we have } \|(\at\nab)^i\o\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}=0.$$ For the two terms involving $\rho$, we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|(\at\nab)^{i}\rho\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'+\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|\sum_{i_1+i_2+i_3+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nab^{i_3}\rho\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\bigg(\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|(\at\nab)^{i}\rho\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'\bigg)^{\f12} \bigg(\int_{{u_{\infty}}}^u\f{a}{|u'|^2} du'\bigg)^{\f12}+\int_{{u_{\infty}}}^u \f{a}{|u'|^2}\cdot \f{\at}{|u'|}\cdot O^2\, du'\\ =&\|(\at\nab)^{i}\rho\|_{L^2_{sc}(\Hb_{{\underline{u}}}^{({u_{\infty}},u)})} \cdot \f{\at}{|u|^{\f12}}+\f{a^{\f32}}{|u|^2}O^{2}\\\leq& \underline{\M R}[\rho]\cdot \f{\at}{|u|^{\f12}}+\f{a^{\f32}}{|u|^2}O^{2}\leq \f{\at}{|u|^{\f12}}\big(\underline{\M R}[\rho]+1\big). \end{split}$$ where Hölder’s inequality and (\[4.6\]) are used in the first inequality; the definition in (\[scale invariant norms 2\]) is used in the identity; (\[R i Hb\]) is employed in the second inequality. For the last two terms, we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^2} \|\sum_{i_1+i_2+i_3+i_4=i} (\at)^i \nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nab^{i_4}\p\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u}\f{\at}{|u'|}\|\sum_{i_1+i_2+i_3+i_4=i} (\at)^i \nabla^{i_1}\p^{i_2}\nabla^{i_3}(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nab^{i_4}\p\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u}\f{\at}{|u'|}\cdot \f{O^2}{|u'|} du' \leq \f{\at}{|u|} O^{2} \leq \f{\at}{|u|^{\f12}}, \end{split}$$ where we use $\at/|u|\leq 1/\at$ and (\[4.3\]) in the second inequality. And $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i-1}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} \p\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u}\at\|\sum_{i_1+i_2+i_3+i_4=i-1}(\at)^{i-1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}})\nabla^{i_4} \p\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq& \int_{u_{\infty}}^{u} \at\cdot\f{O^3}{|u'|^2}du' \leq \f{\at}{|u|}\cdot O^3 \leq \f{\at}{|u|^{\f12}}.\\ \end{split}$$ Here we appeal to (\[4.4\]) for the second inequality. Gather all the estimates and let $a$ to be sufficient large, we then derive $$\sum_{i\leq 10}\|(\at\nab)^i\o\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls \f{\at}{|u|^{\f12}}\big(\underline{\M R}[\rho]+1\big).$$ We then move to estimate ${\hat{{\underline{\chi}}}}$. \[chibh.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10}\f{\at}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls 1.$$ We use the null structure equation $$\nab_3{\hat{{\underline{\chi}}}}+{\mbox{tr}}{\underline{\chi}}\,{\hat{{\underline{\chi}}}}=\ab+\p\cdot{\hat{{\underline{\chi}}}}.$$ Commuting this equation with $i$ angular derivatives, by Proposition \[commute\] we have $$\begin{split} &\nab_3 \nab^i{\hat{{\underline{\chi}}}}+\frac{i+2}{2}{\mbox{tr}}{\underline{\chi}}\nab^i{\hat{{\underline{\chi}}}}\\ = &\nab^{i}\ab+\sum_{i_1+i_2+i_3=i-1}\nab^{i_1}\p^{i_2+1}\nab^{i_3}\ab+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} {\hat{{\underline{\chi}}}}\\ &+\sum_{i_1+i_2+i_3+i_4=i-1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} {\hat{{\underline{\chi}}}}.\\ \end{split}$$ Rewrite the above equation as $$\nab_3 \nab^i{\hat{{\underline{\chi}}}}+\frac{i+2}{2}{\mbox{tr}}{\underline{\chi}}\nab^i{\hat{{\underline{\chi}}}}=F.$$ Applying Proposition \[el\], we have $$\label{chibh F} \begin{split} |u|^{i+1}\|\nab^i{\hat{{\underline{\chi}}}}\|_{L^2(\S)}\leq&|u_{\infty}|^{i+1}\|\nab^i{\hat{{\underline{\chi}}}}\|_{L^2(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u|u'|^{i+1}\|F\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ By signature consideration, we have $$s_2(\nab^i{\hat{{\underline{\chi}}}})=s_2({\hat{{\underline{\chi}}}})+i\cdot\f{1}{2}=\f{i}{2}+1, \quad s_2(F)=s_2(\nab_3\nab^i{\underline{\chi}})=\f{i}{2}+2.$$ Using the definition of $L^2_{sc}(\S)$ norms $$\|\phi\|_{L_{sc}^{2}(\S)}:=a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^{2}(\S)},$$ we have $$\|\nab^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(\S)}=a^{-\f{i}{2}-1}|u|^{i+2}\|\nab^i{\hat{\chi}}\|_{L^2(\S)}, \quad \|F\|_{L^2_{sc}(\S)}=a^{-\f{i}{2}-2}|u|^{i+4}\|F\|_{L^2(\S)},$$ which are equivalent to $$|u|^{i+1}\|\nab^i{\hat{\chi}}\|_{L^2(\S)}=\f{a}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(\S)}, \quad |u|^{i+1}\|F\|_{L^2(\S)}=\f{a^2}{|u|^3}\|(\at)^iF\|_{L^2_{sc}(\S)}.$$ Rewrite (\[chibh F\]) in $L^2_{sc}(\S)$ norms, it follows $$\f{a}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(\S)}\leq \f{a}{|{u_{\infty}}|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u\f{a^2}{|u'|^3}\|(\at)^i F\|_{L^{2}_{sc}(S_{u',\underline{u}})}du'.$$ Multiplying $a^{-\f12}$ on both sides, with the expression of $F$ we have $$\begin{split} \f{\at}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}&\|_{L^2_{sc}(\S)}\leq\f{\at}{|{u_{\infty}}|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|(\at\nab)^{i}\ab\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|\sum_{i_1+i_2+i_3=i-1}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\ab\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} {\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4=i-1}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} {\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'. \end{split}$$ For initial data, we have $$\f{\at}{|u_{\infty}|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}({\underline{u}}) \lesssim 1.$$ For $\ab$ terms, we have $$\begin{split} \int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|(\at\nab)^{i}\ab\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\leq& \bigg(\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|(\at\nab)^{i}\ab\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du' \bigg)^{\f12} \bigg(\int_{{u_{\infty}}}^{u}\f{a^2}{|u'|^4} du'\bigg)^{\f12}\\ \leq&\|(\at\nab)^i\ab\|_{{L^2_{sc}(\underline{H}_{{\underline{u}}}^{(u_{\infty},u)})}}\cdot \f{a}{|u|^{\f32}}\leq \f{{\underline{\mathcal{R}}}}{\at}\leq 1. \end{split}$$ And by (\[4.6\]) $$\begin{split} &\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|\sum_{i_1+i_2+i_3=i-1}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\ab\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\cdot \f{\at}{|u'|}\cdot O^2 \,du'\\ \leq&\f{a^2}{|u|^3}\cdot O^2\leq \f{O^2}{a}\leq 1. \end{split}$$ We then move to $$\begin{split} &\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3} \|\sum_{i_1+i_2+i_3+i_4=i} (\at)^i \nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nab^{i_4}{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u}\f{a^{\f12}}{|u'|}\|\sum_{i_1+i_2+i_3+i_4=i} (\at)^i \nabla^{i_1}\p^{i_2}\nabla^{i_3}(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nab^{i_4}(\f{\at}{|u'|}{\hat{{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u}\f{\at}{|u'|}\cdot \f{O^2}{|u'|} du' \leq \f{\at}{|u|} O^{2} \leq \f{O^2}{\at}\leq 1, \end{split}$$ where we use (\[4.3\]) in the second inequality. We then deal with the last term $$\begin{split} &\int_{u_{\infty}}^{u}\f{a^{\f32}}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4=i-1}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\mbox{tr}}{\underline{\chi}}\nabla^{i_4} {\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u} \at\|\sum_{i_1+i_2+i_3+i_4=i-1}(\at)^{i-1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}})\nabla^{i_4} (\f{\at}{|u'|}{\hat{{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq& \int_{u_{\infty}}^{u} \at\cdot\f{O^3}{|u'|^2}du' \leq \f{\at}{|u|}\cdot O^3 \leq \f{O^3}{\at}\leq 1.\\ \end{split}$$ Here we appeal to (\[4.4\]) for the second inequality. Gathering all the estimates, and letting $a$ to be sufficiently large we have obtained $$\sum_{i\leq 10}\f{\at}{|u|}\|(\at\nab)^i{\hat{{\underline{\chi}}}}\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls 1.$$ Next, we deal with ${\hat{\chi}}$. \[chih.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10} \f{1}{\at}\|(\at\nab)^i {\hat{\chi}}\|_{L^2_{sc}(\S)}\ls \M R[\a]+1.$$ We employ the null structure equation $$\nab_4 {\hat{\chi}}=\a+\p\cdot{\hat{\chi}}.$$ Commuting this equation with $i$ angular derivatives, by Proposition \[commute\] we have $$\begin{split} \nab_4 \nab^i{\hat{\chi}}=& \nab^{i}\a+\sum_{i_1+i_2+i_3=i-1}\nab^{i_1}\p^{i_2+1}\nab^{i_3}\a+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} {\hat{\chi}}.\\ \end{split}$$ Applying Proposition \[transport\] and multiplying $(\at)^{i-1}$ on both sides of equation, we have $$\begin{split} &\f{1}{\at}\|(\at\nab)^{i}{\hat{\chi}}\|_{L^2_{sc}(\S)}\\ \leq &\f{1}{\at}\int_0^{{\underline{u}}}\|(\at\nab)^i \a\|_{L^2_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'+\sum_{i_1+i_2+i_3=i-1}\int_0^{{\underline{u}}} \f{1}{\at}\|(\at)^{i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3} \a\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}} \f{1}{\at}\|(\at)^{i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\p, {\hat{\chi}})\nab^{i_4}{\hat{\chi}}\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq& \f{1}{\at}\bigg(\int_0^{{\underline{u}}}\|(\at\nab)^i \a\|^2_{L^2_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12}\bigg(\int_0^{{\underline{u}}}1 \, d{\underline{u}}'\bigg)^{\f12}\\ &+\sum_{i_1+i_2+i_3=i-1}\int_0^{{\underline{u}}} \|(\at)^{i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3} (\f{\a}{\at})\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}} \at\|(\at)^{i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\f{\p}{\at}, \f{{\hat{\chi}}}{\at})\nab^{i_4}(\f{{\hat{\chi}}}{\at})\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq& \f{1}{\at}\|(\at\nab)^i\a\|_{{L^2_{sc}(H_{u}^{(0,{\underline{u}})})}}+\f{\at}{|u|}\cdot O^2+\at\cdot \f{O^2}{|u|}\\ \leq& \M R[\a]+\f{O^2}{\at}\leq \M R[\a]+1, \end{split}$$ where we use (\[4.6\]) and (\[4.3\]) in the third inequality.\ In the same fashion, we derive estimate for ${\underline{\omega}}$. \[omb.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10} \|(\at\nab)^i {\underline{\omega}}\|_{L^2_{sc}(\S)}\ls \M R[\rho]+1.$$ We have the schematic null structure equation $$\nab_4{\underline{\omega}}=\rho+\p\cdot\p.$$ Commuting this equation with $i$ angular derivatives, by Proposition \[commute\] we have $$\begin{split} \nab_4 \nab^i{\underline{\omega}}=& \nab^{i}\rho+\sum_{i_1+i_2+i_3=i-1}\nab^{i_1}\p^{i_2+1}\nab^{i_3}\rho+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} \p\\ \end{split}$$ Applying Proposition \[transport\] and multiplying $(\at)^{i}$ on both sides of equation, we have $$\begin{split} &\|(\at\nab)^{i}{\underline{\omega}}\|_{L^2_{sc}(\S)}\\ \leq &\int_0^{{\underline{u}}}\|(\at\nab)^i \rho\|_{L^2_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'+\sum_{i_1+i_2+i_3=i-1}\int_0^{{\underline{u}}}\|(\at)^{i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3} \rho\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}}\|(\at)^{i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\p, {\hat{\chi}})\nab^{i_4}\p\|_{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq&\bigg(\int_0^{{\underline{u}}}\|(\at\nab)^i \rho\|^2_{L^2_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12}\bigg(\int_0^{{\underline{u}}}1 \, d{\underline{u}}'\bigg)^{\f12}\\ &+\sum_{i_1+i_2+i_3=i-1}\int_0^{{\underline{u}}} \|(\at)^{i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3} \rho\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}} \at\|(\at)^{i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\f{\p}{\at}, \f{{\hat{\chi}}}{\at})\nab^{i_4}\p\|_{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq& \|(\at\nab)^i\rho\|_{{L^2_{sc}(H_{u}^{(0,{\underline{u}})})}}+\f{\at}{|u|}\cdot O^2+\at\cdot\f{O^2}{|u|}\leq \M R[\rho]+1.\\ \end{split}$$ We use (\[4.6\]) and (\[4.3\]) in the third inequality.\ Similarly, for $\eta$ we have \[eta.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10} \|(\at\nab)^i \eta\|_{L^2_{sc}(\S)}\ls \M R[\b]+1.$$ We have the schematic null structure equation $$\nab_4\eta=\b+\p\cdot{\hat{\chi}}.$$ Commuting this equation with $i$ angular derivatives, by Proposition \[commute\] we have $$\begin{split} \nab_4 \nab^i\eta=& \nab^{i}\b+\sum_{i_1+i_2+i_3=i-1}\nab^{i_1}\p^{i_2+1}\nab^{i_3}\b+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}\p\nabla^{i_4} {\hat{\chi}}\\ \end{split}$$ Applying Proposition \[transport\] and multiplying $(\at)^{i}$ on both sides of equation, we have $$\begin{split} &\|(\at\nab)^{i}\eta\|_{L^2_{sc}(\S)}\\ \leq &\int_0^{{\underline{u}}}\|(\at\nab)^i \b\|_{L^2_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'+\sum_{i_1+i_2+i_3=i-1}\int_0^{{\underline{u}}}\|(\at)^{i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3} \b\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}}\at\|(\at)^{i}\nab^{i_1}\p^{i_2}\nab^{i_3}\p\nab^{i_4}(\f{\p}{\at}, \f{{\hat{\chi}}}{\at})\|_{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq& \|(\at\nab)^i\b\|_{{L^2_{sc}(H_{u}^{(0,{\underline{u}})})}}+\f{\at}{|u|}\cdot O^2+\at\cdot\f{O^2}{|u|}\leq \M R[\b]+1.\\ \end{split}$$ In the third inequality, (\[4.6\]) and (\[4.3\]) are used.\ We move to estimate ${\mbox{tr}}\chi$ \[trchi.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10} \|(\at\nab)^i {\mbox{tr}}\chi\|_{L^2_{sc}(\S)}\ls (\M R[\a]+1)^2.$$ From (\[eqn 2 trchib\]), we have the schematic null structure equation: $$\nab_4{\mbox{tr}}\chi={\hat{\chi}}\cdot{\hat{\chi}}+\p\cdot\p.$$ Commuting this equation with $i$ angular derivatives, by Proposition \[commute\] we have $$\begin{split} \nab_4 \nab^i {\mbox{tr}}\chi=&\sum_{i_1+i_2=i}\nab^{i_1}{\hat{\chi}}\nab^{i_2}{\hat{\chi}}+\sum_{i_1+i_2+i_3+i_4+1=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}{\hat{\chi}}\nabla^{i_4} (\p, {\hat{\chi}})\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nab^{i_1}\p^{i_2}\nab^{i_3}\p\nab^{i_4}\p. \end{split}$$ Applying Proposition \[transport\] and multiplying $(\at)^{i}$ on both sides of equation, we calculate as above $$\begin{split} &\|(\at\nab)^{i}{\mbox{tr}}\chi\|_{L^2_{sc}(\S)}\\ \leq& \sum_{i_1+i_2=i}\int_0^{{\underline{u}}}a\|(\at)^{i}\nab^{i_1}(\f{{\hat{\chi}}}{\at})\nab^{i_2}(\f{{\hat{\chi}}}{\at})\|_{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4+1=i}\int_0^{{\underline{u}}}a\|(\at)^{i}\nab^{i_1}\p^{i_2+1}\nab^{i_3}(\f{{\hat{\chi}}}{\at})\nab^{i_4}(\f{\p}{\at} ,\f{{\hat{\chi}}}{\at})\|_{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}}\|(\at)^{i}\nab^{i_1}\p^{i_2}\nab^{i_3}\p\nab^{i_4}\p\|_{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'.\\ \leq& \f{a}{|u|}O[{\hat{\chi}}]\cdot O[{\hat{\chi}}]+\f{a}{|u|^2}O^3+\f{1}{|u|}O^2\\ \leq&O[{\hat{\chi}}]\cdot O[{\hat{\chi}}]+1\leq (\M R[\a]+1)^2. \end{split}$$ For the last inequality, we use Proposition \[chih.bd\].\ We derive estimates for ${\mbox{tr}}{\underline{\chi}}$. \[trchib.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10} \f{a}{|u|}\|(\at\nab)^i ({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})\|_{L^2_{sc}(\S)}\ls \M R[\rho]+\underline{\M R}[\rho]+1, \quad \sum_{i\leq 10} \f{a}{|u|^2}\|(\at\nab)^i {\mbox{tr}}{\underline{\chi}}\|_{L^2_{sc}(\S)}\ls 1.$$ From (\[eqn 2 trchib\]), we have the schematic null structure equation: $$\begin{split} \nab_3{\widetilde{{\mbox{tr}}{\underline{\chi}}}}+{\mbox{tr}}{\underline{\chi}}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}=\f{2}{|u|^2}(\O^{-1}-1)+{\widetilde{{\mbox{tr}}{\underline{\chi}}}}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}+\p{\mbox{tr}}{\underline{\chi}}-|{\hat{{\underline{\chi}}}}|^2. \end{split}$$ Commuting this equation with $i$ angular derivatives, by Proposition \[commute\] we have $$\begin{split} &\nab_3\nab^i{\widetilde{{\mbox{tr}}{\underline{\chi}}}}+\f{i+2}{2}{\mbox{tr}}{\underline{\chi}}\nab^i{\widetilde{{\mbox{tr}}{\underline{\chi}}}}=\sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}\nab^{i_3}\bigg(\f{2}{|u|^2}(\O^{-1}-1)+{\widetilde{{\mbox{tr}}{\underline{\chi}}}}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}+\p{\mbox{tr}}{\underline{\chi}}-|{\hat{{\underline{\chi}}}}|^2 \bigg)\\ &\quad\quad+\sum_{i_1+i_2+i_3+i_4=i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nab^{i_4} {\widetilde{{\mbox{tr}}{\underline{\chi}}}}+\sum_{i_1+i_2+i_3+i_4=i-1}\nab^{i_1}\p^{i_2+1}\nab^{i_3}{\mbox{tr}}{\underline{\chi}}\nab^{i_4}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}. \end{split}$$ Denote the RHS of above equation to be $\tilde{F}$. As proceeded in Proposition \[chibh.bd\], applying Proposition \[el\] and rewriting everything in scale invariant norms, we arrive at $$\begin{split} \f{a}{|u|}\|(\at\nab)^i{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^2_{sc}(\S)}\leq& \f{a}{|u_{\infty}|}\|(\at\nab)^i{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u \f{a^2}{|u'|^3}\|(\at)^i\tilde{F}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ =& \f{a}{|u_{\infty}|}\|(\at\nab)^i{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+I_1+I_2+I_3+I_4, \end{split}$$ where $$\f{a}{|u_{\infty}|}\|(\at\nab)^i{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}\leq \mathcal{I}_0\lesssim 1,$$ $$\begin{split} I_1=&\int_{u_{\infty}}^u \f{a^2}{|u'|^3}\|(\at)^i \sum_{i_1+i_2+i_3+i_4=i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\p, {\widetilde{{\mbox{tr}}{\underline{\chi}}}}, {\hat{{\underline{\chi}}}})\nab^{i_4}(\p,{\widetilde{{\mbox{tr}}{\underline{\chi}}}}, {\hat{{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ =&\int_{u_{\infty}}^u \f{a}{|u'|}\|(\at)^{i} \sum_{i_1+i_2+i_3+i_4=i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}})\nab^{i_4}(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq& \int_{u_{\infty}}^u \f{a}{|u'|}\cdot\f{1}{|u'|}\cdot \big (O^2[{\hat{{\underline{\chi}}}}]+1\big) \, du' \quad (\mbox{by Proposition \ref{4.3} and letting}\,\, a \,\, \mbox{to be sufficiently large})\\ \ls& O^2[{\hat{{\underline{\chi}}}}]+1\ls 1 \quad (\mbox{by Proposition \ref{chibh.bd}}), \end{split}$$ $$\begin{split} I_2=&\int_{u_{\infty}}^u \f{a^2}{|u'|^3}\|(\at)^i \sum_{i_1+i_2+i_3+i_4=i}\nab^{i_1}\p^{i_2}\nab^{i_3}{\underline{\omega}}\nab^{i_4}{\mbox{tr}}{\underline{\chi}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ =&\int_{u_{\infty}}^u \f{a}{|u'|^2}\|(\at)^{i} \sum_{i_1+i_2+i_3+i_4+1=i}\nab^{i_1}\p^{i_2}\nab^{i_3}{\underline{\omega}}\nab^{i_4+1}(\f{a}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^u \f{a}{|u'|}\|(\at)^{i} \sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}\nab^{i_3}{\underline{\omega}}\cdot(\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq& \int_{u_{\infty}}^u \f{a}{|u'|}\cdot\f{1}{|u'|}\cdot \big (O[{\underline{\omega}}]+1\big) \, du' \quad (\mbox{by Proposition \ref{4.3} and letting}\,\, a \,\, \mbox{to be sufficiently large})\\ \ls& O[{\underline{\omega}}]+1\ls \mathcal{R}[\rho]+1 \quad (\mbox{by Proposition \ref{omb.bd}}), \end{split}$$ $$\begin{split} I_3=&\int_{u_{\infty}}^u \f{a^2}{|u'|^3}\|(\at)^i \sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\f{\O^{-1}-1}{|u'|^2})\|_{L^2_{sc}{(S_{u',{\underline{u}}})}}du'\\ =&\int_{u_{\infty}}^u |u'|^{i+1} \|\sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\f{\O^{-1}-1}{|u'|^2})\|_{L^2{(S_{u',{\underline{u}}})}}du' \quad (\mbox{in standard norms})\\ =&\int_{u_{\infty}}^u |u'|^{i+1} \|\sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}\nab^{i_3}(\f{\O^{-1}-1}{|u'|^2})\|_{L^2{(S_{u',{\underline{u}}})}}du' \,\, (\mbox{Using} \,\, \f{\partial}{\partial {\underline{u}}}\O^{-1}=2\o \Leftrightarrow \nab_4\O^{-1}=2\O^{-1}\o)\\ =&\int_{u_{\infty}}^u |u'|^{i+1} \|\sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}\nab^{i_3}[\f{1}{|u'|^2}\cdot \int_0^{{\underline{u}}}2\o(u',{\underline{u}}',\theta^1, \theta^2)d{\underline{u}}']\|_{L^2{(S_{u',{\underline{u}}})}}du' \\ =&\int_{u_{\infty}}^u |u'|^{i+1} \|\sum_{i_1+i_2+i_3=i}\nab^{i_1}\p^{i_2}[\f{1}{|u'|^2}\cdot \int_0^{{\underline{u}}}2\nab^{i_3}\o(u',{\underline{u}}',\theta^1, \theta^2)d{\underline{u}}']\|_{L^2{(S_{u',{\underline{u}}})}}du' \\ \leq& |u'|^{i+1} \|\sum_{i_1+i_2+i_3=i}\f{1}{|u'|^{i_1+i_2}}\cdot\f{1}{|u'|^2}\cdot\f{1}{|u'|^{i_3}}\cdot \f{\at}{|u'|^{\f12}}\cdot \big(\underline{\mathcal{R}}[\rho]+1\big)\, du' (\mbox{by Proposition \ref{o.bd}})\\ \leq& \int_{u_{\infty}}^u \f{\at}{|u'|^{\f32}} \big(\underline{\mathcal{R}}[\rho]+1\big)\,du' \ls \underline{\mathcal{R}}[\rho]+1, \end{split}$$ $$\begin{split} I_4=&\int_{u_{\infty}}^u \f{a^2}{|u'|^3}\|(\at)^i \sum_{i_1+i_2+i_3=i-1}\nab^{i_1}\p^{i_2+1}\cdot{\mbox{tr}}{\underline{\chi}}\cdot\nab^{i_3}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ =&\int_{u_{\infty}}^u \at \|(\at)^{i-1} \sum_{i_1+i_2+i_3=i-1}\nab^{i_1}\p^{i_2+1}\cdot\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}}\cdot\nab^{i_3}(\f{a}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq& \int_{u_{\infty}}^u \at\cdot \f{O^3}{|u'|^2} du'\leq 1 \quad (\mbox{by Proposition \ref{4.4}}). \end{split}$$ In summary, we have obtained $$\sum_{i\leq 10} \f{a}{|u|}\|(\at\nab)^i ({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})\|_{L^2_{sc}(\S)}\ls \M R[\rho]+\underline{\M R}[\rho]+1.$$ This implies $$\sum_{i\leq 10} \f{a}{|u|^2}\|(\at\nab)^i {\mbox{tr}}{\underline{\chi}}\|_{L^2_{sc}(\S)}\ls 1.$$ We move to the last term ${\underline{\eta}}$. \[etab.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 10}\|(\at\nab)^i\etab\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls \underline{\M R}[{\underline{\beta}}]+\M R[\b]+1.$$ We use the following schematic null structure equation for $\o$: $$\nab_3\etab+\f12{\mbox{tr}}{\underline{\chi}}\,\etab={\underline{\beta}}+{\mbox{tr}}{\underline{\chi}}\eta+{\hat{{\underline{\chi}}}}\cdot\p.$$ Commuting it with angular derivative for $i$ times, we have $$\begin{split} &\nab_3 \nab^i\etab +\frac {i+1}{2} {\mbox{tr} {\underline{\chi}}}\nab^i\etab\\ = &\nab^i{\underline{\beta}}+\sum_{i_1+i_2+i_3+1=i}\nab^{i_1}\p^{i_2+1}\nab^{i_3}{\underline{\beta}}+{\mbox{tr}}{\underline{\chi}}\nab^i\eta+\sum_{i_1+i_2+1=i}\nab^{i_1+1}{\mbox{tr}}{\underline{\chi}}\nab^{i_2}(\eta,\etab)\\ &+\sum_{i_1+i_2+i_3+i_4+1=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\p\nabla^{i_4}{\mbox{tr}}{\underline{\chi}}+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}\p\nabla^{i_4}({\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}}).\\ \end{split}$$ Denote the above equality as $$\nab_3 \nab^i\etab +\f{i+1}{2} {\mbox{tr} {\underline{\chi}}}\nab^i\etab=G.$$ Applying Proposition \[el\], it holds $$\begin{split} |u|^{i}\|\nab^i\etab\|_{L^2(\S)}\leq&|u_{\infty}|^{i}\|\nab^i\etab\|_{L^2(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u|u'|^{i}\|G\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ Times $a^{-\f12}$ on both sides and using $|u|\leq |u'|, \, |u|\leq |u_{\infty}|$ we have $$\label{etab G} \begin{split} a^{-\f12}|u|^{i}\|\nab^i\etab\|_{L^2(\S)}\leq&a^{-\f12}|u_{\infty}|^{i}\|\nab^i\etab\|_{L^2(S_{u_{\infty},{\underline{u}}})}+a^{-\f12}\int_{u_{\infty}}^u|u'|^{i}\|G\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ From signature table and property (\[signature of derivative\]), we have $$s_2(\nab^i\etab)=s_2(\etab)+i\cdot\f{1}{2}=\f12+\f{i}{2}=\f{i+1}{2}.$$ By conversation of signatures in each equation and property (\[signature of derivative\]), it holds $$s_2(G)=s_2(\nab_3 \nab^i\etab)=s_2(\nab^i {\underline{\eta}})+1=\f{i+3}{2}.$$ Using the definition of $L^2_{sc}(\S)$ norms $$\|\phi\|_{L_{sc}^{2}(\S)}:=a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^{2}(\S)},$$ we have $$\|\nab^i\etab\|_{L^2_{sc}(\S)}=a^{-\f{i+1}{2}}|u|^{i+1}\|\nab^i\etab\|_{L^2(\S)}, \quad \|G\|_{L^2_{sc}(\S)}=a^{-\f{i+3}{2}}|u|^{i+3}\|G\|_{L^2(\S)}.$$ That is equivalent to $$a^{-\f12}|u|^{i}\|\nab^i\etab\|_{L^2(\S)}=\f{1}{|u|}\|(\at\nab)^i\etab\|_{L^2_{sc}(\S)}, \quad a^{-\f12}|u|^{i}\|G\|_{L^2(\S)}= \f{a}{|u|^3}\|(\at)^iG\|_{L^2_{sc}(\S)}.$$ We then rewrite (\[etab G\]) in $L^2_{sc}(\S)$ norms $$\begin{split} \f{1}{|u|}\|(\at\nab)^i\etab\|_{L^2_{sc}(\S)}\leq&\f{1}{|{u_{\infty}}|}\|(\at\nab)^i\etab\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|(\at\nab)^{i}{\underline{\beta}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+i_3+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nab^{i_3}{\underline{\beta}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|{\mbox{tr}}{\underline{\chi}}(\at\nab)^{i}\eta\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+1=i}(\at\nab)^{i_1+1}{\mbox{tr}}{\underline{\chi}}(\at\nab)^{i_2}\eta\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\p\nabla^{i_4} {\mbox{tr}}{\underline{\chi}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}\p\nabla^{i_4} ({\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'. \end{split}$$ For the first term, we have $$\f{1}{|{u_{\infty}}|}\|(\at\nab)^i\etab\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}\leq \f{\M I^{(0)}({\underline{u}})}{|{u_{\infty}}|}\ls \f{1}{|{u_{\infty}}|}.$$ For the terms involving ${\underline{\beta}}$, we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|(\at\nab)^{i}{\underline{\beta}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+i_3+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nab^{i_3}{\underline{\beta}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\bigg(\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|(\at\nab)^{i}{\underline{\beta}}\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'\bigg)^{\f12} \bigg(\int_{{u_{\infty}}}^u\f{a}{|u'|^4} du'\bigg)^{\f12}+\int_{{u_{\infty}}}^u \f{a}{|u'|^3}\cdot \f{\at}{|u'|}\cdot O^2\, du'\\ =&\|(\at\nab)^{i}{\underline{\beta}}\|_{L^2_{sc}(\Hb_{{\underline{u}}}^{({u_{\infty}},u)})} \cdot \f{\at}{|u|^{\f32}}+\f{a^{\f32}}{|u|^3}O^{2}\\\leq& \underline{\M R}[{\underline{\beta}}]\cdot \f{\at}{|u|^{\f32}}+\f{a^{\f32}}{|u|^3}O^{2}\leq \f{\underline{\M R}[{\underline{\beta}}]}{|u|}+\f{O^2}{\at|u|}\leq \f{\underline{\M R}[{\underline{\beta}}]+1}{|u|}. \end{split}$$ Here we employ (\[4.6\]), (\[scale invariant norms 2\]) and (\[R i Hb\]). For next two terms, we use (\[Holder’s\]), bootstrap assumption (\[BA.0\]) and Proposition \[eta.bd\] to obtain $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|{\mbox{tr}}{\underline{\chi}}(\at\nab)^{i}\eta\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'+\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+1=i}(\at\nab)^{i_1+1}{\mbox{tr}}{\underline{\chi}}(\at\nab)^{i_2}\eta\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &\leq \int_{u_{\infty}}^{u}\f{1}{|u'|^2}\cdot\f{a}{|u'|^2}\|{\mbox{tr}}{\underline{\chi}}\|_{L^{\infty}_{sc}(S_{u',{\underline{u}}})}\|(\at\nab)^{i}\eta\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &\quad+\int_{u_{\infty}}^{u}\f{1}{|u'|^2}\|\sum_{i_1+i_2+1=i}(\at\nab)^{i_1+1}\bigg(\f{a}{|u'|}({\mbox{tr}}{\underline{\chi}}+\f{2}{|u'|})\bigg)(\at\nab)^{i_2}\eta\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &\leq\f{O[\eta]}{|u|}+\int_{{u_{\infty}}}^u \f{O^2}{|u'|^3}du'\leq \f{\M R[\b]+1}{|u|}+\f{O^2}{|u|^2}\leq \f{\underline{\M R}[{\underline{\beta}}]+1}{|u|}. \end{split}$$ As calculated above, for the sixth term we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\p\nabla^{i_4} {\mbox{tr}}{\underline{\chi}}\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ =&\int_{u_{\infty}}^{u}\f{\at}{|u'|}\|\sum_{i_1+i_2+i_3+i_4+1=i}(\at)^{i-1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\p\nabla^{i_4} (\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u} \f{\at}{|u'|}\cdot\f{O^3}{|u'|^2} du' \leq \f{\at \cdot O^3}{|u|^2}\leq \f{1}{|u|}. \end{split}$$ And for the last term, with (\[4.3\]) we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{a}{|u'|^3}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}\p\nabla^{i_4} ({\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ =&\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}\p\nabla^{i_4} (\f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\int_{u_{\infty}}^{u} \f{\at}{|u'|^2}\cdot\f{O^2}{|u'|} du' \leq \f{\at \cdot O^2}{|u|^2}\leq \f{1}{\at}\cdot\f{O^2}{|u|}\leq \f{1}{|u|}. \end{split}$$ Combining all the estimates derived, we have $$\f{1}{|u|}\|(\at\nab)^i\etab\|_{L^2_{sc}(\S)}\leq \f{1}{|u_{\infty}|}+\f{\underline{\M R}[{\underline{\beta}}]+\M R[\b]+1}{|u|}.$$ Multiplying $|u|$ on both sides, we obtain $$\|(\at\nab)^i\etab\|_{L^2_{sc}(\S)}\ls 1+\underline{\M R}[{\underline{\beta}}]+\M R[\b].$$ $L^2(\S)$ ESTIMATE FOR CURVATURE {#secCurvatureL2} ================================ For $i\leq 9$, we have \[a.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 9}\f{1}{\at}\|(\at\nab)^i\a\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls \underline{\M R}[\b]+1.$$ We have systematical null Bianchi equation: $$\nab_3 \a+\f12{\mbox{tr}}{\underline{\chi}}\a=\nab\b+\p\a+(\p,{\hat{\chi}})\Psi.$$ Commuting it with angular derivative for $i$ times, we have $$\begin{split} &\nab_3 \nab^i\a +\frac {i+1}{2} {\mbox{tr} {\underline{\chi}}}\nab^i\a\\ = &\nab^{i+1}\b+\sum_{i_1+i_2+i_3+1=i}\nab^{i_1}\p^{i_2+1}\nab^{i_3+1}\b+\sum_{i_1+i_2+1=i}\nab^{i_1+1}{\mbox{tr}}{\underline{\chi}}\nab^{i_2}\a+\sum_{i_1+i_2=i}\nab^{i_1}{\hat{{\underline{\chi}}}}\nab^{i_2}\a\\ &+\sum_{i_1+i_2+i_3+i_4+1=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\p, {\mbox{tr}}{\underline{\chi}}, {\hat{{\underline{\chi}}}})\nabla^{i_4}\a+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{\chi}})\nabla^{i_4}\Psi.\\ \end{split}$$ Denote the above equality as $$\nab_3 \nab^i\a +\f{i+1}{2} {\mbox{tr} {\underline{\chi}}}\nab^i\a=G.$$ Applying Proposition \[el\], it holds $$\begin{split} |u|^{i}\|\nab^i\a\|_{L^2(\S)}\leq&|u_{\infty}|^{i}\|\nab^i\a\|_{L^2(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u|u'|^{i}\|G\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ Multiplying $a^{-\f12}$ on both sides, we have $$\label{a G} \begin{split} a^{-\f12}|u|^{i}\|\nab^i\a\|_{L^2(\S)}\leq&a^{-\f12}|u_{\infty}|^{i}\|\nab^i\a\|_{L^2(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^u a^{-\f12}|u'|^{i}\|G\|_{L^{2}(S_{u',\underline{u}})}du'. \end{split}$$ From signature table and property (\[signature of derivative\]), we have $$s_2(\nab^i\a)=s_2(\a)+i\cdot\f{1}{2}=0+\f{i}{2}=\f{i}{2}.$$ By conversation of signature in each equation and property (\[signature of derivative\]), it holds $$s_2(G)=s_2(\nab_3 \nab^i\a)=s_2(\nab^i \a)+1=\f{i+2}{2}.$$ Using the definition of $L^2_{sc}(\S)$ norms $$\|\phi\|_{L_{sc}^{2}(\S)}:=a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^{2}(\S)},$$ we have $$\|\nab^i\a\|_{L^2_{sc}(\S)}=a^{-\f{i}{2}}|u|^{i}\|\nab^i\a\|_{L^2(\S)}, \quad \|G\|_{L^2_{sc}(\S)}=a^{-\f{i+2}{2}}|u|^{i+2}\|G\|_{L^2(\S)}.$$ That is equivalent to $$a^{-\f12}|u|^{i}\|\nab^i\a\|_{L^2(\S)}=a^{-\f12}\|(\at\nab)^i\a\|_{L^2_{sc}(\S)}, \quad a^{-\f12}|u|^{i}\|G\|_{L^2(\S)}= \f{\at}{|u|^2}\|(\at)^iG\|_{L^2_{sc}(\S)}.$$ We then rewrite (\[a G\]) in $L^2_{sc}(\S)$ norms $$\begin{split} &a^{-\f12}\|(\at\nab)^i\a\|_{L^2_{sc}(\S)}\\ \leq&a^{-\f12}\|(\at\nab)^i\a\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|(\at\nab)^{i+1}\b\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+1=i}(\at)^i\nab^{i_1}\p^{i_2+1}\nab^{i_3+1}\b\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+1=i}(\at)^i\nab^{i_1+1}{\mbox{tr}}{\underline{\chi}}\nab^{i_2}\a\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2=i}(\at)^i\nab^{i_1}{\hat{{\underline{\chi}}}}\nab^{i_2}\a\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\p, {\mbox{tr}}{\underline{\chi}}, {\hat{{\underline{\chi}}}})\nabla^{i_4}\a\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{\chi}})\nabla^{i_4}\Psi\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'. \\ \end{split}$$ For the first term, we have $$a^{-\f12}\|(\at\nab)^i\a\|_{L^2_{sc}(S_{u_{\infty},{\underline{u}}})}\leq {\M I^{(0)}({\underline{u}})}\ls 1.$$ For the terms involving $\b$, we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|(\at\nab)^{i+1}\b\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ &+\int_{u_{\infty}}^{u}\f{1}{|u'|^2}\|\sum_{i_1+i_2+i_3+1=i}(\at)^{i+1}\nabla^{i_1}\p^{i_2+1}\nab^{i_3+1}\b\|_{L^2_{sc}(S_{u',{\underline{u}}})}du'\\ \leq&\bigg(\int_{u_{\infty}}^{u}\f{a}{|u'|^2}\|(\at\nab)^{i+1}\b\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'\bigg)^{\f12} \bigg(\int_{{u_{\infty}}}^u\f{1}{|u'|^2} du'\bigg)^{\f12}+\int_{{u_{\infty}}}^u \f{1}{|u'|^2}\cdot \f{\at}{|u'|}\cdot O^2\, du'\\ \leq&\|(\at\nab)^{i+1}\b\|_{L^2_{sc}(\Hb_{{\underline{u}}}^{({u_{\infty}},u)})} \cdot \f{1}{|u|^{\f12}}+\f{a^{\f12}}{|u|^2}O^{2}\\ \leq&a^{-\f12}\|(\at\nab)^{i+1}\b\|_{L^2_{sc}(\Hb_{{\underline{u}}}^{({u_{\infty}},u)})} \cdot \f{\at}{|u|^{\f12}}+\f{a^{\f12}}{|u|^2}O^{2}\\ \leq& a^{-\f12}\|(\at\nab)^{i+1}\b\|_{L^2_{sc}(\Hb_{{\underline{u}}}^{({u_{\infty}},u)})}+1 \leq \underline{\M R}[{\underline{\beta}}]+1, \end{split}$$ where we employ (\[4.6\]), (\[scale invariant norms 2\]) and (\[R i Hb\]). For the next two terms, we use (\[4.6\]) and obtain $$\begin{split} &\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+1=i}(\at)^i\nab^{i_1+1}{\mbox{tr}}{\underline{\chi}}\nab^{i_2}\a\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2=i}(\at)^i\nab^{i_1}{\hat{{\underline{\chi}}}}\nab^{i_2}\a\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ \leq& \int_{u_{\infty}}^{u}\f{a^{-\f12}}{|u'|}\|\sum_{i_1+i_2+1=i}(\at)^{i+1}\nab^{i_1+1}\bigg(\f{a}{|u'|}({\mbox{tr}}{\underline{\chi}}+\f{2}{|u'|})\bigg)\nab^{i_2}(\f{\a}{\at})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{1}{|u'|}\|\sum_{i_1+i_2=i}(\at)^{i+1}\nab^{i_1}(\f{\at}{|u'|}{\hat{{\underline{\chi}}}})\nab^{i_2}(\f{\a}{\at})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ \leq& \int_{{u_{\infty}}}^{u}\f{a^{-\f12}}{|u'|}\cdot \f{\at}{|u'|}O^2 \,du'+ \int_{{u_{\infty}}}^{u}\f{1}{|u'|}\cdot \f{\at}{|u'|}O^2 \,du'\\ \leq& \f{O^2}{|u|}+\f{\at\cdot O^2}{|u|}\leq 1. \end{split}$$ For the last two terms, we have $$\begin{split} &\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4+1=i}(\at)^i\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\p, {\mbox{tr}}{\underline{\chi}}, {\hat{{\underline{\chi}}}})\nabla^{i_4}\a\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^i\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{\chi}})\nabla^{i_4}\Psi\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ \leq&\int_{u_{\infty}}^{u}a^{-\f12}\|\sum_{i_1+i_2+i_3+i_4+1=i}(\at)^{i+1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\f{a}{|u'|^2}\p, \f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}}, \f{a}{|u'|^2}{\hat{{\underline{\chi}}}})\nabla^{i_4}(\f{\a}{\at})\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ &+\int_{u_{\infty}}^{u}\f{\at}{|u'|^2}\|\sum_{i_1+i_2+i_3+i_4=i}(\at)^{i+1}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\f{\p}{\at},\f{{\hat{\chi}}}{\at})\nabla^{i_4}\Psi\|_{L^2_{sc}(S_{u',{\underline{u}}})}du' \\ \leq& \int_{{u_{\infty}}}^{u}a^{-\f12} \f{a}{|u|^2} O^3 \,du'+ \int_{{u_{\infty}}}^{u}\f{\at}{|u'|^2}\cdot \f{\at}{|u'|}O^2 \,du'\\ \leq& \f{\at\cdot O^2}{|u|}+\f{a\cdot O^2}{|u|^2}\leq 1. \end{split}$$ Gathering all the estimates, for sufficiently large $a$ we have showed that $$\sum_{i\leq 9}\f{1}{\at}\|(\at\nab)^i\a\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls \underline{\M R}[\b]+1.$$ Let $\Psi\in \{\b, \rho, \sigma, {\underline{\beta}}, \ab\}$, we proceed to prove \[etab.bd\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\sum_{i\leq 9}\|(\at\nab)^i\Psi\|_{L^2_{sc}(S_{u,{\underline{u}}})} \ls \M R[\a]+\underline{\M R}[\b]+1.$$ For $\Psi\in \{\b, \rho, \sigma, {\underline{\beta}}, \ab\}$, we have the systematical null Bianchi equations: $$\nab_4 \Psi=\nab\Psi+\nab\alpha+({\hat{{\underline{\chi}}}},\p) (\Psi, \a).$$ By Proposition \[commute\], we commute this equations with $i$ angular derivatives $$\begin{split} \nab_4\nab^i\Psi=&\nab^{i+1}\Psi+\nab^{i+1}\a+\sum_{i_1+i_2+i_3+1=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3+1}(\Psi, \a)\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nab^{i_3}(\p,{\hat{\chi}})\nabla^{i_4}\Psi+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}})\nab^{i_4}(\Psi, \a). \end{split}$$ Applying Proposition \[transport\] and multiplying $(\at)^i$ on both sides, we have $$\begin{split} &\|(\at\nab)^{i}\Psi\|_{L^2_{sc}(\S)}\\ \leq&\int_0^{{\underline{u}}}\|(\at)^{i}\nab^{i+1}\Psi\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'+\int_0^{{\underline{u}}}\|(\at)^{i}\nab^{i+1}\a\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+1=i}\int_0^{{\underline{u}}}\|(\at)^{i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3+1}(\Psi, \a)\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}}\|(\at)^{i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{\chi}})\nab^{i_4}\Psi\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}}\|(\at)^{i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p,{\hat{{\underline{\chi}}}})\nab^{i_4}(\Psi, \a)\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq&a^{-\f12}\int_0^{{\underline{u}}}\|(\at)^{i+1}\nab^{i+1}\Psi\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'+a^{-\f12}\int_0^{{\underline{u}}}\|(\at)^{i+1}\nab^{i+1}\a\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+1=i}\int_0^{{\underline{u}}}\|(\at)^{i+1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3+1}(\f{\Psi}{\at}, \f{\a}{\at})\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4=i}\int_0^{{\underline{u}}}\|(\at)^{i+1}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\f{\p}{\at},\f{{\hat{\chi}}}{\at})\nab^{i_4}\Psi\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2+i_3+i_4+1=i}\int_0^{{\underline{u}}}a^{-\f12}|u|\|(\at)^{i+1}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\f{\at}{|u|}\p,\f{\at}{|u|}{\hat{{\underline{\chi}}}})\nab^{i_4}(\f{\Psi}{\at}, \f{\a}{\at})\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ &+\sum_{i_1+i_2=i}\int_0^{{\underline{u}}}a^{-\f12}|u|\|(\at)^{i+1}\nabla^{i_1}(\f{\at}{|u|}\p,\f{\at}{|u|}{\hat{{\underline{\chi}}}})\nab^{i_2}(\f{\Psi}{\at}, \f{\a}{\at})\| _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\\ \leq& a^{-\f12}\bigg(\int_0^{{\underline{u}}}\|(\at)^{i+1}\nab^{i+1}\Psi\|^2 _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12}+a^{-\f12}\bigg(\int_0^{{\underline{u}}}\|(\at)^{i+1}\nab^{i+1}\a\|^2 _{L^{2}_{sc}(S_{u,{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12}\\ &+\f{\at}{|u|}O^2+a^{-\f12}|u|\f{a}{|u|^2}O^3+a^{-\f12}|u|\f{\at}{|u|}(O[{\hat{{\underline{\chi}}}}]\cdot O[\a]+1)\\ \leq& a^{-\f12}\|(\at\nab)^{i+1}\Psi\|_{{L^2_{sc}(H_{u}^{(0,{\underline{u}})})}}+a^{-\f12}\|(\at\nab)^{i+1}\a\|_{{L^2_{sc}(H_{u}^{(0,{\underline{u}})})}}+O[{\hat{{\underline{\chi}}}}]\cdot O[\a]+1\\ \leq& R[\a]+O[{\hat{{\underline{\chi}}}}]\cdot O[\a]+1\ls \M R[\a]+O[\a]+1\ls \M R[\a]+\underline{\M R}[\b]+1 , \end{split}$$ where we use (\[4.6\]), (\[4.7\]), Proposition \[chibh.bd\] and Proposition \[a.bd\]. ENERGY ESTIMATE {#energy estimate} =============== In this section with scale invariant norms we will derive energy estimates for curvature components and their angular derivatives. Our goal is to show that $$\label{R Rb} \M R+{\underline{\mathcal{R}}}\ls \M I^{(0)}+(\M I^{(0)})^2+1.$$ Together with the estimates derived in previous sections, we hence improve all the bootstrap assumptions in . And Theorem \[main.thm1\] is therefore established. Integration by Parts Formula ---------------------------- We first state and prove several useful formula. For $D_{u,{\underline{u}}}=(u_{\infty}, u)\times (0,{\underline{u}})$ by direct computations, we have \[ee 1\] Suppose $\phi_1$ and $\phi_2$ are $r$ tensorfields, then $$\begin{split} &\int_{D_{u,{\underline{u}}}} \phi_1 \nabla_4\phi_2+\int_{D_{u,{\underline{u}}}}\phi_2\nabla_4\phi_1\\ =& \int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}} \phi_1\phi_2-\int_{\Hb_0^{({u_{\infty}},u)}} \phi_1\phi_2+\int_{D_{u,{\underline{u}}}}(2\omega-\trch)\phi_1\phi_2. \end{split}$$ \[ee 2\] Suppose we have an $r$ tensorfield $^{(1)}\phi$ and an $r-1$ tensorfield $^{(2)}\phi$, then $$\begin{split} &\int_{D_{u,{\underline{u}}}}{ }^{(1)}\phi^{A_1A_2...A_r}\nabla_{A_r}{ }^{(2)}\phi_{A_1...A_{r-1}}+\int_{D_{u,{\underline{u}}}}\nabla^{A_r}{ }^{(1)}\phi_{A_1A_2...A_r}{ }^{(2)}\phi^{A_1...A_{r-1}}\\ =& -\int_{D_{u,{\underline{u}}}}(\eta+\etab){ }^{(1)}\phi{ }^{(2)}\phi. \end{split}$$ We will also need the following analogue of Proposition \[ee 1\] with $u$ weights incorporated. \[ee 3\] Suppose $\phi$ is an $r$ tensorfield and let $\lambda_1=2(\lambda_0-\f12)$. Then $$\begin{split} &2\int_{D_{u,{\underline{u}}}} |u'|^{2\lambda_1}\phi (\nabla_3+\lambda_0{\mbox{tr} {\underline{\chi}}})\phi\\ =& \int_{H_u^{(0,{\underline{u}})}} |u|^{2\lambda_1}|\phi|^2-\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} |{u_{\infty}}|^{2\lambda_1}|\phi|^2+\int_{D_{u,{\underline{u}}}}|u'|^{2\lambda_1}f|\phi|^2, \end{split}$$ where $f$ obeys bound $$|f|\ls \f{O}{|u|^2}.$$ Slightly modifying , we have $$\begin{split} \frac{d}{du}(\int_{\S}&|u|^{2\lambda_1}\Omega|\phi|^2)={\underline{L}}\bigg(\int_{\S}|u|^{2\lambda_1}\O|\phi|^2 \bigg)\\ =&\int_{\S}\Omega^2 \l 2|u|^{2\lambda_1}<\phi, \nab_3\phi+\lambda_0{\mbox{tr} {\underline{\chi}}}\phi>\r\\ &+\int_{\S}\Omega^2\l |u|^{2\lambda_1} (-\f{2\lambda_1 (e_3u)}{|u|}+(1-2\lambda_0){\mbox{tr} {\underline{\chi}}}-2{\underline{\omega}})|\phi|^2\r. \end{split}$$ Here we use ${\underline{L}}=\O e_3=\f{\partial}{\partial u}+b^A\f{\partial}{\partial \theta^A}$. By (\[trchib additional\]) and bootstrap assumption \[BA.0\], we have $$|-\f{2\lambda_1 (e_3u)}{|u|}+(1-2\lambda_0){\mbox{tr} {\underline{\chi}}}-2{\underline{\omega}}|\ls \f{O}{|u|^2}.$$ The proposition then follows via integrating with respect to $du\,d{\underline{u}}$ and applying the fundamental theorem of calculus in $u$.\ Observe that for $(\Psi_1, \Psi_2)\in \{ (\a,\b), \big(\b, (\rho, \sigma)\big), \big( (\rho, \sigma), {\underline{\beta}}\big), ({\underline{\beta}}, \ab)\}$, we can rewrite null Bianchi equations into the systematic forms: $$\label{snb 1} \nab_3 \Psi_1+\bigg(\f{1}{2}+s_2(\Psi_1)\bigg){\mbox{tr}}{\underline{\chi}}\Psi_1-\M D\Psi_2=(\p, {\hat{\chi}})\Psi,$$ $$\label{snb 2} \nab_4 \Psi_2-\M D^*\Psi_1=\p\Psi+{\hat{\chi}}(\Psi, \a),$$ where $\Psi\in \{\b, \rho, \sigma, {\underline{\beta}}, \ab\}$. Here we denote $\M D$ to be a differential operator on $\S$, and $\M D^*$ is the $L^2(\S)$ adjoint operator of $\M D$. We further commute (\[snb 1\]) and (\[snb 2\]) with angular derivative $\nab$ for $i$ times and get $$\label{snb 3} \nab_3\nab^i\Psi_1+\big(\f{1+i}{2}+s_2(\Psi_1)\big){\mbox{tr}}{\underline{\chi}}\nab^i\Psi_1-\M D \nab^i\Psi_2=P,$$ $$\label{snb 4} \nab_4\nab^i\Psi_2-\M D^*\nab^{i}\Psi=1+Q.$$ The forms of $P$ and $Q$ will be specified later. Check signature $s_2$, we have $$\begin{split} s_2(\nab^i \Psi_1)=\f{i}{2}+s_2(\Psi_1), \quad &s_2(\nab^i \Psi_2)=\f{i+1}{2}+s_2(\Psi_1)\\ s_2(P)=s_2(\nab_3\nab^i\Psi_1)=\f{i+2}{2}+s_2(\Psi_1), \quad &s_2(Q)=s_2(\M D^*\nab^i\Psi_1)=\f{i+1}{2}+s_2(\Psi_1). \end{split}$$ The Hodge structure will play a crucial role: for pair $(\Psi_1, \Psi_2)$ or pair $(\nab^i\Psi_1, \nab^i\Psi_2)$, the angular derivative operator $\M D$ and its $L^2$ adjoint operator $\M D^*$ form a Hodge system. Through Proposition \[ee 2\], we have $$\label{snb 5} \begin{split} &\int_{\S}\Psi_1\,\M D \Psi_2+\Psi_2\,\M D^*\Psi_1=-\int_{\S}(\eta+{\underline{\eta}})\Psi_1 \Psi_2,\\ &\int_{\S}\nab^i\Psi_1\,\M D \nab^i\Psi_2+\nab^i\Psi_2\,\M D^*\nab^i\Psi_1=-\int_{\S}(\eta+{\underline{\eta}})\nab^i\Psi_1\nab^i\Psi_2. \end{split}$$ We now move forward and apply Proposition \[ee 3\] for $\nab^i\Psi_1$. With $$\lambda_0=\f{1+i}{2}+s_2(\Psi_1), \quad \lambda_1:=2\lambda_0-1=i+2s_2(\Psi_1), \mbox{ we get}$$ $$\label{6.6} \begin{split} &2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_1 \bigg(\nabla_3+\big(\f{1+i}{2}+s_2(\Psi_1)\big){\mbox{tr} {\underline{\chi}}}\bigg)\nab^i\Psi_1\\\ =& \int_{H_u^{(0,{\underline{u}})}} |u|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2-\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} |{u_{\infty}}|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}f|\nab^i\Psi_1|^2, \end{split}$$ where $|f|\leq O/|u'|^2$. We also use Proposition \[ee 1\] with substitution $\phi_1=\phi_2=|u|^{i+2s_2(\Psi_1)}\nab^i\Psi_2$ $$\label{6.7} \begin{split} &2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)} \nab^i \Psi_2 \nabla_4\nab^i\Psi_2\\ =& \int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}} |u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2-\int_{\Hb_0^{({u_{\infty}},u)}} |u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ &+\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}(2\omega-\trch)|\nab^i\Psi_2|^2. \end{split}$$ Add (\[6.6\]) and (\[6.7\]), we hence obtain $$\begin{split} &2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_1 \bigg(\nabla_3+\big(\f{1+i}{2}+s_2(\Psi_1)\big){\mbox{tr} {\underline{\chi}}}\bigg)\nab^i\Psi_1\\ &+2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)} \nab^i\Psi_2 \nabla_4\nab^i\Psi_2\\ =& \int_{H_u^{(0,{\underline{u}})}} |u|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2-\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} |{u_{\infty}}|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}f|\nab^i\Psi_1|^2\\ &+\int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2-\int_{\Hb_0^{({u_{\infty}},u)}} |u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ &+\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}(2\omega-\trch)|\nab^i\Psi_2|^2. \end{split}$$ Apply (\[snb 3\]) and (\[snb 4\]). Wtih the help of (\[snb 5\]), we then arrive at $$\begin{split} &\int_{H_u^{(0,{\underline{u}})}} |u|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ =&\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} |{u_{\infty}}|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{\Hb_{0}^{({u_{\infty}},u)}}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ &+2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_1 \cdot P+2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_2 \cdot Q\\ &-2\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}(\eta+{\underline{\eta}})\nab^i\Psi_1\nab^i\Psi_2\\ &+\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}f|\nab^i\Psi_1|^2+\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}(2\omega-\trch)|\nab^i\Psi_2|^2. \end{split}$$ Using $|(\eta+{\underline{\eta}})\nab^i\Psi_1\nab^i\Psi_2|\leq |\eta+{\underline{\eta}}|(\nab^i\Psi_1)^2+|\eta+{\underline{\eta}}|(\nab^i\Psi_2)^2,$ and the fact $$|\eta+{\underline{\eta}}|\leq \at O/{|u'|^2}, \quad |f|\leq O/|u'|^2, \quad |2\o-{\mbox{tr}}\chi|\leq {O}/{|u'|},$$ by applying Grönwall’s inequality twice (one for $du$, one for $d{\underline{u}}$), we obtain $$\begin{split} &\int_{H_u^{(0,{\underline{u}})}} |u|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ \ls&\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} |{u_{\infty}}|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{\Hb_{0}^{({u_{\infty}},u)}}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ &+2\int_{D_{u,{\underline{u}}}} |u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_1 \cdot P+2\int_{D_{u,{\underline{u}}}}|u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_2 \cdot Q. \end{split}$$ Multiply $a^{-i-2s_2(\Psi_1)}$ on both sides $$\label{6.8} \begin{split} &\int_{H_u^{(0,{\underline{u}})}} a^{-i-2s_2(\Psi_1)}|u|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}}a^{-i-2s_2(\Psi_1)}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ \ls&\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} a^{-i-2s_2(\Psi_1)}|{u_{\infty}}|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_1|^2+\int_{\Hb_{0}^{({u_{\infty}},u)}}a^{-i-2s_2(\Psi_1)}|u'|^{2i+4s_2(\Psi_1)}|\nab^i\Psi_2|^2\\ &+2\int_{D_{u,{\underline{u}}}} a^{-i-2s_2(\Psi_1)}|u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_1 \cdot P+2\int_{D_{u,{\underline{u}}}}a^{-i-2s_2(\Psi_1)}|u'|^{2i+4s_2(\Psi_1)}\nab^i\Psi_2 \cdot Q. \end{split}$$ With signature identities $$\begin{split} s_2(\nab^i \Psi_1)=\f{i}{2}+s_2(\Psi_1), \quad &s_2(\nab^i \Psi_2)=\f{i+1}{2}+s_2(\Psi_1)\\ s_2(P)=s_2(\nab_3\nab^i\Psi_1)=\f{i+2}{2}+s_2(\Psi_1), \quad &s_2(Q)=s_2(\M D^*\nab^i\Psi_1)=\f{i+1}{2}+s_2(\Psi_1), \end{split}$$ and definitions $$\|\phi\|_{L^2_{sc}(\S)}=a^{-s_2(\phi)}|u|^{2s_2(\phi)}\|\phi\|_{L^2(\S)},$$ $$\|\phi\|_{L^1_{sc}(\S)}=a^{-s_2(\phi)}|u|^{2s_2(\phi)-1}\|\phi\|_{L^1(\S)},$$ we rewrite (\[6.8\]) as $$\begin{split} &\int_{H_u^{(0,{\underline{u}})}} \|\nab^i\Psi_1\|^2_{L^2_{sc}(\S)}+\int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}}\f{a}{|u'|^2}\|\nab^i\Psi_2\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}\\ \ls&\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} \|\nab^i\Psi_1\|^2_{L^2_{sc}(S_{{u_{\infty}},{\underline{u}}})}+\int_{\Hb_{0}^{({u_{\infty}},u)}}\f{a}{|{u_{\infty}}|^2}\|\nab^i\Psi_2\|^2_{L^2_{sc}(S_{{u_{\infty}},{\underline{u}}})}\\ &+2\int_{D_{u,{\underline{u}}}} \f{a}{|u'|}\|\nab^i\Psi_1 \cdot P\|_{L^1_{sc}(S_{u',{\underline{u}}'})}+2\int_{D_{u,{\underline{u}}}}\f{a}{|u'|}\|\nab^i\Psi_2 \cdot Q\|_{L^1_{sc}(S_{u',{\underline{u}}'})}. \end{split}$$ Recall the definition in (\[scale invariant norms 2\]) $$\begin{split} \|\phi\|^2_{L^2_{sc}(H_u^{(0,\underline{u})})}:=& \int_0^{\underline{u}}\|\phi\|^2_{L^2_{sc}(S_{u,\underline{u}'})}d\underline{u}',\\ \|\phi\|^2_{L^2_{sc}(\underline{H}_{\underline{u}}^{(u_{\infty},u)})}:=& \int_{u_{\infty}}^{u}{\frac{a}{|u'|^2}}\|\phi\|^2_{L^2_{sc}(S_{u',\underline{u}})}du', \end{split}$$ we therefore arrive at \[EE\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , assume pair $(\nab^i\Psi_1, \nab^i\Psi_2)$ satisfying $$\nab_3\nab^i\Psi_1+\big(\f{1+i}{2}+s_2(\Psi_1)\big){\mbox{tr}}{\underline{\chi}}\nab^i\Psi_1-\M D \nab^i\Psi_2=P,$$ $$\nab_4\nab^i\Psi_2-\M D^*\nab^{i}\Psi_1=Q,$$ where $\M D^*$ is the $L^2$ adjoint operator of $\M D$, then it follows $$\label{EE1} \begin{split} &\int_{H_u^{(0,{\underline{u}})}} \|\nab^i\Psi_1\|^2_{L^2_{sc}(\S)}+\int_{\Hb_{{\underline{u}}}^{({u_{\infty}},u)}}\f{a}{|u'|^2}\|\nab^i\Psi_2\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}\\ \ls&\int_{H_{{u_{\infty}}}^{(0,{\underline{u}})}} \|\nab^i\Psi_1\|^2_{L^2_{sc}(S_{{u_{\infty}},{\underline{u}}})}+\int_{\Hb_{0}^{({u_{\infty}},u)}}\f{a}{|{u_{\infty}}|^2}\|\nab^i\Psi_2\|^2_{L^2_{sc}(S_{{u_{\infty}},{\underline{u}}})}\\ &+\int_{D_{u,{\underline{u}}}} \f{a}{|u'|}\|\nab^i\Psi_1 \cdot P\|_{L^1_{sc}(S_{u',{\underline{u}}'})}+\int_{D_{u,{\underline{u}}}}\f{a}{|u'|}\|\nab^i\Psi_2 \cdot Q\|_{L^1_{sc}(S_{u',{\underline{u}}'})}. \end{split}$$ Energy Estimate in Scale Invariant Norms ---------------------------------------- We are now ready to prove $$\M R+{\underline{\mathcal{R}}}\ls \M I^{(0)}+(\M I^{(0)})^2+1.$$ Let’s start with the pair $(\a, \b)$. \[ee 4\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\begin{split} &\f{1}{\at}\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\f{1}{\at}\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ \leq&\f{1}{\at}\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\f{1}{\at}\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+\f{1}{a^{\f13}} . \end{split}$$ For pair $(\Psi_1, \Psi_2)=(\a, \b)$ we have $$\nab_4 \b-{\M D}^* \a=\p(\b, \a),$$ $$\nab_3 \a+\f12{\mbox{tr}}{\underline{\chi}}\a-\M D\b=(\p, {\hat{\chi}})(\Psi, \b, \a).$$ Commuting the above equations with $\nab$ for i times, we get $$\begin{split} &\nab_3\nab^i\a+\f{1+i}{2}{\mbox{tr}}{\underline{\chi}}\nab^i\a-\M D\nab^i\b\\ =&\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\b+\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}\bigg((\p,{\hat{\chi}})(\Psi, \b, \a)\bigg)\\ &+\sum_{i_1+i_2+i_3+i_4+1=i} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\a+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \a\\ =&\sum_{i_1+i_2+i_3+i_4+1=i} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\a+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} (\Psi, \b, \a)\\ =&F_1. \end{split}$$ And $$\begin{split} &\nab_4\nab^i\b-{\M D}^*\nab^i\a\\ =&\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\b +\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3} \bigg(\p(\b, \a)\bigg)\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} \b\\ =&\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} (\b, \a)\\ =&G_1. \end{split}$$ Applying Proposition \[ee 1\] to Proposition \[ee 3\], it follows $$\begin{split} &\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ \leq & \|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+N_1+M_1. \end{split}$$ where $$\begin{split} N_1=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iF_1\cdot(\at\nab)^i\a\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}',\\ M_1=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iG_1\cdot(\at\nab)^i\b\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}',\\ \end{split}$$ By Hölder’s inequalities in scale invariant norms, we have $$\begin{split} N_1=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iF_1\cdot(\at\nab)^i\a\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}',\\ \leq&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^iF_1\|_{L^2_{sc}(S_{u',{\underline{u}}'})} \|(\at\nab)^i\a\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'\\ \leq&\int_{u_{\infty}}^u \frac{a}{|u'|^2}\bigg(\int_0^{{\underline{u}}}\|(\at)^iF_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12} \|(\at\nab)^i\a\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})} du'.\\ \leq&\int_{u_{\infty}}^u \frac{a}{|u'|^2}\bigg(\int_0^{{\underline{u}}}\|(\at)^iF_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12}du'\cdot \sup_{u'}\|(\at\nab)^i\a\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})}, \end{split}$$ where $$\begin{split} F_1=&\sum_{i_1+i_2+i_3+i_4+1=i} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\a\\ &+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} (\Psi, \b, \a). \end{split}$$ Denote $$H_1:=\int_0^{{\underline{u}}}\|(\at)^i F_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}d{\underline{u}}'.$$ We further have $$\begin{split} H_1=& \int_{0}^{{\underline{u}}} \|(\at)^iF_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})} d{\underline{u}}'\\ \leq& \int_{0}^{{\underline{u}}} a^{-1}\|\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}'})} \|(\at\nab)^i(\Psi, \b, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})} d{\underline{u}}'\\ &+ \sum_{i_1+i_2+i_3+i_4+1=i} \int_{0}^{{\underline{u}}} \f{|u'|^4}{a^2}\|(\at)^{i+1} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\f{a}{|u'|^2}{\hat{{\underline{\chi}}}},\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}})\nab^{i_4}(a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})} d{\underline{u}}'\\ &+\sum_{\substack{i_1+i_2+i_3+i_4=i\\ i_4\leq i-1}} \int_{0}^{{\underline{u}}} \f{|u'|^2}{a}\|(\at)^{i+1} \nabla^{i_1}\p^{i_2}\nabla^{i_3}\big(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\big)\\ &\quad\quad\quad \quad \quad\quad\quad\quad \quad \quad \quad \quad\quad\quad\quad\quad \times\nab^{i_4}(a^{-\f12}\Psi, a^{-\f12}\b, a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})} d{\underline{u}}'\\ \leq& \int_{0}^{{\underline{u}}} \|(\at\nab)^i (a^{-\f12}\Psi, a^{-\f12}\b, a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})} d{\underline{u}}'\cdot \big(O^2[{\hat{\chi}}]+O^2[{\hat{{\underline{\chi}}}}]\big)\\ &+\int_0^{{\underline{u}}} \f{|u'|^4}{a^2}\f{a^2}{|u'|^4} d{\underline{u}}'\cdot O^{6}+\int_0^{{\underline{u}}}\f{|u'|^2}{a}\f{a}{|u'|^2}\cdot O^{4}\\ \leq&R^2[\a]\cdot \big(O^2[{\hat{\chi}}]+O^2[{\hat{{\underline{\chi}}}}]\big)+O^6+O^4. \end{split}$$ Therefore, for $N_1$ we have $$N_1\ls (R[\a]\cdot O[{\hat{\chi}}]+R[\a]\cdot O[{\hat{{\underline{\chi}}}}]+O^3+O^2)\cdot \sup_{u'}\|(\at\nab)^i\a\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})}.$$ We treate $M_1$ in a similar way. $$\begin{split} M_1=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iG_1\cdot(\at\nab)^i\b\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'\\ \leq&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_1\|_{L^2_{sc}(S_{u',{\underline{u}}'})} \|(\at\nab)^i\b\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'\\ \leq&\int_0^{{\underline{u}}}\bigg(\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du'\bigg)^{\f12} \bigg(\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at\nab)^i \b\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du'\bigg)^{\f12} d{{\underline{u}}}'\\ \leq&\int_0^{{\underline{u}}}\bigg(\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du'\bigg)^{\f12}\|(\at\nab)^i\b\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}d{{\underline{u}}}'\\ \leq&\bigg(\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du' d{\underline{u}}'\bigg)^{\f12}\cdot \|(\at\nab)^i\b\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}, \end{split}$$ where $$\begin{split} G_1=&\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} (\b, \a). \end{split}$$ Let $$J_1:=\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_1\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'.$$ Then, by (\[4.6\]) $$\begin{split} J_1\leq& \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i (\p,{\hat{\chi}})\nab^i(\b, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \f{a}{|u'|^2} \|(\at)^i\sum_{\substack{i_1+i_2+i_3+i_4=i\\i_4\leq i-1}}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} (\b, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})} du'd{\underline{u}}'\\ \leq& \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^4}\|\p, {\hat{\chi}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})} \|(\at)^i\nab^i (\b, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \f{a^2}{|u'|^2}\|(\at)^{i+1}\sum_{\substack{i_1+i_2+i_3+i_4=i\\i_4\leq i-1}}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(a^{-\f12}\p, a^{-\f12}{\hat{\chi}})\nabla^{i_4} (a^{-\f12}\b, a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})} du'd{\underline{u}}'\\ \leq& \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a^2}{|u'|^4}\|a^{-\f12}\p, a^{-\f12}{\hat{\chi}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})} \|(\at)^i\nab^i (\b, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \f{a^2}{|u'|^2}\|(\at)^{i+1}\sum_{\substack{i_1+i_2+i_3+i_4=i\\i_4\leq i-1}}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(a^{-\f12}\p, a^{-\f12}{\hat{\chi}})\nabla^{i_4} (a^{-\f12}\b, a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})} du'd{\underline{u}}'\\ \leq&O^2\cdot \sup_{u'}\int_0^{{\underline{u}}}\|(\at)^i\nab^i (\b, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}d{\underline{u}}'\cdot \int_{u_{\infty}}^{u}\f{a^2}{|u'|^4}du'+ \int_{u_{\infty}}^{u} \f{a^3}{|u'|^4}du'\cdot O^{4}\\ \leq&a^{-1}\sup_{u'}\|(\at\nab)^i(\b,\a)\|^2_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})}\cdot O^2+O^{4}\leq R^2\cdot O^2+O^4. \end{split}$$ Hence, $$\begin{split} M_1\leq&\bigg(\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du' d{\underline{u}}'\bigg)^{\f12}\cdot \|(\at\nab)^i\b\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}\\ \leq& (R\cdot O+O^2)\cdot \|(\at\nab)^i\b\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}\\ \end{split}$$ Combining all the above estimates together, we obtain $$\begin{split} &\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ \leq & \|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+N_1+M_1\\ \leq& \|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}\\ &+(R\cdot O+O^3+O^2)\cdot \sup_{u'}\|(\at\nab)^i\a\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})}\\ &+\at(R\cdot O+O^2)\cdot a^{-\f12}\|(\at\nab)^i\b\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}. \end{split}$$ Hence, for sufficiently large $a$, we have $$\begin{split} &a^{-1}\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+a^{-1}\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ \leq& a^{-1}\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+a^{-1}\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}\\ &+a^{-\f12}(R\cdot O+O^{3}+O^2)\cdot \sup_{u'}a^{-\f12}\|(\at\nab)^i\a\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})}\\ &+a^{-\f12}(R\cdot O+O^2)\cdot a^{-\f12}\|(\at\nab)^i\b\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}\\ \leq& a^{-1}\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+a^{-1}\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}\\ &+a^{-\f12}(R\cdot O+O^3+O^2)\cdot R+a^{-\f12}(R\cdot O+O^2)\cdot R\\ \leq& a^{-1}\|(\at\nab)^i\a\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+a^{-1}\|(\at\nab)^i\b\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+\f{1}{a^{\f13}}. \end{split}$$ This further implies $$\label{eee 4} \M R^2[\a]+{\underline{\mathcal{R}}}^2[\b]\leq \M R^2_0[\a]+{\underline{\mathcal{R}}}^2_0[\b]+\f{1}{a^{\f14}} \quad \mbox{ and }$$ $$\label{eee 5} \M R[\a]+{\underline{\mathcal{R}}}[\b]\leq 2\,\M R_0[\a]+2\,{\underline{\mathcal{R}}}_0[\b]+\f{1}{a^{\f18}}.$$ We next derive estimates for other pairs. \[ee 5\] Under the assumptions of Theorem \[main.thm1\] and the bootstrap assumptions , we have $$\begin{split} &\|(\at\nab)^i\b\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i\rho\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}\\ &\quad \quad \quad \quad \quad+\|(\at\nab)^i\sigma\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i{\underline{\beta}}\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}\ls \M I^{(0)}+(\M I^{(0)})^2+1, \end{split}$$ $$\begin{split} &\|(\at\nab)^i\rho\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}+\|(\at\nab)^i\sigma\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ &\quad \quad \quad \quad \quad+\|(\at\nab)^i{\underline{\beta}}\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}+\|(\at\nab)^i\ab\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})} \ls \M I^{(0)}+(\M I^{(0)})^2+1. \end{split}$$ For $(\Psi_1, \Psi_2)\in \{\big(\b, (\rho, \sigma)\big), \big( (\rho, \sigma), {\underline{\beta}}\big), ({\underline{\beta}}, \ab)\}$ and $\Psi\in \{\b, \rho, \sigma, {\underline{\beta}}, \ab\}$, we have the systematic null Bianchi equations: $$\nab_3 \Psi_1+\bigg(\f{1}{2}+s_2(\Psi_1)\bigg){\mbox{tr}}{\underline{\chi}}\Psi_1-\M D\Psi_2=(\p, {\hat{\chi}})\Psi,$$ $$\nab_4 \Psi_2-{\M D}^* \Psi_1=\p\Psi+{\hat{\chi}}(\Psi, \a).$$ Commuting these equations with $\nab$ for i times, we have $$\begin{split} &\nab_3\nab^i\Psi_1+\bigg(\f{1+i}{2}+s_2(\Psi_1)\bigg){\mbox{tr}}{\underline{\chi}}\nab^i\Psi_1-\M D\nab^{i}\Psi_2\\ =&\sum_{i_1+i_2+i_3+1=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\nab\Psi+\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3} \bigg((\p,{\hat{\chi}})\Psi\bigg)\\ &+\sum_{i_1+i_2+i_3+i_4+1=i} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\Psi+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \Psi\\ =&\sum_{i_1+i_2+i_3+i_4+1=i} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\Psi+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \Psi\\ =&F_2.\\ \end{split}$$ And $$\begin{split} &\nab_4\nab^i\Psi_2-{\M D}^*\nab^{i}\Psi_1\\ =&\sum_{i_1+i_2+i_3+1=i}\nabla^{i_1}\p^{i_2+1}\nabla^{i_3}\nab\Psi+\sum_{i_1+i_2+i_3=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3} \bigg(\p\Psi+{\hat{\chi}}(\Psi, \a)\bigg)+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}})\nabla^{i_4} \Psi\\ =&\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}})\nabla^{i_4} (\Psi, \a)\\ =&G_2. \end{split}$$ Applying Proposition \[ee 1\] and Proposition \[ee 2\], we have $$\label{N2+M2} \begin{split} &\|(\at\nab)^i\Psi_1\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i\Psi_2\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ \leq & \|(\at\nab)^i\Psi_1\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\|(\at\nab)^i\Psi_2\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+N_2+M_2. \end{split}$$ Here $$\begin{split} N_2=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iF_2\cdot(\at\nab)^i\Psi_1\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}',\\ M_2=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iG_2\cdot(\at\nab)^i\Psi_2\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'.\\ \end{split}$$ Employing (\[Holder’s\]) and bootstrap assumption (\[BA.0\]), we hence have $$\begin{split} N_2=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iF_2\cdot(\at\nab)^i\Psi\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}',\\ \leq&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^iF_2\|_{L^2_{sc}(S_{u',{\underline{u}}'})} \|(\at\nab)^i\Psi\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'\\ \leq&\int_{u_{\infty}}^u \frac{a}{|u'|^2}\bigg(\int_0^{{\underline{u}}}\|(\at)^iF_2\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}d{\underline{u}}'\bigg)^{\f12} \|(\at\nab)^i\Psi\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})} du'.\\ \leq& \int_{u_{\infty}}^u \frac{a}{|u'|^2}\sup_{{\underline{u}}'}\|(\at)^iF_2\|_{L^2_{sc}(S_{u',{\underline{u}}'})}\cdot\sup_{{\underline{u}}}\|(\at\nab)^i\Psi\|_{L^2_{sc}(H_{u'}^{(0,{\underline{u}})})} du'\\ \leq& \int_{u_{\infty}}^u \frac{a}{|u'|^2}\sup_{{\underline{u}}'}\|(\at)^iF_2\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'\cdot R, \end{split}$$ where $R$ is bootstrap constant for curvature estimates. And recall $F_2$ is of the form: $$\begin{split} F_2=&\sum_{i_1+i_2+i_3+i_4+1=i} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}({\hat{{\underline{\chi}}}},{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\Psi+\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}}, {\widetilde{{\mbox{tr}}{\underline{\chi}}}})\nabla^{i_4} \Psi\\ \end{split}$$ Thus, by using (\[Holder’s\]), (\[4.6\]), (\[4.7\]) and letting $a$ to be sufficiently large, we have $$\label{N2} \begin{split} &N_2\leq \int_{u_{\infty}}^u \frac{a}{|u'|^2}\sup_{{\underline{u}}'}\|(\at)^iF\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'\cdot R\\ \leq& \int_{u_{\infty}}^u \frac{a^{\f12}}{|u'|^2}\sup_{{\underline{u}}'}\|\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\|_{L^{\infty}_{sc}(S_{u',{\underline{u}}'})} \|(\at\nab)^i\Psi\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'\cdot R\\ &+ \sum_{i_1+i_2+i_3+i_4+1=i} \int_{u_{\infty}}^u a^{-\f12}\sup_{{\underline{u}}'}\|(\at)^{i+1} \nabla^{i_1}\p^{i_2+1}\nabla^{i_3}(\f{a}{|u'|^2}{\hat{{\underline{\chi}}}},\f{a}{|u'|^2}{\mbox{tr}}{\underline{\chi}})\nab^{i_4}\Psi \|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'\cdot R\\ +&\sum_{\substack{i_1+i_2+i_3+i_4=i\\i_4\leq i-1}} \int_{u_{\infty}}^u \f{1}{|u'|}\sup_{{\underline{u}}'}\|(\at)^{i+1} \nabla^{i_1}\p^{i_2}\nabla^{i_3}\big(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}, \f{\at}{|u'|}{\widetilde{{\mbox{tr}}{\underline{\chi}}}}\big)\nab^{i_4}\Psi \|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'\cdot R\\ \leq& \sup_{{\underline{u}}'}\bigg(\int^u_{u_{\infty}}\f{1}{|u'|^2}du' \bigg)^{\f12}\cdot O\cdot \sup_{{\underline{u}}'}\bigg(\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at\nab)^i\Psi\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})} du'\bigg)^{\f12}\cdot R\\ &+a^{-\f12}\int^u_{u_{\infty}}\f{a}{|u'|^2} du'\cdot O^{3}\cdot R+\int_{u_{\infty}}^u \f{a^{\f12}}{|u'|^2}\cdot O^{2}\cdot du'\cdot R.\\ \leq&\f{1}{|u|^{\f12}}\cdot O \cdot R \cdot R+a^{-\f12}\cdot \f{a}{|u|} \cdot O^{3}\cdot R+\f{\at}{|u|}\cdot O^{2} \cdot R\leq \f{1}{a^{\f13}}. \end{split}$$ We then treat $M_2$ in the same fashion. By (\[Holder’s\]) and Hölder’s inequality, we obtain $$\begin{split} M_2=&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|}\|(\at)^iG_2\cdot(\at\nab)^i\Psi\|_{L^1_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'\\ \leq&\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_2\|_{L^2_{sc}(S_{u',{\underline{u}}'})} \|(\at\nab)^i\Psi\|_{L^2_{sc}(S_{u',{\underline{u}}'})} du'd{{\underline{u}}}'\\ \leq&\int_0^{{\underline{u}}}\bigg(\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_2\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du'\bigg)^{\f12} \bigg(\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at\nab)^i \Psi\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du'\bigg)^{\f12} d{{\underline{u}}}'\\ \leq&\bigg(\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_2\|^2_{L^2_{sc}(S_{u',{\underline{u}}'})}du' d{\underline{u}}'\bigg)^{\f12}\cdot \|(\at\nab)^i\Psi\|_{L^2_{sc}(\Hb_{{\underline{u}}'}^{(u_{\infty},u)})}, \end{split}$$ where $$\begin{split} G_2=&\sum_{i_1+i_2+i_3+i_4=i}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{{\underline{\chi}}}}, {\hat{\chi}})\nabla^{i_4} (\Psi, \a). \end{split}$$ Denote $$J_2:=\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i G_2\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'.$$ Then by (\[Holder’s\]), (\[4.3\]), (\[4.4\]), Proposition \[a.bd\], Proposition \[ee 4\] and letting $a$ to be sufficiently large, we have $$\begin{split} J_2\leq& \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^2}\|(\at)^i (\p,{\hat{\chi}}, {\hat{{\underline{\chi}}}})\nab^i(\Psi, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \f{a}{|u'|^2} \|(\at)^i\sum_{\substack{i_1+i_2+i_3+i_4=i\\i_4\leq i-1}}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}})\nabla^{i_4} (\Psi, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})} du'd{\underline{u}}'\\ \leq& \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \frac{a}{|u'|^4}\|\p, {\hat{\chi}}, {\hat{{\underline{\chi}}}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})} \|(\at)^i\nab^i (\Psi, \a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+\int_0^{{\underline{u}}}\int_{u_{\infty}}^u \|(\at)^{i+1}\sum_{\substack{i_1+i_2+i_3+i_4=i\\i_4\leq i-1}}\nabla^{i_1}\p^{i_2}\nabla^{i_3}(\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}},\f{\at}{|u'|}{\hat{{\underline{\chi}}}})\nabla^{i_4} (a^{-\f12}\Psi, a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})} du'd{\underline{u}}'\\ \leq& \int_0^{{\underline{u}}}\int_{u_{\infty}}^u a^{-1}\|\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})} \f{a}{|u'|^2} \|(\at)^i\nab^i\Psi\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+ \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \|\f{\at}{|u'|} \p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|} {\hat{{\underline{\chi}}}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})} \f{a}{|u'|^2} \|(\at)^i\nab^i (a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'd{\underline{u}}'\\ &+ \int_0^{{\underline{u}}}\int_{u_{\infty}}^u \bigg(\f{a}{|u'|^2}O^2[{\hat{{\underline{\chi}}}}]O^2[\a]+\f{a^{-\f12}\cdot a\cdot O^4}{|u'|^2}\bigg) du' d{\underline{u}}' \\ \leq& \int_0^{{\underline{u}}} a^{-1}\sup_{u'}\bigg(\|\f{\at}{|u'|}\p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|}{\hat{{\underline{\chi}}}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})}\bigg) \bigg(\int_{u_{\infty}}^u \f{a}{|u'|^2} \|(\at)^i\nab^i\Psi\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}du'\bigg)d{\underline{u}}'\\ &+\int_{u_{\infty}}^u \sup_{u'}\bigg(\|\f{\at}{|u'|} \p, \f{\at}{|u'|}{\hat{\chi}}, \f{\at}{|u'|} {\hat{{\underline{\chi}}}}\|^2_{L^{\infty}_{sc}(S_{u',{\underline{u}}})}\bigg) \f{a}{|u'|^2} \bigg(\int_0^{{\underline{u}}}\|(\at)^i\nab^i (a^{-\f12}\a)\|^2_{L^2_{sc}(S_{u',{\underline{u}}})}d{\underline{u}}'\bigg) du'\\ &+\f{a}{|u|}\bigg(O^2[{\hat{{\underline{\chi}}}}]+O^2[{\hat{\chi}}]+1\bigg)O^2[\a]+\f{\at}{|u|}\cdot O^4\\ \leq&a^{-1}\cdot O^2 \cdot R^2+\bigg(O^2[{\hat{{\underline{\chi}}}}]+O^2[{\hat{\chi}}]+1\bigg) \cdot \big({\underline{\mathcal{R}}}^2[\b]+\f{1}{a}\big)\\ &+\f{a}{|u|}\bigg(O^2[{\hat{{\underline{\chi}}}}]+O^2[{\hat{\chi}}]+1\bigg) \cdot \bigg(\M R^2[\a]+O^2[\a]\bigg)+\f{\at}{|u|}\cdot O^4\\ \ls& \f{1}{a^{\f13}}+(1+\M R^2[\a])\cdot({\underline{\mathcal{R}}}^2[\b]+\M R^2[\a]+1), \end{split}$$ where we use $\M O[{\hat{{\underline{\chi}}}}]\ls 1$, $\M O[{\hat{\chi}}]\ls \M R[\a]+1$ and $O[\a]\ls {\underline{\mathcal{R}}}[\b]+1$. This implies $$M_2\leq J_2^{\f12}\|(\at\nab)^i\Psi\|_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\leq J_2+\f{1}{4}\|(\at\nab)^i\Psi\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}.$$ From (\[N2\]), we have $N_2\leq 1/a^{\f13}$. Together with (\[N2+M2\]), we have $$\begin{split} &\|(\at\nab)^i\Psi_1\|^2_{L^2_{sc}(H_u^{(0,{\underline{u}})})}+\|(\at\nab)^i\Psi_2\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}\\ \leq & \|(\at\nab)^i\Psi_1\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\|(\at\nab)^i\Psi_2\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+N_2+M_2\\ \leq & \|(\at\nab)^i\Psi_1\|^2_{L^2_{sc}(H_{u_{\infty}}^{(0,{\underline{u}})})}+\|(\at\nab)^i\Psi_2\|^2_{L^2_{sc}(\Hb_{0}^{(u_{\infty},u)})}+\f{1}{a^{\f13}}+J_2+\f{1}{4}\cdot\|(\at\nab)^i\Psi\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}. \end{split}$$ The last term $1/4 \|(\at\nab)^i\Psi\|^2_{L^2_{sc}(\Hb_{{\underline{u}}}^{(u_{\infty},u)})}$ could be absorbed by the left. Recall $$J_2\ls \f{1}{a^{\f13}}+(1+\M R^2[\a])\cdot({\underline{\mathcal{R}}}^2[\b]+\M R^2[\a]+1).$$ We hence derive $$\begin{split} &\M R^2[\b]+\M R^2[\rho]+\M R^2[\sigma]+\M R^2[{\underline{\beta}}]+{\underline{\mathcal{R}}}^2[\rho]+{\underline{\mathcal{R}}}^2[\sigma]+{\underline{\mathcal{R}}}^2[{\underline{\beta}}]+{\underline{\mathcal{R}}}^2[\ab]\\ \ls& \M R^2_0[\b]+\M R^2_0[\rho]+\M R^2_0[\sigma]+\M R^2_0[{\underline{\beta}}]+{\underline{\mathcal{R}}}^2_0[\rho]+{\underline{\mathcal{R}}}^2_0[\sigma]+{\underline{\mathcal{R}}}^2_0[{\underline{\beta}}]+{\underline{\mathcal{R}}}^2_0[\ab]\\ &+\big(1+\M R^2_0[\a]+{\underline{\mathcal{R}}}^2_0[\b]\big)^2+\f{1}{a^{\f14}}\ls \M (I^{(0)})^2+(\M I^{(0)})^4+1+\f{1}{a^{\f14}}, \end{split}$$ where we use (\[eee 4\]) in the last step. This implies $$\label{eee 6} \begin{split} &\M R[\b]+\M R[\rho]+\M R[\sigma]+\M R[{\underline{\beta}}]+{\underline{\mathcal{R}}}[\rho]+{\underline{\mathcal{R}}}[\sigma]+{\underline{\mathcal{R}}}[{\underline{\beta}}]+{\underline{\mathcal{R}}}[\ab]\ls \M I^{(0)}+(\M I^{(0)})^2+1. \end{split}$$ Recall (\[eee 5\]) $$\M R[\a]+{\underline{\mathcal{R}}}[\b]\leq 2\,\M R_0[\a]+2\,{\underline{\mathcal{R}}}_0[\b]+\f{1}{a^{\f18}}.$$ These together conclude $$\label{bootstrap curvature conclusion} \M R(u,{\underline{u}})+{\underline{\mathcal{R}}}(u,{\underline{u}})\ls \M I^{(0)}+(\M I^{(0)})^2+1.$$ By estimates in Section \[secRicci\] and Section \[secCurvatureL2\], we have $$\M O(u,{\underline{u}})\ls \M R(u,{\underline{u}})+{\underline{\mathcal{R}}}(u,{\underline{u}})+1.$$ Hence, with (\[bootstrap curvature conclusion\]) we also deduce $$\label{bootstrap Ricci conclusion} \M O(u,{\underline{u}})\ls \M I^{(0)}+(\M I^{(0)})^2+1.$$ \[bootstrap openness\] Conclusions in (\[bootstrap curvature conclusion\]) and (\[bootstrap Ricci conclusion\]) are improvements of bootstrap assumption (\[BA.0\]): $$\M O(u, {\underline{u}})\leq O, \quad \M R(u,{\underline{u}}) +\underline{\M R}(u,{\underline{u}}) \leq R, \mbox{ where } O \mbox{ and } R \mbox{ are large numbers satisfying}$$ $$\M I^{(0)}+(\M I^{(0)})^2+1\ll O, \quad \M I^{(0)}+(\M I^{(0)})^2+1\ll R, \quad (O+R)^{20}\leq {a^{\f{1}{16}}}.$$ And these improvements prove openness in the bootstrap argument. FORMATION OF TRAPPED SURFACES {#TSF} ============================= In this section, we will prove\ (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-a/4}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{1}({\underline{u}}=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); [**Theorem \[main.thm2\]** ]{} Given $\M I^{(0)}$, there exists a sufficiently large $a_0=a_0(\M I^{(0)})$. For $0<a_0<a$, for Einstein vacuum equations with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|\nab^{k}_4(|u_{\infty}|\nab)^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}$ along $u=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}=0$, - $\int_0^1|u_{\infty}|^2|{\hat{\chi}}_0|^2(u_{\infty}, {\underline{u}}')d{\underline{u}}'\geq a$ for every direction along $u=u_{\infty}$, we have that $S_{-a/4,1}$ is a trapped surface. Proof. We first derive pointwise estimates for $|{\hat{\chi}}|^2_{\gamma}$. Fix $(\theta^1, \theta^2)\in S^2$. We consider the following null structure equation $$\nab_3{\hat{\chi}}+\frac 12 {\mbox{tr}}{\underline{\chi}}{\hat{\chi}}-2\omegab {\hat{\chi}}=\nab\widehat{\otimes} \eta-\frac 12 {\mbox{tr}}\chi {\hat{{\underline{\chi}}}}+\eta\widehat{\otimes} \eta.$$ We contract this $2$-tensor with another $2$-tensor ${\hat{\chi}}$ and get $$\label{chih square} \f12 \nab_3|{\hat{\chi}}|^2_{\gamma}+\f12{\mbox{tr}}{\underline{\chi}}|{\hat{\chi}}|^2_{\gamma}-2{\underline{\omega}}|{\hat{\chi}}|^2_{\gamma}={\hat{\chi}}(\nab\widehat{\otimes}\eta-\f12{\mbox{tr}}\chi{\hat{{\underline{\chi}}}}+\eta\widehat{\otimes}\eta).$$ Employing the fact ${\underline{\omega}}=-\f12\nab_3(\log \Omega)=-\f12\O^{-1}\nab_3\O$, we rewrite (\[chih square\]) as $$\begin{split} \nab_3(\O^2|{\hat{\chi}}|^2_{\gamma})+\O^2{\mbox{tr}}{\underline{\chi}}|{\hat{\chi}}|^2_{\gamma}=2\O^2{\hat{\chi}}(\nab\widehat{\otimes}\eta-\f12{\mbox{tr}}\chi{\hat{{\underline{\chi}}}}+\eta\widehat{\otimes}\eta). \end{split}$$ Using $\nab_3=\f{1}{\O}(\f{\partial}{\partial u}+b^A\f{\partial}{\partial \theta^A})$, we rewrite the above equation as $$\begin{split} \f{\partial}{\partial u}(\O^2|{\hat{\chi}}|^2_{\gamma})+\O{\mbox{tr}}{\underline{\chi}}\cdot \O^2|{\hat{\chi}}|^2_{\gamma}=&2\O^3{\hat{\chi}}(\nab\widehat{\otimes}\eta-\f12{\mbox{tr}}\chi{\hat{{\underline{\chi}}}}+\eta\widehat{\otimes}\eta)-b^A\f{\partial}{\partial \theta^A}(\O^2|{\hat{\chi}}|^2_{\gamma}). \end{split}$$ Substitute $\O{\mbox{tr}}{\underline{\chi}}$ with $$\Omega {\mbox{tr}}{\underline{\chi}}=\Omega({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})-\Omega\f{2}{|u|}=\Omega({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})-(\Omega-1)\f{2}{|u|}-\f{2}{|u|}$$ we have $$\begin{split} \f{\partial}{\partial u}(\O^2|{\hat{\chi}}|^2_{\gamma})-\f{2}{|u|}\O^2|{\hat{\chi}}|^2_{\gamma}=&2\O^3{\hat{\chi}}(\nab\widehat{\otimes}\eta-\f12{\mbox{tr}}\chi{\hat{{\underline{\chi}}}}+\eta\widehat{\otimes}\eta)-b^A\f{\partial}{\partial \theta^A}(\O^2|{\hat{\chi}}|^2_{\gamma})\\ &-\O({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})(\O^2|{\hat{\chi}}|^2_{\gamma})+(\O-1)\cdot\f{2}{|u|}\cdot(\O^2|{\hat{\chi}}|^2_{\gamma}). \end{split}$$ This gives $$\label{chih square 2} \begin{split} \f{\partial}{\partial u}\bigg(u^2\O^2|{\hat{\chi}}|^2_{\gamma}\bigg)=&2\cdot|u|^2\cdot\O^3{\hat{\chi}}(\nab\widehat{\otimes}\eta-\f12{\mbox{tr}}\chi{\hat{{\underline{\chi}}}}+\eta\widehat{\otimes}\eta)-|u|^2\cdot b^A\f{\partial}{\partial \theta^A}(\O^2|{\hat{\chi}}|^2_{\gamma})\\ &-|u|^2\cdot\O({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})(\O^2|{\hat{\chi}}|^2_{\gamma})+|u|^2\cdot(\O-1)\cdot\f{2}{|u|}\cdot(\O^2|{\hat{\chi}}|^2_{\gamma}). \end{split}$$ For $b^{A}$, we have equation $$\f{\partial b^{A}}{\partial {\underline{u}}}=-4\Omega^2\zeta^{A},$$ which is from $$[L,{\underline{L}}]=\f{\partial b^{A}}{\partial {\underline{u}}}\f{\partial}{\partial \theta^{A}}.$$ Applying the identity $\zeta_A=\f12\eta_A-\f12{\underline{\eta}}_A$, Propositions \[Omega\], derived estimates of $\eta, {\underline{\eta}}$, it holds in $D_{u,{\underline{u}}}$ $$\|b^{A}\|_{L^{\infty}(\S)}\leq \f{\at}{|u|^2}.$$ For the right hand side of (\[chih square 2\]), we have $$\|2\cdot|u|^2\cdot\O^3{\hat{\chi}}(\nab\widehat{\otimes}\eta-\f12{\mbox{tr}}\chi{\hat{{\underline{\chi}}}}+\eta\widehat{\otimes}\eta)\|_{L^{\infty}(\S)}\leq |u|^2\cdot\f{\at}{|u|}\cdot(\f{\at}{|u|^3}+\f{a}{|u|^4})\leq\f{a}{|u|^2},$$ $$\||u|^2\cdot b^A\f{\partial}{\partial \theta^A}(\O^2|{\hat{\chi}}|^2_{\gamma})\|_{L^{\infty}(\S)}\leq |u|^2\cdot\f{\at}{|u|^2}\cdot\f{a}{|u|^2}\leq\f{a^{\f32}}{|u|^2},$$ $$\|-|u|^2\cdot\O({\mbox{tr}}{\underline{\chi}}+\f{2}{|u|})(\O^2|{\hat{\chi}}|^2_{\gamma})\|_{L^{\infty}(\S)}\leq |u|^2\cdot\f{1}{|u|^2}\cdot\f{a}{|u|^2}\leq\f{a}{|u|^2},$$ $$\||u|^2\cdot(\O-1)\cdot\f{2}{|u|}\cdot(\O^2|{\hat{\chi}}|^2_{\gamma}) \|_{L^{\infty}(\S)}\leq |u|^2\cdot\f{1}{|u|}\cdot\f{2}{|u|}\cdot\f{a}{|u|^2}\leq\f{a}{|u|^2}.$$ In summary, for (\[chih square 2\]) we have $$\f{\partial}{\partial u}\bigg(u^2\O^2|{\hat{\chi}}|^2_{\gamma}\bigg)=M, \mbox{ and } |M|\ls \f{a^{\f32}}{|u|^2}\ll \f{a^{\f74}}{|u|^2},$$ which [implies]{} $$-\f{a^{\f74}}{|u|}+\f{a^{\f74}}{|{u_{\infty}}|} \leq |u|^2\O^2|{\hat{\chi}}|^2_{\gamma}(u,{\underline{u}},\theta^1,\theta^2)-|u_{\infty}|^2\O^2|{\hat{\chi}}|^2_{\gamma}(u_{\infty},{\underline{u}},\theta^1,\theta^2).$$ Recall $\O(u_{\infty},{\underline{u}},\theta^1,\theta^2)=1$. We hence have $$|u|^2\O^2|{\hat{\chi}}|^2_{\gamma}(u,{\underline{u}},\theta^1,\theta^2)\geq|u_{\infty}|^2|{\hat{\chi}}|^2_{\gamma}(u_{\infty},{\underline{u}},\theta^1,\theta^2)-\f{a^{\f74}}{|u|}.$$ Together with the assumption in Theorem \[main.thm2\], we further have for ${u_{\infty}}\leq u\leq -a/4$ $$\begin{split} \int_0^1 |u|^2\O^2|{\hat{\chi}}|^2_{\gamma}(u,{\underline{u}}',\theta^1,\theta^2)d{\underline{u}}'\geq \int_0^1 |u_{\infty}|^2|{\hat{\chi}}|^2_{\gamma}(u_{\infty},{\underline{u}}',\theta^1,\theta^2)d{\underline{u}}'-\f{a^{\f74}}{|u|}\geq a-\f{a^{\f74}}{|u|}\geq a-\f{4a^{\f74}}{a} \geq \f{7a}{8}. \end{split}$$ Pick $u=-a/4$. With the fact $\|\O-1\|_{L^{\infty}(\S)}\ls {1}/{a}$, for sufficiently large $a$, we hence have $$\begin{split} (-\f{a}{4})^2\int_0^1 |{\hat{\chi}}|^2_{\gamma}(-\f{a}{4},{\underline{u}}',\theta^1,\theta^2)d{\underline{u}}'\geq& \f67\cdot \int_0^1 (-\f{a}{4})^2\O^2|{\hat{\chi}}|^2_{\gamma}(-\f{a}{4},{\underline{u}}',\theta^1,\theta^2)d{\underline{u}}'\\ \geq&\f67\cdot \f{7a}{8}=\f{3a}{4}. \end{split}$$ This implies $$\label{int chih square} \begin{split} \int_0^1 |{\hat{\chi}}|^2_{\gamma}(-\f{a}{4},{\underline{u}}',\theta^1,\theta^2)d{\underline{u}}'\geq&\f{3a}{4}\cdot \f{16}{a^2}= \f{12}{a} \end{split}$$ Now we consider the other null structure equation $$\nab_4 {\mbox{tr}}\chi+\f12({\mbox{tr}}\chi)^2=-|{\hat{\chi}}|^2_{\gamma}-2\o{\mbox{tr}}\chi.$$ Using $\o=-\f12\nab_4(\log \Omega)$, we have $$\begin{split} \nab_4 {\mbox{tr}}\chi+\f12({\mbox{tr}}\chi)^2=&-|{\hat{\chi}}|^2-2\o{\mbox{tr}}\chi\\ =&-|{\hat{\chi}}|^2_{\gamma}+\nab_4(\log\Omega){\mbox{tr}}\chi=-|{\hat{\chi}}|^2_{\gamma}+\f{1}{\Omega}\nab_4\Omega\cdot {\mbox{tr}}\chi. \end{split}$$ Hence, $$\begin{split} \nab_4(\Omega^{-1} {\mbox{tr}}\chi)=&-\O^{-2}\nab_4\O\cdot{\mbox{tr}}\chi+\O^{-1}\nab_4{\mbox{tr}}\chi\\ =&\O^{-1}(\nab_4{\mbox{tr}}\chi-\O^{-1}\cdot\nab_4\O\cdot{\mbox{tr}}\chi)=\O^{-1}\bigg(-\f12({\mbox{tr}}\chi)^2-|{\hat{\chi}}|^2_{\gamma}\bigg). \end{split}$$ With the fact $e_4=\Omega^{-1}\f{\partial}{\partial {\underline{u}}}$, we have $$\label{O trchi} \begin{split} \f{\partial}{\partial {\underline{u}}}(\Omega^{-1} {\mbox{tr}}\chi)=&-\f12({\mbox{tr}}\chi)^2-|{\hat{\chi}}|^2_{\gamma}. \end{split}$$ For every $(\theta^1,\theta^2)\in \mathbb{S}^2$, along $\Hb_0$ we have $$(\Omega^{-1}{\mbox{tr}}\chi)(-\f{a}{4}, 0, \theta^1, \theta^2)=1^{-1}\cdot \f{2}{a/4}=\f{8}{a}.$$ We then integrate (\[O trchi\]). Using (\[int chih square\]) we obtain $$\begin{split} &(\Omega^{-1}{\mbox{tr}}\chi)(-\f{a}{4},1, \theta^1, \theta^2)\\ \leq & (\Omega^{-1}{\mbox{tr}}\chi)(-\f{a}{4}, 0, \theta^1, \theta^2)-\int_0^{1}|{\hat{\chi}}|^2_{\gamma}(-\f{a}{4},{\underline{u}}',\theta^1,\theta^2)d{\underline{u}}'\\ \leq & \f{8}{a}-\f{12}{a}<0. \end{split}$$ Recall in $D_{u,{\underline{u}}}$ the following estimate holds $$\|{\mbox{tr}}{\underline{\chi}}+\f{2}{|u|}\|_{L^{\infty}(S_{u,{\underline{u}}})}\leq \f{1}{|u|^2}.$$ In particular, this implies $${\mbox{tr}}{\underline{\chi}}(-\f{a}{4},1, \theta^1, \theta^2)<0 \mbox{ for every} (\theta^1, \theta^2)\in \mathbb{S}^2.$$ Therefore, we conclude that $S_{-\f{a}{4},1}$ is a trapped surface. A Scaling Argument and a connection to \[An-Luk\] {#sec rescale} ================================================= In this article, we use coordinate system $(u, {\underline{u}}, \theta^1, \theta^2)$ based on double null foliations, where $(\theta^1, \theta^2)$ are stereographic coordinates on $\mathbb{S}^2$. In these coordinates, we study spacetime region $$u_{\infty} \leq u \leq -\f{a}{4}, \, \, \quad 0\leq {\underline{u}}\leq 1.$$ The Lorentzian metric $g$ satisfies ansatz $$g=-2\O^2(du\otimes d{\underline{u}}+d{\underline{u}}\otimes du)+\gamma_{AB}(d\theta^A-d^A du)\otimes(d\theta^B-d^B du).$$ In the below, by exploring a rescaling, we will find an interesting connection between the results above and the results in [@A-L] proved by An-Luk. A Spacetime Rescaling --------------------- We introduce a new coordinate system $(u', {\underline{u}}', \theta^{1'}, \theta^{2'})$, where $$\label{rescale} u'=\d u, \, \, \, {\underline{u}}'=\d{\underline{u}}, \, \, \, \theta^{1'}=\d \theta^1, \, \, \,\theta^{2'}=\d \theta^2.$$ Note that coordinates $(\theta^1, \theta^2)$ on $\S$ are set up through stereographic projection. Assume $(x_1, x_2, x_3)$ satisfying $x_1^2+x_2^2+x_3^2=a^2$ and lying on the upper hemisphere of $S_{-a,0}$ (with radius $a$). It then has stereographic projection $(\zeta_1, \zeta_2)=(\f{ax_1}{a+x_3}, \f{ax_2}{a+x_3})$. Scale down the length by a factor $\d$, we then have $x_1'=\d x_1, \, x_2'=\d x_2, \, x_3'=\d x_3, \,\, (x_1')^2+(x_2')^2+(x_2')^2=\d^2 a^2$ and $(x_1', x_2', x_3')$ has stereographic projection $$(\zeta_1', \zeta_2')=(\f{\d a x_1'}{\d a+x_3'}, \f{\d ax_2'}{\d a+x_3'})=(\f{\d a\cdot \d x_1}{\d a+\d x_3}, \f{\d a \cdot \d x_2}{\d a+\d x_3})=(\f{\d ax_1}{a+x_3}, \f{\d ax_2}{a+x_3})=(\d\zeta_1, \d\zeta_2).$$ Therefore, the rescaled coordinates $(\theta^{1'}, \theta^{2'})=(\d \theta^1, \d \theta^2)$ on $S_{u', {\underline{u}}'}$ make perfect sense since $2$-sphere $S_{u',{\underline{u}}'}=S_{\d u, \d {\underline{u}}}$ is scaled down from $\S$ by a factor $\delta$. We then rewrite Theorem \[main.thm1\] and Theorem \[main.thm2\] in coordinate system $(u', {\underline{u}}', \theta^{1'}, \theta^{2'})$: (3,-1)– node\[midway, sloped, below,black\][$H'_{\d u_{\infty}}(u'=\d u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H'_{-\d a/4}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb'_{1}({\underline{u}}'=\d)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb'_{0}({\underline{u}}'=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); With an open set of characteristic initial data (corresponding to the initial data in Theorem \[main.thm1\] and Theorem \[main.thm2\]), - Einstein vacuum equations (\[1.1\]) admit a unique smooth solution in the colored region: $$\d\cdot u_{\infty}\leq u' \leq -{\d\cdot a}/{4}, \quad 0\leq {\underline{u}}' \leq \d.$$ - $S'_{-{\d a}/{4}, \,\d}:=\{u'=-\d a/4\}\cap\{{\underline{u}}'=\d\}$ is a trapped surface.\ The above conclusion is very similar to the main theorem in [@A-L]. In the following, we will verify that this conclusion is indeed an extension of [@A-L]. In particular, we will show that all the rescaled Ricci coefficients $\Gamma'$ and rescaled curvature components $R'$ would obey the same apriori estimates as in [@A-L]. Under the rescaling (\[rescale\]), it follows $$g'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^2\cdot g(u,{\underline{u}}, \theta^1, \theta^2).$$ In $(u',{\underline{u}}',\theta^{1'}, \theta^{2'})$ coordinates, we let $$g'(u', {\underline{u}}', \theta^1, \theta^2)=-2{\O'}^2(du'\otimes d{\underline{u}}'+d{\underline{u}}'\otimes du')+\gamma'_{A'B'}(d\theta^{A'}-d^{A'} du')\otimes(d\theta^{B'}-d^{B' }du').$$ Compare with $$g(u, {\underline{u}}, \theta^1, \theta^2)=-2\O^2(du\otimes d{\underline{u}}+d{\underline{u}}\otimes du)+\gamma_{AB}(d\theta^A-d^A du)\otimes(d\theta^B-d^B du).$$ Here we have $$du'=\d\cdot du, \quad d{\underline{u}}'=\d\cdot d{\underline{u}}, \quad d\theta^{A'}=\d \cdot d \theta^A \,\mbox{ for } A=1, 2 ,$$ $${\O'}^2(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\O^2(u,{\underline{u}}, \theta^1, \theta^2), \quad \gamma'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\gamma_{AB}(u,{\underline{u}}, \theta^1, \theta^2),$$ $$d^{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=d^A(u,{\underline{u}}, \theta^1, \theta^2),$$ $$e'_3(u',{\underline{u}}', \theta^{1'}, \theta^{2'})={\O'}^{-1}(\f{\partial}{\partial u'}+d^{A'}\f{\partial}{\partial {\underline{u}}'})=\d^{-1}{\O}^{-1}(\f{\partial}{\partial u}+d^{A}\f{\partial}{\partial {\underline{u}}})=\d^{-1}\cdot e_3(u, {\underline{u}}, \theta^1, \theta^2),$$ $$\label{e4 e4'} e'_4(u',{\underline{u}}', \theta^{1'}, \theta^{2'})={\O'}^{-1}\f{\partial}{\partial{\underline{u}}'}=\d^{-1}{\O}^{-1}\f{\partial}{\partial{\underline{u}}}=\d^{-1}\cdot e_4(u,{\underline{u}}, \theta^1, \theta^2),$$ $$\label{eA eA'} e'_A(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot e_A(u,{\underline{u}}, \theta^1, \theta^2), \mbox{ for } A=1,2.$$ Rescaled Geometric Quantities ----------------------------- As usual, with frame $\{e'_3, e'_4, e'_A, e'_B\}$, we define $$\begin{split} &\chi'_{A'B'}=g'(D'_{A'} e'_4,e'_B),\, \,\, \quad {\underline{\chi}}'_{A'B'}=g'(D'_{A'} e'_3,e'_B),\\ &\eta'_{A'}=-\frac 12 g'(D'_{3'} e'_A,e'_4),\quad \etab'_{A'}=-\frac 12 g'(D'_{4'} e'_A,e'_3),\\ &\omega'=-\frac 14 g'(D'_{4'} e'_3,e'_4),\quad\,\,\, \omegab'=-\frac 14 g'(D'_{3'} e'_4,e'_3),\\ &\zeta'_{A'}=\frac 1 2 g'(D'_{A'} e'_4,e'_3). \end{split}$$ With $\gamma'_{A'B'}$ being the induced metric on $\S'$, we further decompose $\chi', {\underline{\chi}}'$ into $$\chi'_{A'B'}=\f12{\mbox{tr}}\chi'\cdot \gamma'_{A'B'}+{\hat{\chi}}'_{A'B'}, \quad {\underline{\chi}}'_{A'B'}=\f12{\mbox{tr}}{\underline{\chi}}'\cdot \gamma'_{A'B'}+{\hat{{\underline{\chi}}}}'_{A'B'}.$$ $$\mbox{Here} \quad D'_{e'_{\mu}}e'_{\nu}:=\Gamma'^{\lambda}_{\mu'\nu'}e'_{\lambda} \quad \mbox{ and } \quad \Gamma'^{\lambda}_{\mu'\nu'}:=\f12g'^{\lambda'\kappa'}(\f{\partial g'_{\kappa'\mu'}}{\partial x'_{\nu}}+\f{\partial g'_{\kappa' \nu'}}{\partial x'_{\mu}}-\f{\partial g'_{\mu'\nu'}}{\partial x'_{\kappa}}).$$ Note by rescaling $$\label{rescaling 2} g'=\d^2\cdot g, \mbox{ and } g'^{-1}=\d^{-2}\cdot g^{-1},$$ we then have $$\label{Gamma Gamma'} \Gamma'^{\lambda'}_{\mu'\nu'}:=\f12g'^{\lambda'\kappa'}(\f{\partial g'_{\kappa'\mu'}}{\partial x'_{\nu}}+\f{\partial g'_{\kappa' \nu'}}{\partial x'_{\mu}}-\f{\partial g'_{\mu'\nu'}}{\partial x'_{\kappa}})=\f12g^{\lambda'\kappa'}(\f{\partial g_{\kappa'\mu'}}{\partial x'_{\nu}}+\f{\partial g_{\kappa' \nu'}}{\partial x'_{\mu}}-\f{\partial g_{\mu'\nu'}}{\partial x'_{\kappa}})=\Gamma^{\lambda'}_{\mu'\nu'},$$ which implies $$\label{D D'} D'_{e'_{\mu}}e'_{\nu}=\Gamma'^{\lambda}_{\mu'\nu'}e'_{\lambda}=\Gamma^{\lambda}_{\mu'\nu'}e'_{\lambda}=D_{e'_{\mu}}e'_{\nu}.$$ We are ready to prove \[Prop rescale 1\] For $\Gamma\in \{{\hat{\chi}}, {\mbox{tr}}\chi, {\hat{{\underline{\chi}}}}, {\mbox{tr}}{\underline{\chi}}, \eta, {\underline{\eta}}, \zeta, \o,{\underline{\omega}}\}$ written in two different coordinates $(u',{\underline{u}}',\theta^{1'}, \theta^{2'})$ and $(u, {\underline{u}}, \theta^1, \theta^2)$, it holds that $$\Gamma'(u',{\underline{u}}',\theta^{1'}, \theta^{2'})=\d^{-1}\cdot \Gamma (u,{\underline{u}},\theta^{1}, \theta^{2}).$$ We first calculate $\chi'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})$. With definition of $\chi'$ in the first step, (\[D D’\]) in the second step, (\[rescaling 2\]) in the third step, (\[e4 e4’\]) and (\[eA eA’\]) in the fourth step, we have $$\begin{split} &\chi'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=g'(D'_{e'_A} e'_4,e'_B)\\ =&g'(D_{e'_A}e'_4, e'_B)=\d^2\cdot g(D_{e'_A}e'_4, e'_B)\\ =&\d^2\cdot \d^{-1}\cdot \d^{-1}\cdot \d^{-1} g(D_{e_A} e_4, e_B)=\d^{-1}\cdot\chi_{AB}(u, {\underline{u}}, \theta^1, \theta^2). \end{split}$$ In the same fashion, we conclude $${\underline{\chi}}'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot{\underline{\chi}}_{AB}(u, {\underline{u}}, \theta^1, \theta^2), \,\, \zeta'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot\zeta_{AB}(u, {\underline{u}}, \theta^1, \theta^2),$$ $$\eta'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot\eta_{A}(u, {\underline{u}}, \theta^1, \theta^2), \quad {\underline{\eta}}'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot{\underline{\eta}}_{A}(u, {\underline{u}}, \theta^1, \theta^2),$$ $$\o'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot\o(u, {\underline{u}}, \theta^1, \theta^2), \quad {\underline{\omega}}'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot{\underline{\omega}}(u, {\underline{u}}, \theta^1, \theta^2).$$ We then focus on ${\mbox{tr}}\chi'$ and ${\hat{\chi}}'_{A'B'}$. As calculated above, we have $$\begin{split} &{\mbox{tr}}\chi'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})\\ =&g'^{A'B'}g'(D'_{A'} e'_4,e'_B)=g'^{A'B'}g'(D_{A'} e'_4,e'_B)=g^{A'B'}g(D_{A'} e'_4,e'_B)\\ =&\d^{-1}g^{AB}g(D_{A} e_4,e_B)=\d^{-1}\cdot{\mbox{tr}}\chi(u,{\underline{u}},\theta^1, \theta^2), \quad \quad \mbox{and} \end{split}$$ $$\begin{split} &{\hat{\chi}}'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})\\ =&\chi'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})-\f12{\mbox{tr}}\chi'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})\cdot\gamma'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})\\ =&\chi'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})-\f12{\mbox{tr}}\chi'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})\cdot\d^2\cdot \d^{-1} \cdot \d^{-1}\cdot \gamma_{AB}(u,{\underline{u}}, \theta^{1}, \theta^{2})\\ =&\d^{-1}\cdot\chi_{AB}(u,{\underline{u}}, \theta^1, \theta^2)-\f12\cdot\d^{-1}\cdot{\mbox{tr}}\chi(u,{\underline{u}}, \theta^1, \theta^2)\cdot \gamma_{AB}(u,{\underline{u}}, \theta^1, \theta^2)\\ =&\d^{-1}\cdot {\hat{\chi}}_{AB}(u, {\underline{u}}, \theta^1, \theta^2). \end{split}$$ Similarly, it also holds $${\mbox{tr}}{\underline{\chi}}'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot{\mbox{tr}}{\underline{\chi}}(u,{\underline{u}},\theta^1, \theta^2), \quad {\hat{{\underline{\chi}}}}'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-1}\cdot {\hat{{\underline{\chi}}}}(u, {\underline{u}}, \theta^1, \theta^2).$$ We then conclude this proposition.\ For curvature components, we further have \[Prop rescale 2\] For $\Psi\in \{\a, \b, \rho, \sigma, {\underline{\beta}}, \ab\}$ written in coordinates $(u',{\underline{u}}',\theta^{1'}, \theta^{2'})$ and $(u, {\underline{u}}, \theta^1, \theta^2)$, the following identity is true $$\Psi'(u',{\underline{u}}',\theta^{1'}, \theta^{2'})=\d^{-2}\cdot\Psi(u,{\underline{u}},\theta^{1}, \theta^{2}).$$ We first write Riemann curvature in $(u',{\underline{u}}', \theta^{1'}, \theta^{2'})$ coordinate. $$R'^{l}_{ijk}=\partial_i\Gamma'^{l}_{jk}-\partial_j \Gamma'^{l}_{ik}+\Gamma'^{p}_{jk}\Gamma'^l_{ip}-\Gamma'^{p}_{ik}\Gamma'^{l}_{ip}$$ With the help of (\[Gamma Gamma’\]), we obtain $$\begin{split} R'^{l}_{ijk}=&\partial_i\Gamma'^{l}_{jk}-\partial_j \Gamma'^{l}_{ik}+\Gamma'^{p}_{jk}\Gamma'^l_{ip}-\Gamma'^{p}_{ik}\Gamma'^{l}_{ip}\\ =&\partial_i\Gamma^{l}_{jk}-\partial_j \Gamma^{l}_{ik}+\Gamma^{p}_{jk}\Gamma^l_{ip}-\Gamma^{p}_{ik}\Gamma'^{l}_{ip}=R^{l}_{ijk}. \end{split}$$ This [implies]{} $${R'}_{ijkl}={R'}^m_{ijk} g'_{ml}=R^m_{ijk}\cdot\d^2\cdot g_{ml}=\d^2\cdot R_{ijkl}.$$ Therefore, we obtain $$\begin{split} a'_{A'B'}(u', {\underline{u}}', \theta^{1'}, \theta^{2'}):=&R'(e'_A, e'_4, e'_B, e'_4)=\d^2\cdot R(e'_A, e'_4, e'_B, e'_4)\\ =&\d^{2}\cdot\d^{-4}\cdot R(e_A, e_4, e_B, e_4)=\d^{-2}\cdot\a_{AB}(u,{\underline{u}},\theta^1, \theta^2). \end{split}$$ In the same manner, we have $$\b'_{A'}(u', {\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-2}\cdot\b_{A}(u,{\underline{u}},\theta^1, \theta^2), \quad \rho'(u', {\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-2}\cdot\rho(u,{\underline{u}},\theta^1, \theta^2),$$ $$\sigma'_{A'}(u', {\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-2}\cdot\sigma(u,{\underline{u}},\theta^1, \theta^2), \quad {\underline{\beta}}'_{A'}(u', {\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-2}\cdot{\underline{\beta}}_{A}(u,{\underline{u}},\theta^1, \theta^2),$$ $$\ab'_{A'B'}(u', {\underline{u}}', \theta^{1'}, \theta^{2'})=\d^{-2}\cdot\ab_{AB}(u,{\underline{u}},\theta^1, \theta^2).$$ These conclude the proposition. Rescaled Uniform Bounds {#rescale bounds} ----------------------- Applying Proposition \[Prop rescale 1\] and Proposition \[Prop rescale 2\], next we establish the connection to [@A-L]. Take ${\hat{\chi}}$ as an example. Applying Proposition \[Prop rescale 1\], estimates derived for $\mathcal{O}_{i,\infty}[{\hat{\chi}}]$ and $u'=\d u$, we have $$\begin{split} &|{\hat{\chi}}'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |{\hat{\chi}}_{AB}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{\at}{|u|}=\f{\at}{\d|u|}=\f{\at}{|u'|}. \end{split}$$ In the same fashion, we have $$|{\hat{{\underline{\chi}}}}'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |{\hat{{\underline{\chi}}}}_{AB}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{\at}{|u|^2}=\f{\d\at}{\d^2|u|^2}=\f{\d\at}{|u'|^2},$$ $$|{\mbox{tr}}\chi'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |{\mbox{tr}}\chi(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{1}{|u|}=\f{1}{\d|u|}=\f{1}{|u'|},$$ $$|\eta'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |\eta_{A}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{\at}{|u|^2}=\f{\d\at}{\d^2|u|^2}=\f{\d\at}{|u'|^2},$$ $$|{\underline{\eta}}'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |{\underline{\eta}}_{A}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{\at}{|u|^2}=\f{\d\at}{\d^2|u|^2}=\f{\d\at}{|u'|^2},$$ $$|\o'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |\o(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{1}{|u|}=\f{1}{\d|u|}=\f{1}{|u'|},$$ $$|{\underline{\omega}}'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-1}\cdot |{\underline{\omega}}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-1}\cdot \f{a}{|u|^3}=\f{\d^2 a}{\d^3|u|^3}=\f{\d^2 a}{|u'|^3}\boxed{\leq \f{\d\at}{|u'|^2}} \, ,$$ $$|{\mbox{tr}}{\underline{\chi}}'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})+\f{2}{|u'|}|=\d^{-1}\cdot |{\mbox{tr}}{\underline{\chi}}(u, {\underline{u}}, \theta^1, \theta^2)+\f{2}{|u|}|\leq \d^{-1}\cdot \f{a}{|u|^3}=\f{\d^2 a}{\d^3|u|^3}=\f{\d^2 a}{|u'|^3}\boxed{\leq \f{\d\at}{|u'|^2}} \, .$$ For the estimates of ${\underline{\omega}}'$ and ${\mbox{tr}}{\underline{\chi}}'$, we use $|u'|\geq \d\at$. In the same manner, by Proposition \[Prop rescale 2\] and with the help that $|u'|\geq \d a/4$ we have $$\begin{split} &|\b'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-2}\cdot |\b_{A}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-2}\cdot \f{\at}{|u|^2}=\f{\at}{\d^2|u|^2}=\f{\at}{|u'|^2}, \end{split}$$ $$\begin{split} &|\rho'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-2}\cdot |\rho(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-2}\cdot \f{a}{|u|^3}=\f{\d a}{\d^3|u|^3}=\f{\d a}{|u'|^3}, \end{split}$$ $$\begin{split} &|\sigma'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-2}\cdot |\sigma(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-2}\cdot \f{a}{|u|^3}=\f{\d a}{\d^3|u|^3}=\f{\d a}{|u'|^3}, \end{split}$$ $$\begin{split} &|{\underline{\beta}}'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-2}\cdot |{\underline{\beta}}_{A}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-2}\cdot \f{a^{\f32}}{|u|^4}=\f{\d^2 a^{\f32}}{\d^4|u|^4}=\f{\d^2 a^{\f32}}{|u'|^4}\boxed{\leq \f{\d\at}{|u'|^3}} \, , \end{split}$$ $$\label{rescale ab} \begin{split} &|\ab'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-2}\cdot |\ab_{AB}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-2}\cdot \f{a^2}{|u|^5}=\f{\d^3 a^2}{\d^5|u|^5}=\f{\d^3 a^2}{|u'|^5}, \end{split}$$ $$\label{rescale alpha} \begin{split} &|\a'_{A'B'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=\d^{-2}\cdot |\a_{AB}(u, {\underline{u}}, \theta^1, \theta^2)|\leq \d^{-2}\cdot \f{\at}{|u|}=\f{\d^{-1}\at}{\d |u|}=\f{\d^{-1}\at}{|u'|}. \end{split}$$ \[systematical improvement\] By above rescaling argument, we hence transfer the bounds derived in preceding sections into new bounds, holding in the spacetime region $$\d a\leq |u'| \leq \d|u_{\infty}|, \quad 0\leq {\underline{u}}' \leq \d.$$ If we focus on the region $$\d a\leq |u'| \leq 1, \quad 0\leq {\underline{u}}' \leq \d,$$ *these bounds encode peeling properties, and they systematically sharpen the a priori bounds in [@A-L]*: - For $\{{\underline{\omega}}',\, {\mbox{tr}}{\underline{\chi}}'+\f{2}{|u'|},\, {\underline{\beta}}'\}$, we improve their estimates in [@A-L]. For comparison, the terms boxed are the estimates obtained in [@A-L]. - For $\{\ab', \a'\}$, their estimates are avoid in [@A-L] by several geometric renormalizations. *But for here we have estimates for $\ab'$ and $\a$, and they respect peeling properties*. In summary, in [@A-L] we have $1\ll b\leq \at \leq \d^{-\f12}$. If we focus on the scenario $1\ll b=\at \leq \d^{-\f12}$, compared with [@A-L] the new approach in this paper completely avoids geometric renormalizations and still gives a systematical improvement encoding intrinsic[^15] peeling properties. Since the estimates obtained above are uniform for $|u_{\infty}|$, we could keep $\d$ and let $|u_{\infty}|\rightarrow +\infty$. This extends [@A-L] and establish an existence result from pass null infinity. By repeating the arguments as in Section \[TSF\], we hence obtain one of the main conclusions in [@A-L] by An-Luk [on formation of a small trapped surface]{}: (3,-1)– node\[midway, sloped, below,black\][$H_{u_{\infty}}(u=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{-\f{\d a}{4}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{\d}({\underline{u}}=\d)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); [**Theorem \[thm3\]**]{} We solve Einstein vacuum equations. Given $\M I^{(0)}$, for a fixed $\d$ there exists a sufficiently large $a_0=a_0(\M I^{(0)},\d)$. For $0<a_0<a$, with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|(\d\nab_4)^k(|u_{\infty}|\nab)^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}})}\leq \M I^{(0)}$ along $u=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}=0$, - $\int_0^{\delta}|u_{\infty}|^2|{\hat{\chi}}_0|^2(u_{\infty}, {\underline{u}}')d{\underline{u}}'\geq \d a$ for every direction along $u=u_{\infty}$, we have that $S_{-\d a/4,\d}$ is a trapped surface. If we further choose $a=4c\cdot\d^{-1}$, where $0< c\leq 1$ being of size $1$, we then obtain Corollary \[Corollary1.5\]. Rescale from [@A-L] and Comparison {#rescale and comparison} ================================== Since the results in [@A-L] by An and Luk are scaling-critical, first choosing $\d=1/|u_{\infty}|$ and then scaling up the length by a factor $|u_{\infty}|$, we could change the main conclusion in [@A-L] into a conclusion similar to Theorem \[main.thm2\]. The full strength of the proof in [@A-L] ensures all the constants in the inequalities of [@A-L] are independent of $\d$; hence the constants in the new result after scaling up would be independent of $|u_{\infty}|$. This is similar to the proof of Theorem \[main.thm2\]. In the below, we will demonstrate this approach and make the comparsion. Fix $|u_{\infty|}$ to be a large positive constant. And set $\d=1/|u_{\infty}|$. By applying the main conclusion in [@A-L], we have (3,-1)– node\[midway, sloped, below,black\][$H_{-1}(u=-1)$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H_{\f{-a}{4|u_{\infty}|}}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb_{\f{1}{|u_{\infty}|}}({\underline{u}}=\f{1}{|u_{\infty}|})$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb_{0}({\underline{u}}=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); \[Prop 9.1\] We solve Einstein vacuum equations. Given $\M I^{(0)}>0$, for a fixed $1/|u_{\infty}|$ there exists a sufficiently large $a_0=a_0(\M I^{(0)}, 1/|u_{\infty}|)$. For $0<a_0<a$, with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|(\f{1}{|u_{\infty}|}\nab_4)^k\nab^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{-1,{\underline{u}}})}\leq \M I^{(0)}$ along $u=-1$, - Minkowskian initial data along ${\underline{u}}=0$, - $\int_0^{\f{1}{|u_{\infty}|}}|{\hat{\chi}}_0|^2(-1, {\underline{u}}')d{\underline{u}}'\geq \f{a}{|u_{\infty}|}$ for every direction along $u=-1$, we have that $S_{\f{-a}{4|u_{\infty}|},\f{1}{|u_{\infty}|}}$ is a trapped surface.\ Note: In [@A-L], in the proof of Proposition \[Prop 9.1\] we only use the largeness of $a$ and hence all the constants in the inequalities are independent of $|u_{\infty}|$. By applying [@A-L] we also derive the following bounds $$\label{9.1} \begin{split} &|{\hat{\chi}}_{AB}(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u|,\quad |\o(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u|, \quad |{\mbox{tr}}\chi(u,{\underline{u}},\theta^1, \theta^2)|\leq 1/|u|,\\ &|{\hat{{\underline{\chi}}}}_{AB}(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u_{\infty}u^2|, \quad |\eta_A(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u_{\infty}u^2|, \quad |{\underline{\eta}}_A(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u_{\infty}u^2|,\\ &|{\underline{\omega}}(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u_{\infty}u^2|, \quad |{\mbox{tr}}{\underline{\chi}}(u,{\underline{u}},\theta^1, \theta^2)+\f{2}{|u|}|\leq \at/|u_{\infty}u^2|,\\ &|\b_A(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u^2|, \quad |\rho(u,{\underline{u}},\theta^1, \theta^2)|\leq a/|u_{\infty}u^3|,\\ &|\sigma(u,{\underline{u}},\theta^1, \theta^2)|\leq a/|u_{\infty}u^3|, \quad |{\underline{\beta}}_A(u,{\underline{u}},\theta^1, \theta^2)|\leq \at/|u_{\infty}u^3|. \end{split}$$ We then rescale the length from small scale to large scale: $$\mbox{Set} \, \,\, u'=|u_{\infty}|u, \,\, {\underline{u}}'=|u_{\infty}|{\underline{u}}, \,\, {\theta^1}'=|u_{\infty}|\theta^1, \,\, {\theta^2}'=|u_{\infty}|\theta^2.$$ We have the rescaled result: (3,-1)– node\[midway, sloped, below,black\][$H'_{u_{\infty}}(u'=u_{\infty})$]{}(4,0); (1,1)– node\[midway,sloped,above,black\][$H'_{-a/4}$]{}(2,2); (2,2)–node \[midway,sloped,above,black\] [$\Hb'_{1}({\underline{u}}'=1)$]{}(4,0); (1,1)–node \[midway,sloped, below,black\] [$\Hb'_{0}({\underline{u}}'=0)$]{}(3,-1); (0, 4)–(0, -4); (0, -4)–(4,0)–(0,4); (0,0)–(2,2); (0,-4)–(4,0); (0,2)–(3,-1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (1,1)–(3,-1)–(4,0)–(2,2)–(1,1); (3.3,-0.6)– node\[midway, sloped, above,black\][$e_4$]{}(3.6,-0.3); (1.4,1.3)– node\[midway, sloped, below,black\][$e_4$]{}(1.7,1.6); (3.3,0.6)– node\[midway, sloped, below,black\][$e_3$]{}(2.7,1.2); (2.4,-0.3)– node\[midway, sloped, above,black\][$e_3$]{}(1.7,0.4); \[Prop 9.2\] Given $\M I^{(0)}$, there exists a sufficiently large $a_0=a_0(\M I^{(0)})$. For $0<a_0<a$, for Einstein vacuum equations with initial data: - $\sum_{i\leq 10, k\leq 3}a^{-\frac12}\|\nab^{k}_4(|u_{\infty}|\nab')^{i}{\hat{\chi}}_0\|_{L^{\infty}(S_{u_{\infty},{\underline{u}}'})}\leq \M I^{(0)}$ along $u'=u_{\infty}$, - Minkowskian initial data along ${\underline{u}}'=0$, - $\int_0^1|u_{\infty}|^2|{\hat{\chi}}_0|^2(u_{\infty}, {\underline{u}}'')d{\underline{u}}''\geq a$ for every direction along $u'=u_{\infty}$, we have that $S'_{-a/4,1}$ is a trapped surface. Note: The statements in Proposition \[Prop 9.2\] are the same as in Theorem \[main.thm2\]. In the below, we will explain that the derived bounds by these two approaches would be slightly different. Proceed the same as in Section \[sec rescale\], via \[9.1\], for $\psi'\in\{{\hat{\chi}}'_{A'B'}, \o'\}$ we have $$\begin{split} &|\p'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-1}\cdot |\p(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-1}\cdot \f{\at}{|u|}=\f{\at}{|u_{\infty}u|}=\f{\at}{|u'|}. \end{split}$$ Similarly, for $\q\in\{{\hat{{\underline{\chi}}}}'_{A'B'}, \eta'_{A'}, {\underline{\eta}}'_{A'}, {\underline{\omega}}', {\mbox{tr}}{\underline{\chi}}'+\f{2}{|u'|}\}$, we have $$|\q'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-1}\cdot |\q(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-1}\cdot \f{\at}{|u_{\infty}u^2|}=\f{\at}{|u_{\infty}|^2|u|^2}=\f{\at}{|u'|^2}.$$ And $$|{\mbox{tr}}\chi'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-1}\cdot |{\mbox{tr}}\chi(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-1}\cdot \f{1}{|u|}=\f{1}{|u_{\infty} u|}=\f{1}{|u'|}.$$ For curvature components, with the same method as in Section \[sec rescale\], via \[9.1\] we obtain $$\begin{split} &|\b'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-2}\cdot |\b_{A}(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-2}\cdot \f{\at}{|u|^2}=\f{\at}{|u_{\infty}u|^2}=\f{\at}{|u'|^2}, \end{split}$$ $$\begin{split} &|\rho'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-2}\cdot |\rho(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-2}\cdot \f{a}{|u_{\infty}u^3|}=\f{a}{|u_{\infty}u|^3}=\f{a}{|u'|^3}, \end{split}$$ $$\begin{split} &|\sigma'(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-2}\cdot |\sigma(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-2}\cdot \f{a}{|u_{\infty}u^3|}=\f{a}{|u_{\infty}u|^3}=\f{a}{|u'|^3}, \end{split}$$ $$\begin{split} &|{\underline{\beta}}'_{A'}(u',{\underline{u}}', \theta^{1'}, \theta^{2'})|=|u_{\infty}|^{-2}\cdot |{\underline{\beta}}_{A}(u, {\underline{u}}, \theta^1, \theta^2)|\leq |u_{\infty}|^{-2}\cdot \f{\at}{|u_{\infty}u^3|}=\f{\at}{|u_{\infty}u|^3}=\f{\at}{|u'|^3}. \end{split}$$ Note that these estimates are slight different from the estimates we obtained in previous sections. We don’t have the proved peeling property and the $a$-weights are quite different. [99]{} X. An, *Formation of Trapped Surfaces from Past Null Infinity*, preprint (2012), arXiv:1207.5271. X. An, *Emergence of Apparent Horizon in Gravitational Collapse*, preprint (2017), arXiv:1703.00118. X. An, *A scale-critical trapped surface formation criterion II: Applications*, in preparation. X. An, N. Athanasiou, *A scale-critical trapped surface formation criterion for the Einstein-Maxwell equations*, in preparation. X. An, J. Luk, *Trapped surfaces in vacuum arising dynamically from mild incoming radiation*, Advances in Theoretical and Mathematical Physics 21 (2017), 1-120. D. Christodoulou, *The formation of black holes and singularities in spherically symmetric gravitational collapse*, Comm. Pure Appl. Math. 44 (1991), no. 3, 339–373. D. Christodoulou, *Bounded variation solutions of the spherically symmetric Einstein-scalar field equations*, Comm. Pure Appl. Math. 46 (1993), no. 8, 1131-1220. D. Christodoulou, *Examples of naked singularity formation in the gravitational collapse of a scalar field*, Ann. of Math. (2) 140 (1994), no. 3, 607–653. D. Christodoulou, *The instability of naked singularities in the gravitational collapse of a scalar field*, Ann. of Math. (2) 149 (1999), no. 1, 183–217. D. Christodoulou, *The Formation of Black Holes in General Relativity*, Monographs in Mathematics, European Mathematical Soc. (2009). D. Christodoulou, S. Klainerman,*The global nonlinear stability of the Minkowski space*, Princeton mathematical series 41, (1993). M. Dafermos, *The formation of black holes in general relativity*, Astrisque 352 (2013). M. Dafermos, G. Holzegel, I. Rodnianski, *A scattering theory construction of dynamical vacuum black holes*, preprint (2013), arXiv: 1306.5364. G. Holzegel, *Ultimately Schwarzschildean spacetimes and the black hole stability problem*, preprint (2010), arXiv:1010.3216. S. Klainerman, F. Nicolo, *The evolution problem in General Relativity*, Progress in Mathematical Physics, Birkhaüser (2003). S. Klainerman, F. Nicolo, *Peeling properties of asymptotically flat solutions to the Einstein vacuum equations*, Class. Quantum Grav. 20 (2003), 3215-3257. S. Klainerman, J. Luk, I. Rodnianski, *A fully anisotropic mechanism for formation of trapped surfaces in vacuum*, Invent. Math. 198 (2014), no.1, 1-26. S. Klainerman, I. Rodnianski, *Causal geometry of Einstein-Vacuum spacetimes with finite curvature flux*, Invent. Math. 159 (2005), 437-529. S. Klainerman, I. Rodnianski, *A geometric approach to the Littlewood-Paley theory*, Geom. Funct. Anal. 16, (2006) no. 1, 126-163. S. Klainerman, I. Rodnianski, *On emerging scarred surfaces for the Einstein vacuum equations*, Discrete Contin. Dyn. Syst. , 28 (2010), no. 3, 1007-1031. S. Klainerman, I. Rodnianski, *On the formation of trapped surfaces*, Acta Math. 208 (2012), no.2, 211-333. P. Le, *The intersection of a hyperplane with a lightcone in the Minkowski spacetime*, J. Differential Geom., 109, (2018), no.3, 497-507. J. Li, J. Liu, *Instability of spherical naked singularities of a scalar field under gravitational perturbations*, preprint, (2017), arXiv: 1710.02422. J. Li, P. Yu, *Construction of Cauchy data of vacuum Einstein field equations evolving to black holes*, Ann. of Math. 181 (2015), no.2, 699-768. J. Luk, *On the local existence for the characteristic initial value problem in general relativity*, Int. Mat. Res. Notices 20, (2012), 4625-4678. J. Luk, *Weak null singularity in general relativity*, Journal of AMS. 31 (2018) 1-63. J. Luk, I. Rodnianski, *Local propagation of impulsive gravitational waves*, Comm. Pure Appl. Math., 68 (2015), no.4, 511-624. J. Luk, I. Rodnianski, *Nonlinear interactions of impulsive gravitational waves for the vacuum Einstein equations*, Cambridge Journal of Math. 5(4): 435-570, 2017. R. Penrose, *Gravitational collapse and space-time singularities*, Phys. Rev. Lett. 14 (1965), 57-59. M. Reiterer, E. Trubowitz, *Strongly focused gravitational waves*, Comm. Math. Phys. 307 (2011), no. 2, 275-313. I. Rodnianski, Y. Shlapentokh-Rothman, *The asymptotically self-similar regime for the Einstein vacuum equations*, Geom. Funct. Anal. Vol. 28 (2018) 755–878. P. Yu, *Energy estimates and gravitational collapse*, Comm. Math. Phys. 317 (2013), no. 2, 275-316. P. Yu, *Dynamical formation of black holes due to the condensation of matter field*, preprint (2011), arXiv: 1105.5898. [^1]: The detailed construction of double null foliation will be explained in Section \[secdnf\]. [^2]: A 2-surface is called a trapped surface if its area element is infinitesimally decreasing along both families of null geodesics emanating from the surface [^3]: Metric of Minkowskian spacetime in spherical coordinates: $g_{M}=-dt^2+dr^2+r^2(d\theta^2+\sin\theta^2 d\phi^2)$. [^4]: Metric of Minkowskian spacetime in stereographic coordinates: $g_{M}=-dt^2+dr^2+\f{4r^4}{(r^2+\theta_1^2+\theta_2^2)^2}(d\theta_1^2+d\theta_2^2)$. In Section \[sec rescale\], we will do a scaling argument in stereographic coordinates. [^5]: Metric of Schwarzschild spacetime: $g_S=-(1-\f{2M}{r})dt^2+({1-\f{2M}{r}})^{-1}dr^2+r^2(d\theta^2+\sin\theta^2 d\phi^2)$. Here $M$ is a [constant]{}. In $g_S$ all metric components are independent of $t$, Schwarzschild spacetime is static, i.e. not changing with $t$. [^6]: The classic approach to deriving energy estimates with Bel-Robinson tensors as in [@Chr:book] and [@KR:Trapped] is avoid, since for higher order energy estimates there are many more terms (including borderline terms) from deformation tensors would appear. [^7]: Letting $a=\d^{-1}$, in a finite region they recover Christodoulou’s main result of [@Chr:book]. [^8]: [It comes from spacetime conformal compactification. See [@KN:peeling].]{} [^9]: More details will be provided in Section \[Signatures\]. [^10]: For a general double null foliation, we have the gauge freedom of choosing how to extend $\O$ along $H_{u_{\infty}}$ and $\Hb_0$. In this paper, we extend $\O\equiv 1$ on both $H_{u_{\infty}}$ and $\Hb_0$. [^11]: That’s because (\[1.1\]) is a geometric PDE system and it respects some natural scalings. [^12]: See full details in Chapter 16 of [@Chr:book] or [@Luk] or Section 10 of [@Dafermos] for a beautiful exposition. [^13]: See Remark \[bootstrap openness\]. [^14]: Recall ${\underline{L}}=\O e_3$. [^15]: From conformal compactification.

ArXiv

=-0.5cm =16.cm =22.cm         [**[THE ANALYSIS OF $\pi^{-}$ MESONS PRODUCED IN NUCLEUS-NUCLEUS COLLISIONS AT A MOMENTUM OF 4.5 GeV/c/NUCLEON IN LIGHT FRONT VARIABLES]{}**]{}           [**[ M.Anikina$^{a}$, L.Chkhaidze$^{b}$, T.Djobava$^{b}$,]{}**]{} [**[ V.Garsevanishvili $^{c}$, L.Kharkhelauri$^{b}$]{}**]{}   ** $^{a}$ Joint Institute for Nuclear Research, 141980 Dubna, Russia $^{b}$ High Energy Physics Institute, Tbilisi State University, University Str. 9, 380086 Tbilisi, Republic of Georgia $^{c}$ Mathematical Institute of the Georgian Academy of Sciences M.Alexidze Str. 1 , 380093 Tbilisi, Republic of Georgia     E-mail: [email protected]     **[ ABSTRACT ]{}**   The light front analysis of $\pi^{-}$ mesons in He(Li,C), C-Ne, C-Cu and O-Pb collisions is carried out. The phase space of secondary pions is divided into two parts in one of which the thermal equilibrium assumption seems to be in a good agreement with the data. Corresponding temperatures $T$ are extracted and their dependence on $(A_{P}*A_{T})^{1/2}$ is studied. The results are compared with the predictions of the Quark-Gluon String Model (QGSM). The QGSM satisfactorily reproduces the experimental data for light and intermediate-mass nuclei.         PACS: 25.70.-z     [*[Keywords]{}*]{}: NUCLEAR REACTION He(Li,$\pi^{-}$,X), He(C,$\pi^{-}$,X), C(Ne,$\pi^{-}$,X), C(Cu,$\pi^{-}$,\ X), O(Pb,$\pi^{-}$,X) at 4.5 GeV/c/nucleon; measured pion disributions: deduced thermal equilibrium, characterized by the temperature $T$. Analysis in light-front variables. Comparison with the Quark Gluon String Model. Light-front variables; Phase space; Thermal equilibrium. **[ 1. INTRODUCTION ]{}**   In the present paper we continue the study of $\pi^{-}$ mesons produced in the relativistic nucleus-nucleus collisions in terms of the light front variables. The choice of the light front variables is due to the fact, that these variables seem to be more sensitive to the dynamics of interaction as compared to the well-known Feynman variables $x_{F}$ and rapidity $Y$. In the previous publication \[1\] we performed the light front analysis of $\pi^{-}$ mesons produced in Mg-Mg collisions at 4.3 GeV/c. On the basis of this analysis we were able to separate the region of the phase space, where the thermal equilibrium seems to be reached. It is interesting in this connection to perform the same analysis for other pairs of nuclei ( He(Li,C), C-Ne, C-Cu, O-Pb) in order to study the dependence of the corresponding characteristics on the atomic number of colliding pairs of nucleus . The data were obtained on the SKM-200-GIBS facility of the Dubna Joint Institute for Nuclear Research.     **[ 2. EXPERIMENT ]{}**   SKM-200-GIBS consists of a 2m streamer chamber, placed in a magnetic field of $\sim$ 0.8 T and a triggering system. The streamer chamber was exposed by beams of He , C , O , Ne and Mg nuclei accelerated in the synchrophasotron up to a momentum of 4.5 GeV/c per incident nucleon. The solid targets in the form of thin discs with thickness 0.2$\div$0.4 g/cm$^{2}$ ( for Li the thickness was 1.59 g/cm$^{2}$ and for Mg 1.56 g/cm$^{2}$) were mounted within the fiducial volume of the chamber. Neon gas filling of the chamber served also as a nuclear target. The triggering system allowed the selection of “inelastic” and “central” collisions. The central trigger was selecting events with no charged and neutral projectile spectator fragments (with $P/Z>3$ GeV/c ) within a cone of half angle $\Theta_{ch}$, $\Theta_{n}$ = 2$^{0}$ or 3$^{0}$. The trigger mode for each exposure is defined as T ($\Theta_{ch}$ ,$\Theta_{n}$ ). Thus T(0,0) corresponds to all inelastic collisions. For inelastic collisions He-Li and He-C all charged secondaries were measured and central subsamples T(2,0) were selected. The details of the experimental setup and the logic of the triggering systems are presented in \[1,2\]. Primary results of scanning and measurements were biased due to several experimental effects and appropriate corrections were introduced. The biases and correction procedures were discussed in detail in \[2,3\]. Average measurement errors of the momentum and production angle determination for $\pi^{-}$ mesons were $<\Delta P/P >$= 5$\%$, $\Delta$$\Theta$ =0.5$^{0}$ for He-Li, He-C, C-Ne, C-Cu, O-Pb.     **[ 3. THE DATA ANALYSIS IN TERMS OF THE LIGHT FRONT VARIABLES ]{}**   The light front variables $\xi^{\pm}$ in the centre of mass frame for the inclusive reaction $a+b\to c+X$ are defined as follows \[1\]: $$\begin{aligned} \xi^{\pm}&=&\pm {E\pm p_z\over{\sqrt{s}}}=\pm {E+|p_z|\over{ \sqrt{s}}} \label{eq1}\end{aligned}$$ where $s$ is the usual Mandelstam variable, $E=\sqrt{p^{2}_z+p^{2}_T+m^{2}}$ and $p_{z}$ are the energy and the $z$ - component of the momentum of produced particle. The $z$ -axis is taken to be the collision axis i.e. $p_{z}=p_{3}$. The upper sign in Eq. (1) is used for the right hand side hemisphere and the lower sign for the left hand side hemisphere. In order to enlarge the scale in the region of small $\xi^{\pm}$, it is convenient also to introduce the variables $$\begin{aligned} \zeta^{\pm}=\mp{\rm ln}|\xi^{\pm}| \label{eq2}\end{aligned}$$ The invariant differential cross section in terms of these variables looks as follows: $$\begin{aligned} E{d\sigma\over{d\vec p}}={|\xi^{\pm}|\over{\pi}}\ {d\sigma\over{ d\xi^{\pm}dp^{2}_T}} = {1\over{\pi}}\ {d\sigma\over{ d\zeta^{\pm}dp^{2}_T}} \label{eq3}\end{aligned}$$ The light front variables have been introduced by Dirac \[4\] and they are widely used now in the treatment of many theoretical problems (see, e.g. original and review papers \[5-10\]). The analysis has been carried out for the $\pi^{-}$ mesons from He(Li,C), C-Ne, C-Cu and O-Pb collisions in the nucleus-nucleus centre of mass system. The number of events for all pairs of nuclei and corresponding trigger modes are listed in Table 1. Due to the small statistics and average multiplicities, the data of He-Li and He-C collisions has been united and thus He(Li,C) represents this sample of the data. In Fig. 1 the $\xi^{\pm}$ – distributions of $\pi^{-}$ mesons from He(Li,C), C-Ne, C-Cu and O-Pb interactions are presented. These distributions are similar for all analysed pairs of nuclei. The principal differences of $\xi^{\pm}$ distributions as compared to the corresponding $x_{F}$ – distributions (Fig. 2) are the following: (1) existence of some forbidden region around the point $\xi^{\pm}=0$; (2) existence of maxima at some $\tilde{\xi^{\pm}}$ in the region of relatively small $|\xi^{\pm}|$. The experimental data for invariant distributions $(1/\pi) \cdot dN/d\zeta^{\pm}$ are shown in Fig. 3. The curves are the result of the polynomial approximation of the experimental distributions. The maxima at $\tilde{\zeta}^{\pm}$ are also observed in the invariant distributions $(1/\pi) \cdot dN/d\zeta^{\pm}$. However, the region $|\xi^{\pm}|>|\tilde{\xi}^{\pm}|$ goes over to the region $|\zeta^{\pm}|<|\tilde{\zeta}^{\pm}|$ and vice versa (see Eqs. (1) and (2)). The value of maxima are observed at $\tilde{\zeta^{\pm}}=2.0\pm0.1$ for all pairs of nuclei. The $\tilde{\zeta^{\pm}}$ is the function of the energy (see Eqs. (1), (2)) and does not depend on the mass numbers of the projectile ($A_{P}$) and target ($A_{T}$). In order to study the nature of these maxima we have divided the phase space into two regions $|\zeta^{\pm}|>|\tilde{\zeta}^{\pm}|$ ($\tilde{\zeta^{+}}$= 2.0) and $|\zeta^{\pm}|<|\tilde{\zeta}^{\pm}|$ and studied the $p_{T}^{2}$ and the angular distributions of $\pi^{-}$ mesons in these regions separetely. The numbers of pions in these two regions are approximately equal. For example in C-Cu interactions in the region $|\zeta^{\pm}|>|\tilde{\zeta}^{\pm}|$ the number of pions is equal to –1987 and in $|\zeta^{\pm}|<|\tilde{\zeta}^{\pm}|$ — 2212. In Figs. 4, 5 the $p_{T}^{2}$ and the angular distributions of $\pi^{-}$ mesons from He(Li,C), C-Ne, C-Cu and O-Pb interactions in different regions of $\zeta^{+}$ ( $\zeta^{+} > \tilde{\zeta^{+}}$ and $\zeta^{+} < \tilde{\zeta^{+}}$) in the forward hemisphere are presented. One can see from Figs. 4, 5, that the $p_{T}^{2}$ and the angular distributions of $\pi^{-}$ mesons differ significantly in $\zeta^{+} > \tilde{\zeta^{+}}$ and $\zeta^{+} < \tilde{\zeta^{+}}$ regions. The angular distribution of pions in the region $\zeta^{+} < \tilde{\zeta^{+}}$ (Figs. 5.b and 5.c) is sharply anisotropic in contrast to the almost flat distribution in the region $\zeta^{+} > \tilde{\zeta^{+}}$ (Figs. 5.a and 5.c). The flat behaviour of the angular distribution allows one to think that one observes a partial thermal equilibrium \[11\] in the region $|\zeta^{\pm}| > |\tilde{\zeta^{\pm}}|$ ($|\xi^{\pm}| < |\tilde{\xi^{\pm}}|$) of phase space. The slopes of $p_{T}^{2}$ – distributions differ greatly in different regions of $\zeta^{\pm}$ (Fig. 4). Thus the values of $\tilde{\zeta^{\pm}}$ are the boundaries of the two regions with significantly different characteristics of $\pi^{-}$ mesons. The validity of this statement can be seen from the momentum distributions of $\pi^{-}$ mesons in the Laboratory frame. Fig. 6 presents the momentum disribution of pions from C-Cu collisions in the laboratory frame. The shaded area corresponds to the region of $\zeta^{+} > \tilde{\zeta^{+}}$ and the non-shaded one to the region of $\zeta^{+} < \tilde{\zeta^{+}}$. One can see from the Fig. 6, that these two regions almost do not overlap in the momentum space unlike to the cms case (overlap $\sim 45\%$). The pions from the region $\zeta^{+} > \tilde{\zeta^{+}}$ have small momentum, approximately up to 0.6 GeV/c as compared to the pions from $\zeta^{+} < \tilde{\zeta^{+}}$ (the momentum of pions ranges from $\sim$ 0.6 GeV/c to 3 GeV/c). The similar results have been also obtained for the other pairs of nuclei. Fig. 7 presents the dependence of $<P>_{lab}$ on $\Theta_{lab}$ for all analysed pairs of nuclei (He(Li,C) and C-Ne data are presented with the same symbol because of the similarity of their dependences) and Mg-Mg interactions in the $\zeta^{+} > \tilde{\zeta^{+}}$ and $\zeta^{+} < \tilde{\zeta^{+}}$ regions. The shapes of these dependences are different in two regions of $\zeta^{+}$. The curves are the result of polynomial approximation. $<P>_{lab}$ decreases and $<\Theta>_{lab}$ increases with the increasing of $A_{P}$, $A_{T}$. To describe the spectra in the region $\zeta^{+} > \tilde{\zeta^{+}}$ the Boltzmann $ f(E)\sim e^{-E/T} $ distribution have been used. The distributions $1/\pi \cdot dN/d\zeta^{+}$, $dN/dp_{T}^{2}$, $dN/dcos\Theta$ look in this region as follows : $$\begin{aligned} {1\over{\pi}}\ {dN\over{d\zeta^+}}&\sim&\int_0^{p^2_{T,max}}Ef(E)dp^2_T \label{eq4}\\ {dN\over{dp^2_T}}&\sim&\int_0^{p^{}_{z,max}}f(E)dp^{}_z \label{eq5}\\ {dN\over{d\cos\theta}}&\sim&\int_0^{p^{}_{max}}f(E)p^2dp \label{eq6}\\ E&=&\sqrt{\vec p\, ^2+m^2_{\pi}}\ ,\ \vec p\, ^2=p^2_z+p^2_T \label{eq7}\end{aligned}$$ where: ${p^{2}_{T,max}} =(\tilde{\xi^{+}}\sqrt{s})^{2} - m_{\pi}^{2}$ ${p_{z,max}} =[p_{T}^{2}+m^{2}-(\tilde{\xi^{+}}\sqrt{s})^{2}]/ (-2\tilde{\xi^{+}}\sqrt{s})$ $p_{max}=(-\tilde{\xi^{+}}\sqrt{s}cos\Theta + \sqrt{( \tilde{\xi^{+}}\sqrt{s})^{2}- m_{\pi}^{2} sin^{2}\Theta})/sin^{2}\Theta$ The experimental distributions in the region $\zeta^{+} >\tilde{\zeta^{+}}$ have been fitted by the expressions (4), (5), (6), respectively. The results of the joint fit of the distributions $1/\pi \cdot dN/d\zeta^{+}$, $dN/dp_{T}^{2}$, $dN/dcos\Theta$ are given in Table 1 and Figs. 4, 5, 8. They show a rather good agreement with experiment. In Table 1 the values of the parameter $T$ obtained by fitting the data with Boltzmann distribution are presented. In order to determine how the characteristics vary the analysis have been carried out also for $\tilde{\zeta^{+}}$=1.9 and 2.1. The results are similar, but the joint fit of the distributions is better for $\tilde{\zeta^{+}}$=2.0 (presented on Figures). The spectra of $\pi^{-}$ mesons in the region $\zeta^{+} > \tilde{\zeta^{+}}$ are satisfactorily described by the formulae which follow from the thermal equilibrium. The same formulae when extrapolated to the region $\zeta^{+} < \tilde{\zeta^{+}}$ (Fig. 8) deviate significantly from the data. Therefore in the region $\zeta^{+} < \tilde{\zeta^{+}}$ the $p_{T}^{2}$ – distributions has been fitted by the formula $$\begin{aligned} \frac {dN}{dp_{T}^{2}} \sim \alpha \cdot e^{-\beta_{1}P_{T}^{2}} + (1-\alpha) \cdot e^{-\beta_{2}p_{T}^{2}} \label{eq8}\end{aligned}$$ and the $\zeta^{+}$ – distributions by the formula $$\begin{aligned} \frac{1}{\pi}\cdot\frac {dN}{d\zeta^{+}} \sim (1 - \xi^{+})^{n}= (1 - e^{-\vert \zeta^{+}\vert})^{n} \label{eq9}\end{aligned}$$ which is an analogue of the $(1-x_{F})^{n}$ dependence – the result of the well-known quark-parton model consideration (see, e.g. \[12\]), which for $\pi^{-}$ mesons gives the value n=3. The dependence $(1 - e^{-\vert \zeta^{+}\vert})^{n}$ is in good agreement with experiment in the region $\zeta^{+} < \tilde{\zeta^{+}}$ and deviates from it in the region $\zeta^{+} > \tilde{\zeta^{+}}$ ( Fig. 8). The results of the fit are given in Table 1 and Figs. 4 and 8. Thus in the $\zeta^{\pm}$ ($\xi^{\pm}$) distributions we have singled out points $\tilde{\zeta^{\pm}}$ ($\tilde{\xi^{\pm}}$) which seperate in the phase space two groups of particles with significantly different characteristics. There are no such points in the $x_{F}$ and $Y$ - distributions. In this paper the Quark Gluon String Model (QGSM) \[13\] is used for a comparison with experimental data. The QGSM is based on the Regge and string phenomenology of particle production in inelastic binary hadron collisions \[14\]. The QGSM simplifies the nuclear effects (neglects the potential interactions between hadrons, Pauli blocking, coalescence of nucleons and etc.) and concentrates on hadron rescatterings. The QGSM includes only low lying vector mesons and baryons with spin 3/2, mostly $\Delta$ (3/2, 3/2) resonances. A detailed description and comparison of the QGSM with experimental data in a wide energy range can be found in Ref. \[1,15,16\]. We have generated He(Li,C), C-Ne, C-Cu, O-Pb interactions using Monte-Carlo generator COLLI, based on the QGSM. The events had been traced through the detector and trigger filter. In the generator COLLI there are two possibilities to generate events: 1) at not fixed impact parameter $\tilde{b}$ and 2) at fixed $b$. The events have been generated for $\tilde{b}$. From the impact parameter distribution we obtained the mean value of $<b>$. For the obtained value of $<b>$, we have generated a total samples of $A_{P}-A_{T}$ events. The numbers of generated events for all analysed pairs of nuclei are listed in Table 1. The two regimes are consistent and it seems, that in our experiment the following values of b are most probable: b=1.55 fm for He(Li,C); b=2.20 fm for C-Ne; b=2.75 fm for C-Cu; b=3.75 fm for O-Pb. The experimental results have been compared with the predictions of the QGSM for the above mentioned values of $b$ and satisfactory agreement between the experimental data and the model have been found. In Figs 1.b and 3.b the $\xi^{\pm}$ and $\zeta^{\pm}$ - distributions of $\pi^{-}$ mesons from the QGSM calculations are presented together with the experimental ones for C-Cu interactions. One can see, that the QGSM reproduces these distributions well. The similar results have been obtained for all analysed pairs of nuclei. The QGSM also reproduces the $ p_{T}^{2} $ and $ cos\Theta $ distributions (Figs. 4.c and 5.c). The QGSM data show the similar characteristics in the different regions of $\zeta$ as experimental ones: sharply anisotropic angular distributions in the region $\zeta^{+} < \tilde{\zeta^{+}}$ and the almost flat distribution in the region $\zeta^{+} > \tilde{\zeta^{+}}$; the slopes of $p_{T}^{2}$ – distributions differ greatly in different regions of $\zeta^{+}$; the momentum distributions of pions in the laboratory frame in different regions of $\zeta^{+}$ have the similar different shape of spectra as experimental ones (Fig 6). Momentum distributions of QGSM data reproduces the corresponding experimental spactra in both regions of $\zeta^{+}$. The distributions obtained by the QGSM in the region $\zeta^{+} >\tilde{\zeta^{+}}$ have been fitted by the expressions (4), (5), (6). The results of the fit are given in Table 1 and Figs. 4.c, 5.c. In the region $\zeta^{+} < \tilde{\zeta^{+}}$ the $p_{T}^{2}$ and the $\zeta^{+}$ – distributions have been fitted by the formulae (8) and (9), respectively. The results of the fit are given in Table 1 and Fig. 4.c. One can see from the Table 1, that the values of the $T$ extracted from the experimental and QGSM data coincide within the errors. The QGSM does not reproduce satisfactorily the O-Pb data. This is may be caused by the fact, that QGSM simplifies the nuclear effects, which are more pronounced for heavy nuclei. In Ref. \[17\] it has been indicated that the model can be improved by including higher mass baryon resonances and taking into account a possible increase of the pion absorption cross section, $\sigma_{\Delta\Delta \rightarrow NN}$ in dense baryon medium, in comparison with the cross section, obtained from detailed balance relation. In Fig. 9 the dependence of the parameter $T$ from the Table 1 on $(A_{P}*A_{T})^{1/2}$, obtained from the experimental and QGSM data, is presented. The temperature for Mg-Mg interactions is extracted from \[1\]. One can see, that $T$ decreases linearly with the increasing of $(A_{P}*A_{T})^{1/2}$ i.e with the increasing number of participating nucleons. Similar behaviour is predicted by the QGSM. In our previous article \[18\] the temperatures of pions in He-Li, He-C, C-Ne, C-Cu and O-Pb interactions were obtained be means of inclusive kinetic energy and transverse momentum spectra in central rapidity interval (0.5 – 2.1 for light nuclei and 0.1 – 1.8 for heavy ones), which corresponds to the pionization region and with the c.m.s. angles $90 ^{0} \pm 10^{0}$. The pion spectra for He-Li, He-C and C-Ne have been fitted by one exponent and for C-Cu and O-Pb by a sum of two exponents, or two temperatures $T_{1}$ and $T_{2}$ (describing the low and high momentum part of the spectrum). The temperatures extracted by the light front analysis for light pairs of nuclei are less about (15–20) $\%$ as compared to those obtained in Ref. \[18\]. For heavy pairs of nuclei the temperatures are more close to the low temperature $T_{1}$. It seems obvious, as the thermal equilibrium region corresponds to lower momenta. It should be mentioned that the extraction procedures of $T$ in the light-front variables and in Ref. \[18\] are quite different and it seems, that different regions of phase space are seperated by these two methods. It is interesting to compare the temperatures extracted by means of light front analysis with those obtained in the GSI experiments (FOPI, KAON and TAPS- Collaborations, see, e.g. \[19,20\]). The $T$ in the GSI experiments have been obtained in a same manner as in our Ref. \[18\]. The numerical values of the parameter $T$ for pions in Au-Au collisions at 1 A GeV and our values for heaviest colliding pair are close to each other. It seems interesting in this connection to perform the light front analysis of the GSI, AGS and SPS data.     **[ 4. CONCLUSIONS ]{}**   The light front analysis of $\pi^{-}$ – mesons in the relativistic nucleus-nucleus collisions is carried out. The results of this paper confirm the conclusions of our previous publication \[1\], that the phase space of secondary $\pi^{-}$ mesons is divided into two parts, in one of which the thermal equilibrium seems to be reached. Corresponding temperatues are obtained from the fitting of the data by the Boltzmann distribution. The characteristics of the $\pi^{-}$ mesons (the momentum, angular, $p_{T}^{2}$ – distributions) in these two regions differ significantly. The dependence of $T$ on $(A_{P}*A_{T})^{1/2}$ has been studied. The temperature decreases with increase of $(A_{P}*A_{T})^{1/2}$. The experimental results have been compared with the QGSM. The model seems to be in a good agreement with the data excluding O-Pb collisions – the heaviest colliding pair.           ACKNOWLEDGEMENTS     The authors express their deep gratitude to J.-P.Alard, Sh.Esakia, D.Ferenc, G.Kuratashvili, J.-F.Mathiot, Z.Menteshashvili, G.Paic for interesting discussions.       [99]{} , 1990 [(World Scientific, Singapore, 1990)]{};\ [N.Amelin, K.Gudima and V.Toneev, Sov.J. Nucl.Phys. 51 (1990) 327]{} **[FIGURE CAPTIONS]{}**   [**[Fig. 1]{}**]{} [The $\xi^{\pm}$ – distribution of $\pi^{-}$ mesons   a) from $\ast$ – He(Li,C), $\triangle$ – C-Ne, $\circ$ – O-Pb interactions. b) From C-Cu interactions: $\circ$ – the experimental data, $\triangle$ – the QGSM data.\ The curves are the result of polynomial approximation of the experimental data.]{}   [**[Fig. 2]{}**]{} [The $x_{F}$ – distribution of $\pi^{-}$ mesons from $\ast$ – He(Li,C), $\triangle$ – C-Cu, $\circ$ – O-Pb interactions.]{}   [**[Fig. 3]{}**]{} [The $ \zeta^{\pm} $ – distribution of $\pi^{-}$ mesons    a) from $\ast$ – He(Li,C), $\triangle$ – C-Ne, $\circ$ – O-Pb interactions. b) From C-Cu interactions: $\circ$ – the experimental data, $\triangle$ – the QGSM data.\ The curves are the result of polynomial approximation of the experimental data.]{}   [**[Fig. 4]{}**]{} [The $ p_{T}^{2} $ distribution of $\pi^{-}$ mesons from $\ast$ – He(Li,C), $\triangle$ – C-Ne, $\circ$ – O-Pb interactions    a) for $\zeta^{+} > \tilde{\zeta^{+}}$.   b) For $\zeta^{+} < \tilde{\zeta^{+}}$.   c) From C-Cu interactions for $\zeta^{+} > \tilde{\zeta^{+}}$ $\circ$ – experimental and $\bigtriangleup$ – the QGSM data; for $\zeta^{+} < \tilde{\zeta^{+}}$      – experimental and $\ast$ – the QGSM data.\ The solid lines - fit of the experimental data in the regions $\zeta^{+} > \tilde{\zeta^{+}}$ and $\zeta^{+} < \tilde{\zeta^{+}}$ by the Eqs.(5) and (8), correspondingly.]{}   [**[Fig. 5]{}**]{} [The $ cos\Theta $ distribution of $\pi^{-}$ mesons from $\ast$ – He(Li,C), $\triangle$ – C-Ne, $\circ$ – O-Pb interactions    a) for $\zeta^{+} > \tilde{\zeta^{+}}$.   b) For $\zeta^{+} < \tilde{\zeta^{+}}$.   c) From C-Cu interactions for $\zeta^{+} > \tilde{\zeta^{+}}$ $\circ$ – experimental and $\bigtriangleup$ – the QGSM data; for $\zeta^{+} < \tilde{\zeta^{+}}$,      – experimental and $\ast$ – the QGSM data.\ The solid lines - fit of the experimental data in the region $\zeta^{+} > \tilde{\zeta^{+}}$ by the Eq.(6) and in the $\zeta^{+} < \tilde{\zeta^{+}}$ by the polynom.]{}   [**[Fig. 6]{}**]{} [The momentum distribution of $\pi^{-}$ mesons from C-Cu interactions in the laboratory system. The shaded area corresponds to the region of $\zeta^{+} > \tilde{\zeta^{+}}$.]{}   [**[Fig. 7]{}**]{} [The dependence of $<P>_{lab}$ on $\Theta_{lab}$ in the regions $\zeta^{+} > \tilde{\zeta^{+}}$ (bottom data) and $\zeta^{+} < \tilde{\zeta^{+}}$ (top data) for      – He(Li,C) and C-Ne, $\diamond$ – C-Cu, $\ast$ – Mg-Mg, $\circ$ – O-Pb. The curves – result of polynomial approximation.]{}   [**[Fig. 8]{}**]{} [The $ (1/\pi) \cdot dN/d\zeta^{+} $ distribution of $\pi^{-}$ mesons from C-Cu interactions. $\circ$ – experimental data, the solid line – fit of the experimental data in the region $\zeta^{+} > \tilde{\zeta^{+}}$ by the Eq.(4), the dashed line – fit of the experimental data in the region $\zeta^{+} < \tilde{\zeta^{+}}$ by the Eq.(9).]{}   [**[Fig. 9]{}**]{} [The dependence of the parameter $T$ on $(A_{P}*A_{T})^{1/2}$ for He(Li,C), C-Ne, Mg-Mg, C-Cu and O-Pb. $\circ$ – the experimental data, $\bigtriangleup$ – the QGSM data. The dashed line is a result of linear approximation of the experimental data.]{}         **[TABLE CAPTIONS]{}**     [**[Table 1.]{}**]{} [ Number of the events, trigger and the results of the joint fit of the distributions $1/\pi \cdot dN/d\zeta^{+}$, $dN/dp_{T}^{2}$, $dN/dcos\Theta$ of $\pi^{-}$ – mesons by Eqs. (4), (5), (6) in the region $\zeta^{+} >\tilde{\zeta^{+}}$ and $1/\pi \cdot dN/d\zeta^{+}$ distributions by Eq.(9) in the region $\zeta^{+} <\tilde{\zeta^{+}}$.]{} Number of the events, trigger and the results of the joint fit of the distributions $1/\pi \cdot dN/d\zeta^{+}$, $dN/dp_{T}^{2}$, $dN/dcos\Theta$ of $\pi^{-}$ – mesons by Eq. (4), (5), (6) in the region $\zeta^{+} >\tilde{\zeta^{+}}$ and $1/\pi \cdot dN/d\zeta^{+}$ distributions by Eq.(9) in the region $\zeta^{+} <\tilde{\zeta^{+}}$.\     ----------------------------- ------ ----------- -------------------------------- -------------------------------- $A_{p} - A_{T} $ Number of T  (MeV) $n$ $T(\Theta_{ch},\Theta_{n}$) events $\zeta^{+} >\tilde{\zeta^{+}}$ $\zeta^{+} <\tilde{\zeta^{+}}$ $ He(Li,C) $ exp. 6147 81 $\pm$ 2 3.6 $\pm$ 0.2 $T(2,0)$ QGSM 15566 84 $\pm$ 2 3.5 $\pm$ 0.1 $ C-Ne $ exp. 902 79 $\pm$ 3 3.7 $\pm$ 0.2 $T(2,0)$ QGSM 3950 82 $\pm$ 2 3.4 $\pm$ 0.8 $ C-Cu $ exp. 1203 72 $\pm$ 2 3.0 $\pm$ 0.1 $T(3,3)$ QGSM 3463 74 $\pm$ 2 3.2 $\pm$ 0.8 $ O-Pb $ exp. 732 55 $\pm$ 3 2.6 $\pm$ 0.1 ----------------------------- ------ ----------- -------------------------------- --------------------------------

ArXiv

--- abstract: | We present the classical Poisson-Lichnerowicz cohomology for the Poisson algebra of polynomials $\mathbb{C}[X_{1},\ldots ,X_{n}]$ using exterior calculus. After presenting some non homogeneous Poisson brackets on this algebra, we compute Poisson cohomological spaces when the Poisson structure corresponds to a bracket of a rigid Lie algebra. [**2000 MSC:**]{} 17B-XX author: - | Nicolas Goze\ LMIA,\ Université de Haute Alsace,\ 4, rue des Frères Lumière,\ MULHOUSE, 68093. France\ `[email protected]` title: 'Poisson structures on $\mathbb{C[}X_{1},\ldots ,X_{n}]$ associated with rigid Lie algebras.' --- Introduction ============ The first Poisson structures appeared in classical mechanics. In 1809, D. Poisson introduced a bracket of functions, which permits to write the Hamilton’s equations as differential equations. This leaded to define a Poisson manifold, that is, a manifold $M$ whose algebra of smooth functions $F(M)$ is equipped with a skew-symmetric bilinear map $$\{\ ,\ \} : F(M) × F(M) \rightarrow F(M),$$ satisfying the Leibniz rule, $$\{FG, H\} = F\{G, H\} + \{F, H\}G,$$ and the Jacobi identity. In [@Lich], A. Lichnerowicz has also introduced a cohomology, associated to a Poisson structure, called Poisson cohomology. In this paper we study in terms of exterior calculus the Poisson structures on the associative algebra of complex polynomials in $n$ variables. We apply this approach to the determination of non homogeneous quadratic Poisson brackets and to the computation of the Poisson cohomology. The linear Poisson structures are naturally related to the $n$-dimensional Lie algebras. Recall that a complex Lie algebra $\frak g$ is rigid when its orbit in the algebraic variety of $n$-dimensional complex Lie algebra defined by the Jacobi relations is Zariski open. Such an algebra admits a non trivial Malcev torus and it is graded by the roots of the torus. We study the Poisson structure on $\mathbb{C}[X_{1},\ldots ,X_{n}]$ whose Poisson brackets correspond to a solvable rigid Lie bracket with non zero roots. In a generic example we compute the corresponding Poisson cohomology. Poisson structures on $\mathbb{C}[X_{1},\ldots ,X_{n}]$ and exterior calculus ============================================================================= Poisson brackets and differential forms --------------------------------------- Let $\mathcal{A}^{n}$ be the commutative associative algebra $\mathbb{C}[X_{1},\ldots ,X_{n}]$ of complex polynomials in $X_{1},\cdots ,X_{n}$. We define a Poisson structure on $\mathcal{A}^{n}$ as a bivector $$\mathcal{P}=\sum_{1\leq i<j\leq n}P_{ij}\partial _{i}\wedge \partial _{j},$$where $\partial _{i}=\frac{\partial }{\partial X_{i}}$ and $P_{ij}\in \mathcal{A}^{n}$, satisfying the axiom $$\lbrack \mathcal{P},\mathcal{P}]_{S}=0,$$where $[,]_{S}$ denotes the Schouten’s bracket. If $\mathcal{P}$ is a Poisson structure on $\mathcal{A}^{n}$, then $$\{P,Q\}=\mathcal{P}(P,Q)$$defines a Lie bracket on $\mathcal{A}^{n}$ which satisfies the Leibniz identity $$\{PQ,R\}=P\{Q,R\}+Q\{P,R\},$$for any $P,Q,R\in \mathcal{A}^{n}.$ We denote by $Sh_{p,q}$ the set of $(p,q)$-shuffles where a $(p,q)$-shuffle is a permutation $\sigma $ of the symmetric group $\Sigma_{p+q}$ of degree $p+q$ such that $\sigma (1)<\sigma (2)<\cdots <\sigma (p)$ and $\sigma (p+1)<\sigma (p+2)<\cdots <\sigma (p+q)$. For any bivector $\mathcal{P}$ we consider the $(n-2)$-exterior form $$\Omega =\sum_{\sigma \in S_{2,n-2}}(-1)^{\varepsilon (\sigma )}P_{\sigma (1)\sigma (2)}dX_{\sigma (3)}\wedge \cdots \wedge dX_{\sigma (n)},$$where $(-1)^{\varepsilon (\sigma )}$ is the signature of the permutation $\sigma $. If $n>3$, we consider the Pfaffian form $\alpha _{i_{1},\cdots ,i_{n-3}}$ given by $$\begin{array}{l} \alpha _{i_{1},\cdots ,i_{n-3}}(Y)=\Omega (\partial _{i_{1}},\partial _{i_{2}},\ldots ,\partial _{i_{n-3}},Y)\end{array}$$with $Y=\sum_{i=1}^{n}Y_{i}\partial _{i},$ $Y_{i}\in \mathcal{A}^{n}.$ A bivector $\mathcal{P}$ on $\mathcal{A}^{n}$ satisfies $[\mathcal{P},\mathcal{P}]_{S}=0$ if and only if - for $n>3$, $$d\alpha _{i_{1},\cdots ,i_{n-3}}\wedge \Omega =0,$$for every ${i_{1},\cdots ,i_{n-3}}$ such that $1\leq i_1 < \cdots < i_{n-3}\leq n$. - for $n=3$ $$d\Omega\wedge \Omega =0.$$ *Proof.* The integrability condition $[\mathcal{P},\mathcal{P}]_{S}=0$ writes$$\sum_{r=1}^{n}P_{ri}\partial _{r}P_{jk}+P_{rj}\partial _{r}P_{ki}+P_{rk}\partial _{r}P_{ij}=0,$$for any $1\leq i,j,k\leq n.$ But$$\alpha _{i_{1},\cdots ,i_{n-3}}=\sum (-1)^{N}P_{jk}dX_{l},$$summing over all triples $(j,k,l)$ such that $(j,k,i_{1},\ldots ,l,\ldots i_{n-3})$ is a permutation of $S_{2,n-2}$ and $N=\varepsilon (\sigma )+p-3$ where $(-1)^\varepsilon(\sigma )$ is the signum of $\sigma$. Then $$d\alpha _{i_{1},\cdots ,i_{n-3}}=\sum (-1)^{N}dP_{jk}\wedge dX_{l}$$and $d\alpha _{i_{1},\cdots ,i_{n-3}}\wedge \Omega =0$ corresponds to $[\mathcal{P},\mathcal{P}]_{S}=0.$ The proof is similar if $n=3$. Lichnerowicz-Poisson cohomology ------------------------------- We denote by ${\mathcal{A}}^n_{\mathcal{P}}$ the algebra $\mathcal{A}^n=\mathbb{C[}X_{1},\ldots ,X_{n}]$ provided with the Poisson structure $\mathcal{P}. $ For $k \geq 1$, let $\chi ^{k}({\mathcal{A}}^n_{\mathcal{P}})$ be the vector space of $k$-derivations that is of $k$-skew linear maps on ${\mathcal{A}}^n_{\mathcal{P}}$ satisfying$$\varphi (P_{1}Q_{1},P_{2},\ldots ,P_{k})=P_{1}\varphi (Q_{1},P_{2},\ldots ,P_{k})+Q_{1}\varphi (P_{1},P_{2},\ldots ,P_{k}),$$for all $Q_{1},P_{1},\ldots ,P_{k}\in {\mathcal{A}}^n_{\mathcal{P}}.$ For $k=0$ we put $\chi ^{0}({\mathcal{A}}^n_{\mathcal{P}})={\mathcal{A}}^n_{\mathcal{P}}.$ Let $\delta ^{k}$ be the linear map$$\delta ^{k}:\chi ^{k}({\mathcal{A}}^n_{\mathcal{P}})\longrightarrow \chi ^{k+1}({\mathcal{A}}^n_{\mathcal{P}})$$given by$$\begin{array}{ll} \delta ^{k}\varphi (P_{1},P_{2},\ldots ,P_{k+1})=& \displaystyle\sum_{i=1}^{k+1}(-1)^{i-1}\{P_{i},\varphi (P_{1},\ldots ,\widehat{P_{i}},\ldots P_{k+1})\} \\ \medskip &+\displaystyle\sum_{1\leq i<j\leq k+1}(-1)^{i+j}\varphi (\{P_{i},P_{j}\},P_{1},\ldots ,\widehat{P_{i}},\ldots ,\widehat{P_{j}},\ldots P_{k+1}), \end{array}$$where $\widehat{P_{i}}$ means that the term $P_{i}$ does not appear. We have $\delta ^{k+1}\circ \delta ^{k}=0$ and the Lichnerowicz-Poisson cohomology corresponds to the complex $(\chi ^{k}(\mathcal{A_P}),\delta ^{k})_{k}$. Let us note that $\chi ^{k}({\mathcal{A}}^n_{\mathcal{P}})$ is trivial as soon as $k>n.$ A description of the cocycle $\delta ^{k}\varphi $ is presented in [@Pi] for $n=3$ using the vector calculus. We will describe these formulae using exterior calculus for $n>3$. Let us begin with some notations: - To any element $P\in {\mathcal{A}}^n_{\mathcal{P}}=\chi ^{0}({\mathcal{A}}^n_{\mathcal{P}}),$ we associate the $n$-exterior form $$\Phi _{n}(P)=PdX_{1}\wedge \ldots \wedge dX_{n}.$$ - To any $\varphi \in \chi ^{k}({\mathcal{A}}_{\mathcal{P}}^{n})$ for $1\leq k<n,$ we associate the $(n-k)$-exterior form$$\Phi _{n-k}(\varphi )=\sum_{\sigma \in S_{k,n-k}}(-1)^{\varepsilon (\sigma )}\varphi (X_{\sigma (1)},\ldots ,X_{\sigma (k)})dX_{\sigma (k+1)}\wedge \ldots \wedge dX_{\sigma (n)}.$$ - To any $\varphi \in \chi ^{n}({\mathcal{A}}_{\mathcal{P}}^{n}),$ we associate the function $\Phi _{0}(\varphi )=\varphi .$ Finally, if $\theta $ is an $k$-exterior form and $Y=\displaystyle\sum_{i=1}^{n}Y_{i}\partial _{i}$ is a vector field with $Y_{i}\in {\mathcal{A}}_{\mathcal{P}}^{n}$, then the inner product $i(Y)\theta $ is the $(k-1)$-exterior form given by $$i(Y)\theta (Z_{1},\cdots ,Z_{k-1})=\theta (Y,Z_{1},\cdots ,Z_{k-1}),$$for every vector fields $Z_{1},\cdots ,Z_{k-1}$. Assume that $n=3$. Then we have 1\. For all $P\in {\mathcal{A}}_{\mathcal{P}}^{3},$ $$\Phi _{2}(\delta ^{0}P)=-\Omega \wedge dP.$$ 2\. For all $f\in \chi ^{1}({\mathcal{A}}_{\mathcal{P}}^{3}),$$$\begin{array}{ll} \Phi _{1}(\delta ^{1}f) = & -i(\partial _{1},\partial _{2})[\Omega \wedge d(i(\partial _{3})\Phi _{2}(f))+d(i(\partial _{3})\Omega )\wedge \Phi _{2}(f)] \\ & +i(\partial _{1},\partial _{3})[\Omega \wedge d(i(\partial _{2})\Phi _{2}(f))+d(i(\partial _{2})\Omega )\wedge \Phi _{2}(f)] \\ & -i(\partial _{2},\partial _{3})[\Omega \wedge d(i(\partial _{1})\Phi _{2}(f))+d(i(\partial _{1})\Omega )\wedge \Phi _{2}(f)],\end{array}$$where $i(X,Y)$ denotes the composition $i(X)\circ i(Y).$ 3\. For all $\varphi \in \chi ^{2}({\mathcal{A}}^3_{\mathcal{P}}),$$$\Phi _{0}(\delta ^{2}\varphi )=i(\partial_{1},\partial_{2},\partial_{3})(d\Omega \wedge \Phi _{1}(\varphi )+\Omega \wedge d\Phi _{1}(\varphi )).$$ *Proof.* If $n=3$ we have $$\Omega =P_{12}dX_{3}-P_{13}dX_{2}+P_{23}dX_{1}.$$Then the integrability of $\mathcal{P}$ is equivalent to $\Omega \wedge d\Omega =0.$ The theorem results of a direct computation and of the following general formula: $$\forall \varphi \in \chi ^{k}({\mathcal{A}}_{\mathcal{P}}^{n}),\ \varphi (P_{1},\ldots ,P_{k})=\sum_{1\leq i_{1}\leq \ldots \leq i_{k}\leq n}\varphi (X _{i_{1}},\ldots ,X _{i_{k}})\partial _{i_{1}}P_{1}\ldots \partial _{i_{k}}P_{k}.$$ **Example.** We consider the Poisson algebra ${\mathcal{A}}_{\mathcal{P}_{1}}^{n}=(\mathbb{C[}X_{1},X_{2},X_{3}],\mathcal{P}_{1})$ where $\mathcal{P}_{1}$ is given by $$\left\{ \begin{array}{l} \mathcal{P}_{1}(X_{1},X_{2})=X_{2}, \\ \mathcal{P}_{1}(X_{1},X_{3})=2X_{3}, \\ \mathcal{P}_{1}(X_{2},X_{3})=0.\end{array}\right.$$Then $$\dim H^{0}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})=1,\ \dim H^{1}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})=3,\ \\dim H^{2}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})=2,\ H^{3}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})=\{0\}.$$In this case $\Omega =X_{2}dX_{3}-2X_{3}dX_{2}$ and $d\Omega =3dX_{2}\wedge dX_{3}$. Let us compute $\dim H^{2}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})$ . Let $\varphi \in \chi ^{2}({{\mathcal{A}}_{\mathcal{P}}^{n}}_{1}).$ Then $\Phi _{0}(\delta ^{2}\varphi )=0$ implies$$d\Omega \wedge \Phi _{1}(\varphi )+\Omega \wedge d\Phi _{1}(\varphi ))=0,$$ that is $$\begin{array}{l} X_{2}(\partial _{1}\varphi (X_{1},X_{3})+\partial _{2}\varphi (X_{2},X_{3}))+2X_{3}(-\partial _{1}\varphi (X_{1},X_{2})+\partial _{3}\varphi (X_{2},X_{3})) \\ +3\varphi (X_{2},X_{3})=0.\end{array}$$Now, if $f\in \chi ^{1}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})$ then $$\begin{array}{ll} \Phi _{1}(\delta f)= & [X_{2}(-\partial _{2}f(X_{2})-\partial _{1}f(X_{1}))-2X_{3}(\partial _{3}f(X_{2}))+f(X_{2})]dX_{3} \\ & -[2X_{3}(\partial _{1}f(X_{1})+\partial _{3}f(X_{3}))+X_{2}(\partial _{2}f(X_{3}))-2f(X_{3})]dX_{2} \\ & -[X_{2}(-\partial _{1}f(X_{3}))-2X_{3}(\partial _{1}f(X_{2}))]dX_{1}.\end{array}$$Comparing these two relations we obtain that $H^{2}({\mathcal{A}}_{\mathcal{P}_{1}}^{n})$ is generated by the two cocycles $$\left\{ \begin{array}{l} \Phi _{1}(\varphi _{1})=X_{3}dX_{2}, \\ \Phi _{1}(\varphi _{2})=X_{2}^{2}dX_{2}. \\ \end{array}\right.$$ Now consider the general case. Let $\mathcal{A}=\mathbb{C[}X_{1},\ldots ,X_{n}]$ be provided with the Poisson structure $\mathcal{P}.$ Let $\varphi \in \chi ^{k}(\mathcal{A_{P}})$. Then, we have $$\begin{array}{ll} \Phi _{n-k-1}(\delta ^{k}\varphi )=&\varepsilon \sum i(\partial _{\sigma (1)},\cdots ,\partial _{\sigma (k+1)})[d(i(\partial _{\sigma (k+2)},\cdots ,\partial _{\sigma (n)})\Omega )\wedge \Phi _{n-k}(\varphi )\\ \medskip\\ &+\Omega \wedge d(i(\partial _{\sigma (k+2)},\cdots ,\partial _{\sigma (n)})\Phi _{n-k}(\varphi ))], \end{array}$$ for all $\sigma \in S_{k+1,n-k-1}$, where $\varepsilon =\varepsilon (n,k)=(-1)^{\frac{(n-k)(n-k+1)}{2}}$. *Proof.* To simplify we write $d_{i}$ in place of $dX_{i}$. We have seen that for every $P\in \mathcal{A_{P}}$ we have $\delta ^{0}P=-\Omega \wedge dP.$ But $$\Phi _{n-1}(\delta P)=\displaystyle\sum_{k=1}^{n}(-1)^{k-1}\{X_{k},P\}d_{1}\wedge \cdots \wedge \hat{d_{k}}\wedge \cdots \wedge d_{n},$$where $\hat{d_{i}}$ means that this factor does not appear, with $\{P,X_{i}\}=\displaystyle\sum_{j=1}^{n}P_{ji}\partial _{j}P$ with $P_{ji}=-P_{ij}$ when $j>i$. But $$\begin{array}{l} \smallskip i(\partial_1)[\Omega \wedge d(i(\partial_2,\cdots,\partial_{n})\Phi _{n}(P))+d(i(\partial_2,\cdots,\partial_{n})\Omega )\wedge \Phi _{n}(P) \\ \smallskip \ \ \ =i(\partial_1)[\Omega \wedge d(i(\partial_2,\cdots,\partial_{n})\Phi_{n}(P)] =(-1)^{\frac{n(n-1)}{2}}i(\partial_1)[\Omega \wedge dP \wedge d_1]\\ \smallskip \ \ \ = -(-1)^{\frac{n(n-1)}{2}} \displaystyle\sum_{i=2}^nP_{1i}\partial_iPd_2\wedge\cdots\wedge d_n =(-1)^{\frac{n(n-1)}{2}}\Phi_{n-1}(P)(\partial_2,\cdots,\partial_n). \end{array}$$ Similarly $$\begin{array}{l} \smallskip i(\partial_j)[\Omega \wedge d(i(\partial_1,\cdots,\hat{\partial_j},\cdots , \partial_{n})\Phi _{n}(P))+d(i(\partial_1,\cdots,\hat{\partial_j},\cdots ,\partial_{n})\Omega )\wedge \Phi _{n}(P)] \\ \smallskip \ \ \ = i(\partial_j)[\Omega \wedge d(i(\partial_1,\cdots,\hat{\partial_j},\cdots , \partial_{n})\Phi _{n}(P))=(-1)^{j-1+\frac{n(n-1)}{2}}i(\partial_j)[\Omega \wedge dP \wedge dX_j]\\ \smallskip \ \ \ = (-1)^{j-1+\frac{n(n-1)}{2}}i(\partial_j)(\displaystyle\sum_{l=1}^{l=j-1}P_{1j}\partial_lP-\displaystyle\sum_{l=j+1}^{l=n}P_{jl}\partial_lP)d_1\wedge\cdots \wedge d_n \\ \smallskip \ \ \ = (-1)^{\frac{n(n-1)}{2}}(\sum_{l=1}^{l=j-1}P_{1j}\partial_lP-\displaystyle\sum_{l=j+1}^{l=n}P_{jl}\partial_lP)d_1\wedge\cdots \wedge \hat{d_j}\cdots\wedge d_n \\ \smallskip \ \ \ =(-1)^{\frac{n(n-1)}{2}}\{P,X_i\}d_1\wedge \cdots\wedge \hat{d_i}\wedge \cdots \wedge d_n. \end{array}$$ We deduce $$\Phi _{n-1}(\delta ^{0}P)=(-1)^{\frac{n(n-1)}{2}}\sum_{j=1}^{n}(-1)^{j-1}i(\partial _{j})[\Omega \wedge d(i(\partial _{1},\cdots ,\hat{\partial _{j}},\cdots ,\partial _{n})\Phi _{n}(P))$$which proves the theorem for $k=0$. The proof is similar for any $k$. **Application.** We consider the $n$-dimensional complex Lie algebra defined by the brackets $$\lbrack X_{1},X_{i}]=(i-1)X_{i},$$for $i=2,\cdots ,n.$ Let $\mathcal{P}_{2}$ be the corresponding Poisson bracket on $\mathbb{C}[X_{1},\cdots ,X_{n}]$. Let $\chi _{2}^{k}(\mathcal{A}_{\mathcal{P}_{2}})$ be the subspace of $\chi ^{k}(\mathcal{A}_{\mathcal{P}_{2}})$ whose elements are homogeneous of degree $2$. We denote by $H_{2}^{2}(\mathcal{A}_{\mathcal{P}_{2}})=Z_{2}^{2}/B_{2}^{2}$ the corresponding subspace of $H^{2}(\mathcal{A}_{\mathcal{P}_{2}})$. Define $N:=\frac{n(n-1)}{2}$. - If $n$ is even, then $$\dim B^2_2=N+(N-1)+\cdots+N-n/2+1=\displaystyle \frac{n(2n^2-3n+2)}{8}.$$ - If $n$ is odd, $$\dim B^2_2=N+(N-1)+\cdots+(N-(n-1)/2)= \displaystyle \frac{(n^2-1)(2n-1)}{8}.$$ In fact, if $f \in \chi^1_2(\mathcal{A}_{\mathcal{P}})$, then $f(X_i)=P_i=\Sigma a_{i_{1},\cdots,i_{n}}^{i}X_{1}^{i_{1}}X_{2}^{i_{2}}\cdots X_{n}^{i_{n}} $ is homogeneous of degree $2$, then: 1\. In $\delta f(X_1,X_{2l})$ we find $N-l$ independent coefficients of $P_{2l}$. The coefficients which do not appear are: $$a^{2l}_{1,0,0,\cdots,0,1,0,\cdots,0}, \ \ a^{2l}_{0,1,0,\cdots,0,1,0,\cdots,0},\cdots, a^{2l}_{0,0,\cdots,1,1,0,\cdots,0},$$ where the second $1$ is respectively in the place $2l,2l-1,\cdots,l+1$. 2\. In $\delta f(X_1,X_{2l+1})$ we find $N-l-1$ independent coefficients of $P_{2l+1}$. The coefficients which do not appear are: $$a^{2l+1}_{1,0,0,\cdots,0,1,0,\cdots,0}, \ \ a^{2l+1}_{0,1,0,\cdots,0,1,0,\cdots,0}, \ \ \cdots, a^{2l+1}_{0,0,..,0,2,0,\cdots,0},$$ where the second $1$ is in place $2l+1,2l,\cdots,l+2$ and in the last case the $2$ is in place $l+1$. 3\. For $i \geq 2$ and $j>i$, $\delta f(X_i,X_j)$ is defined by the $(n-2)$ coefficients $a^i_{1,0,0,\cdots,0,1,0,\cdots,0}$. Now we can to find the generators of $H_{2}^{2}(\mathcal{A}_{\mathcal{P}})$. We can choose $\phi \in \chi _{2}^{2}$ such that $$\left\{ \begin{array}{l} \phi (X_{1},X_{2})=0, \\ \phi (X_{1},X_{3})=a_{1,3}^{1,3}X_{1}X_{3}+a_{1,3}^{2,2}X_{2}^{2}, \\ \cdots \\ \phi (X_{1},X_{2l})=a_{1,2l}^{1,2l}X_{1}X_{2l}+a_{1,2l}^{2,2l-1}X_{2}X_{2l-1}+\cdots +a_{1,2l}^{l,l+1}X_{l}X_{l+1}, \\ \phi (X_{1},X_{2l+1})=a_{1,2l+1}^{1,2l+1}X_{1}X_{2l+1}+a_{1,2l+1}^{2,2l}X_{2}X_{2l}+\cdots +a_{1,2l+1}^{l,l}X_{l}^{2}, \\ \cdots \\ \phi (X_{1},X_{n})=a_{1,n}^{1,n}X_{1}X_{n}+a_{1,n}^{2,n-1}X_{2}X_{n-1}+\cdots , \\ \phi (X_{i},X_{j})=A_{i,j},\end{array}\right.$$where $A_{i,j}$ is a degree $2$ homogeneous polynomial without monomial of type $X_{1}X_{k}$ and $X_{i}X_{j}$. By solving $\Phi _{n-2}(\delta \phi )=0$ we obtain the generators of $H_{2}^{2}(\mathcal{A}_{\mathcal{P}})$. They are given by $$\left\{ \begin{array}{l} \phi (X_{1},X_{2})=0, \\ \phi (X_{1},X_{3})=a_{1,3}^{2,2}X_{2}^{2}, \\ \cdots \\ \phi (X_{1},X_{2l})=a_{1,2l}^{2,2l-1}X_{2}X_{2l-1}+\cdots +a_{1,2l}^{l,l+1}X_{l}X_{l+1}, \\ \phi (X_{1},X_{2l+1})=a_{1,2l+1}^{2,2l}X_{2}X_{2l}+\cdots +a_{1,2l+1}^{l+1,l+1}X_{l+1}^{2}, \\ \cdots \\ \phi (X_{1},X_{n})=a_{1,n}^{2,n-1}X_{2}X_{n-1}+\cdots +a_{1,n}^{m,m+1}X_{m}X_{m+1},\ \ \mbox{\rm if}\ n=2m, \\ \phi (X_{i},X_{j})=A_{i,j},\end{array}\right.$$or $\phi (X_{1},X_{n})=a_{1,n}^{2,n-1}X_{2}X_{n-1}+\cdots +a_{1,n}^{m+1,m+1}X_{m}X_{m+1},\ \ \mbox{\rm if}\ n=2m+1.$ For example: - if $n=2$ , $\dim H^2_2( \mathcal{A}_{\mathcal{P}_2}, \mathcal{A}_{\mathcal{P}_2})=1$, - if $n=3$, $\dim H^2_2( \mathcal{A}_{\mathcal{P}_2}, \mathcal{A}_{\mathcal{P}_2})=3$, - if $n=4$, $\dim H^2_2( \mathcal{A}_{\mathcal{P}_2}, \mathcal{A}_{\mathcal{P}_2})=8$, - if $n=5$, $\dim H_{2}^{2}(\mathcal{A}_{\mathcal{P}_{2}},\mathcal{A}_{\mathcal{P}_{2}})=16$. Poisson structures of degree $2$ on $\mathbb{C}[X_1,X_2,X_3]$ ============================================================= Let $\mathcal{P} $ be a Poisson structure on $\mathcal{A}^3=\mathbb{C}[X_{1},X_{2},X_{3}]$ with $P_{ij}$ of degree $2$. Then $\mathcal{P}$ writes $$\mathcal{P}=\mathcal{P}_{0}+\mathcal{P}_{1}+\mathcal{P}_{2},$$where $\mathcal{P}_{i}$ is homogeneous of degree $i$. The associated form $\Omega $ is decomposed in homogeneous parts $\Omega =\Omega _{0}+\Omega _{1}+\Omega _{2}$ and, since $d\Omega_0=0$, the condition $\Omega \wedge d\Omega=0$ is equivalent to $$\left\{ \begin{array}{l} \medskip \Omega _{2}\wedge d\Omega _{2}=0, \\ \medskip \Omega _{0}\wedge d\Omega _{1}+\Omega _{1}\wedge d\Omega _{0}=0, \\ \medskip \Omega _{0}\wedge d\Omega _{2}+\Omega _{2}\wedge d\Omega _{0}+\Omega _{1}\wedge d\Omega _{1}=0, \\ \medskip \Omega _{1}\wedge d\Omega _{2}+\Omega _{2}\wedge d\Omega _{1}=0.\end{array} \right.$$If $\Omega _2=0$, then ${\mathcal{P}}$ is a linear Poisson structure on ${\mathcal{A}}^3$ ([@JPD]). If $\Omega _2 \neq 0$ and $\Omega_0=\Omega _1=0$, then ${\mathcal{P}}$ is a quadratic homogeneous Poisson structure and the classification is given in [@Mo]. In this section we will study the remaining cases $\Omega_0 \neq 0$ or $\Omega _1\neq 0$. The associative algebra ${\mathcal{A}}^3$ admits a natural grading ${\mathcal{A}}^3= \oplus _{n\geq 0}V_n$ where $V_n$ is the space of degree $n$ homogeneous polynomial of ${\mathcal{A}}^3$. A linear isomorphism $$f:\oplus _{n\geq 0}V_{n}\rightarrow \oplus _{n\geq 0}V_{n}$$is called equivalence of order $2$ if it satisfies - $f(V_{1})\subset V_{1}\oplus V_{2},$ - $f(V_{0})=V_{0},$ - $f\mid _{\oplus _{n\geq 2}V_{n}}=Id.$ Moreover if $V_1$ is provided with a Lie algebra structure, then - $\pi_{1}\circ f$ is a Lie automorphism of $V_{1}$, where $\pi _{1}$ is the projection on $V_{1}.$ Such a map writes$$\left\{ \begin{array}{l} f(X_{i}) =\displaystyle \sum_{j=1}^n a_{i}^{j}X_{j}+\sum_{j,k=1}^n b_{i}^{jk}X_{j}X_{k}, \\ f(X_{i}X_{j}) =X_{i}X_{j}. \end{array} \right.$$ Thus, if ${\mathcal{P}}$ is a degree $2$ Poisson structure on ${\mathcal{A}}^3$, putting $Y_{i}=f(X_{i})$ and $$\{Y_{i},Y_{j}\}=f^{-1}(\{f(X_{i}),f(X_{j})\}),$$we obtain a new Poisson structure of degree $2$. These two Poisson structures are called equivalent. In the following, we classify the non homogeneous Poisson structure of degree $2$ up to an equivalence of order $2$. Note that the quadratic homogeneous Poisson structures are classified in [@Ha]. We assume also that these Poisson structures are not trivial extensions of Poisson structures on $\mathcal{A}^2$, that is, Poisson structures which do not depend only of two variables. First case: $\Omega = \Omega _2 + \Omega _1, \Omega _1 \neq 0$ -------------------------------------------------------------- The integrability condition of $\Omega$ reduces to $$\left\{ \begin{array}{l} \medskip \Omega _{1}\wedge d\Omega _{1}=0,\\ \medskip \Omega _{1}\wedge d\Omega _{2}+\Omega _{2}\wedge d\Omega _{1} =0,\\ \medskip \Omega _{2}\wedge d\Omega _{2} =0. \end{array}\right.$$ As $\Omega ^{1}\wedge d\Omega ^{1}=0$, $\Omega _1$ defines on ${\mathcal{A}}^3$ a linear Poisson structure. Then this form is isomorphic to one of the following $$\left\{ \begin{array}{l} \Omega ^{1}_1=X_{3}dX_{3}, \\ \Omega ^{2}_1=X_{2}dX_{3}+X_{3}dX_{2}+X_{1}dX_{1}, \\ \Omega ^{3}_1=X_{2}dX_{3}-\alpha X_{3}dX_{2}, \\ \Omega ^{4}_1=(X_{2}+X_{3})dX_{3}-X_{3}dX_{2}. \\ \end{array} \right.$$ Consider $ \Omega _{2}=A_{3}dX_{3}-A_{2}dX_{2}+A_{3}dX_{1} $ with$$\left\{ \begin{array}{lll} A_{1} &=&a_{1}X_{1}^{2}+a_{2}X_{2}^{2}+a_{3}X_{3}^{2}+a_{4}X_{1}X_{2}+a_{5}X_{1}X_{3}+a_{6}X_{2}X_{3}, \\ A_{2} &=&b_{1}X_{1}^{2}+b_{2}X_{2}^{2}+b_{3}X_{3}^{2}+b_{4}X_{1}X_{2}+b_{5}X_{1}X_{3}+b_{6}X_{2}X_{3}, \\ A_{3} &=&c_{1}X_{1}^{2}+c_{2}X_{2}^{2}+c_{3}X_{3}^{2}+c_{4}X_{1}X_{2}+c_{5}X_{1}X_{3}+c_{6}X_{2}X_{3}. \end{array} \right.$$ ### $d\Omega _{2}=0$ If $ \Omega _1 = \Omega _1 ^1$ or $\Omega _1^2$, then $d\Omega _1 =0$ and (3.2) is satisfied. An equivalence of order $2$ of type $Y_1=X_1,Y_2=X_2,Y_3=X_3+B$ where B is an homogeneous polynomial of degree $2$, allows to reduce the form $\Omega_2$ to a form with $A_1=0$. We obtain the following Poisson structure associated to $$\begin{array}{l} \Omega(1)= (aX_{1}^{2} -\frac{b}{2}X_{2}^{2}-2cX_{1}X_{2})dX_1 -(cX_{1}^{2}+eX_{2}^{2}+bX_{1}X_{2})dX_2+X_3dX_3 \end{array}$$ corresponding to $ \Omega _1 = \Omega _1 ^1$, and $$\begin{array}{ll} \Omega(2)= &(X_1+aX_{1}^{2}-\frac{b}{2}X_{2}^{2}-2cX_{1}X_{2})dX_1+(X_3-cX_{1}^{2}-eX_{2}^{2} -bX_{1}X_{2})dX_2\\ & + X_3dX_3 \end{array}$$ corresponding to $ \Omega _1 = \Omega _1 ^2$. If $ \Omega _1 = \Omega _1 ^3$ or $\Omega _1^4$, then $d\Omega _1 =kdX_2\wedge dX_3$ with $k \neq 0$. Then (3.2) implies $\Omega_2 \wedge dX_2\wedge dX_3=0$ that is $A_3=0$. Such a structure is a Poisson structure on $\mathcal{A}^2$. ### $d\Omega _{2}\neq 0$, $\Omega _1=\Omega_1^1$ As $d\Omega _{1}= 0$, then (3.2) is equivalent to $$\left\{ \begin{array}{l} \Omega _{1}\wedge d\Omega _{2}=0, \\ \Omega _{2}\wedge d\Omega _{2}=0. \end{array}\right.$$This implies $Pd\Omega _{2}=\Omega _{1}\wedge \Omega _{2}$ where $P$ is an homogeneous polynomial of degree $2$. The equivalence of order $2$ given by $Y_{i}=X_{i} $ for $i=1,2$ and $Y_{3}=X_{3}+B$ with $B\in V_{2}$ enables to consider $A_1=0$. In this case, $\Omega _{1}\wedge \Omega _{2}=Pd\Omega _{2}$ is equivalent to $$\left\{ \begin{array}{l} \partial _{1}A_{2}+\partial _{2}A_{3}=0, \\ P\partial _{3}A_{3}=X_{3}A_{3}, \\ P\partial _{3}A_{2}=X_{3}A_{2}.\end{array}\right.$$If $X_3$ is not a factor of $P$, then $\partial_{3}A_{2}=\alpha X_{3}$ and $\partial _{3}A_{3}=\beta X_{3}.$ If $\alpha =\beta =0,$ then $\Omega_2=0$. The case $\alpha \beta \neq 0$ reduces by a change of variables to the case $\alpha \neq 0$ and $\beta =0$, then $A_3=0$. Thus we obtain $$\Omega= X_3dX_3 - (aX_{2}^{2}+bX_{3}^{2})dX_2.$$ This structure is a trivial extension of a Poisson structure on $\mathbb{C}[X_2,X_3]$. If $P=X_{3}Q$ and $Q$ is a degree $1$ homogeneous polynomial, then $Q$ satisfies $$\left\{ \begin{array}{l} Q(\partial _{1}A_{2}+\partial _{2}A_{3})=0, \\ Q\partial _{3}A_{3}=A_{3}, \\ Q\partial _{3}A_{2}=A_{2}.\end{array}\right.$$We deduce the following structures $$\begin{array}{l} \Omega= (aX_1^2+bX_1X_3)dX_1 +X_3dX_3, \\ \Omega= (aX_1+X_3/2)^2dX_1+X_3dX_3, \\ \Omega= (aX_1X_3+bX_2X_3)dX_1+(bX_1X_3+cX_2X_3)dX_2+X_3dX_3,\\ \end{array}$$ The two first ones depends only of two variables. Then we obtain the following Poisson structure $$\Omega(3)= (aX_1X_3+bX_2X_3)dX_1+(bX_1X_3+cX_2X_3)dX_2+X_3dX_3.$$ ### $d\Omega _{2}\neq 0$, $\Omega _1=\Omega_1^2$ By an equivalence of degree $2$, we can consider that $A_3=0$. Then $Pd\Omega_2=\Omega_1 \wedge \Omega_2$ gives $$\left\{ \begin{array}{l} P\partial _{1}A_{2}=X_{1}A_{2}, \\ P\partial _{1}A_{1}=X_{1}A_{1}, \\ P(\partial _{2}A_{1}+\partial _{3}A_{2})=(A_{2}X_{2}+A_{1}X_{3}). \end{array} \right.$$ Solving these equations, we obtain: $$\begin{array}{l} \Omega(4)= X_1dX_1+(X_3-aX_1X_3)dX_2+(X_2+aX_1X_2)dX_3,\\ \Omega(5)= X_1dX_1 +(X_3-aX_1^2-2aX_2X_3)dX_2+X_2dX_3. \end{array}$$ ### $d\Omega _{2}\neq 0$, $\Omega _1=\Omega_1^3=X_{2}dX_{3}-\alpha X_{3}dX_{2}$ Assume that $\alpha \neq 0$ and $\alpha \neq -1$. The equivalence given by $Y_{2}=X_{2}+B_{2},Y_{i}=X_{i}$ for $i=1,3$ and $B_{2}\in V_{2}$ shows that the structure corresponding to $\Omega=\Omega_1$ is equivalent to a structure of degree $2$ defines as follow $$\left\{ \begin{array}{l} A_{1}=a_{2}X_{2}^{2}+a_{3}X_{3}^{2}+\frac{c_{6}}{\alpha }X_{1}X_{2}+c_{3}X_{1}X_{3}, \\ A_2=0, \\ A_{3}=c_{3}X_{3}^{2}+c_{5}X_{1}X_{3}+c_{6}X_{2}X_{3}. \end{array} \right.$$ Thus we can assume that in $\Omega _2$ we have $c_{3}=c_{5}=c_{6}=a_{2}=a_{3}=a_6=0$. The new equivalence of degree $2$ given by $Y_{3}=X_{3}+B_{3},Y_{i}=X_{i}$ for $i=1,2$ and $B_{3}\in V_{2}$ gives a Poisson structure of degree $2$ equivalent to the structure of degree $1$ with $$\left\{ \begin{array}{l} A_1=0, \\ A_{2}=b_{2}X_{2}^{2}+b_{3}X_{3}^{2}-c_{2}X_{1}X_{2}+\frac{c_{6}}{\alpha }X_{1}X_{3}, \\ A_{3}=c_{2}X_{2}^{2}+c_{4}X_{1}X_{2}+c_{6}X_{2}X_{3}. \end{array} \right.$$ Thus we can assume that $$\Omega_2=(a_{1}X_{1}^{2}+a_{4}X_{1}X_{2}+a_{5}X_{1}X_{3})dX_1+(b_{1}X_{1}^{2}+b_{4}X_{1}X_{2}+b_{5}X_{1}X_{3})dX_2+c_{1}X_{1}^{2}dX_3.$$ As $\Omega_{1}\wedge d\Omega _{2}+\Omega _{2}\wedge d\Omega _{1}=0$ we obtain the following Poisson structure $$\begin{array}{l} \Omega(6)=aX_1X_3dX_1-\alpha X_3dX_2+(X_{2}-\frac{a}{2\alpha}X_1^2)dX_3. \end{array}$$ with $\alpha \neq 0$ and $\alpha \neq -1$. If $\alpha =-1$, then $d\Omega _1 =0$ and this case has already been studied. If $\alpha =0,$  by equivalence of degree $2$ we can assume that $$\left\{ \begin{array}{l} A_1=a_1X_1^2+a_3X_3^2+a_4X_1X_2+a_5X_1X_3, \\ A_{2}=b_{1}X_{1}^{2}+b_{3}X_{3}^{2}+b_5X_1X_3, \\ A_{3}=c_{1}X_{1}^{2}+c_{3}X_3^2+c_{5}X_{1}X_{3}. \end{array} \right.$$ Then we have $A_3=0$ and the Poisson structure concerns only two variables. ### $d\Omega_2 \neq 0, \Omega_1=\Omega_1^4=(X_{2}+X_{3})dX_{3}-X_{3}dX_{2}$ By equivalence of degree $2$, we can assume that $A_{1}=0,c_{5}=0$ and $b_{4}=0. $ The equation $\Omega _{1}\wedge d\Omega_{2}+\Omega _{2}\wedge d\Omega _{1}=0$ implies that $c_{1}=c_{4}=b_{1}=0,c_{6}=-b_{5}=2c_{2}.$ The equation $\Omega _{2}\wedge d\Omega _{2}=0$ implies that $c_{2}=0$ and $b_{2}c_{3}=b_{6}c_{3}=0.$ Then we obtain the following Poisson structure: $$\Omega(7)= aX_3^2dX_1 - (X_3+bX_3^2)dX_2+(X_2+X_3)dX_3, \\$$ with $a \neq 0$. Second case: $\Omega = \Omega _2 + \Omega _1 +\Omega _0 , \Omega _0 \neq 0$ --------------------------------------------------------------------------- The form $\Omega_0 \oplus \Omega_1$ provides the vector space $V_{0}\oplus V_{1}$ with a linear Poisson structure. Then $V_{0}\oplus V_{1}$ is a Lie algebra such that $V_0$ is in the center. This implies $\Omega _{1}\wedge d\Omega _{1}=0.$ We deduce that $\Omega_0 + \Omega_1$ is equivalent to $$\left\{ \begin{array}{l} \medskip dX_{3}-X_{3}dX_{2}, \\ \medskip X_{3}dX_{3}-dX_{2}, \\ \medskip X_{2}dX_{3}+X_{3}dX_{2}+dX_{1}. \end{array} \right.$$ ### $\Omega_0 + \Omega _1=dX_{3}-X_{3}dX_{2} $ By equivalence, we can assume that $a_{3}=a_{5}=b_{3}=c_{5}=0.$ The equation $\Omega _{0}\wedge d\Omega _{2}=0$ implies $b_{4}=-2c_{2},c_{4}=-2b_{1},c_{6}=-b_{5}$, $\Omega _{1}\wedge d\Omega _{2}+\Omega _{2}\wedge d\Omega _{1}=0 $ implies that $ c_{1}=c_{2}=c_{3}=c_{4}=0,a_{1}=a_{4}=0$ and $\Omega _{2}\wedge d\Omega _{2}=0$ gives $b_{5}b_{2}=b_{5}a_{2}=b_{5}a_{6}=0.$ Thus we obtain the following Poisson structures given by $$\Omega(8)= aX_2X_3dX_1-(X_3-aX_1X_3+bX_2X_3)dX_2+dX_3 \\$$ with $a \neq 0$. ### $\Omega_0 + \Omega _1=-dX_{2}+X_{3}dX_{3} $ We can assume that $A_{2}=b_{2}X_{2}^{2}+b_{4}X_{1}X_{2}.$ As $d\Omega _{1}=0$, the system reduces to $\Omega _{0}\wedge d\Omega _{2}=\Omega _{1}\wedge d\Omega _{2}=0$. This gives $c_{4}=c_{6}=a_{4}=0$ and $b_{4}+2c_{2}=a_{5}-2c_{3}=2a_{1}-c_{5}=0.$ Thus $\Omega _{2}\wedge d\Omega_{2}=0$ is equivalent to $(2a_{2}X_{2}+a_{6}X_{3})A_{3}=0.$ We obtain the following Poisson structures $$\begin{array}{ll} \Omega(9)= &- (X_2+aX_2^2+bX_1X_2)dX_2+(1+cX_1^2+eX_3^2+fX_1X_3)dX_3 \\ &+(gX_1^2-\frac{b}{2}X_2^2 +\frac{f}{2}X_3^2+2cX_1X_3)dX_1. \\ \end{array}$$ ### $\Omega_0+\Omega _1=dX_1+X_{3}dX_{2}+X_{2}dX_{3}$ By equivalence, we can assume $b_{5}=b_{2}=a_{3}=a_{5}=c_{2}=c_{5}=0.$ As $d\Omega _{1}=0$, the equation $\Omega _{0}\wedge d\Omega _{2}=\Omega _{1}\wedge d\Omega _{2}=0$ implies that $b_{6}+2a_{2}=a_{6}+2b_{3}=a_{4}=a_{1}=b_{1}=b_{4}=c_{3}=0 $ . In this case $\Omega _{2}\wedge d\Omega _{2}=0$ is equivalent to $c_{6}(X_{2}A_{2}+X_{3}A_{1})=0.$ We obtain $$\Omega(10)= (1+aX_1^2)dX_1+X_3dX_2+X_2dX_3,$$ and $$\Omega(11)= (1+aX_1^2)dX_1+(X_3+bX_3^2+cX_2X_3)dX_2+(X_2+\frac{c}{2}X_2^2+2bX_2X_3)dX_3.$$ Poisson algebras associated to rigid Lie algebras ================================================= Rigid Lie algebras ------------------ Let us fix a basis of $\mathbb{C}^{n}.$ With respect to this basis, a multiplication $\mu $ of a $n$-dimensional complex Lie algebra is determined by its structure constants $C_{ij}^{k}$. We denote by $L_{n}$ the algebraic variety  $\mathbb{C}[C_{ij}^{k}]/I$ where $I$ is the ideal generated by the polynomials:$$\left\{ \begin{array}{l} C_{ij}^{k}+C_{ji}^{k}=0, \\ \displaystyle\sum_{l=1}^{n}C_{ij}^{l}C_{lk}^{s}+C_{jk}^{l}C_{li}^{s}+C_{ki}^{l}C_{li}^{s}=0, \end{array} \right.$$for all $1\leq i,j,k,s\leq n.$ Then every multiplication $\mu $ of a $n$-dimensional complex Lie algebra is identified to one point of $L_{n}.$ We have a natural action of the algebraic group $Gl(n,\mathbb{C)}$ on $L_{n}$ whose orbits correspond to the classes of isomorphic multiplications:$$\mathcal{O(\mu )=}\left\{ f^{-1}\circ \mu \circ (f\times f),\quad f\in Gl(n,\mathbb{C)}\right\} .$$ Let $\mathfrak{g}=(\mathbb{C }^n,\mu) $ be a $n$-dimensional complex Lie algebra. We denote also by $\mu $ the corresponding point of $L_n$. The Lie algebra $\mathfrak{g}$ is rigid if its orbit $\mathcal{O}(\mu )$ is open (for the Zariski topology) in $L_n$. Among rigid complex Lie algebras, there are all simple and semi-simple Lie algebras, all Borel algebras and parabolic Lie algebras. Concerning the classification of rigid Lie algebras, we know the classification up the dimension $8$ ([@G.A]), the classification in any dimension of solvable rigid Lie algebras whose nilradical is filiform ([@G.A]). Recall two interesting tools to study rigidity of a given Lie algebra. Let $\mathfrak{g}=(\mathbb{C}^n,\mu)$ be a $n$-dimensional complex Lie algebra. Then 1\. $\mathfrak{g}$ is rigid if and only if any valued deformation $\mathfrak{g}^{\prime }$ is ($K^*$)-isomorphic to $\mathfrak{g}$ where $K^*$ is the fraction field of the valuation ring $R$ containing the structure constants of $\mathfrak{g}^{\prime }$. 2\. (Nijenhuis-Richardson Theorem) If $H^2(\mathfrak{g},\mathfrak{g})=0$, then $\mathfrak{g}$ is rigid. The notion of valued deformation, which extends in a natural way the classical notion of Gerstenhaber deformations, is developed in [@G.R1]. In the Nijenhuis-Richardson theorem, the second cohomological space $H^2(\mathfrak{g},\mathfrak{g})$ of the Chevalley cohomology of ${\mathfrak{g}}$ is trivial. Let us recall that the converse of this theorem is not true. There exists solvable rigid Lie algebras with $H^{2}(\mathfrak{g,g})\neq 0$ (see for example [@G.A]). In this case there exists a $2$-cocycle $\varphi_1 \in H^2(\mathfrak{g},\mathfrak{g})$ which is not the first term of a valued (or formal) deformation $$\mu _{t}=\mu +\sum_{i\geq 1}t^{i}\varphi _{i}$$ of the Lie multiplication $\mu$ of ${\mathfrak{g}}$. Finite dimensional Poisson algebras whose Lie bracket is rigid -------------------------------------------------------------- We recall in this section some results of [@G.R2] which precise the structure of a finite dimensional complex Poisson algebra with rigid underlying Lie bracket. Let ${\mathcal{P}}=({\mathbb{C}}^n,{\mathcal{P}})$ be a finite dimensional complex Poisson algebra. We denote by $\{X,Y\}$ and $X \cdot Y$ the corresponding Lie bracket and associative multiplication, by ${\mathfrak{g}}_{{\mathcal{P}}}$ the Lie algebra $({\mathcal{P}},\{,\})$ and by $\mathcal{A}_{{\mathcal{P}}}$ the associative algebra $({\mathcal{P}},\cdot)$. ([@G.R2]). If the Lie algebra ${\mathfrak{g}}_{{\mathcal{P}}}$ is a simple complex Lie algebra, then the associative product is trivial that is $X\cdot Y=0$ for every $X$, $Y$ in $\mathcal{P}$. Let us assume now that ${\mathfrak{g}}_{{\mathcal{P}}}$ is a complex rigid solvable Lie algebra. Then $\mathfrak{g}$ is written: $$\mathfrak{g}=\mathfrak{t} \oplus \mathfrak{n},$$ where $\mathfrak{n}$ is the nilradical of $\mathfrak{g}$ and $\mathfrak{t}$ a maximal abelian subalgebra such that the adjoint operators $adX$ are diagonalizable for every $X \in \mathfrak{t}$. This subalgebra $\mathfrak{t}$ is usually called a Malcev torus. All these maximal torus are conjugated and their common dimension is called the rank of $\mathfrak{g}$. \[lemma2\] If there is a non-zero vector $X \in \frak{g}_{\mathcal{P}}$ such that $adX$ is diagonalizable with $0$ as a simple root, then $\mathcal{A}_\mathcal{P}.\mathcal{A}_\mathcal{P}=\{0\}$. [*Proof.*]{} Let $\left\{ e_1, ..., e_n \right\}$ be a basis of $\frak{g}_\mathcal{P}$ such that $ad e_1$ is diagonal with respect to this basis. By assumption, $\{e_1,e_i\}= \lambda_i e_i$ with $\lambda_i \neq 0$ for $i \geq 2$. Since $\{ e_1^2,e_1\}=2e_1 \cdot \{e_1,e_1\}= 0$, it follows that $e_1^2=ae_1.$ But for any $i \neq 1,$ $\{e_1^2,e_i\}=2e_1 \cdot \{e_1,e_i\}=2 \lambda_i e_1 \cdot e_i$ and $\{e_1^2,e_i\}=a\lambda_ie_i$, thus $e_1\cdot e_i=\frac{a}{2}e_i$. The associativity of the product $X\cdot Y$ implies that $(e_1 \cdot e_1)\cdot e_i=ae_1 \cdot e_i=\frac{a^2}{2}e_i =e_1 \cdot (e_1 \cdot e_i)=\frac{a^2}{4}e_i$. Therefore $a=0$ and $e_1^2=0=e_1 \cdot e_i$ for any $i.$ Finally, $0=\{e_1 \cdot e_j,e_i\}=e_1\cdot \{ e_j,e_i\}+e_j \cdot \{e_1 ,e_i\}=\lambda_i e_j \cdot e_i$, which implies $ e_i\cdot e_j=0, \ \forall i,j \geq 1.$ $\Box$ Let $\frak{g}$ be a rigid solvable Lie algebra of rank 1 with non-zero roots. Then there is only one Poisson algebra ${\mathcal{P}}$ such that $\frak{g}_\mathcal{P}=\frak{g}$. It corresponds to $$X\cdot Y = 0,$$ for any $X,Y \in {\mathcal{P}}$. [*Proof.*]{} By hypothesis we have $\dim \, \frak{t}=1$ and for $X \in \mathfrak{g}_\mathcal{P}$, $X \neq 0$, as the roots of $\frak{g}$ are non zero, the restriction of the operator $ad X$ on $\frak{n}$ is invertible (all known solvable rigid Lie algebras satisfy this hypothesis). By the previous lemma, the associated algebra $\mathcal{A}_\mathcal{P}$ satisfies $\mathcal{A}_\mathcal{P}.\mathcal{A}_\mathcal{P}= \{0\}$. \[theo22\] Let $\mathcal{P}$ a complex Poisson algebra such that $\frak{g}_\mathcal{P}$ is rigid solvable of rank 1 (i.e $\dim \, \frak{t}=1$) with non-zero roots. Then $\mathcal{P}$ is a rigid Poisson algebra. [*Proof.* ]{} See [@G.R2] Linear Poisson structures on $\mathcal{A}^{n+1}={\mathbb{C}}[X_0,X_1,\cdots,X_n]$ given by a rigid Lie bracket -------------------------------------------------------------------------------------------------------------- In this section we consider a linear Poisson bracket on ${\mathbb{C}}[X_0,\cdots,X_n]$ such that the brackets $\{X_i,X_j\}={\mathcal{P}}(X_i,X_j)$ corresponds to a solvable rigid Lie algebra ${\mathfrak{g}}$ of rank $1$. We assume that the roots (see [@G.A]) of this rigid Lie algebras are $1,\cdots,n$. In this case we have $$\left\{ \begin{array}{l} \{X_0,X_i\}=iX_i, \ \ i=1,\cdots,n \\, \{X_1,X_i\}=X_{i+1}, \ \ i=2,\cdots,n-1 \\, \{X_2,X_i\}=X_{i+2}, \ \ i=3, \cdots, n-2. \end{array} \right.$$ We denote this $(n+1)$-dimensional Poisson algebra by $\mathcal{P}({\mathfrak{g}})$. This algebra is a deformation of the Poisson algebra studied in Section 1.2. The corresponding $(n-1)-$exterior form is $$\begin{array}{ll} \Omega & =\sum\limits_{i=1}^{n}(-1)^{i-1}X_{i}d_{1}\wedge \cdots\wedge \hat{d_{i}}\wedge \cdots\wedge d_{n}+\sum\limits_{i=2}^{n-1}(-1)^{i}X_{i+1}d_{0}\wedge d_{2}\wedge \cdots\wedge \hat{d_{i}}\wedge \cdots\wedge d_{n}\\ &+\sum\limits_{i=3}^{n-2}(-1)^{i+1}X_{i+2}d_{0}\wedge d_{1}\wedge d_{3}\wedge \cdots\wedge \hat{d_{i}}\wedge \cdots\wedge d_{n}, \end{array}$$ where $d_{i}$ denotes $dX_{i}$ and $\hat{d_{i}}$ means that this term does not appear. Let $\varphi$ be a $2$-cochain. We denote by $\varphi(i,j)$ the vector $\varphi(X_i,X_j)$. Then $\varphi $ is a $2$ cocycle if and only if$$\begin{array}{lll} \Phi _{n-1}(\varphi ) & = & (-1)^{n-2}\varphi (1,i)d_{0}\wedge d_{2}\wedge \cdots\wedge \hat{d_{i}}\wedge \cdots\wedge d_{n} \\ && + \sum\limits_{i=3}^{n}(-1)^{i-1}\varphi (2,i)d_{0}\wedge d_{1}\wedge d_{3}\wedge \cdots\wedge \hat{d_{i}}\wedge \cdots\wedge d_{n} \\ && + \sum\limits_{3\leq i<j\leq n}(-1)^{j-i-1}\varphi (i,j)d_{0}\wedge \cdots\wedge \hat{d_{i}}\wedge \cdots\wedge \hat{d_{j}}\wedge \cdots\wedge d_{n}\end{array}$$satisfies$$d[i(\partial _{\sigma (1)},\cdots,\partial _{\sigma (n-2)})\Omega ]\wedge \Phi _{n-1}(\varphi )+\Omega \wedge d[i(\partial _{\sigma (1)},\cdots,\partial _{\sigma (n-2)})\Phi _{n-2}(\varphi )]=0,$$ for any $\sigma \in S_{3,n-2}.$ As ${\mathfrak{g}}=\frak{t}\oplus \frak{n}$, we have the decomposition $\mathcal{P}({\mathfrak{g}})=\mathcal{P}(\frak{t})\oplus \mathcal{P}(\frak{n})$ where $\mathcal{P}(\frak{t})$ and $\mathcal{P}(\frak{n})$) are the Poisson algebras $({\mathbb{C}}[X_0],{\mathcal{P}})$ and $({\mathbb{C}}[X_{1},\cdots,X_{n}],{\mathcal{P}})$. From the Hochschild-Serre factorization theorem, we assume that the cocycles are $\frak{t}$-invariant and with values in $\mathcal{P}(\frak{n})$. We denote this space by $\chi ^{k}(\mathcal{P}(\frak{g}),\mathcal{P}(\frak{g}))^{\frak{t}}.$ If $\ f\in \chi ^{1}(\mathcal{P}(\frak{g}),\mathcal{P}(\frak{g}))^{\frak{t}}$ then $$\{X_{0},f(X_{i})\}=if(X_{i}),$$ and we obtain $$f(X_{1})=a_{1}^{1}X_{1},f(X_{2})=a_{1}^{11}X_{1}^{2}+a_{2}^{2}X_{2},\cdots,f(X_{i})=\sum\limits_{l_{1}+\cdots+l_{k}=i}a_{i}^{l_{1} \cdots l_{k}}X_{1}^{l_{1}}\cdots X_{k}^{l_{k}}.$$ Thus $\delta f(X_{1},X_{i})=a_{1}^{1}\{X_{1},X_{i}\}+\{X_{1},f(X_{i})\}-f(X_{i+1})$ and we can reduce any element $\varphi \in Z^{2}(\mathcal{P}(\frak{g}),\mathcal{P}(\frak{g}))^{\frak{t}}$ to a $2$-cocycle satisfying$$\varphi (X_{1},X_{i})=0\text{ for }i=2,\cdots,n-1.$$ We denote by $Z_{k}^{\ast }(\mathcal{P}(\frak{g}),\mathcal{P}(\frak{g}))^{\frak{t}}$ the subspace of homogeneous cocycles of degree $k.$ Let us look the system on the $\varphi(i,j)$ which is deduced from Equation(4.1). - If ($\sigma (1),\cdots,\sigma (n-2))=(3,4,\cdots,n)$ then Condition (4.1) is trivial. - If ($\sigma (1),\cdots,\sigma (n-2))=(2,3,\cdots,\hat l,\cdots,n)$ then Condition (4.1) is trivial as soon as $l \neq n$. If $l=n$ we obtain $$n\varphi (X_{1},X_{n})+(-1)^{n-1}\sum iX_{i}\partial _{i}\varphi(X_{1},X_{n})=0$$ and $\varphi (X_{1},X_{n})$ is of weight $n+1.$ - If ($\sigma (1),\cdots,\sigma (n-2))=(1,2,\cdots,\hat i,\cdots,\hat j,\cdots,n)$ we obtain $$(i+j)\varphi (X_{i},X_{j})=\sum kX_{k}\partial _{k}\varphi (X_{i},X_{j})$$ and $\varphi (X_{i},X_{j})$ is of weight $i+j.$ Other relations show that the space of cocycles of degree $2$ is generated by $\varphi(X_1,X_n)$ and $\varphi(X_2,X_{2k+1})$ with $k=1,\cdots,l$ where $n=2l+1$ or $n=2l$. The relations between these generators leads to study two cases: $k=1$ and $k=2.$ #### Case $k=1$. As $\varphi (X_{1},X_{n})$ is of weight $n+1$, then $\varphi (X_{1},X_{n})=0.$ We have also $\varphi (X_{i},X_{j})=a_{ij}^{i+j}X_{i+j}$ if $i+j\leq n.$ If ($\sigma (1),\cdots,\sigma (n-2))=(1,2,\cdots,\hat i,\cdots,\hat j,\cdots,n)$ we obtain $$(i+j)\varphi (X_{i},X_{j})=\sum kX_{k}\partial _{k}\varphi (X_{i},X_{j})$$ and $\varphi (X_{i},X_{j})$ is of weight $i+1.$ Then $$\varphi (X_{i},X_{j})=a_{ij}^{i+j}X_{i+j}$$ if $i+j\leq n.$ If ($\sigma (1),\cdots,\sigma (n-2))=(0, 1,2,\cdots,\hat i,\cdots,\hat j,\cdots,\hat k,\cdots,n)$ with $i\geq 3$, then the related conditions are always satisfied. If ($\sigma (1),\cdots,\sigma (n-2))=(0,3,\cdots, \hat i,\cdots, n)$, $i \geq 3$, we obtain relation between $\varphi (3,l)$ and $\varphi (2,l+1).$ We deduce that $$a_{3,i}=-a_{2,i+1} + a_{2,i}$$ and $a_{2,3}=a_{2,4}$. If ($\sigma (1),\cdots,\sigma (n-2))=(0, 1,2,\cdots,\hat i,\cdots,\hat j,\cdots,\hat k,\cdots,n)$ with $i\geq 3$ then $$a_{4,i}=a_{2,i+2}-2a_{2,i+1}+a_{2,i}$$ and $$a_{3,4}=a_{3,5}.$$ If ($\sigma (1),\cdots,\sigma (n-2))=(0,2,3,\cdots, \hat i,\cdots, \hat j,\cdots, n)$, $i \geq 4$, then we have $$a_{i+1,j}=-a_{i,j+1}+a_{i,j}$$ and $$a_{i,i+2}=a_{i,i+1}.$$ If ($\sigma (1),\cdots,\sigma (n-2))=(0,1,3,\cdots, \hat i,\cdots, \hat j,\cdots, n)$, $i \geq 4$, then we have $$a_{i+2,j}=-a_{i,j+2}+a_{i,j}$$ and $$a_{3,j}=a_{2,j}-a{2,j+1}.$$ If we solve this linear system, we obtain If $n\geq 7$, then $H_{1}^{2}(A_{p},A_{p})$ is of dimension $1$ and generated by the cocycle given by$$\left\{ \begin{array}{l} \varphi (X_{2},X_{i})=(4-i)X_{2+i}\text{ \ }i=5,\cdots,n-2, \\ \varphi (X_{3},X_{i})=X_{3+i}\text{ \ }i=4,\cdots,n-3, \\ \varphi (X_{i},X_{j})=0\text{ \ in other cases}.\end{array}\right.$$ #### Case $k=2$. The set of generators is of dimension $\frac{p^2+5p}{2}$ if $n=2p+1$ and $\frac{p^2+3p-2}{2}$ if $n=2p$. The number of independent relations concerning these parameters is greater than the dimension of the set of generators as soon as $n \geq 6$. For $n=5$ the dimension is equal to $2$ and for $n=6$, this dimension is $0$. We deduce that $\dim H^2_2=0$ when $n \geq 7$. [**Remark: Deformations of the Enveloping algebra of a rigid Lie algebra**]{} Let $\mathfrak{g}$ be a finite dimensional complex Lie algebra. We denote by $\mathcal{U}(\mathfrak{g})$ its enveloping algebra. One of the most important problem in this time is to look the deformations of the associative algebra $\mathcal{U}(\mathfrak{g})$. The theory of quantum groups comes from the deformation of $\mathcal{U}(sl(2))$. In this case, $\mathfrak{g}=sl(2)$ is a rigid Lie algebra and $\mathcal{U}(sl(2))$ is a rigid associative algebra. Thus we have to look what happens for any rigid Lie algebra. The aim of this section is to study the deformations of $\mathcal{U}(\mathfrak{g})$ when $\mathfrak{g}$ is the rigid Lie algebras studied in the previous paragraph. We denote by $S(\mathfrak{g)}$ the symmetric algebra on the vector space $\mathfrak{g.}$ This associative commutative algebra is interpreted as the algebra of polynomials on the dual vector space $\mathfrak{g}^{\ast }$ of $\mathfrak{g}$ that is $\mathbb{C}[\alpha _{1},\cdots ,\alpha _{n}]$ where $\{\alpha _{1},\cdots ,\alpha _{n}\}$ is a basis of $\mathfrak{g}^{\ast }$. But the Lie structure of $\mathfrak{g}$ induces a Linear Poisson structure (or of degree $1$), $\mathcal{P}$, on $\mathfrak{g}^{\ast }.$ In fact, if $\{X_{1},\cdots ,X_{n}\}$ is the (dual) basis of $\mathfrak{g}$, this Poisson structure corresponds to the Poisson structure of degree $1$ on $\mathbb{C}[X_{1},\ldots ,X_{n}]$ associated to $\mathfrak{g.}$ From the formality theorem of Kontsevich, $\mathcal{U}(\mathfrak{g})$ is a deformation of the Poisson algebra $(\mathbb{C}[X_{1},\ldots ,X_{n}],\mathcal{P)}.$ In his thesis, Toukaidine Petit ([@Pe]) shows that every nontrival deformation of the Poisson structure $\mathcal{P}$ on $\mathbb{C}[X_{1},\ldots ,X_{n}]$ induces a nontrivial deformation of the associative algebra $\mathcal{U}(\mathfrak{g}).$ As a consequence, we have that if $\mathfrak{g}$ is a nonrigid Lie algebra, then there is a nontrivial deformation of $\mathcal{U}(\mathfrak{g}).$ If we consider the rigid Lie algebra $\frak{g}_{n+1}$ studied in the previous paragraph, we have determinate a non trivial cocycle of degree one for the corresponding Poisson algebra which is not integrable. Thus we cannot define a deformation of its enveloping algebra. But the Lie algebra $\frak{g}_{n+1}$ admit a deformation in the following nonLie algebra which is written $$\left\{ \begin{array}{l} \mu (X_{0},X_{i})=iX_{i}\text{ \ }i=1,\cdots,n, \\ \mu (X_{1},X_{i})=X_{i+1}\text{ \ }i=2,\cdots,n-1, \\ \mu (X_{2},X_{3})=X_{5}, \\ \mu (X_{2},X_{i})=(5-i)X_{2+i}\text{ \ }i=4,\cdots,n-2, \\ \mu (X_{3},X_{i})=X_{3+i}\text{ \ \ }i=4,\cdots,n-3.\end{array}\right.$$ [9]{} Dufour Jean-Paul. Formes normales de structures de Poisson. Symplectic geometry and mathematical physics (Aix-en-Provence, 1990), 129–135, Progr. Math., **99**, Birkhäuser Boston, Boston, MA, 1991. Goze Michel, Ancochea Bermudez Jose Maria. On the classification of rigid Lie algebras. *J. Algebra*, **245**, (2001), 68-91. Goze Michel. Algèbres de Lie. Classifications, Déformations et Rigidité, Géométrie différentielle. In [*Algèbre, dynamique et analyse pour la géométrie: aspects récents*]{}. Editions Ellipse, (2009), 39-99. Goze Michel, Remm Elisabeth. Valued deformations of algebras. *J. Algebra Appl.*, **3**, (2004), no. 4, 345–365. Goze Michel, Remm Elisabeth. Poisson algebras in terms of non-associative algebras. *J. Algebra*, **320**, (2008), no. 1, 294–317. Haraki Abdeljalil. Quadratisation de certaines structures de Poisson. J. London Math. Soc. (2) **56** (1997), no. 2, 384–394. Lichnerowicz André. Les variétés de Poisson et leurs algèbres de Lie associées. J. Differential Geometry, **12**(2):253–300, 1977. Markl Martin, Remm Elisabeth, Algebras with one operation including Poisson and other Lie-admissible algebras, *J. Algebra*, **299**, (2006), 171-189. Monnier Philippe. Formal Poisson cohomology of quadratic Poisson structures. *Lett. Math. Phys*. **59**, (2002), no. 3, 253–267. Petit Toukaidine. Sur les algèbres enveloppantes des algèbres de Lie rigides. Thèse Université de Mulhouse, (2001). Pichereau Anne. Poisson (co)homology and isolated singularities. *J. Algebra*. **299**, (2006), no. 2, 747–777. Remm Elisabeth. Opérades Lie-admissibles. *C. R. Math. Acad. Sci. Paris*. **334**, (2002)

ArXiv

--- author: - | [^1],  Pedro Carrasco Millán, Ana Porras Riojano and Esteban M. Sánchez García\ Departamento de Fisica Teorica I, Plaza de las Ciencias 1, Fac. CC. Fisicas; Universidad Complutense de Madrid, 28040, Madrid, Spain.\ E-mail: - | M. Ángeles García Ferrero\ Instituto de Ciencias Matemáticas, C/ Nicolás Cabrera, 13-15, 28049 Madrid, Spain title: Shannon entropy and hadronic decays --- In this note we deploy Shannon’s entropy [@Shannon:1948zz], a concept belonging to the field of information theory (as a measure of the uncertainty associated with a random variable, or on ignoring the value that this takes, of the average missing information content) to analyze particle decay distributions. This is granted because the sum of branching ratios $BR_i=\Gamma_i/\Gamma$, $i=1\dots N$ is a probability distribution since $\sum_i \Gamma_i=\Gamma$ so that $\sum_i BR_i= 1$. We exemplify with actual data from meson and gauge boson decays, all taken from [@Patrignani:2016xqp]. Often part of the decaying particle width is unaccounted for the known decay channels. We may then assign that unknown width to one last channel, and Shannon’s entropy for the decay reads $$S(i)=-\sum_{f_{\rm known}}BR_{(i\longrightarrow f)}\log_{k}BR_{(i\longrightarrow f)} -(1-\sum_{f_{\rm known}}BR_{(i\longrightarrow f)})\log_{k}(1-\sum_{f_{\rm known}}BR_{(i\longrightarrow f)}) \ .$$ A more sophisticated treatment could go along the lines of the coarse-graining methods of [@Alonso-Serrano:2017poc]. Figure \[fig:K2\] represents the entropy accrued, from left to right, upon adding each decay channel of the $K_2^*(1430)$ meson. ![Entropy of the decay distribution of the strange $K_{2}^{*}(1430)$ meson as function of the number of included channels (left) or the accumulated fraction of the partial width after accounting for each channel in order of decreasing branching fraction. The error bars in this and the following figures are propagated from the PDG uncertainties in the measurement of $\Gamma_i/\Gamma$.\[fig:K2\]](FIGS1.DIR/k2channels.pdf "fig:"){width="45.00000%"}   ![Entropy of the decay distribution of the strange $K_{2}^{*}(1430)$ meson as function of the number of included channels (left) or the accumulated fraction of the partial width after accounting for each channel in order of decreasing branching fraction. The error bars in this and the following figures are propagated from the PDG uncertainties in the measurement of $\Gamma_i/\Gamma$.\[fig:K2\]](FIGS1.DIR/k2branching.pdf "fig:"){width="45.00000%"}\ The entropy saturates after the channels with largest branching fraction have been incorporated, indicating that additional channels, of small branching fraction, barely add any information to the decay distribution. Figure \[fig:upsilon\] shows the effect of including or not the $\gamma$-channels of the $\Upsilon(3S)$ decays. It is immediately visible that increasing the number of channels increases the entropy. In fact, the *maximum possible* entropy of a decay distribution with $N$ channels is $\log N$ [@Inprep]. ![Shannon entropy of the decay distribution of $\Upsilon (3s)$. We compare the result of including all channels with a calculation that ignores radiative ones (denoted NRC, Non-Radiative Channels).\[fig:upsilon\]](FIGS1.DIR/rsrup.png){width="45.00000%"} Additionally, we find that a determining parameter that influences the total entropy of a decay distribution is the largest branching fraction of all possible decay channels. This strong correlation is shown in figure \[fig:correlation\]. ![This scatter plot, where each point coresponds to a light meson, shows a clear correlation between the entropy $S$ for a decay distribution and the maximum branching fraction among the channels through which each of the particles can decay. \[fig:correlation\]](FIGS1.DIR/unflavoredSvsmaxBR.jpg){width="47.00000%"} An advantage of employing the entropy with its logarithm of the branching fraction is its *additivity* property upon subdividing a channel into subchannels. For example, one can think of grouping all channels including a pion into a semiinclusive channel, and later examine them exclusively one by one. Then, $S({\rm both}) = w S(1) + (1-w) S(2)$ where $w$ is the weight of channel 1, that is, $w=BR(1)/(BR(1)+BR(2))$ in terms of the branching ratios. The smallest conceivable subdivision of the decay tree is at the level of individual final quantum states. A practical way of counting these is with the two-body phase space, and figure \[fig:W\] shows the entropy against phase space for all two-body and quasi-two-body decays of the electroweak $W$ boson. ![Entropy against the phase space $\sum_i \rho_i$ accrued upon including further two–body decay channels of the electroweak $W$ boson, taken from larger to smaller branching fraction. \[fig:W\]](FIGS1.DIR/W.png){width="45.00000%"} We wish to propose a simple criterion to quantify what information the discovery of a new branching fraction provides. The first obvious effect is that, since the maximum entropy grows as the logarithm of the number of channels, if all were equally weighted, the actual importance of a new channel can be obtained by studying the separation of the entropy from this maximum value. Therefore, we propose two measures of this added information. One is the normalized entropy increment, that is plotted in figure \[fig:normincrement\], as defined by $$\label{normincrement} \frac{\Delta S(N)}{\Delta log(N)}=\frac{S(N+1)-S(N)}{log(N+1)-log(N)}\ .$$ ![ [**Left**]{}: Increase of the normalized entropy for the decay distribution of the $W$ boson as function of the number of channels. Additional ones can be seen to contribute decreasingly less, so this quantity could be taken as a measure of the amount of information contained in each newly reported channel. [**Right**]{}: Change in the degree of likeness $\Theta$ upon discovering each new channel (so that the last, unknown channel, splits off part of its probability to that new one, and $N$ increases by one unit) for the $W$ boson. []{data-label="fig:normincrement"}](FIGS1.DIR/discretw.png "fig:"){width="0.45\columnwidth"}   ![ [**Left**]{}: Increase of the normalized entropy for the decay distribution of the $W$ boson as function of the number of channels. Additional ones can be seen to contribute decreasingly less, so this quantity could be taken as a measure of the amount of information contained in each newly reported channel. [**Right**]{}: Change in the degree of likeness $\Theta$ upon discovering each new channel (so that the last, unknown channel, splits off part of its probability to that new one, and $N$ increases by one unit) for the $W$ boson. []{data-label="fig:normincrement"}](FIGS1.DIR/gradimpw.png "fig:"){width="0.45\columnwidth"} Another possibility is to employ a certain “degree of likeness” which can be simply given by $\frac{S(N)}{log(N)}\in (0,1)$. Its increment upon adding one new channel would then be $$\Theta = \frac{S(N+1)}{log(N+1)}-\frac{S(N)}{log(N)}\ .$$ A positive $\Theta$ says that the entropy of distribution gets closer to the maximum possible upon introducing the new channel; therefore, this new channel has a branching fraction similar to the others. If $\Theta$ is negative, the entropy decreases relative to its maximum possible value, and the new channel is very dissimilar from the others. This function is also plotted in figure \[fig:normincrement\]. In conclusion, Shannon’s entropy is a promising way of assessing and quantifying the information gained upon discovering a new decay channel of a particle. [99]{} C. E. Shannon, Bell Syst. Tech. J.  [**27**]{} (1948) 379 \[Bell Syst. Tech. J.  [**27**]{} (1948) 623\]; Michael A. Nielsen and Isaac L. Chuang, *Quantum Computation and Quantum Information*, Cambridge Univ. Press, UK, 2000. C. Patrignani [*et al.*]{} \[Particle Data Group\], Chin. Phys. C [**40**]{} (2016) no.10, 100001. doi:10.1088/1674-1137/40/10/100001 A. Alonso-Serrano and M. Visser, Entropy [**19**]{} (2017) no.5, 207 doi:10.3390/e19050207 \[arXiv:1704.00237 \[quant-ph\]\]. P. Carrasco *et al.*, work in preparation, will give further detail. [^1]: Work supported by grants from MINECO FPA2011-27853-C02-01 and FPA2016-75654-C2-1-P, and carried out in the inspiring atmosphere of the theoretical physics department and UPARCOS.

ArXiv

--- abstract: 'In the quasi-stationary states of the Hamiltonian Mean-Field model, we numerically compute correlation functions of momenta and diffusion of angles with homogeneous initial conditions. This is an example, in a N-body Hamiltonian system, of anomalous transport properties characterized by non exponential relaxations and long-range temporal correlations. Kinetic theory predicts a striking transition between weak anomalous diffusion and strong anomalous diffusion. The numerical results are in excellent agreement with the quantitative predictions of the anomalous transport exponents. Noteworthy, also at statistical equilibrium, the system exhibits long-range temporal correlations: the correlation function is inversely proportional to time with a logarithmic correction instead of the usually expected exponential decay, leading to weak anomalous transport properties.' author: - 'Yoshiyuki Y. Yamaguchi$^{1}$[^1], Freddy Bouchet$^{2}$[^2], Thierry Dauxois$^3$[^3]' title: Algebraic Correlation Function and Anomalous Diffusion in the HMF model --- Introduction ============ Recently, a new light was shed on long-range interacting systems [@Les_Houches]. The first reason is that a mathematical characterization [@ellis_turkington] and the study of several simple models have completely clarified the inequivalence of ensembles that might exists between the microcanonical and the canonical ensembles [@barre01; @hugo]. The second is the appearance of a very useful technique, namely the large deviation theory, to compute the microcanonical number of microstates and thus the associated microcanonical entropy [@barre-05]. The third is a classification of all possible situations of ensemble inequivalence [@Classification]. The last, but not the least, reason is the understanding that the broad spectrum of applications (self-gravitating [@Chavanis_96] and Coulomb systems, vortices in two-dimensional fluid mechanics, wave-particles interaction, trapped charged particles, ...) [@Les_Houches] should be considered simultaneously since significant advances were performed independently in the different domains. However as usual in Physics, the study of simple models is of particular interest not only for pedagogical properties, but also for testing ideas that might be derived analytically and verified numerically without very expensive simulations. We consider here the Hamiltonian Mean Field (HMF) model, which is considered as the paradigmatic dynamical model for long-range interacting systems. This model [@zaslavsky_HMF; @konishi-kaneko; @pichon; @antoni-95] consists of $N$ particles moving on the unit circle, and is described by the Hamiltonian $$\label{eq:hamiltonian} H = \dfrac{1}{2} \sum_{j=1}^{N} p_{j}^{2} + \dfrac{1}{2N} \sum_{j=1}^{N} \sum_{k=1}^{N} [ 1-\cos(\theta_{j}-\theta_{k}) ],$$ where $\theta_{j}$ is the angle of $j$-th particle and $p_{j}$ its conjugate momentum. Using a change of the time unit, the prefactor $1/N$ of the second term is added in order to get an extensive energy [@barre-05]. Thus, in the limit $N\rightarrow \infty$, the appropriate mean-field scaling is obtained for the statistical mechanics. Studies of the HMF model have been recently reinforced by the discovery of its link with the Colson-Bonifacio model for the single-pass free electron laser [@barre-05]. Within this model, a striking disagreement was reported between the canonical statistical mechanics predictions and time averages of constant energy molecular dynamics simulations [@antoni-95; @lrt2001]. As the model has only a second order phase transition [@barre-05] at the critical energy density $U_{c}=3/4$, the possibility that the origin might lead to an inequivalence between canonical and microcanonical statistical mechanics can be excluded [@Classification]. Moreover, recently, it has been shown unambiguously that the microcanonical entropy leads to the same predictions than the canonical free energy [@barre-05]. Very interesting results about the behavior of such a system in contact with a thermal bath has however been recently reported [@fulvio1; @fulvio2]. The origin of the apparent disagreement comes from a particularly slow dynamical evolution of this long-range system. Indeed, in Hamiltonian systems with long-range interactions, systems are sometimes trapped in quasi-stationary states (QSS) before going to equilibrium. Examples of such QSS were found in a 1-dimensional self-gravitating system [@tsuchiya-96] and in the HMF model [@yamaguchi-03]. The trapping time diverges algebraically in the limit $N\to\infty$ and, hence, time averages disagree with canonical averages if the computing time is not long enough. To understand the dynamics during such a long period, QSS were interpreted as stable stationary states of the Vlasov equation [@yamaguchi-04; @barre-06] that can be derived from the Hamiltonian dynamics. The Vlasov equation, which governs 1-particle distribution function is indeed exact [@braunhepp] in the limit $N\to\infty$, but only approximate for a finite system: finite size effects drive indeed the system from the Vlasov stable stationary state to the Boltzmann equilibrium. Recently, Caglioti and Rousset [@caglioti-04] proved for a wide class of potentials which includes the HMF case, that $N$ particles starting close to a Vlasov stable stationary state remain close to it during a time scale proportional at least to $N^{1/8}$. The result is consistent with numerical results which state that the lifetime of QSS scales like $N^{1.7}$ [@yamaguchi-03]. Using a kinetic approach which goes beyond the above Vlasov interpretation, the correlation function of momenta was recently derived [@Bouchet-Dauxoisa; @Bouchet-Dauxois] with the following assumptions: (i) a finite but large enough number of particles, (ii) a homogeneous distribution of angles, and (iii) a system in a (quasi-)stationary state. As shown in Refs. [@yamaguchi-04; @barre-06], the latter condition amounts to consider initial distributions of momenta $f_{0}(p)$ satisfying the inequality $$\label{eq:stability-condition} 1 + \dfrac{1}{2} \int_{-\infty}^{+\infty} \dfrac{f'_{0}(p)}{p} \mbox{d}p > 0.$$ This condition has been derived for the linear [@Inagaki] and formal [@yamaguchi-04] stability of the distribution $f_0$ (see also Ref. [@chavanisvattevillebouchet] for another derivation). This condition defines a critical energy $U_{c}^{\ast}$ which is, in general, different from the critical energy $U_{c}=3/4$ where the second order phase transition is located. However, as expected, both values coincide for a gaussian distribution $f_{0}(p)$. Above theory is expected to be valid in the time interval $1\ll \tau\ll N$, where $\tau=t/N$ is the appropriate rescaled time. Among the main predictions resumed in Table \[tab:Cp\], one might emphasize that distributions $f_0(p)$ with algebraic tails were proved to have a correlation function of momenta $C_{p}(\tau)$ with an algebraic decay in the long-time regime. Striking algebraic large time behaviors for momentum autocorrelations had been first numerically observed in Refs. [@lrt2001; @Pluchino]. On the contrary, distributions with stretched exponential tails correspond to correlation functions inversely proportional to time with a logarithmic correction. It is also important to stress that gaussian distributions, which corresponds to $\delta=2$ in the stretched exponential case, leads to a long-time correlation of $\ln\tau/\tau$ instead of the usual exponential decay in the stable, supercritical energy regime $U>U_{c}^{\ast}=U_{c}$, although the initial distribution is at equilibrium. The origin of this unusual long-time momentum correlations does not depend on the center part of the momentum distributions $f_{0}(p)$ but on its tails. One might understand physically this behavior, by the fact that particles located in these tails move almost freely, and hence yield long-time correlations. ----------------------- --------------------------------- -------------------------------------- ------------------------------- Tails $f_{0}(p)$ $C_{p}(\tau)$ $\sigma_{\theta}^{2}(\tau)$ Power-law $|p|^{-\nu}$ $\tau^{-\alpha}$ $\tau^{2-\alpha}$ Stretched exponential $\quad\exp(-\beta|p|^{\delta})$ $\dfrac{(\ln\tau)^{2/\delta}}{\tau}$ $\tau (\ln\tau)^{2/\delta+1}$ ----------------------- --------------------------------- -------------------------------------- ------------------------------- : Asymptotic forms of initial distributions $f_{0}(p)$, and theoretical predictions of correlation functions $C_{p}(\tau)$ and the diffusion $\sigma_{\theta}^{2}(\tau)$ in the long-time regime. Asymptotic forms of the distribution and the predictions are assumed and predicted in the limits $|p|\to\infty$ and $\tau\to\infty$ respectively, where $\tau=t/N$ is a rescaled time. The exponent $\alpha$ is given as $\alpha=(\nu-3)/(\nu+2)$. See Ref. [@Bouchet-Dauxois] for details.[]{data-label="tab:Cp"} In these (quasi-)stationary states, the theoretical law for the diffusion of angles $\sigma_{\theta}^{2}(\tau)$ has been also derived. The predictions [@Freddy_PRE; @Bouchet-Dauxois] for the diffusion properties are listed in Table \[tab:Cp\]. They clarify the highly debated disagreement between different numerical simulations reporting either anomalous [@latora-99] or normal [@yamaguchi-03] diffusion, in particular by delineating the time regime for which such anomalous behavior should occur. We briefly recall that when the moment of order $n$ of the distribution scales like $\tau^{n/2}$ at large time, such a transport is called [*normal*]{}. However, [*anomalous*]{} transport [@bouchaud; @majda; @Castiglione], where moments do not scale as in the diffusive case, were reported in some stochastic models, in continuous time random walks (Levy walks), and for systems with a lack of stationarity of the corresponding stochastic process [@Fermi]. When the distribution $f_0(p)$ is changed within the HMF model, a transition between weak anomalous diffusion (normal diffusion with logarithmic corrections) and strong anomalous diffusion is thus predicted. From the physical point of view, as particles with large momentum $p$ fly very fast in comparison to the typical time scales of the fluctuations of the potential, they are subjected to a very weak diffusion and thus maintain their large momentum during a very long time. A thick distribution of waiting time with a large momentum explains the anomalous diffusion. From a mathematical point of view, these behaviors are linked to the non exponential relaxation of the Fokker-Planck equation describing the diffusion of momenta, leading to long-range temporal correlations [@Bouchet-Dauxois]. This mechanism is new in the context of kinetic theory. However, similar Fokker-Planck equations, with a rapidly vanishing diffusion coefficients obtained by other physical mechanisms, have been studied in several frameworks [@farago; @micciche; @lutz]. The first purpose of this article is to numerically check the theoretical predicted correlation functions for power-tail and gaussian distributions by using accurate numerical simulations. The other is to clarify whether diffusion is normal or anomalous, which depends on the choice of the initial distribution $f_0(p)$. The article is organized as follows. Some useful quantities are introduced in Section \[sec:quantities\]. In Sections \[sec:power-tail\] and \[sec:gaussian\], we respectively focus on initial distributions with power-law and gaussian tails. In each section, we first check the stationarity and the stability following the method developed in Ref. [@yamaguchi-04] and determine the time region of the QSS. We also study carefully the correlation function and the diffusion comparing them with theoretical predictions. Finally, section \[sec:summary\] concludes the discussion. Quantities of interest and numerical protocol {#sec:quantities} ============================================= In order to check the stationarity and the stability of an initial distribution $f_{0}(p)$, we study the temporal evolutions of several macrovariables: - the magnetization defined as the modulus $M$ of the vector ${\bf M}=(M_{x},M_{y})$, where both components are defined as $ M_{x} = {\left\langle \cos\theta \right\rangle}_{N}$ and $ M_{y} = {\left\langle \sin\theta \right\rangle}_{N} $. The bracket ${\left\langle \cdot \right\rangle}_{N}$ represents the average over all particles, for instance ${\left\langle \cos\theta \right\rangle}_{N} =( \sum_{j=1}^{N} \cos\theta_{j})/N$. Note that the magnetization $M$ is constant if the system is stable stationary. - the moments of the $1$-body distribution function $f(\theta,p,t)$. As explained in details in Ref. [@yamaguchi-04], the stationarity of the $1$-body distribution $f(\theta,p,t)$ implies the stationarity of the individual energy distribution $f_{e}(e,t)$, where $e=p^{2}/2-M_{x}\cos\theta-M_{y}\sin\theta$. Moreover, the stationarity of $f_{e}(e,t)$ implies the stationarity of all moments $\mu_{n}={\left\langle e^{n} \right\rangle}_{N}$. As the stationarity of the moment is a necessary condition for stability, vanishing derivatives $\dot{\mu}_{n}=\mbox{d}\mu_{n}/\mbox{d}t$, for $n=1,2$ and $3$, would suggest that the system is in a (quasi-)stationary state, while large derivatives clearly indicate a non-stationary state. In addition, the stability is suggested if a state stays stationary for a long period. While checking the stationarity and the stability, we identify a time region where the system is in the QSS, during which we observe the correlation function of momenta $C_{p}(\tau)={\left\langle p(\tau)p(0) \right\rangle}_{N} $ and the diffusion of angles $ \sigma_{\theta}^{2}(\tau)={\left\langle [\theta(\tau)-\theta(0)]^{2} \right\rangle}_{N}. $ The latter quantity can be rewritten as follows $$\dfrac{\sigma_{\theta}^{2}(\tau)}{N^{2}} = \int_{0}^{\tau} \mbox{d}\tau_{1} \int_{0}^{\tau} \mbox{d}\tau_{2} {\left\langle p(\tau_{1})p(\tau_{2}) \right\rangle}_{N} = 2 \int_{0}^{\tau} \mbox{d}s \int_{0}^{\tau-s} \mbox{d}\tau_{2} {\left\langle p(s+\tau_{2})p(\tau_{2}) \right\rangle}_{N},$$ where the factor $1/N^{2}$ comes from the time rescaling $\tau=t/N$, while the new variable $s=\tau_{1}-\tau_{2}$ was introduced to take advantage of the division of the square domain into two isoscale triangles corresponding to $s>0$ and $s<0$. In the (quasi-)stationary states, the integrand ${\left\langle p(s+\tau_{2})p(\tau_{2}) \right\rangle}_{N}$ does not depend on $\tau_{2}$ (the QSS evolve on a time scale much larger than $N$) and hence diffusion can be simplified [@yamaguchi-03] by using the correlation function as $$\label{eq:sigma_Cp} \dfrac{\sigma^{2}_{\theta}(\tau)}{N^{2}} = 2 \int_{0}^{\tau} (\tau-s) C_{p}(s)~\mbox{d}s.$$ We numerically performed the temporal evolution of the canonical equations of motion by using a $4$-th order symplectic integrator [@yoshida-90; @yoshida-93] with a time step $\Delta t=0.2$ and a total momentum set to zero. Initial values of angles being randomly chosen from a homogeneous distribution, the magnetization $M$ is hence of order $1/\sqrt{N}$. Omitting this vanishing value of $M$, the energy density $U= K + ({1-M^{2}})/{2}$ where $K$ is the kinetic energy density can thus be well approximated by the kinetic energy density $K$ as $U=K+1/2$. To characterize the simulations, the only remaining point is the initial distribution of momenta: in the following sections, as anticipated, we will carefully study distributions with power-law and gaussian tails. Power-law tails {#sec:power-tail} =============== Initial distribution -------------------- In this section, we consider the initial distribution $$f_{0}(p) = \dfrac{A_{\nu}}{1+|p/p_{0}|^{\nu}},$$ whose power-law tails are characterized by the exponent $\nu$. The unity, added in the denominator to avoid the divergence at the origin $p=0$, does not affect neither the asymptotic form, nor the theoretical predictions. The parameter $p_{0}$ is directly determined by the kinetic energy density as $ p_{0} = \left( {2K\sin(3\pi/\nu)}/{\sin(\pi/\nu)} \right)^{1/2}$, while the normalization factor is $$A_{\nu} = \dfrac{\nu}{2\pi} \left( \dfrac{\sin^{3}(\pi/\nu)}{2K\sin(3\pi/\nu)} \right)^{1/2} .$$ From the stability criterion (\[eq:stability-condition\]), one gets that this initial state is Vlasov stable when the kinetic energy density satisfies the condition $$K > \dfrac{1}{4} \dfrac{\sin(\pi/\nu)}{\sin(3\pi/\nu)}.$$ One thus gets a dynamical critical energy $U_{c}^{\ast}=0.75,~0.625$ and $0.60355\dots$ for $\nu=4,~6$ and $8$ respectively. In the rest of this section, we set the exponent $\nu$ to $8$. Stationarity and stability checks {#sec:power-stationary} --------------------------------- Let us numerically check the stationarity and the stability of these states; in particular, it will clarify the time region of existence of the QSS. Figure \[fig:stationarity-power\] presents the temporal evolution of $M$ and $\dot{\mu}_{n}~(n=1,2,3)$ for the unstable ($U=0.6<U_{c}^{\ast}$) and stable ($U=0.7>U_{c}^{\ast}$) cases. In both cases, the magnetization $M$ eventually goes toward the equilibrium value $M_{eq}$, indicated by horizontal lines. The three quantities $\dot{\mu}_{n}$ have vanishing small fluctuations around zero except during the time interval $0.0005<\tau<0.003$ for the unstable case. In the unstable case, the system is first in an unstable stationary state ($\tau<0.0005$), before becoming non-stationary ($0.0005<\tau<0.003$) and finally reaches stable stationary states ($\tau>0.003$). On the other hand, in the stable case, the stable stationarity holds throughout the computed time. In the stable case, the magnetization $M$ stays around zero before taking off around $\tau=20$ to reach the equilibrium value $M_{eq}$. The fluctuation level of $\dot{\mu}_{n}$ increases around the take-off time $\tau=20$, but the increase does not imply any non-stationarity of the system, since the fluctuation level is $10$ times smaller than the corresponding one in the non-stationary time region of the unstable case. The nonzero magnetization might be at the origin of the larger fluctuations than in the zero magnetization cases, since the former has a phase and an individual energy $e$ which depends not only on the modulus $M$ but also on the phase. Check of the theoretical prediction {#sec:power-theory} ----------------------------------- In the stable case ($U=0.7$), we perform numerical computations for $N=10^{3},10^{4},2.10^{4}$ and $5.10^{4}$, and average over 20, 20, 10 and $5$ sample orbits respectively. Temporal evolutions of magnetization $M$ are shown in Fig. \[fig:Cp-power\](a), and $M$ takes off toward the equilibrium value $M_{eq}$ around $\tau_{2}=1,20$ and $50$ for $N=10^{3},10^{4}$ and $2.10^{4}$ respectively. The take-off time defines the end of applicable time region of the theory since the homogeneous assumption (ii) breaks. Note that no take-off time appears in the case $N=5.10^{4}$, within the computed time interval. The theory predicts (see Table \[tab:Cp\] for $\nu=8$) that the correlation function decays algebraically with the exponent $-1/2$, i.e. $C_{p}(\tau)\sim\tau^{-1/2}$, up to the take-off time $\tau_{2}$. According to Fig. \[fig:Cp-power\](b), the theoretical prediction agrees well with numerical computations in the intermediate time region $\tau_{1}<\tau<\tau_{2}$, where $\tau_{1}=2$ for any value $N$. This is expected since, on the one hand, the short-time region $\tau<\tau_{1}$ is out of the time domain of application since the theory gives asymptotic estimates. The time $\tau_{1}$ is marked as a long vertical line in Fig. \[fig:Cp-power\](b) to clearly indicate the start of the applicable time domain. Although the quantity ${\tau_{\text{s}}}\simeq 0.005$ is not derived theoretically, the straight lines with the slope $-1/2$ in Fig. \[fig:Cp-power\](b), representing $( \tau / {\tau_{\text{s}}})^{-1/2}$, emphasizes the agreement of the predicted exponent. Introducing the expression of the correlation function in relation (\[eq:sigma\_Cp\]) leads to the law $\sigma_{\theta}^{2}(\tau)\sim\tau^{3/2}$: Figure \[fig:Cp-power\](c), in which the four curves for the four different values of $N$ almost collapse, attests also the validity of this prediction in the intermediate time region $\tau_{1}<\tau<\tau_{2}$. It is possible to confirm more precisely that the diffusion exponent is $3/2$ by introducing the instantaneous exponent $\gamma$ [@moyano-06] defined as $$\gamma = \dfrac{\mbox{d}\ln\sigma_{\theta}^{2}(\tau)}{\mbox{d}\ln\tau} = \dfrac{1}{\sigma_{\theta}^{2}(\tau)} \dfrac{\mbox{d}\sigma_{\theta}^{2}(\tau)}{\mbox{d}\ln\tau} .\label{instantexponent}$$ The instantaneous exponent $\gamma$, shown in Fig. \[fig:Cp-power\](d), goes down and once crosses $3/2$. However, $\gamma$ comes back and stays around $3/2$ in the time interval $\tau_{1}<\tau<\tau_{2}$. Above result confirms therefore unambiguously that the diffusion is anomalous, namely superdiffusive, in the intermediate QSS time interval as predicted by the theory [@Bouchet-Dauxois]. The temporal evolution of $\gamma$ was also recently discussed by Antoniazzi et al. [@antoniazzi-06], and was shown to monotonically decrease toward 1. The difference has two different origins: First, Antoniazzi et al considered non-homogeneous initial distribution of angles, which are out of the applicable range of the theory tested here. Second they considered a waterbag initial distribution of momenta, which does not have tails initially, although tails develop of course as soon as the time is slightly positive. As the theory states that the asymptotic law for diffusion is determined by the tails of the initial distribution of momenta, there is no contradiction that the temporal evolution of $\gamma$ is different. A similar remark applied with the out-of-equilibrium initial distribution discussed by Moyano and Anteneodo [@moyano-06]. For the power-tail initial distributions, the theoretical predictions are essentially good, but not exact. We first note that the increase of $N$ does not affect neither the correlation function, nor the diffusion, at least for $10^{4}\leq N\leq 5.10^{4}$ (the case $N=10^{3}$ has been excluded since no validity time region $\tau_{1}<\tau<\tau_{2}$ appears). In the numerical results, the slope of the diffusion $\gamma$ is not $1.5$ but belongs to $[1.44, 1.48]$. The relative discrepancy is thus at most of 4 percents. There are two possibilities to understand this small discrepancy: (a) the lack of the samples, or (b) the lack of stationarity which is the assumption (iii) of the theory. We will discuss on the origin of these discrepancies at the end of the next section. Gaussian distribution {#sec:gaussian} ===================== Initial distribution -------------------- In this section, we consider the gaussian initial distribution $$f_{0}(p) = \dfrac{1}{\sqrt{2\pi T}} e^{-p^{2}/2T},$$ where the initial temperature $T$ is determined from the energy density as $T=2K=2U-1$. The dynamical critical energy of this gaussian distribution coincides with the critical energy of the second order phase transition $U_{c}=3/4$. As the distribution of angles is homogeneous, the system is therefore at equilibrium for any supercritical energy $U>U_{c}$. Stationarity and stability checks {#stationarity-and-stability-checks} --------------------------------- The stationarity and stability are checked as in Sec. \[sec:power-stationary\] by considering temporal evolutions of magnetization and the derivatives of moments $\mu_n$ shown in Fig. \[fig:stationary-gaussian\]. The scenario of relaxation of this initial distribution with power-law tails is very similar. In the unstable case $(U=0.7<U_{c})$, the system reaches stable stationary states after experiencing unstable stationary and non-stationary states. In the stable case $(U=0.8>U_{c})$, the state is stable stationary in the whole time domain since it is initially at equilibrium. Check of the theoretical prediction {#check-of-the-theoretical-prediction} ----------------------------------- Let us focus on the stable case $U=0.8$ with $N=10^{4}$. The correlation function obtained numerically, and shown in Fig. \[fig:Cp-gaussian\](a), is in good agreement with the theoretical prediction $(\ln \tau)/\tau$ in the long-time region $\tau>\tau_{1}=1$ if we accept the second scaling of time as $\tau\to\tau/{\tau_{\text{s}}}$ with ${\tau_{\text{s}}}=0.2$. As already mentioned in Sec. \[sec:power-theory\], the second scaling is not provided by the theory, while the asymptotic theoretical estimate is out of applicability in the short time domain $\tau<\tau_{1}$. We would like also to stress that the logarithmic correction makes the prediction more precise rather than a simple algebraic decay $1/\tau$. The correlation function can thus be approximated as $$C_{p}(\tau) = \left\{ \begin{array}{ll} C_{p}(0) & \quad \mbox{if}\ \tau<\tau_{1} \\ &\\ \dfrac{c{\tau_{\text{s}}}}{\tau} \ln \dfrac{\tau}{{\tau_{\text{s}}}} & \quad\mbox{if}\ \tau>\tau_{1} \end{array} \right. ,$$ where the short time value has to be $C_{p}(0)={\left\langle p^2(0) \right\rangle}_{N}=2K=0.6$, while $c=0.85$ is obtained by a fitting procedure. This approximation of the correlation function and the relation (\[eq:sigma\_Cp\]) leads to the following expression for the diffusion $$\label{eq:gaussian-sigma-fitting} \dfrac{\sigma_{\theta}^{2}(\tau)}{N^{2}} = \left\{ \begin{array}{ll} C_{p}(0)\tau^{2}, & \quad\mbox{if}\ \tau<\tau_{1} \\ &\\ 2C_{p}(0)\left(\tau_{1}\tau-\dfrac{\tau_{1}^{2}}{2} \right) + c{\tau_{\text{s}}}\tau\left[ \left( \ln\dfrac{\tau}{{\tau_{\text{s}}}} \right)^{2} - \left( \ln\dfrac{\tau_{1}}{{\tau_{\text{s}}}} \right)^{2} \right] \\ \hspace*{4.8em}- 2c{\tau_{\text{s}}}\left[ \tau \left( \ln\dfrac{\tau}{{\tau_{\text{s}}}}-1 \right) - \tau_{1} \left( \ln\dfrac{\tau_{1}}{{\tau_{\text{s}}}}-1 \right) \right] & \quad\mbox{if}\ \tau >\tau_{1} \end{array} \right. .$$ Figure \[fig:Cp-gaussian\](b) presents the diffusion obtained numerically. The two straight lines indicating the short- and long-time regions shows a very good agreement. The diffusion seems anomalous with an exponent $1.35$ in the long-time region. Similarly, the instantaneous exponent $\gamma$ seems to converge toward $1.35$ as shown by Fig. \[fig:Cp-gaussian\](d). However, these observations are not accurate, and only due to a long transient, induced by the logarithmic correction. Diffusion is essentially proportional to the time $\tau$, and hence must be normal in the asymptotic time region. Expression (\[eq:gaussian-sigma-fitting\]) provides the asymptotic form of the instantaneous exponent $$\label{eq:asymptotic-gamma} \gamma = 1 + \dfrac{2}{\ln(\tau/{\tau_{\text{s}}})}.$$ This prediction is in good agreement with numerical results as attested by Fig. \[fig:Cp-gaussian\](d). In the limit of $\tau\to\infty$, the exponent $\gamma$ goes logarithmically toward unity, and we therefore conclude that diffusion is normal although a long transient time is necessary to observe it. This is an excellent illustration of the difficulty to get reliable numerical estimates for the diffusion exponent. Such a case explains [*a posteriori*]{} the reason of previous disagreement [@latora-99; @yamaguchi-03]. As predicted by Table \[tab:Cp\], the logarithmic correction of the correlation function yields a logarithmic correction of the diffusion, so that its asymptotic form should be $ \sigma_{\theta}^{2}(\tau)/N^{2} \sim \tau(\ln\tau)^{2}$. Figure \[fig:Cp-gaussian\](c) confirms this prediction by plotting $\sqrt{\sigma_{\theta}^{2}(\tau)/(\tau N^{2})}$ as a function of $\ln\tau$: one gets a linear behavior in the long time region $\tau>1$. We thus have confirmed the existence of weak anomalous diffusion, i.e. normal diffusion with logarithmic corrections. Let us return to the origin of discrepancies for $\gamma$, discussed at the end of the previous section for the power-law tails. It seems natural to exclude the possibility (a), lack of samples, since the same number of orbits, 20, has been used in the case $N=10^{4}$, for both the power-tails and the gaussians, while the latter case agrees extremely well with the theoretical predictions, even including the logarithmic correction. This excellent agreement comes from the absence of any breaking of theoretical assumptions, since the state is at equilibrium and stationary accordingly. Consequently, we can consider that the possibility (b), lack of stationarity, explains the discrepancies of $\alpha$ and $\gamma$ for power-law tails. Summary {#sec:summary} ======= We have numerically confirmed the theoretical predictions proposed in Ref. [@Bouchet-Dauxois] for initial distributions with power-law or gaussian tails: correlation function and diffusion are in good agreement with numerical results. Diffusion is indeed [*anomalous superdiffusion*]{} in the case of power-law tails, while [*normal*]{} when gaussian. In the latter case, the system is at equilibrium, but the diffusion exponent shows a logarithmically slow convergence to unity due to a logarithmic correction of the correlation function. This long transient time to observe normal diffusion, even for gaussian distribution and at equilibrium, suggests that one should be very careful to decide whether diffusion is anomalous or not [@Correlation; @Correlationb; @Correlationc]. For the power-law tails initial distribution, the numerically obtained exponent of diffusion is slightly different from the theoretical prediction (few percents). As discussed above, this discrepancy comes from the breaking of the stationary assumption. The state is only approximately stationary, explaining that the theoretical predictions are not exact but only approximate. We stress that in the limit of large $N$, these states become stationary because their living times diverge much faster than $N$. For the gaussian initial distribution, the state is in equilibrium from the start, and stationary even with finite $N$: hence the theoretical predictions agree extremely well with numerical results. In addition, above numerical computations clarify two new points: (i) the time region where the theory is applicable, (ii) the second time scaling to fit the correlation function and the diffusion. Both might depend on the degrees of freedom, but the latter, (ii), appears to be not the case for the power-law tails. Obtaining the dependence for the gaussian is a future work. Finally, let us remark that the scenario of the relaxation described in Refs. [@yamaguchi-04; @barre-06] is confirmed even for initial distributions with power-law tails: this had never been tested previously. The scenario claims that the system with long-range interactions experiences first a violent relaxation, before the so-called collisional relaxation which drives the system toward Boltzmann’s equilibrium. In the simulations reported here, non-stationary and stable stationary states correspond respectively to the violent and the collisional relaxations. One might also remark that distributions with power-law tails induce quasi-stationary states above the dynamical critical energy, while being not a member of $q$-distributions [@tsallis-88]. The latter might be a sufficient condition of QSS, but is definitely not a necessary condition. To conclude let us remark that if the results discussed here concerns the simple HMF model, let us mention that it is believed to be general for long-range interacting systems [@Chavanis_Lemou; @relatxationtest]. [**Acknowledgment**]{}\ YYY has been supported by the Ministry of Education, Science, Sports and Culture, Grant-in-Aid for Young Scientists (B), 16740223, 2006. [99]{} T. Dauxois, S. Ruffo, E. Arimondo, M. Wilkens Eds., *Dynamics and Thermodynamics of Systems with Long Range Interactions,* Lecture Notes in Physics Vol. 602, Springer (2002). R.S. Ellis, K. Haven and B. Turkington, [*J. Stat. Phys.*]{} [**101**]{}, 999 (2000) J. Barré, D. Mukamel, S. Ruffo, [*Physical Review Letters*]{} [**87**]{}, 030601 (2001). R.S. Ellis, H. Touchette, B. Turkington, [*Physica A*]{} [**335**]{}, 518 (2004). J. Barré, F. Bouchet, T. Dauxois, S. Ruffo, [*Journal of Statistical Physics*]{} [**119**]{}, 677 (2005). F. Bouchet, J. Barré, *Journal of Statistical Physics,* **118**, 1073 (2005). P.H. Chavanis, J. Sommeria and R. Robert, [*ApJ.*]{} [**471**]{}, 385 (1996) G. M. Zaslavsky, V. F. Shabanov, K. S. Aleksandrov, I. P. Aleksandrov, [*Sov. Phys. JETP*]{} **45**, 315-321 (1977). T. Konishi, K. Kaneko, [*Journal of Physics A*]{} **25**, 6283 (1992). C. Pichon, PhD thesis, Cambridge (1994). M. Antoni, S. Ruffo, [*Physical Review E*]{} [**52**]{}, 2361 (1995). , [*Physical Review E*]{} [**64**]{}, [056134]{} ([2001]{}). F. Baldovin, E. Orlandini, [*Physical Review Letters*]{} [**96**]{}, 240602 (2006). F. Baldovin, E. Orlandini, [*Physical Review Letters*]{} [**97**]{}, 100601 (2006). T. Tsuchiya, N. Gouda, T.Konishi, [*Physical Review E*]{} [**53**]{}, 2210 (1996). Y.Y. Yamaguchi, [*Physical Review E*]{} [**68**]{}, 066210 (2003). Y.Y. Yamaguchi, J. Barré, F. Bouchet, T. Dauxois, S. Ruffo, [*Physica A*]{} [**337**]{}, 36 (2004). J. Barré, F. Bouchet, T. Dauxois, S. Ruffo, Y.Y. Yamaguchi, [*Physica A*]{} [**365**]{}, 177 (2006). W. Braun, K. Hepp, [*Communication in Mathematical Physics*]{} [**56**]{}, 101 (1977). E. Caglioti, F. Rousset, [*Long time estimates in the mean field limit*]{}, Preprint (2004). F. Bouchet, T. Dauxois, [*Journal of Physics: Conference Series*]{} [**7**]{}, 34 (2005). F. Bouchet, T. Dauxois, [*Physical Review E*]{} [**72**]{}, 045103(R) (2005). S. Inagaki, T. Konishi, [*Publ. Astron. Soc. Japan*]{} [**45**]{}, 733 (1993). P.H Chavanis, J. Vatteville, F. Bouchet, [*European Physical Journal B*]{} [**46**]{}, 61 (2005). , [*Physical Review E*]{} [**69**]{}, [056113]{} ([2004]{}). F. Bouchet, *Physical Review E* **70**, 036113 (2004). V. Latora, A. Rapisarda, S. Ruffo, [*Physical Review Letters*]{} [**83**]{}, 2104 (1999). J. P. Bouchaud, A. Georges, [*Physics Reports*]{} **195**, 127 (1990). A. J. Majda, P. R. Kramer, [*Physics Reports*]{} **314**, 238 (1999). , [*Physica D*]{} **134**, [75]{} ([1999]{}). F. Bouchet, F. Cecconi, A. Vulpiani, [*Physical Review Letters*]{} [**92**]{}, 040601 (2004). , [*Europhysics Letters*]{} [**52**]{}, [379]{} ([2000]{}). , submitted to [*Phys. Rev. Letters*]{} E. Lutz, Phys. Rev. Lett. [**93**]{}, 190602 (2004). H. Yoshida, [*Physics Letters A*]{} [**150**]{}, 262 (1990). H. Yoshida, [*Cele. Mech. Dyn. Astro.*]{} [**56**]{}, 27 (1993). L.G. Moyano, C. Anteneodo, [*Phys. Rev. E*]{} [**74**]{}, 021118 (2006). A. Antoniazzi, D. Fanelli, J. Barré, P.-H. Chavanis, T. Dauxois, S. Ruffo, cond-mat/0603813 v3. A. Pluchino, V. Latora, A. Rapisarda, [*Physica A*]{} [**340**]{}, 187 (2004). A. Pluchino, V. Latora, A. Rapisarda, [*Physica D*]{} [**193**]{}, 315-328 (2004). , [*Physical Review E*]{} [**67**]{}, [031105]{} ([2003]{}). C. Tsallis, [*Journal of Statistical Physics*]{} [**52**]{}, 479 (1988). P.H. Chavanis, M. Lemou, *Physical Review E* **72**, 061106 (2005). P.H. Chavanis, [*European Physical Journal*]{} B 52, 61 (2006). [^1]: E-mail: [email protected] [^2]: E-mail: [email protected] [^3]: E-mail: [email protected]

ArXiv

--- abstract: 'We construct a Fock space representation and the action of the two-parameter quantum algebra $U_{r,s}(\frak{gl}_{\infty})$ using extended Young diagrams. In particular, we obtain an integrable representation of the two-parameter quantum affine algebra of type $C_n^{(1)}$ which is a two-parameter generalization of Kang-Misra-Miwa’s realization.' address: - 'School of Mathematical Sciences, South China University of Technology, Guangzhou 510640, China and Department of Mathematics, North Carolina State University, Raleigh, NC 27695, USA' - 'Department of Mathematics, Shanghai University, Shanghai 200444, China' author: - Naihuan Jing - 'Honglian Zhang$^{\star}$' title: 'Fermionic realization of two-parameter quantum affine algebra $U_{r,s}({C}_l^{(1)})$ ' --- [^1] [**Résumé**]{}. Nous construisons une représentation de l’espace de Fock et l’action de la 2-paramètre quantique algèbre $U_{r,s}(\frak{gl}_{\infty})$ en utilisant diagrammes de Young prolongées. Dans particulier, on obtient une représentation intégrable de la 2-paramètres quantique algèbre affine pour le type $C_n^{(1)}$ qui est un 2-paramètres généralisation de la réalisation de Kang-Misra-Miwa. Introduction ============= Quantum groups, introduced independently by Drinfeld [@Dr] and Jimbo [@Jb], are deformations of the universal enveloping algebras of the Kac-Moody Lie algebras. Among the most important classes of quantum groups, quantum affine algebras have a rich representation theory and broad applications in mathematics and physics. In particular they are expected to provide the mathematical foundation for $q$-conformal field theory. Two-parameter quantum groups associated to ${\frak{gl_n}}$ and ${\frak{sl_n}}$ were studied in [@BW1; @BW2; @BW3] by Benkart and Witherspoon (see also earlier work by Takeuchi [@T]). Other classical types and some exceptional types of two-parameter quantum groups and their representations have been investigated in [@BGH1; @BGH2; @HS] (see references therein). The two-parameter quantum affine algebras were introduced in [@HRZ] and their Drinfeld realization and vertex operator representations were also known with help of Lyndon bases for type $A$. More recently these structures have been generalized to all classical untwisted types in [@HZ1; @GHZ], which are analog of the basic representations of the quantum affine algebras [@FJ]. The latter builds upon certain quantization of the so-called bosonic fields. From the other angle aimed toward a categorification, [@JZ2] provided a group-theoretic realization of two-parameter quantum toroidal algebras using finite subgroups of $SL_2(\mathbb{C})$ via the McKay correspondence. It is well known that quantum affine algebras also admit fermionic realizations [@H; @MM; @KMM; @LT] that have played an important role in integrable systems and representation theory. In [@JZ1] such a fermionic realization of the two-parameter quantum affine algebra was constructed for type $A$ using Young diagrams. The combinatorial model gives rise to a natural interpretation of the deforming parameters $r$ and $s$. In this paper, we construct a fermionic realization of the two-parameter quantum affine algebra of type $C$ along the same line. We have taken a slightly different presentation from [@JZ1] to use the approach of Kang-Misra-Miwa [@KMM]. We expect that this model will also work for other 2-parameter twisted quantum affine algebras. The Fock Space of $U_{r, s}(gl(\infty))$ ======================================== In this section, we first define the two-parameter quantum algebra $U_{r, s}(gl(\infty))$, and obtain an irreducible integrable representation using extended Young diagrams. Let $\{ \ep_i, |i\in \mathbb{Z} \}$ be an orthonormal basis of a Euclidean space $E$ with an inner product $(\,,)$. Let $\{\alpha_i|i \in \mathbb{Z}\}$ be the simple roots of the Lie algebra $\mathfrak{g}=gl(\infty)$. We assume that the ground field $\mathbb{K}$ is the field $\mathbb{Q}(r,s)$ of rational functions in $r, s$. Similar to the definition of $U_{r, s}(gl_n)$ (cf. [@BW1]), we define $U_{r, s}(gl(\infty))$ as follows. Let $U_{r, s}(gl(\infty))$ be the unital associative algebra over $\mathbb{K}$ generated by the elements $e^{\infty}_i, f^{\infty}_i, \omega^{\infty}_i, {\omega'}^{\infty}_i$ for $i \in \mathbb{Z}$ satisfying the following defining relations: $$\begin{aligned} &(R1) \quad & {(\omega^{\infty}_i)}^{\pm 1}, {({{\omega'}^{\infty}_j})}^{\pm 1}~~\hbox{all commute with each another and}\\ &&~~~~~~~~~~~~{\omega^{\infty}_i}{({\omega^{\infty}_i})}^{-1}={{\omega'}^{\infty}_j}{({{\omega'}^{\infty}_j})}^{-1}=1, \\ &(R2) \quad & {\omega_i^{\infty}} {e^{\infty}_j}=r^{(\varepsilon_{i},\,\alpha_j)}\, {e^{\infty}_j} {\omega_i^{\infty}} \qquad ~~~~\hbox{and}~~~~ {\omega_i^{\infty}} {f^{\infty}_j}=r^{-(\varepsilon_{i},\,\alpha_j)} \,{f^{\infty}_j} {\omega_i^{\infty}},\\ &(R3) \quad & {{\omega'}^{\infty}_i} {e^{\infty}_j}=s^{(\varepsilon_{i},\,\alpha_j)}\, {e^{\infty}_j} {{\omega'}^{\infty}_i} \qquad~~~~\hbox{and}~~~~ {{\omega'}^{\infty}_i} {f^{\infty}_j}=s^{-(\varepsilon_{i},\,\alpha_j)}\, {f^{\infty}_j} {{\omega'}^{\infty}_i},\\ &(R4) \quad & [\,{e^{\infty}_i},\, {f^{\infty}_j}\,]=\frac{\delta_{ij}}{r-s}({\omega^{\infty}_i{\omega'}^{\infty}_{i+1}}-{\omega^{\infty}_{i+1}{\omega'}^{\infty}_i}),\\ &(R5) \quad &[\,{e^{\infty}_i},\, {e^{\infty}_j}\,]=[\,{f^{\infty}_i},\, {f^{\infty}_j}\,]=0~~~\hbox{if}~~|i-j|>1,\\ &(R6) \quad & {({e^{\infty}_i})}^2{e^{\infty}_{i+1}}-(r+s)\,{e^{\infty}_i}{e^{\infty}_{i+1}}{e^{\infty}_i} +rs \,{e^{\infty}_{i+1}}{({e^{\infty}_{i}})}^2=0,\\ & &e^{\infty}_i{(e^{\infty}_{i+1})}^2-(r+s)\,e^{\infty}_{i+1}e^{\infty}_{i}e^{\infty}_{i+1} + rs \,{(e^{\infty}_{i+1})}^2e^{\infty}_{i}=0,\\ &(R7) \quad & {(f^{\infty}_i)}^2f^{\infty}_{i+1}-(r^{-1}+s^{-1})\,f^{\infty}_i f^{\infty}_{i+1} f^{\infty}_i +{(rs)}^{-1}\, f^{\infty}_{i+1}{(f^{\infty}_{i})}^2=0,\\ & &f^{\infty}_i{(f^{\infty}_{i+1})}^2-(r^{-1}+s^{-1})\,f^{\infty}_{i+1}f^{\infty}_{i}f^{\infty}_{i+1}+ {(rs)}^{-1} {(f^{\infty}_{i+1})}^2f^{\infty}_{i}=0.\end{aligned}$$ Now we construct a Fock space representation for the two-parameter quantum algebra $U_{r,s}(\frak{gl}_{\infty})$, which generalizes the fermionic representation of the usual quantum algebra given in [@KMM]. We begin with the definition of extended Young diagram given in [@JMMO]. An extended Young diagram $Y$ is a sequence $(y_k)_{k\geq 0}$ such that $$\begin{aligned} &&(i)\quad y_k\in \mathbb{Z},\quad y_k\leq y_{k+1} ~~\hbox{for~~ all} ~~k,\\ &&(ii)\quad \hbox{there~~ exists~~ fixed~~ integer}~~ y_{\infty}~~ \hbox{such~~ that}~~ y_k=y_\infty ~~\hbox{for}~~ k\gg 0.\end{aligned}$$ The integer $y_{\infty}$ is called the charge of $Y$. Another way to identify an extended Young diagram is by specifying the fourth quadrant of the xy-plane with sites $\{(i,j)\in \mathbb{Z}\times \mathbb{Z}|i\geq 0,\, j\leq 0\}$. Thus an extended Young diagram $Y=(y_k)_{k\geq 0}$ is an infinite Young diagram drawn on the lattice in the right half plane with sites $\{(i,j)\in \mathbb{Z}\times \mathbb{Z}|i\geq 0,\, j\leq 0\}$, where $y_k$ denotes the “depth” of the $k$-th column. Note that if $y_k\neq y_{k+1}$ for some $k$, then we will have corners in the extended Young diagram $Y=(y_k)_{k\geq o}$. A corner is either “concave” or “convex”. A corner located at site $(i, j)$ is called a $d-$diagonal corner (or corner with diagonal number d), where $d=i+j$. For more details please see [@JMMO] and [@KMM]. For any fixed integer $n$, let $\phi_n$ denote the “empty” diagram $(n, n, n, \cdots )$ of charge $n$. Let $Y_n$ denote the set of all extended Young diagrams of charge $n$. The Fock space of charge $n$ $$\mathscr{F}_n=\bigoplus_{Y\in y_n}\mathbf{Q}(r, s)Y$$ denotes the $\mathbf{Q}(r, s)$-vector space having all $Y\in Y_n$ as base vectors. The algebra $U_{r, s}(gl(\infty))$ acts on the Fock space as follows: $\mathscr{F}_n$ is an irreducible integrable $U_{r, s}(gl(\infty))$- module under the action defined as follows. For $Y\in y_n$, $$\begin{aligned} {e^{\infty}_i}Y&=Y', \qquad & \hbox{if Y has an i-diagonal convex corner then},\\ && \hbox{$Y'$ is the same as Y except that the $i$-diagonal }\\ && \hbox{convex corner is replaced by a concave corner,}\\ &=0, \qquad & otherwise;\\ {f^{\infty}_i}Y&=Y'', \qquad & \hbox{if Y has an i-diagonal concave corner then},\\ && \hbox{$Y''$ is the same as Y except that the $i$-diagonal }\\ && \hbox{concave corner is replaced by a convex corner,}\\ &=0, \qquad & otherwise;\\ {\omega^{\infty}_i}Y &=s^{-1}Y, \qquad & \hbox{if Y has an i-diagonal concave corner },\\ &=r^{-1}, \qquad & \hbox{if $Y$ has an i-diagonal convex corner },\\ &=Y, \qquad & otherwise;\\ {{\omega'}^{\infty}_i}Y&=rY, \qquad & \hbox{if Y has an i-diagonal concave corner },\\ &=s, \qquad & \hbox{if $Y$ has an i-diagonal convex corner },\\ &=Y, \qquad & otherwise.\end{aligned}$$ It is straightforward to verify the relations $(R1)-(R7)$ for the action on $\mathscr{F}_n$ for all generators. We remark that this is very much the same as in type $A$ situation [@JZ1]. Fock Space Representations of $U_{r,s}({C}_l^{(1)})$ ==================================================== Having constructed the Fock space representation of the two-parameter quantum affine algebra $U_{r, s}(gl(\infty))$, we can build the Fock space representation of $U_{r,s}({C}_l^{(1)})$ by generalizing the well–known embedding of the latter inside $U_{r, s}(gl(\infty))$. First let us recall the definition of the two-parameter quantum affine algebra $U_{r,s}({C}_l^{(1)})$ from [@GHZ]. Let $I_0=\{0,\,1,\,2,\,\cdots, n\}$, and $(a_{ij}),\, i,\,j\in I_0$ be the Cartan matrix of type $C_l^{(1)}$. We take the normalization $(\alpha_0,\, \alpha_0)=(\alpha_l,\, \alpha_l)=1$ and $(\alpha_i,\, \alpha_i)=\frac{1}{2}$ for $1\leq i\leq l-1$. Let $r_i=r^{\frac{(\al_i, \al_i)}{2}}$ and $s_i=s^{\frac{(\al_i, \al_i)}{2}}$. Denote by $c$ the canonical central element of $\mathfrak{g}(C_l^{(1)})$ and let $\delta_{ij}$ denote the Kronecker symbol. *The [*two-parameter quantum affine algebra $U_{r,s}(\mathrm{C}_n^{(1)})$*]{} is the unital associative algebra over $\mathbb{K}$ generated by the elements $e_j,\, f_j,\, \omega_j^{\pm 1},\, \omega_j'^{\,\pm 1}\, (j\in I_0),\, \gamma^{\pm\frac{1}2},\,\gamma'^{\pm\frac{1}2},$ $D^{\pm1}, D'^{\,\pm1}$, satisfying the following relations:* $(\hat{C}1)$ $\gamma^{\pm\frac{1}2},\,\gamma'^{\pm\frac{1}2}$ are central with $\gamma=\om_\delta$, $\gamma'=\om'_\delta$, $\gamma\gamma'=(rs)^c$, such that $\omega_i\,\omega_i^{-1}=\omega_i'\,\omega_i'^{\,-1}=1 =DD^{-1}=D'D'^{-1}$, and $$\begin{split}[\,\omega_i^{\pm 1},\omega_j^{\,\pm 1}\,]&=[\,\omega_i^{\pm1}, D^{\pm1}\,]=[\,\omega_j'^{\,\pm1}, D^{\pm1}\,] =[\,\omega_i^{\pm1}, D'^{\pm1}\,]=0\\ &=[\,\omega_i^{\pm 1},\omega_j'^{\,\pm 1}\,]=[\,\om_j'^{\,\pm1}, D'^{\pm1}\,]=[D'^{\,\pm1}, D^{\pm1}]=[\,\omega_i'^{\pm 1},\omega_j'^{\,\pm 1}\,]. \end{split}$$ $(\hat{C}2)$  *For* $\,0 \leqslant i,\, j \leqslant l$, $$\begin{array}{lll} &D\,e_i\,D^{-1}=r_i^{\delta_{0i}}\,e_i,\qquad\qquad\quad &D\,f_i\,D^{-1}=r_i^{-\delta_{0i}}\,f_i,\\ &\omega_je_i\omega_j^{-1}=\langle i,\, j \rangle\,e_i,~~~~~~~~~~~ &\omega_jf_i\omega_j^{-1}=\langle j,\, i \rangle^{-1}\,f_i. \end{array}$$ $(\hat{C}3)$  *For* $\,0 \leqslant i,\, j \leqslant l$, $$\begin{array}{lll} &D'\,e_i\,D'^{-1}=s_i^{\delta_{0i}}\,e_i,\qquad\qquad\quad &D'\,f_i\,D'^{-1}=s_i^{-\delta_{0i}}\,f_i,\\ &\omega'_je_i{\omega'}_j^{-1}=\langle i,\, j \rangle^{-1}\,e_i,~~~~~~~~~~~ &\omega'_jf_i{\omega'}_j^{-1}=\langle j,\, i \rangle\,f_i. \end{array}$$ $(\hat{C}4)$  *For* $\,0 \leqslant i,\, j \leqslant l$, $$[\,e_i, f_j\,]=\frac{\delta_{ij}}{r_i-s_i}(\omega_i-\omega'_i).$$ $(\hat{C}5)$ *For all* $1\leqslant i\ne j\leqslant l$ *but* $(i, j)\not\in\{(0, l), (l, 0)\}$ *such that* $a_{ij}=0$, $$[e_i, e_j]=[f_i, f_j]=0,$$ $$e_le_0=rs\,e_0e_l,\qquad f_0f_l=rs\,f_lf_0.$$ $(\hat{C}6)$  *For* $1\leqslant i\leqslant l-2$, * the* $(r,s)$-Serre *relations for $e_i's$:* $$\begin{gathered} e_0^2e_{1}-(r{+}s)\,e_0e_{1}e_0+rs\,e_{1}e_0^2=0,\\ e_i^2e_{i+1}-(r_i{+}s_i)\,e_ie_{i+1}e_i+(r_is_i)\,e_{i+1}e_i^2=0,\\ e_{i+1}^2e_i-(r_{i+1}^{-1}{+}s_{i+1}^{-1})\,e_{i+1}e_ie_{i+1}+ (r_{i+1}^{-1}s_{i+1}^{-1})\,e_ie_{i+1}^2=0,\\ e_l^2e_{l-1}-(r^{-1}{+}s^{-1})\,e_le_{l-1}e_l+(r^{-1}s^{-1})\,e_{l-1}e_l^2=0,\\ e_{l-1}^3e_l-(r{+}(rs)^{\frac{1}{2}}{+}s)\,e_{l-1}^2e_le_{l-1}\\ +(rs)^{\frac{1}{2}}\, (r{+}(rs)^{\frac{1}{2}}{+}s)\,e_{l-1}e_le_{l-1}^2- (rs)^{\frac{3}{2}}\,e_le_{l-1}^3=0\\ e_{1}^3e_0-(r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,e_{1}^2e_0e_{1}\\ +(rs)^{-\frac{1}{2}}\, (r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,e_{1}e_0e_{1}^2- (rs)^{-\frac{3}{2}}\,e_0e_{1}^3=0.\end{gathered}$$ $(\hat{C}7)$  *For $1\leqslant i\leqslant l-2$,  the $(r,s)$-Serre relations for $f_i's$ are obtained from $(\hat{C}6)$ by replacing $e_i$ for $f_i$ and $r, s$ by $r^{-1}, s^{-1}$ respectively.* In the above $\langle i,\, j\rangle$ are the matrix entries of the two-parameter quantum Cartan matrix for type $\mathrm{C}_{l}^{(1)}$: $$\left(\begin{array}{cccccc} rs^{-1}& r^{-1}& 1 & \cdots & 1 & rs \\ s & r^{\frac{1}{2}}s^{-\frac{1}{2}} & r^{-\frac{1}{2}} & \cdots & 1 & 1\\ \cdots &\cdots &\cdots & \cdots & \cdots & \cdots\\ 1 & 1 & 1 & \cdots & r^{\frac{1}{2}}s^{-\frac{1}{2}} & r^{-1}\\ (rs)^{-1} & 1 & 1 & \cdots & s & rs^{-1} \end{array}\right)$$ We now desribe the integrable representation of the two-parameter quantum affine algebra $U_{r,s}({C}_l^{(1)})$. We start with the folding map $$\pi: \qquad \{0,\,1,\cdots, 2l-1 \}\rightarrow \{0,\,1,\cdots, l \}$$ where $\pi(0)=0, \pi(l)=l$ and $\pi(i)=\pi(2l-i)=i$ for $1 \leq i \leq l-1$. Extend $\pi$ to a map from $\mathbb{Z}$ into $\{0,\,1,\cdots, l \}$ by periodicity $2l$. For any $Y=(y_k)_{k\geq 0}\in Y_n$ define the operators: $$t_kY=r^aY, \qquad t'_k=s^a Y$$ where $a=|\{p\in\mathbb{Z}|y_k<p\leq n \quad \hbox{and}\quad \pi(p+k)=0\}|$ which depends on k. As we still act on the space $\mathscr{F}_n$, so we continue to use the same notation for the new Fock space representation. The following theorem is proved by direct verification (see [@JZ1]). For $k=0,\,1,\,\cdots, l$, the algebra $U_{r,s}(\mathrm{C}_l^{(1)})$ acts on $\mathscr{F}_n$ by the equations: $$\begin{aligned} &e_i= \sum\limits_{\substack{j\\ \pi(j)=i}} \Big( \prod\limits_{\substack{k>j\\ \pi(k)=i}}{\omega}_{k}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}e_j^{\infty},\\ &f_i=\sum\limits_{\substack{j\\ \pi(j)=i}}f_j^{\infty}\Big( \prod\limits_{\substack{k<j\\ \pi(k)=i}}{{{\omega'}_k^{\infty}}}\Big)^{(\alpha_i,\,\alpha_i)},\\ &\omega_i=\Big( \prod\limits_{\substack{j\\ \pi(j)=i}}\omega_{j}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)},\\ &\omega'_i=\Big( \prod\limits_{\substack{j\\ \pi(j)=i}}{\omega'}_{j}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)},\\ &D=\prod\limits_{k\geq 0}t_k, \qquad D'=\prod\limits_{k\geq 0}t'_k.\end{aligned}$$ Under the above action $\mathscr{F}_n$ is an integrable $U_{r,s}(\mathrm{C}_l^{(1)})$-module. We proceed in the same way. First we have $$\om'_j\,e_i\,{\om'}_j^{-1}=\Big( \prod\limits_{\substack{k\\ \pi(k)=j}}{\omega'}_{k}^{\infty}\Big)^{(\alpha_j,\,\alpha_j)}\, \sum\limits_{\substack{j\\ \pi(j)=i}} \Big( \prod\limits_{\substack{j'>j\\ \pi(j')=i}}{\omega}_{j'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}e_j^{\infty} \,\Big( \prod\limits_{\substack{k\\ \pi(k)=j}}{\omega'}_{k}^{\infty}\Big)^{-(\alpha_j,\,\alpha_j)}$$ We don’t have to prove anything for $|i-j|\geqslant 2$ due to ${{\omega'}_i^{\infty}} {e^{\infty}_j}={e^{\infty}_j} {{\omega'}_i^{\infty}} $. For $i=j$, we have $e^{\infty}_m\,({{\omega'}_m^{\infty}})^{-1}=r^{-1}s\,({{\omega'}_m^{\infty}})^{-1}\,e^{\infty}_m$, which follows from $\langle i,\, i\rangle^{-1}=r^{-(\alpha_i,\,\alpha_i)}s^{(\alpha_i,\,\alpha_i)}$. For $0 \leqslant i=j-1 \leqslant l-1$, applying $e^{\infty}_m\,({\om'}^{\infty}_{m+1})^{-1}=s^{-1}\,({\om'}^{\infty}_{m+1})^{-1}\,e^{\infty}_m$ and $\langle i+1,\, i\rangle^{-1}=s^{-(\alpha_{i+1},\,\alpha_{i+1})}$, we arrive at the required relation. Finally, when $1 \leqslant i=j+1 \leqslant l $, we have $e^{\infty}_m\,({\om'}^{\infty}_{m-1})^{-1}=r\,({\om'}^{\infty}_{m-1})^{-1}\,e^{\infty}_m$ and $\langle i-1,\, i\rangle^{-1}=r^{(\alpha_{i-1},\,\alpha_{i-1})}$, and this implies the conclusion. For further reference, we need a few useful identities. By direct calculations, we get the actions on $\mathscr{F}_n$, $$\begin{aligned} &f_m^{\infty}\,({\om_{m'}^{\infty}})^{-1}=\langle m, m'\rangle_{\infty}^{-1}({\om_{m'}^{\infty}})^{-1}\,f_m^{\infty},\\ &e_k^{\infty}\,({\om_{m'}^{\infty}})^{-1}=\langle k, m'\rangle_{\infty}({\om_{m'}^{\infty}})^{-1}\,e_k^{\infty},\\ &f_m^{\infty}\,({{\om'}_{k'}^{\infty}})^{-1}=\langle m, k'\rangle_{\infty}({{\om'}_{k'}^{\infty}})^{-1}\,f_m^{\infty}.\end{aligned}$$ where $\langle i,\,j\rangle_{\infty}$ is defined as follows: $$\begin{aligned} \begin{array}{llll} &\langle i,\, j\rangle_{\infty}=\left\{\begin{array}{cl} rs^{-1},& i=j ;\\ r^{-1},& i=j-1 ;\\ s, & i=j+1;\\ 1, & otherwise.\\ \end{array}\right. \end{array}\end{aligned}$$ Now we turn to the relation $(\hat{C}4)$. From definition and Lemma 3.3, it follows that $$\begin{aligned} &&e_if_j-f_je_i\\ &=& \sum\limits_{\substack{k\\ \pi(k)=i}} \Big( \prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}e_k^{\infty}\,\sum\limits_{\substack{m\\ \pi(m)=j}}f_m^{\infty}\Big( \prod\limits_{\substack{m'<m\\ \pi(m')=j}}{{{\omega'}_{m'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\\ &&-\sum\limits_{\substack{m\\ \pi(m)=j}}f_m^{\infty}\Big( \prod\limits_{\substack{m'<m\\ \pi(m')=j}}{{{\omega'}_{m'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\, \sum\limits_{\substack{k\\ \pi(k)=i}} \Big( \prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}e_k^{\infty}\\ &=&\sum\limits_{\substack{k, m\\ \pi(k)=i\\ \pi(m)=j}}\Big[ \Big(\prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}e_k^{\infty}\, f_m^{\infty}\Big( \prod\limits_{\substack{m'<m\\ \pi(m')=j}}{{{\omega'}_{m'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\\ &&- f_m^{\infty}\Big( \prod\limits_{\substack{m'<m\\ \pi(m')=j}}{{{\omega'}_{m'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\, \Big(\prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}e_k^{\infty}\, \Big]\end{aligned}$$ $$\begin{aligned} &=&\sum\limits_{\substack{k>m\\ \pi(k)=i\\ \pi(m)=j}} \Big(\prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}\Big( \prod\limits_{\substack{m'<m\\ \pi(m')=j}}{{{\omega'}_{m'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\,(e_k^{\infty}\, f_m^{\infty}- f_m^{\infty}\,e_k^{\infty})\\ &&+\delta_{i,j}\sum\limits_{\substack{k\\ \pi(k)=i}}\Big(\prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}\Big( \prod\limits_{\substack{k'<k\\ \pi(k')=j}}{{{\omega'}_{k'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\,(e_k^{\infty}\, f_k^{\infty}- f_k^{\infty}\,e_k^{\infty})\\ &&+\sum\limits_{\substack{k<m\\ \pi(k)=i\\ \pi(m)=j}} \Big(\prod\limits_{\substack{k'>k\\ \pi(k')=i}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}\Big( \prod\limits_{\substack{m'<m\\ \pi(m')=j}}{{{\omega'}_{m'}^{\infty}}}\Big)^{(\alpha_j,\,\alpha_j)}\\ &&\hskip2cm\Big(\sum\limits_{m'<m}\langle k, m'\rangle_{\infty}^{(\alpha_j,\,\alpha_j)}e_k^{\infty}\, f_m^{\infty}- \sum\limits_{k'>k}\langle m, k'\rangle_{\infty}^{(\alpha_i,\,\alpha_i)}f_m^{\infty}\,e_k^{\infty}\Big)\end{aligned}$$ Note that if $m=k+1$, then we have $e_k^{\infty}\,f_m^{\infty}=0=f_m^{\infty}\,e_k^{\infty}$, and if $m>k+1$, then $$\begin{gathered} \sum\limits_{\substack{m'<m\\ \pi(k)=i\\ \pi(m)=j=\pi(m')}}\langle k, m'\rangle_{\infty}^{(\alpha_j,\,\alpha_j)}=\sum\limits_{\substack{k'>k\\ \pi(m)=j\\ \pi(k)=i=\pi(k')}}\langle m, k'\rangle_{\infty}^{(\alpha_i,\,\alpha_i)}\end{gathered}$$ On $\mathscr{F}_n$, it is clear that $$\begin{gathered} e_k^{\infty}\, f_m^{\infty}- f_m^{\infty}\,e_k^{\infty}=\delta_{k,m} \Big(\frac{(\om^{\infty}_k)^{(\alpha_i,\,\alpha_i)}-{\om'}^{\infty}_k)^{(\alpha_i,\,\alpha_i)}}{r_i-s_i} \Big)\end{gathered}$$ Consequently, it follows that on $\mathscr{F}_n$, $$\begin{aligned} &&e_if_j-f_je_i\\ &=&\delta_{i,j}(r_i-s_i)^{-1}\sum\limits_{\substack{k\\ \pi(k)=i}}\{\Big(\prod\limits_{k'\geqslant k}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}\Big( \prod\limits_{k'<k}{{{\omega'}_{k'}^{\infty}}}\Big)^{(\alpha_i,\,\alpha_i)}\\ &&-\Big(\prod\limits_{k'> k}{\omega}_{k'}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}\Big( \prod\limits_{k'\leqslant k}{{{\omega'}_{k'}^{\infty}}}\Big)^{(\alpha_i,\,\alpha_i)}\}\\ &=&\delta_{i,j}(r_i-s_i)^{-1}\{ \Big(\prod\limits_{\substack{k\\ \pi(k)=i}}{\omega}_{k}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}-\Big(\prod\limits_{\substack{k\\ \pi(k)=i}}{\omega'}_{k}^{\infty}\Big)^{(\alpha_i,\,\alpha_i)}\}\\ &=&\delta_{i,j}\,\frac{\om_i-{\om'}_i}{r_i-s_i}. \end{aligned}$$ It is straightforward to check the relation $(\hat{C}5)$, $$\begin{aligned} e_le_0 &=&\sum\limits_{\substack{k\\ \pi(k)=l}} \Big( \prod\limits_{\substack{k'>k\\ \pi(k')=l}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_l,\,\alpha_l)}e_k^{\infty}\, \sum\limits_{\substack{m\\ \pi(m)=0}} \Big( \prod\limits_{\substack{m'>m\\ \pi(m')=0}}{\omega}_{m'}^{\infty}\Big)^{(\alpha_0,\,\alpha_0)}e_m^{\infty}\\ &=&r\,\sum\limits_{\substack{k, m\\ \pi(k)=l\\ \pi(m)=0}} \Big( \prod\limits_{\substack{k'>k\\ \pi(k')=l}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_l,\,\alpha_l)}\, \Big( \prod\limits_{\substack{m'>m\\ \pi(m')=0}}{\omega}_{m'}^{\infty}\Big)^{(\alpha_0,\,\alpha_0)}e_k^{\infty}e_m^{\infty}\\ &=&rs\,\sum\limits_{\substack{k, m\\ \pi(k)=l\\ \pi(m)=0}} \Big( \prod\limits_{\substack{m'>m\\ \pi(m')=0}}{\omega}_{m'}^{\infty}\Big)^{(\alpha_0,\,\alpha_0)}e_m^{\infty}\,\Big( \prod\limits_{\substack{k'>k\\ \pi(k')=l}}{\omega}_{k'}^{\infty}\Big)^{(\alpha_l,\,\alpha_l)} e_k^{\infty}\\ &=&rs\, e_0e_l. \end{aligned}$$ The others relations can be proved similarly. The last task is to verify the Serre relations $(\hat{C}6)$ and $(\hat{C}7)$. Here we only check the relation $(\hat{C}6)$ as the other relations are proved exactly in the same way. To show the Serre relations $(\hat{C}6)$, let us begin with the following notation to save space. $$\begin{gathered} p_j=\prod\limits_{\substack{j'>j\\ \pi(j')=\pi(j)}}{\omega}_{j'}^{\infty}\\ p'_j=\prod\limits_{\substack{j'>j\\ \pi(j')=\pi(j)}}{\omega'}_{j'}^{\infty}.\end{gathered}$$ Let us write $i \gg j$ if $i-j>2$. The following lemmas can be checked directly. For all $j$ and $k$, on $\mathscr{F}_n$ then we obtain, $$\begin{gathered} e_k^{\infty}\, e_k^{\infty}=0, \vspace*{3mm}\\ e_k^{\infty}\, e_j^{\infty}\, e_k^{\infty}=0.\end{gathered}$$ If $\pi(k)=0=\pi(j)$, then it holds $$\begin{gathered} e_j^{\infty}\,p_k=\left\{\begin{array}{cl}p_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k;\vspace*{2mm}\\ r^{-1}s\,p_k\,e_j^{\infty}, &\textit{for}\,\, j> k. \end{array}\right.\\ e_j^{\infty}\,p'_k=\left\{\begin{array}{cl}p'_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k; \vspace*{2mm}\\ rs^{-1}\,p'_k\,e_j^{\infty}, &\textit{for}\,\, j> k. \end{array}\right.\end{gathered}$$ If $\pi(j)=0, \pi(k)=1$, then it follows that $$\begin{gathered} e_j^{\infty}\,p_k=\left\{\begin{array}{cl}p_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k;\vspace*{2mm}\\ r\,p_k\,e_j^{\infty}, &\textit{for}\,\, j=k+1;\vspace*{2mm}\\ r^{2}\,p_k\,e_j^{\infty}, &\textit{for}\,\, j\gg k. \end{array}\right.\\ e_j^{\infty}\,p'_k=\left\{\begin{array}{cl}p'_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k;\vspace*{2mm}\\ s\,p'_k\,e_j^{\infty}, &\textit{for}\,\, j=k+1;\vspace*{2mm}\\ s^{2}\,p'_k\,e_j^{\infty}, &\textit{for}\,\, j \gg k. \end{array}\right.\end{gathered}$$ If $\pi(j)=1, \pi(k)=0$, then we have $$\begin{gathered} e_j^{\infty}\,p_k=\left\{\begin{array}{cl}p_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k \,\textit{or}\, j=k+1;\vspace*{2mm}\\ s^{-1}\,p_k\,e_j^{\infty}, &\textit{for}\,\, j \gg k. \end{array}\right.\\ e_j^{\infty}\,p'_k=\left\{\begin{array}{cl}p'_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k \,\textit{or}\, j=k+1;\vspace*{2mm}\\ r^{-1}\,p'_k\,e_j^{\infty}, &\textit{for}\,\, j \gg k. \end{array}\right.\end{gathered}$$ If $\pi(j)=1=\pi(k)$, it is easy to see that $$\begin{gathered} e_j^{\infty}\,p_k=\left\{\begin{array}{cl}p_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k ;\vspace*{2mm}\\ r^{-1}s\,p_k\,e_j^{\infty}, &\textit{for}\,\, j>k. \end{array}\right.\\ e_j^{\infty}\,p'_k=\left\{\begin{array}{cl}p'_k\,e_j^{\infty}, & \textit{for}\,\, j\leqslant k ;\vspace*{2mm}\\ rs^{-1}\,p'_k\,e_j^{\infty}, &\textit{for}\,\, j>k. \end{array}\right.\end{gathered}$$ We now prove the following Serre relation: $$\begin{aligned} e_0^{2}e_1+(r+s)e_0e_1e_0+rse_1e_0^2=0 \end{aligned}$$ We first use definition to simply the left hand side (LHS) of $(3.20)$. $$\begin{aligned} &&LHS\\&=&\sum\limits_{\substack{j, k, m\\ \pi(j)=0=\pi(k)\\ \pi(m)=1}}\Big[ \Big( \prod\limits_{\substack{j'>j\nonumber\\ \pi(j')=0}}{\omega}_{k'}^{\infty}\Big)e_j^{\infty}\,\Big( \prod\limits_{\substack{k'>k\\ \pi(k')=0}}{\omega}_{k'}^{\infty}\Big)e_k^{\infty}\, \Big( \prod\limits_{\substack{m'>m\\ \pi(m')=1}}{\omega}_{m'}^{\infty}\Big)^{\frac{1}{2}}e_m^{\infty}\nonumber \\ &&-(r+s)\,\Big( \prod\limits_{\substack{j'>j\\ \pi(j')=0}}{\omega}_{k'}^{\infty}\Big)e_j^{\infty}\,\Big( \prod\limits_{\substack{m'>m\\ \pi(m')=1}}{\omega}_{m'}^{\infty}\Big)^{\frac{1}{2}}e_m^{\infty}\,\Big( \prod\limits_{\substack{k'>k \\ \pi(k')=0}}{\omega}_{k'}^{\infty}\Big)e_k^{\infty}\nonumber \\ &&+\Big( \prod\limits_{\substack{m'>m\\ \pi(m')=1}}{\omega}_{m'}^{\infty}\Big)^{\frac{1}{2}}e_m^{\infty}\,\Big( \prod\limits_{\substack{j'>j\\ \pi(j')=0}}{\omega}_{k'}^{\infty}\Big)e_j^{\infty}\,\Big( \prod\limits_{\substack{k'>k\\ \pi(k')=0}}{\omega}_{k'}^{\infty}\Big)e_k^{\infty}\Big]\nonumber \\ &=&\Big(\sum\limits_{m\gg j>k}+\sum\limits_{m=j+1\gg k}+\sum\limits_{j\gg m\gg k}+\sum\limits_{j\gg m=k+1}+\sum\limits_{j=m+1>k}+\sum\limits_{j>k\gg m}+\sum\limits_{j>k=m+1}\Big)\nonumber \\ &&\times \{p_je_j^{\infty}\,p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}+p_ke_k^{\infty}\,p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\nonumber\\ &&-(r+s)(p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}+p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_me_m^{\infty})\nonumber\\ &&+(rs)(p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty}\,p_ke_k^{\infty}+p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}\,p_je_j^{\infty})\}.\nonumber \end{aligned}$$ Using Lemma 3.4 through Lemma 3.8, we would like to show that each summand is actually 0. Taking the second summand for example, we get immediately, $$\begin{aligned} &&\sum\limits_{m=j+1\gg k} \{p_je_j^{\infty}\,p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}+p_ke_k^{\infty}\,p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\\ &&-(r+s)(p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}+p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty})\\ &&+(rs)(p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty}\,p_ke_k^{\infty}+p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}\,p_je_j^{\infty})\}\\ &=&\sum\limits_{m=j+1>k}\{(r^{-1}s+1-r^{-1}(r+s))p_jp_kp_m^{\frac{1}{2}}e_j^{\infty}e_m^{\infty}e_k^{\infty}\\ &&\hskip1cm+(-(r+s)s^{-1}+(rs)(r^{-1}s^{-1}+s^{-2})) p_jp_kp_m^{\frac{1}{2}}e_m^{\infty}e_j^{\infty}e_k^{\infty} \}\\ &=&0. \end{aligned}$$ The other summands are seen as zero by the same method. Subsequently Relation $(3.20)$ has been verified. Next we turn to the relation $$\begin{aligned} e_{1}^3e_0-(r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,e_{1}^2e_0e_{1}+ (rs)^{-\frac{1}{2}}\times\nonumber\\ (r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,e_{1}e_0e_{1}^2-(rs)^{-\frac{3}{2}}\,e_0e_{1}^3=0.\end{aligned}$$ Note that by definition, the left hand side (LHS) of $(3.21)$ is equal to $$\begin{aligned} LHS&=&\sum\limits_{\substack{i,j, k, m\\ \pi(i)=\pi(j)=\pi(k)=1\\ \pi(m)=0}}\Big[p_i^{\frac{1}{2}}e_i^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\, p_me_m^{\infty}\\ &&-(r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,p_i^{\frac{1}{2}}e_i^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty} \,p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\\ &&+(rs)^{-\frac{1}{2}}\, (r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,p_i^{\frac{1}{2}}e_i^{\infty}\,p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty} \,p_k^{\frac{1}{2}}e_k^{\infty}\\ &&-(rs)^{-\frac{3}{2}}\,p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty} \,p_k^{\frac{1}{2}}e_k^{\infty} \Big]\end{aligned}$$ Applying Lemma 3.4 through Lemma 3.8, the last relation becomes, $$\begin{aligned} LHS&=&\Big(\sum\limits_{m\gg j>k>i}+\sum\limits_{m=j+1> k>i}+\sum\limits_{j\gg m\gg k\gg i}+\sum\limits_{j\gg m=k+1>i}\\ &&+\sum\limits_{j=m+1\gg k>i}+\sum\limits_{j>k\gg m\gg i}+\sum\limits_{j>k\gg m=i+1}+\sum\limits_{j>k=m+1>i}\\ &&+\sum\limits_{j>k\gg i\gg m}+\sum\limits_{j>k>i=m+1}\Big) \{\Big(p_i^{\frac{1}{2}}e_i^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\, p_me_m^{\infty}\\ &&+p_i^{\frac{1}{2}}e_i^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_me_m^{\infty}+p_j^{\frac{1}{2}}e_j^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_me_m^{\infty}\\ &&+p_j^{\frac{1}{2}}e_j^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\, p_me_m^{\infty}+p_k^{\frac{1}{2}}e_k^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_me_m^{\infty}\\ &&+p_k^{\frac{1}{2}}e_k^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_me_m^{\infty} \Big)-(r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,\\ &&\Big(p_i^{\frac{1}{2}}e_i^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}+p_i^{\frac{1}{2}}e_i^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\, p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\\ &&+p_j^{\frac{1}{2}}e_j^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\, p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}+p_j^{\frac{1}{2}}e_j^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\\ &&+p_k^{\frac{1}{2}}e_k^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}+p_k^{\frac{1}{2}}e_k^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty} \Big)\\ &&+(rs)^{-\frac{1}{2}}\, (r^{-1}{+}(rs)^{-\frac{1}{2}}{+}s^{-1})\,\Big(p_i^{\frac{1}{2}}e_i^{\infty}\, p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\\ &&+p_i^{\frac{1}{2}}e_i^{\infty}\, p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}+p_j^{\frac{1}{2}}e_j^{\infty}\, p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\\ &&+p_j^{\frac{1}{2}}e_j^{\infty}\,p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_k^{\frac{1}{2}}e_k^{\infty}+p_k^{\frac{1}{2}}e_k^{\infty}\,p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_j^{\frac{1}{2}}e_j^{\infty}\\ &&+p_k^{\frac{1}{2}}e_k^{\infty}\,p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_i^{\frac{1}{2}}e_i^{\infty} \Big)-(rs)^{-\frac{3}{2}}\,\Big(p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_j^{\frac{1}{2}}e_j^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\\ &&+p_me_m^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_k^{\frac{1}{2}}e_k^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}+p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_k^{\frac{1}{2}}e_k^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\\ &&+p_me_m^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_k^{\frac{1}{2}}e_k^{\infty}+p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\,p_i^{\frac{1}{2}}e_i^{\infty}\, p_j^{\frac{1}{2}}e_j^{\infty}\\ &&+p_me_m^{\infty}\,p_k^{\frac{1}{2}}e_k^{\infty}\,p_j^{\frac{1}{2}}e_j^{\infty}\, p_i^{\frac{1}{2}}e_i^{\infty} \Big) \}=0.\end{aligned}$$ Every summand of the last relation can be shown to be 0 as before. Finally we check that Serre relation involving in $i$ and $i+1$ holds in the Fock space. For $1\leqslant i \leqslant l-2$, we compute that $$\begin{aligned} &&e_i^2e_{i+1}-(r_i{+}s_i)\,e_ie_{i+1}e_i+(r_is_i)\,e_{i+1}e_i^2\\ &=&\sum\limits_{\substack{j, k, m\\ \pi(j)=i=\pi(k)\\ \pi(m)=i+1}} \{p_je_j^{\infty}\,p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}+p_ke_k^{\infty}\,p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\\ &&-(r_i+s_i)(p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}+p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty})\\ &&+(r_is_i)(p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty}\,p_ke_k^{\infty}+p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}\,p_je_j^{\infty})\}\end{aligned}$$ $$\begin{aligned} &=&\Big(\sum\limits_{m\gg j>k}+\sum\limits_{m=j+1\gg k}+\sum\limits_{j\gg m\gg k}+\sum\limits_{j\gg m=k+1}+\sum\limits_{j=m+1>k}\\ &&+\sum\limits_{j>k\gg m}+\sum\limits_{j>k=m+1}\Big)\{p_je_j^{\infty}\,p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}+p_ke_k^{\infty}\,p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\\ &&-(r_i+s_i)(p_je_j^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}+p_ke_k^{\infty}\,p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty})\\ &&+(r_is_i)(p_m^{\frac{1}{2}}e_m^{\infty}\,p_je_j^{\infty}\,p_ke_k^{\infty}+p_m^{\frac{1}{2}}e_m^{\infty}\,p_ke_k^{\infty}\,p_je_j^{\infty})\}=0.\end{aligned}$$ Therefore we have finished the proof of Theroem 3.2. **ACKNOWLEDGMENT** Jing thanks the support of Simons Foundation grant 198129, NSFC grant 11271138 and NSF grants 1014554 and 1137837. H. Zhang would like to thank the support of NSFC grant 11371238. [99]{} N. Bergeron, Y. Gao and N. Hu, [*Drinfel’d doubles and Lusztig’s symmetries of two-parameter quantum groups*]{}, J. Algebra, [**301**]{}, (2006), 378-405. N. Bergeron, Y. Gao and N. Hu, [*Representations of two-parameter quantum orthogonal and symplectic groups*]{}, In Studies in Advanced Mathematics, Vol 39. Providence: AMS/IP, [**39**]{} (2007), 1-21. G. Benkart and S. Witherspoon, [*Two-parameter quantum groups and Drinfel’d doubles*]{}, Alg. Rep. Theory, [**7**]{} (2004), 261-286. G. Benkart and S. Witherspoon, [*Representatons of two-parameter quantum groups and Schur-Weyl duality*]{}, In Bergen J, Catoiu S, Chin W, eds. Hopf Algebras. Lecture Notes in Pure and Appl Math, 237. New York: Marcel Dekker, (2004), 65-92. G. Benkart and S. Witherspoon, [*Restricted two-parameter quantum groups (of type A)*]{}. Representations of Finite Dimensional Algebras and Related Topics in Lie Theory and Geometry. Fields Institute Communications, Vol 40. Providence: Amer Math Soc, (2004), 293-318. V. G. Drinfel’d, [*Quantum groups*]{}, ICM-Berkeley Proceedings, New York, 1986, pp.798-820. I. B. Frenkel and N. Jing, [*Vertex representations of quantum affine algebras*]{}, Proc. Nat’l. Acad. Sci. USA. [**85**]{} (1998), 9373-9377. Y. Gao, N. Hu and H. Zhang, [*Two-parameter Quantum Affine algebra of Type ${\mathrm G_2^{(1)}}$, Drinfeld realization and Vertex Representation*]{}, J. Math. Phys., [**56**]{} (2015), 011704. T. Hayashi, [*$q$-analogues of Clifford and Weyl algebras–spinor and oscillator representations of quantum enveloping algebras*]{}. Comm. Math. Phys., [**127**]{} (1990), 129-144. N. Hu and Q. Shi, [*The two-parameter quantum group of exceptional type $G_2$ and Lusztig’s symmetries*]{}, Pacific J. Math.,[**230**]{} (2007), 327-345. N. Hu, M. Rosso and H. Zhang, [*Two-parameter quantum affine algebra $U_{r,s}(\widehat{\frak {sl}_n})$, Drinfel’d realization and quantum affine Lyndon basis*]{}, Comm. Math. Phys., [**278**]{} (2008), 453-486. N. Hu and H. Zhang, *Vertex reprsentations of two-parameter quantum affine algebras: the simply-laced cases*, Preprint, 2013 M. Jimbo, K. C. Misra, T. Miwa and M. OKado, [*Combinatorics of Representations of $U_q(\hat{\frak{sl}_n})$ at $q=0$*]{}, Comm. Math. Phys., [**136**]{} (1991), 543-566. M. Jimbo [*A q-difference analogue of $U(g)$ and the Yang-Baxter equation*]{}, Lett. Math. Phys. [**10**]{} (1985), 63-69. N. Jing and H. Zhang, [*Fermionic realization of two-parameter quantum affine algebra $U_{r,s}(\hat{{sl}_n})$*]{}, Lett. Math. Phys., [**89**]{} (2009), 159-170. N. Jing and H. Zhang, [*Two-parameter quantum vertex representations via finite groups and the McKay correspondence*]{}, Trans. AMS., [**363**]{} (2011), 3769-3797. S. Kang, K. Misra and T. Miwa, [*Fock space representations of the quantized universal enveloping algebras $U_q(C_l^{(1)})$, $U_q(A_{2l}^{(2)})$ and $U_q(D_{l+1}^{(2)})$* ]{}, J. Algebra, [**155**]{} (1993), 238-251. B. Leclerc and J.-Y. Thibon, [*Canonical bases of q-deformed Fock spaces*]{}, Int. Math. Res. Notices, [**9**]{} (1996), 447-456. K. C. Misra and T. Miwa, [*Crystal base for the basic representation of $U_q(\hat{sl}(n))$*]{}. Comm. Math. Phys., [**134**]{} (1990), 79-88. M. Takeuchi, [*A two-parameter quantization of $GL(n)$,*]{} Proc. Japan Acad., [**66**]{} (1990), 112-114. [^1]: $^\star$H. Zhang, Corresponding Author

ArXiv

--- abstract: 'We give a truly elementary proof of the convexity of metric adjusted skew information following an idea of Effros. We extend earlier results of weak forms of superadditivity to general metric adjusted skew informations. Recently, Luo and Zhang introduced the notion of semi-quantum states on a bipartite system and proved superadditivity of the Wigner-Yanase-Dyson skew informations for such states. We extend this result to general metric adjusted skew informations. We finally show that a recently introduced extension to parameter values $ 1<p\le 2 $ of the WYD-information is a special case of (unbounded) metric adjusted skew information.' author: - Liang Cai and Frank Hansen bibliography: - 'mathharv.bib' date: | July 30, 2009\ [Revised August 24, 2009]{} title: 'Metric adjusted skew information: Convexity and restricted forms of superadditivity' --- Introduction ============ The theory of measures of quantum information was initiated by Wigner and Yanase [@kn:wigner:1963] who wanted to find a good measure of our knowledge of a difficult-to-measure observable (state) with respect to a conserved quantity. They were motivated by earlier observations [@kn:wigner:1952; @kn:araki:1960] indicating that the obtainable accuracy of the measurement of a physical observable, in the presence of a conservation law, is limited if the operators representing the observable and the conserved quantity do not commute. Wigner and Yanase discussed a number of postulates that such a measure should satisfy and proposed, tentatively, the so-called [*skew information*]{} defined by $$I_{\rho}(A)=-\frac{1}{2}{\rm Tr}([\rho^{\frac{1}{2}},A]^2),$$ where $\rho$ is a state (density matrix) and $A$ is an observable (self-adjoint matrix). The most important requirement to a measure of quantum information is that knowledge decreases under the mixing of states; or equivalently that the measure is convex in the state variable. They proved this property for the skew information, but Dyson suggested that the measures defined by $$\label{WYD information} I_{\rho}(p,A)=-\frac{1}{2}{\rm Tr}([\rho^p,A] \cdot [\rho^{1-p},A]),$$ where $ p $ is a parameter $ (0<p<1) $ may have the same property. This became the celebrated Wigner-Yanase-Dyson conjecture later proved by Lieb [@Lieb73]. In addition to the convexity requirement Wigner and Yanase suggested that the measure should be additive with respect to the aggregation of isolated subsystems and, for an isolated system, independent of time. These requirements are easily seen to be satisfied for both the Wigner-Yanase skew information and for the Dyson generalization in (\[WYD information\]). Finally, Wigner and Yanase conjectured that the skew information should satisfy a certain superadditivity condition coming from thermodynamics, where it is satisfied for both classical and quantum mechanical systems. It reflects the loss of information about statistical correlations between two subsystems when they are only considered separately. Wigner and Yanase conjectured that the skew information also possesses this property and proved it when the state of the aggregated system is pure. The conjecture was generally believed to be true [@kn:luo:2007] until the second author gave a counter example in [@kn:hansen:2007:2]. The first author, Li and Luo [@kn:cai:2008] discussed weak forms of superadditivity for the Wigner-Yanase information, and Luo and Zhang [@Luo072; @LuoZhang08] introduced the notion of semi-quantum states on a bipartite system and proved superadditivity for the Wigner-Yanase-Dyson skew informations for such states. We continue these inquiries and extend the obtained results to general metric adjusted skew informations. Monotone metrics ---------------- We denote by $ {\cal F}_{op} $ the set of functions $ f\colon\mathbb R_{+}\to\mathbb R_{+} $ such that (i) $ f $ is operator monotone, (ii) $ f(t)=tf(t^{-1}) $ for all $ t>0, $ (iii) $ f(1)=1. $ Since $ f $ is increasing, it may be extended to a continuous function defined on $ [0,\infty) $ with $ f(0)\ge 0. $ We say that $ f $ is regular if $ f(0)>0 $ and non-regular if $ f(0)=0. $ A Morozova-Censov function $ c $ is a function of two positive variables given on the form $$c(x,y)=\frac{1}{y f(xy^{-1})}\qquad x,y>0$$ where $ f\in {\cal F}_{op}. $ It defines a Rimannian metric $ K_\rho^c $ on the tangent space of the state manifold (the space of all positive definite density matrices supported by the underlying Hilbert space), and it is given by $$K_\rho^c(A,B)={{\rm Tr\hskip -0.2em}~}A^* c(L_\rho,R_\rho) B$$ where $ L_\rho $ and $ R_\rho $ are the positive definite commuting left and right multiplication operators by $ \rho. $ The metric is customarily extended to all linear operators $ A $ and $ B $ in the underlying Hilbert space. It is decreasing in the sense that $$K^c_{T(\rho)}(T(A),T(A))\leq K^c_\rho(A, A)$$ for any self-adjoint A and completely positive trace preserving map T, cf. [@kn:petz:1996:2; @kn:ruskai:1999]. The combined efforts of Cencov and Petz established that any decreasing Rimannian metric on the state manifold is given in this way [@kn:censov:1982; @kn:morozova:1990; @kn:petz:1996:2]. We say that the metric is regular if the generating operator monotone function is regular. Recently, a correspondence between the regular and the non-regular operator monotone functions in $ {\cal F}_{op} $ has been constructed by Gibilisco, the second author and Isola [@kn:gibilisco:2008]. Metric adjusted skew information -------------------------------- Let $ c $ be the Morozova-Cencov function of a regular decreasing metric. The corresponding [*metric adjusted skew information*]{} is defined by $$\label{metric adjusted skew information} I^c_\rho(A) = \frac{m(c)}{2} K^c_\rho(i[\rho, A], i[\rho, A]),$$ where $ m(c)=f(0) $ is the metric constant. It may be extended [@kn:hansen:2008:1 Theorem 3.8] to positive semi-definite states $ \rho $ and arbitrary operators $ A. $ The following theorem is taken from [@kn:hansen:2008:1]. \[th:hansen:1\] Let $ c $ be a regular Morozova-Chentsov function, and consider the finite-dimensional Hilbert space $ \mathbf C^n $ for some natural number $ n. $ 1. The metric adjusted skew information is a convex function, $ \rho\to I^c_\rho(A), $ on the manifold of states for any $ n\times n $ matrix $ A. $ 2. For $ \rho=\rho_1\otimes\rho_2 $ and $ A=A_1\otimes 1 + 1\otimes A_2 $ we have $$I^c_\rho(A)=I^c_{\rho_1}(A_1)+I^c_{\rho_2}(A_2).$$ 3. If $ A $ commutes with an Hamiltonian operator $ H $ then $$I^c_{\rho_t}(A)=I^c_\rho(A)\qquad t\ge 0,$$ where $ \rho_t=e^{itH}\rho e^{-itH}. $ 4. For any pure state $ \rho $ (one-dimensional projection) we have $$I^c_\rho(A)={\operatorname{Var}}_\rho(A)$$ for any $ n\times n $ matrix $ A. $ 5. For any density matrix $ \rho $ and $ n\times n $ matrix $ A $ we have $$0\le I^c_\rho(A)\le{\operatorname{Var}}_\rho(A).$$ The first item shows that the metric adjusted skew information is decreasing under the mixing of states. The second shows that it is additive with respect to the aggregation of isolated subsystems, and the third that it, for an isolated system, is independent of time. These requirements were considered essential by Wigner and Yanase [@kn:wigner:1963] to an effective measure of quantum information. Notice that the metric adjusted skew information gives no information about observables commuting with the state variable. It would thus appear that the passage from a regular decreasing metric to the metric adjusted skew information is irreversible. Luo [@kn:luo:2003; @kn:luo:2003:2] was the first to notice that the Wigner-Yanase information is bounded by the variance. Recent developments can be found in [@kn:andai:2008; @kn:audenaert:2008; @kn:gibilisco:2007:2]. Hasegawa and Petz [@kn:hasegawa:1996 Theorem 2] proved that the functions $ f_p $ defined by setting $$\label{family of operator monotone functions} f_p(t)=\left\{ \begin{array}{ll} \displaystyle p(1-p)\frac{(t-1)^2}{(t^p-1)(t^{1-p}-1)}\qquad &t>0,\, t\ne 1\\[1ex] 1 &t=1 \end{array}\right.$$ are operator monotone for $ 0<p<1; $ they are in fact regular functions in $ \mathcal{F}_{op}$ with $ \lim_{t\to 0}f_p(t)=p(1-p). $ The corresponding Morozova-Chentsov functions are given by $$c(x,y)=\frac{1}{p(1-p)}\cdot\frac{(x^p-y^p)(x^{1-p}-y^{1-p})}{(x-y)^2}\qquad x,y>0,$$ and the associated metric adjusted skew information is therefore the Wigner-Yanase-Dyson skew information $$\label{WYD skew information} I_\rho^c(A)=-\frac{1}{2}{{\rm Tr\hskip -0.2em}~}[\rho^p,A]\cdot [\rho^{1-p},A].$$ The convexity in the state variable of the metric adjusted skew information is thus a generalization of the convexity of the Wigner-Yanase-Dyson information proved by Lieb [@Lieb73]. Jenčova and Ruskai considered in the paper [@kn:jencova:2009] the extension of the Wigner-Yanase-Dyson information to parameter values $ 1< p\le 2, $ and they proved that also the extension is convex in the state variable. We show that this extension is associated with monotone metrics, and if we relax the definition of a metric adjusted skew information to allow for non-regular metrics, Jenčova and Ruskai’s extension can be understood as an (unbounded) metric adjusted skew information. In particular, the extension automatically enjoys all the well-known properties of a metric adjusted skew information except boundedness. Convexity of skew information ============================= Convexity of the metric adjusted skew information in the state variable was proved in [@kn:hansen:2008:1], but the given proof relies on Löwner’s deep theory of the relationship between operator monotone functions and the theory of analytic functions. Recently, Effros [@Effros Theorem 2.2] put forward a simple idea[^1] that made it possible to prove all the major quantum information inequalities by reduction to simple block matrix manipulations. The same scheme may also be used to prove convexity in the state variable of the metric adjusted skew information. \[effros transform\] Suppose that $h$ is operator convex. When restricted to positive commuting matrices the function $ g $ defined by $$g(L, R)= h \left(\frac{L}{R}\right) R$$ is jointly convex, that is if the commutator $ [L,R]=0 $ and $$L = \lambda L_1+(1-\lambda)L_2 \quad\text{and}\quad R=\lambda R_1+(1-\lambda)R_2$$ where also the commutators $ [L_1, R_1]=0 $ and $ [L_2, R_2]=0, $ then $$g(L,R)\leq \lambda g(L_1, R_1)+(1-\lambda)g(L_2, R_2)$$ for $0\leq \lambda\leq 1.$ By a simple reduction the metric adjusted skew information may be written on the form $$\label{representation of metric adjusted skew information} I^c_\rho(A)=\frac{m(c)}{2}\displaystyle{{\rm Tr\hskip -0.2em}~}A\, \hat c(L_\rho,R_\rho)A,$$ where $$\label{c hat} \hat c(x,y)=(x-y)^2 c(x,y)\qquad x,y>0.$$ It is easy to prove that the metric adjusted skew information is convex in the state variable $ \rho $ if $ \hat c $ is operator convex, see for example [@kn:hansen:2006:3 Theorem 1.1]. The key observation is now to realize that $ \hat c $ is the Effros transform in Theorem \[effros transform\] of the function $$h(t)=\frac{(t-1)^2}{f(t)}\qquad t>0.$$ Operator convexity of $ \hat c $ and convexity of the metric adjusted skew information will thus follow if just $ h $ is operator monotone. It turns out that a proof of this assertion can be accomplished by the same simple matrix manipulations that underlies Effros’s proof of the quantum mechanical inequalities. \[theorem: operator convex functions\] Let $f\colon\mathbf R_+\to \mathbf R_+ $ be an operator monotone function. Then the function $$h(t)=\frac{(t-1)^2}{f(t)}\qquad t>0$$ is operator convex[^2]. We first assume that the left limit $$f(0)=\lim_{t\to 0} f(t)>0$$ and write $ h $ as a sum $ h=h_1+h_2+h_3 $ of the three functions $$h_1(t)=\frac{t^2}{f(t)}\,,\quad h_2(t)=\frac{-2t}{f(t)}\,,\quad h_3(t)=\frac{1}{f(t)}$$ defined on the positive half-axis. We prove the statement of the theorem by showing that each of these three functions is operator convex, and we shall do so by elementary methods without the usage of Löwner’s theorem. 1\. The function $ f $ is strictly positive, defined on the positive half-axis, and operator monotone, hence also the function $ t\to t f(t)^{-1} $ is operator monotone [@HansenPedersen82 2.6 Corollary]. The function $ h_1 $ may be extended to a continuous function defined on $ [0,\infty) $ with $ h_1(0)=0, $ and since the function $$\frac{h_1(t)}{t}=\frac{t}{f(t)}$$ is operator monotone, it follows [@HansenPedersen82 2.4 Theorem] that $ h_1 $ is operator convex. 2\. It follows from the remarks above that the function $ t\to tf(t)^{-1} $ may be extended to a non-negative operator monotone function defined on $ [0,\infty). $ It is consequently also operator concave [@HansenPedersen82 2.5 Theorem], hence $ h_2 $ is operator convex. 3\. The function $ f $ itself is operator concave [@HansenPedersen82 2.5 Theorem] and since inversion is operator decreasing and operator convex we obtain $$\begin{array}{rl} h_3(\lambda x +(1-\lambda)y)&\displaystyle= \frac{1}{f(\lambda x +(1-\lambda)y)}\\[3ex] &\displaystyle\leq\frac{1}{\lambda f(x)+(1-\lambda)f(y)}\\[3ex] &\displaystyle\leq \lambda \frac{1}{f(x)}+(1-\lambda)\frac{1}{f(y)}\\[3ex] &\displaystyle=\lambda h_3(x)+(1-\lambda)h_3(y) \end{array}$$ for $0\leq \lambda \leq 1$ and positive semi-definite operators $x $ and $y.$ From these three observations we obtain the desired conclusion. In the general case we first apply the obtained result to the functions $ f_\varepsilon(t)=f(t+\varepsilon) $ for $ \varepsilon>0. $ By letting $ \varepsilon\to 0, $ we thus obtain $ h $ as a point-wise limit of operator convex functions, hence $ h $ is also operator convex. Notice that the above theorem is of a general nature and does not require $ f $ to be a function in $ {\cal F}_{op}\,. $ In conclusion, we have proved convexity of the metric adjusted skew information by completely elementary methods. Weak forms of superadditivity ============================= Lieb [@Lieb73] noticed that the Wigner-Yanase-Dyson information is superadditive in the special case where the conserved quantity is of the form $ A\otimes 1_2 $ and we may effortlessly extend this result to any metric adjusted skew information by using that the underlying metric is decreasing. \[lieb lemma\] Let $f$ be a regular function in $ \mathcal{F}_{op}$ and let $ c $ be the associated Morozova-Chentsov function. Let furthermore $\rho$ be a bipartite density operator on a tensor product $ H_1\otimes H_2 $ of two parties. If $ A $ is an observable of the first party then $$\begin{aligned} \label{coroweaksup} I_\rho^c(A\otimes {\mathbf 1_2}) \geq I^c_{\rho_1}(A),\end{aligned}$$ where $ \rho_1={{\rm Tr\hskip -0.2em}~}_2\, \rho $ is the partial trace of $ \rho $ on $ H_1. $ We first notice that $ {{\rm Tr\hskip -0.2em}~}_2 \rho(A\otimes 1_2)=\rho_1 A. $ Indeed, for any vectors $ \xi,\eta\in H_1 $ and orthonormal basis $ (e_i)_{i\in I} $ in $ H_2 $ we have $$\begin{array}{rl} (\xi\mid{{\rm Tr\hskip -0.2em}~}_2\rho(A\otimes\mathbf 1_2)\eta) &\displaystyle=\sum_{i\in I}(\xi\otimes e_i\mid \rho(A\otimes\mathbf 1_2)(\eta\otimes e_i))\\[3ex] &=\displaystyle\sum_{i\in I}(\xi\otimes e_i\mid \rho(A\eta\otimes e_i))\\[3ex] &=(\xi\mid \rho_1A\eta). \end{array}$$ It follows that $ {{\rm Tr\hskip -0.2em}~}_2 i[\rho, A\otimes\mathbf 1_2]=i[\rho_1,A] $ for any self-adjoint $ A $ on $ H_1. $ Since the partial trace is completely positive and trace preserving and the metric $ K_\rho^c $ is decreasing we obtain $$\begin{array}{rl} I^c_\rho(A\otimes\mathbf 1_2)&=\displaystyle\frac{m(c)}{2} K_{\rho}^c(i[\rho,A\otimes {\bf 1_2}],i[\rho,A\otimes\mathbf 1_2])\\[2ex] &\geq\displaystyle\frac{m(c)}{2} K_{\rho_1}^c({{\rm Tr\hskip -0.2em}~}_2 i[\rho,A\otimes\mathbf 1_2], {{\rm Tr\hskip -0.2em}~}_2 i[\rho,A\otimes\mathbf 1_2])\\[2ex] &=\displaystyle\frac{m(c)}{2}K_{\rho_1}^c(i[\rho_1,A],i[\rho_1,A])=I^c_{\rho_1}(A) \end{array}$$ and the proof is complete. In the same spirit, we extend two weak forms of superadditivity for the Wigner-Yanase information [@kn:cai:2008] to general metric adjusted skew informations. Let $f$ be a regular function in $ \mathcal{F}_{op}$ and let $ c $ be the associated Morozova-Chentsov function. Let furthermore $\rho$ be a bipartite density operator on a tensor product $ H_1\otimes H_2 $ of two parties, then we obtain the following two weak forms of superadditivity for the metric adjusted skew information: $$\label{weaksup} I^c_\rho(A\otimes {\mathbf 1_2}+{\mathbf 1_1}\otimes B) \geq\frac{1}{2}\big(I^c_{\rho_1}(A)+I^c_{\rho_2}(B)\big)$$ and $$\begin{array}{l} I^c_\rho(A\otimes {\mathbf 1_2}+{\mathbf 1_1}\otimes B)+ I^c_\rho(A\otimes {\mathbf 1_2}-{\mathbf 1_1}\otimes B)\\[1.5ex] \hskip 9.9em\geq 2\big(I^c_{\rho_1}(A)+I^c_{\rho_2}(B)\big), \end{array}$$ where $ A $ is an observable of the first party and $ B $ is an observable of the second party. We begin by noticing that the partial trace of a commutator $${{\rm Tr\hskip -0.2em}~}_2 [\rho, \mathbf 1_1\otimes B]=0.$$ Indeed, if $ \rho=\sigma_1\otimes\sigma_2 $ is a simple tensor, then a simple calculation yields $$(\xi\mid ({{\rm Tr\hskip -0.2em}~}_2 [\sigma_1\otimes\sigma_2, \mathbf 1_1\otimes B])\eta]= (\xi\mid\sigma_1\eta) {{\rm Tr\hskip -0.2em}~}[\sigma_2,B]=0$$ for vectors $ \xi,\eta\in H_1, $ and since a bipartite state $ \rho $ is a sum of simple tensors, the statement follows. We thus obtain $${{\rm Tr\hskip -0.2em}~}_2 [\rho, A\otimes \mathbf 1_2+\mathbf 1_1\otimes B]= {{\rm Tr\hskip -0.2em}~}_2 [\rho, A\otimes \mathbf 1_2]=[\rho_1, A],$$ where $ \rho_1={{\rm Tr\hskip -0.2em}~}_2\rho $ is the partial trace of $ \rho $ on the first party. By using that the metric $ K^c_\rho $ is decreasing and the partial trace is completely positive and trace preserving, we thus obtain $$\begin{array}{rl} I^c_\rho(A\otimes\mathbf 1_2+\mathbf 1_1\otimes B) &= \displaystyle\frac{m(c)}{2} K^c_\rho(i[\rho,X^+], i[\rho,X^+])\\[2ex] &\ge\displaystyle\frac{m(c)}{2} K^c_{\rho_1}(i[\rho_1,A], i[\rho_1,A])\\[2ex] &=I^c_{\rho_1}(A), \end{array}$$ where $ X^+=A\otimes\mathbf 1_2+\mathbf 1_1\otimes B. $ By symmetry, we similarly obtain $$I^c_\rho(A\otimes\mathbf 1_2+\mathbf 1_1\otimes B)\ge I^c_{\rho_2}(B),$$ and the first assertion follows. The second statement is an easy consequence of the parallelogram identity $$I^c_\rho(X^+)+I^c_\rho(X^-) =2\big(I^c_\rho(A\otimes {\mathbf 1_2})+I^c_\rho({\mathbf 1_1}\otimes B)\big),$$ where in addition $ X^-=A\otimes\mathbf 1_2-\mathbf 1_1\otimes B, $ and the inequality in Lemma \[lieb lemma\]. Semi-quantum states =================== A local von Neumann measurement $ P $ of the first party of a bipartite state $ \rho $ on a tensor product $ H_1\otimes H_2 $ is given by $$P(\rho)=\sum_{i\in I}(P_i\otimes \mathbf 1_2)\rho (P_i\otimes \mathbf 1_2)$$ where $ \{P_i\}_{i\in I} $ is a resolution of the identity on $ H_1 $ in one-dimensional projections. We may similarly define a local von Neumann measurement of the second party. The concept of semi-quantum states was introduced by Luo and Zhang [@LuoZhang08]. A bipartite state $\rho$ is called a semi-quantum state if there exists a local von Neumann measurement $P=\{P_i\}_{i\in I}$ of the first (or second) party leaving $\rho$ invariant, i.e. $ P(\rho)=\rho. $ Luo and Zhang [@LuoZhang08] showed that $\rho$ is a semi-quantum state (corresponding to the local von Neumann measurement $P) $ if and only if $$\label{representation of semi-quantum state} \rho=\sum_{i\in I} p_i P_i\otimes\rho_i$$ where $ (p_i)_{i\in I} $ is a probability distribution and $ \rho_i $ for each $ i\in I $ is a state of the second party. They also proved that the Wigner-Yanase-Dyson skew information is superadditive when the bipartite state is semi-quantum. We extend this result to general metric adjusted skew informations. Let $f$ be a regular function in $ \mathcal{F}_{op}$ and let $ c $ be the associated Morozova-Chentsov function. Let furthermore $\rho$ be a semi-quantum state on a tensor product $ H_1\otimes H_2 $ of two parties, then we obtain superadditivity of the metric adjusted skew information $$I^c_\rho(A \otimes\mathbf 1_2+ \mathbf 1_1\otimes B) \geq I^c_{\rho_1}(A)+ I^c_{\rho_2}(B),$$ where $ A $ is an observable of the first party and $ B $ is an observable of the second party. The essential step is to prove that $$\label{vanishing cross terms} K_{\rho}^c(i[\rho, A\otimes\mathbf 1_2], i[\rho, \mathbf 1_1\otimes B])=0$$ for a semi-quantum state $ \rho. $ We first notice that $$K_{\rho}^c(i[\rho , A\otimes\mathbf 1_2], i[\rho, \mathbf 1_1\otimes B])= {{\rm Tr\hskip -0.2em}~}(A\otimes\mathbf 1_2) \hat c(L_\rho,R_\rho) (\mathbf 1_1\otimes B)$$ where $ \hat c $ is defined in (\[c hat\]). We write $ \rho $ on the form (\[representation of semi-quantum state\]) and choose for each $ i\in I $ a spectral resolution $$\rho_i=\sum_{j\in J} \lambda_{ij} Q_{ij}$$ in terms of one-dimensional projections and obtain in this way a spectral resolution $$\rho=\sum_{i\in I;\,j\in J}p_i\lambda_{ij} (P_i\otimes Q_{ij})$$ of the semi-quantum state $ \rho. $ We may then calculate $$\begin{array}{l} \hat c(L_\rho,R_\rho)(\mathbf 1_1\otimes B)\\[3ex] =\displaystyle\sum_{i,i'\in I;\,j,j'\in J} \hat c(p_i\lambda_{ij}, p_{i'}\lambda_{i'j'}) (P_i\otimes Q_{ij})(\mathbf 1_1\otimes B)(P_{i'}\otimes Q_{i'j'})\\[4ex] =\displaystyle\sum_{i\in I;\,j,j'\in J} \hat c(p_i\lambda_{ij}, p_i\lambda_{ij'}) (P_i\otimes Q_{ij} B Q_{ij'}) \end{array}$$ and obtain $${{\rm Tr\hskip -0.2em}~}(A\otimes\mathbf 1_2) \hat c(L_\rho,R_\rho) (\mathbf 1_1\otimes B) =\sum_{i\in I;\,j\in J} \hat c(p_i\lambda_{ij}, p_i\lambda_{ij}) ({{\rm Tr\hskip -0.2em}~}AP_i)({{\rm Tr\hskip -0.2em}~}BQ_{ij}).$$ Since $ \hat c(x,x)=0 $ for $ x>0 $ we derive (\[vanishing cross terms\]) as desired. The metric $ K^c_\rho $ is sesqui-linear, hence $$\begin{array}{l} I^c_\rho(A \otimes\mathbf 1_2+ \mathbf 1_1\otimes B)\\[2ex] =I^c_\rho(A \otimes\mathbf 1_2)+I^c_\rho(\mathbf 1_1\otimes B)+ m(c) K_{\rho}^c(i[\rho,A \otimes\mathbf 1_2], i[\rho, \mathbf 1_1\otimes B])\\[2ex] =I^c_\rho(A \otimes\mathbf 1_2)+I^c_\rho(\mathbf 1_1\otimes B)\\[2ex] \ge I^c_{\rho_1}(A)+ I^c_{\rho_2}(B), \end{array}$$ where we used (\[vanishing cross terms\]) and Lemma \[lieb lemma\]. Unbounded skew information ========================== The definition in (\[metric adjusted skew information\]) of the metric adjusted skew information requires the associated monotone metric to be regular. The monotone metric may thus be extended radially from the state manifold to the state space, and the metric adjusted skew information is then normalized such that it coincides with the variance on pure states. We may also introduce a metric adjusted skew information associated with a non-regular monotone metric by setting $$\label{unbounded metric adjusted skew information} I^c_\rho(A) = K^c_\rho(i[\rho, A], i[\rho, A])$$ if $ c $ is a non-regular Morozova-Chentsov function. This type of metric adjusted skew information is unbounded and can no longer be extended from the state manifold to the state space. However, it enjoys all the same general properties as a bounded metric adjusted skew information; it is non-negative and the convexity in the state variable follows directly from Theorem \[theorem: operator convex functions\]. Similarly, all the statements concerning limited forms of superadditivity appearing in this paper may be extended ad verbatim to unbounded metric adjusted skew informations. The functions $ f_p(t) $ defined in (\[family of operator monotone functions\]) are for $ 1<p\le 2 $ non-regular functions in $ \mathcal{F}_{op}.$ The function $ f_p $ is for $ 1< p\le 2 $ positive since $ p(1-p)<0 $ and exactly one of the terms in the denominator is negative for each $ t\ne 1. $ We notice that $ f_p(t)\to 1 $ for $ t\to 1 $ and that $ f_p(t)\to 0 $ for $ t\to 0. $ We only need to prove that $ f_p $ is operator monotone, and to obtain this we consider the identity $$f_p(t)=-p(1-p)\frac{t-1}{g_p(t)-1}$$ where the function $$g_p(t)=\left\{ \begin{array}{ll} \displaystyle \frac{t^p-1}{t-1}+\frac{t^{1-p}-1}{t-1}\qquad &t>0,\, t\ne 1\\[2ex] 1 &t=1. \end{array} \right.$$ Since both $ t^p $ and $ t^{1-p} $ are operator convex for $ 1<p\le 2, $ it follows from [@kn:bendat:1955 Theorem 3.2] that $ g_p $ is operator monotone, and it is therefore also operator concave (notice that this conclusion does not require $ g_p $ to be positive). By appealing to Bendat and Sherman’s theorem once more and taking inverse we conclude that $ f_p $ is operator monotone. The monotone metrics corresponding to the functions $ f_p $ for $ 1<p< 2 $ have not been studied in the literature[^3]. Hasegawa and Petz [@kn:hasegawa:1996 Theorem 2] proved operator monotonicity for $ 0<p<1 $ by ingeniously constructing integral representations. However, this type of representations only work for $ 0<p<1. $ By symmetry we also obtain that $ f_p $ is operator monotone for $ -1<p<0. $ We also notice that $$f_p(t)\to \frac{t-1}{\log t}\qquad\text{for}\quad p\to 0\quad\text{or}\quad p\to 1,$$ and this is the function generating the Kubo metric. Similarly, $$f_p(t)=\frac{2t}{t+1}\qquad\text{for}\quad p=-1\quad\text{or}\quad p=2,$$ and this is the function generating the minimal monotone metric. Therefore, with these extensions, we obtain monotone metrics for $ -1\le p\le 2. $ For $ 1<p\le 2 $ the (unbounded) metric adjusted skew information associated with the non-regular functions $ f_p\in\mathcal{F}_{op} $ is given by $$\label{extension of WYD} I_\rho^c(A)=\frac{-1}{p(1-p)}{{\rm Tr\hskip -0.2em}~}[\rho^p,A]\cdot[\rho^{1-p},A].$$ In fact, the only difference from (\[WYD skew information\]) is the constant in front of the trace. But this is exactly the extension of the Wigner-Yanase-Dyson skew information to parameter values $ 1<p\le 2 $ studied by Jencǒva and Ruskai in [@kn:jencova:2009]. The above result shows that this extension of the WYD-information is also associated with monotone metrics and can be understood in terms of the notion of metric adjusted skew information. The main difference is that the metric is regular for $ 0<p<1 $ but non-regular for $ 1<p\le 2. $ It is therefore immediate that the extension proposed by Jencǒva and Ruskai is non-negative and convex in the state variable. Furthermore, it satisfies all the restricted forms of monotonicity under partial traces studied in this paper. [30]{} A. Andai. Uncertainty principle with quantum isher information. , 49:012106, 2008. H. Araki and M.M. Yanase. Measurement of quantum mechanical operators. , 120:622–626, 1960. K. Audenaert, L. Cai, and F. Hansen. Inequalities for quantum skew information. , 85:135-146, 2008. J. Bendat and S. Sherman. Monotone and convex operator functions. , 79:58–71, 1955. L. Cai, N. Li and S. Luo. Weak superadditivity of skew information. , 41 (2008), 135301 N.N. Censov. , volume 53. Amer. Math. Soc., Providence, 1982. E.G. Effros. A Matrix Convexity Approach to Some Celebrated Quantum Inequalities. , 106: 1006–1008, 2009. P. Gibilisco, F. Hansen, and T. Isola. On a correspondence between regular and non-regular operator monotone functions. , 430:2225-2232, 2009. P. Gibilisco, D. Imparato, and T. Isola. Uncertainty principle and quantum isher information . , 48:072109, 2007. F. Hansen. Extensions of ieb’s concavity theorem, , 124:87–101, 2006. F. Hansen. The igner-anase entropy is not subadditive. , 126:643–648, 2007. F. Hansen. Metric adjusted skew information. , 105:9909–9916, 2008. F. Hansen, G.K. Pedersen. Jensen’s Inequality for Operators and Löwner’s Theorem. , 258: 229-241, 1982. H. Hasegawa, D. Petz. On the imannian metric of $ \alpha $-entropies of density matrices. , 38:221–225, 1996. H. Hasegawa and D. Petz. Non-commutative extension of the information geometry . In O. Hirota, editor, [*Quantum Communication and Measurement*]{}, pages 109–118. Plenum, New York, 1997. A. Jencǒva and M.B. Ruskai. A unified treatment of convexity of relative entropy and related trace functions, with conditions for equality. A. Lesniewski and M.B. Ruskai. Monotone iemannian metrics and relative entropy on non-commutative probability spaces. , 40:5702–5724, 1999. E.H. Lieb. Convex trace functions and Wigner-Yanase-Dyson conjecture. , 11: 267-288, 1973. S. Luo. Wigner-anase skew information and uncertainty relations. , 91:180403, 2003. S. Luo. Wigner-anase skew information vs. quantum isher information. , 132:885–890, 2003. S. Luo. Using measurement-induced disturbance to characterize correlations as classical or quantum. , 77: 022301, 2008. S. Luo. Notes on superadditivity of Wigner-Yanase-Dyson information. , 128:1177–1188, 2007. S. Luo, Q. Zhang. Superadditivity of Wigner-Yanase-Dyson information revisited. , 131:1169-1177, 2008. E.A. Morozova and N.N. Chentsov. arkov invariant geometry on state manifolds (ussian). , 36:69–102, 1990. Translated in J. Soviet Math. 56:2648-2669, 1991. D. Petz. Quasi-entropies for finite quantum systems. , 23:57–65, 1986. D. Petz. Monotone metrics on matrix spaces. , 244:81–96, 1996. E.P. Wigner. Die essung quantenmechanischer peratoren. , 133:101–108, 1952. E.P. Wigner and M.M. Yanase. Information contents of distributions. , 49:910–918, 1963. Liang Cai: Department of Mathematics, Beijing Institute of Technology, Beijing, China. Frank Hansen: Department of Economics, University of Copenhagen, Øster Farimagsgade 5, building 26, DK-1353 Copenhagen K, Denmark. [^1]: There is some similarity to the approach used in [@kn:petz:1986]. [^2]: Notice that the theorem cannot be inverted. There are operator convex functions $ h\colon\mathbf R_+\to\mathbf R_+ $ such that the function $ f(t)=(t-1)^2/h(t) $ is not operator monotone. [^3]: There is a claim in [@kn:hasegawa:1997] that the functions $ f_p $ are operator monotone also for $ 1<p<2. $ However, it does not seem to be widely noticed that there is an unrecoverable error in the proof. In equation $ (3.3') $ in the reference there is an integral representation of functions written on the form $ (cx+d)/(ax+b)^2, $ and the key argument is that these functions should be operator monotone decreasing for positive coefficients; or equivalently that the functions $ (ax+b)^2/(cx+d) $ should be operator monotone. To this effect there is a reference to a paper [@HansenPedersen82] co-authored by the second order. However, there is no such claim in [@HansenPedersen82] and the assertion is not true. If it were true, and since a point-wise limit of operator monotone functions is again operator monotone, the claim would also be true for non-negative coefficients. Then, taking $ a=1, $ $ b=0, $ $ c=0, $ and $ d=1 $ we would obtain that $ x^2 $ is operator monotone; but that is not the case.

ArXiv

--- abstract: '1.cm The second (unphysical) critical charge in the 3-body quantum Coulomb system of a nucleus of positive charge $Z$ and mass $m_p$, and two electrons, predicted by F Stillinger has been calculated to be equal to $Z_{B}^{\infty}\ =\ 0.904854$ and $Z_{B}^{m_p}\ =\ 0.905138$ for infinite and finite (proton) mass $m_p$, respectively. It is shown that in both cases, the ground state energy $E(Z)$ (analytically continued beyond the first critical charge $Z_c$, for which the ionization energy vanishes, to $Re Z < Z_c$) has a square-root branch point with exponent 3/2 at $Z=Z_B$ in the complex $Z$-plane. Based on analytic continuation, the second, excited, spin-singlet bound state of negative hydrogen ion H${}^-$ is predicted to be at -0.51554 a.u. (-0.51531 a.u. for the finite proton mass $m_p$). The first critical charge $Z_c$ is found accurately for a finite proton mass $m_p$ in the Lagrange mesh method, $Z^{m_p}_{c}\ =\ 0.911\, 069\, 724\, 655$.' author: - 'A.V. Turbiner' - 'J.C. Lopez Vieyra' - 'H. Olivares Pilón' title: 'Three-body quantum Coulomb problem: analytic continuation' --- We consider the 3-body quantum Coulomb system of two electrons $(e, m)$ and a (heavy) positive charge $Z$ of mass $m_p$, $(Z, m_p)$, where $m \ll m_p$. It is one of the most fundamental systems in theoretical physics. In atomic units $|e|=1$. Usually, this system is called the helium isoelectronic sequence, among many different names used. We denote this system as $(Z, e, e)$ and prefer to call it the [*two-electron sequence*]{}. This system is fundamental: it includes the species like H${}^-$, He, Li${}^+$ etc, which play exceptionally important role in Nature. The non-relativistic Hamiltonian has the form $$\label{H} {\cal H}\ =\ -\frac{1}{2} ({\Delta}_1 + {\Delta}_2)\ -\ \frac{{\Delta}}{2 m_p} \ -\ \frac{Z}{r_{1}}\ -\ \frac{Z}{r_{2}} \ +\ \frac{1}{r_{12}}\ ,$$ when written in atomic units with electronic mass $m=1$, where $r_1(r_2)$ is the distance from the charge $Z$ to the first (second) electron, $r_{12}$ is interelectron distance, ${\Delta}_1 ({\Delta}_2)$ is the Laplacian which describes the kinetic energy of the first (second) electron, and ${\Delta}$ is the Laplacian which describes the kinetic energy of charge $Z$. The configuration space of (\[H\]) is ${\bf R}^9$, but after separation of centre-of-mass motion, we arrive at a six-dimensional configuration space of the relative motion. Note that in the static limit $m_p {\rightarrow}\infty$ the non-relativistic Hamiltonian (\[H\]) degenerates to the form $$\label{Hs} {\cal H}\ =\ -\frac{1}{2} ({\Delta}_1 + {\Delta}_2) \ -\ \frac{Z}{r_{1}}\ -\ \frac{Z}{r_{2}} \ +\ \frac{1}{r_{12}}\ .$$ Due to charge quantization, the physical system occurs when $Z$ takes integer values, $Z=1,2,\ldots $. At $Z \geq 1$ the energy of a bound state $E(Z)$ is a smooth function proportional to $(-Z^2)$ at large $Z$. Long ago it was given a rigorous mathematical proof in [@Kato:1980] that for $m_p=\infty$ the function $E(Z)/Z^2$ is the analytic function in vicinity of $1/Z=0$. There is still unresolved challenge to find the radius of convergence in $1/Z$ expansion and establish nature of the singularity(ies) on its circle of convergence. The lower edge of continuum (the threshold energy) is given by the ground state energy of the $Z$-hydrogen atom $(Z,e)$, $E_{th}= - Z^2 \frac{m_p}{2(1+m_p)}$ . It is well-known that there exists a certain charge $Z$ - called the critical charge $Z_c$ - for which the ground state energy of a two-electron sequence degenerates with the lower edge of continuum, thus, spontaneous ionization becomes possible. Quite recently, in the static limit (\[Hs\]), the critical charge was accurately calculated (with triple basis sets containing up to 2276 terms, see [@Drake:2014]) and verified in independent calculation (using the Lagrange mesh method with lattice size $92 \times 92 \times 20$, see [@OT-PLA:2014]) within 12 significant figures $$\label{zcrit} Z^{\infty}_{c}\ =\ 0.911\, 028\, 224\, 077\ .$$ It was proved in [@Simon:1977] that at $Z=Z^{\infty}_{c}$ the Schrödinger equation has normalizable eigenfunction. In [@Drake:2014] this result was verified - it was explicitly stated that the trial function (triple basis set containing 2276 terms) is normalizable and localized, $<r_1> \sim 1/{\alpha}_1 \sim 1$a.u. and $<r_2> \sim 1/{\alpha}_2 \sim 5$a.u. , see below (6)-(7). It implies that the probability to meet electron at infinity is equal to zero, thus, the spontaneous ionization is absent. The critical charge $Z_c$ for the case of the finite mass $m_p$ was studied in [@Kais] for some particular values of $m/m_p$ and in [@Moini:2014] for $m_p \gtrsim 1$ and $m=1$. However, only recently, it was calculated accurately for the mass of proton, $m_p = 1836.152 672 45 m_e$, where $m_e=1$ was set for the electron mass, using a 2856-term trial function (the convergence in energy was checked up to 11 s.d.) [@King:2015] $$\label{zcrit-mp} Z^{m_p}_{c}\ =\ 0.911\, 069\, 7(3)\ ,$$ which now we confirm (and improve) in the Lagrange mesh method [@Baye:2015] (see for details [@TLO:2016]) (the convergence in energy is checked up to 12 s.d. including, the maximal lattice size is $92 \times 92 \times 20$) $$\label{zcrit-mp-us} Z^{m_p}_{c}\ =\ 0.911\, 069\, 724\, 655\ ,$$ Note that finite-mass effects are sufficiently small $\sim 10^{-5}$ changing the fifth figure in the critical charge, see (\[zcrit\]) and (\[zcrit-mp-us\]). It is evident that when Z gets slightly smaller than $Z_c$ the bound (stationary) state becomes quasi-stationary state (or Gamow state), characterized by complex energy, for discussion see [@BZP; @BZP2]. Its wavefunction is complex: it is mostly localized in the Coulomb well and it corresponds to outgoing spherical wave at large $r_2 (r_1)$ at fixed $r_1 (r_2)$, $\sim e^{-ikr_2} \big(e^{-ikr_1}\big) $. The imaginary part of energy is defined by tunneling rate from the Coulomb well to infinity. In principle, it can be found via multidimensional WKB method, it should be exponentially small, $\sim (Z_c -Z)^a e^{-\frac{b}{(Z_c -Z)^c}}$, where $b,c>0$ at $Z {\rightarrow}Z_c-$. We are not aware about any concrete theoretical calculations done so far. Hence, we have two different spectral problems which can be summarized as follows: (i) for $Z > Z_c$ we have a problem of bound states looking in the Schrödinger equation for solutions in the Hilbert space with (real) energies $E=E_{>}(Z)$, (ii) for $Z < Z_c$ we have a problem of quasi-stationary states looking in the Schrödinger equation for solutions corresponding to outgoing spherical wave with (complex) energies $E=E_{<}(Z)$. Thus, the boundary conditions for these two problems are completely different. In principle, there is no reason that these two analytic functions should coincide $E_{>}(Z)=E_{<}(Z)$. Furthermore, the function $E_{<}(Z)$ approaching to $Z {\rightarrow}Z_c-$ is complex with exponentially small imaginary part (it indicates to the essential singularity at $Z = Z_c$), while the $E_{>}(Z)$ is real and approaching from the right to $Z {\rightarrow}Z_c+$ behaves linearly [@Simon:1977]. It looks like a type of Kosterlitz-Thouless phase transition (of the infinite order) occurs at $Z=Z_c$: the functions $E_{<}(Z)$ and $E_{>}(Z)$ and [*all*]{} their derivatives match at $Z=Z_c$. The goal of our study is to find the function $E_{>}(Z)$ - the energy of the ground state - its analytic properties. We will not touch the function $E_{<}(Z)$ either theoretically, or numerically. In remarkable papers of 1966 and 1974, F.H. Stillinger and D.K. Stillinger [@Stillinger:1966; @Stillinger:1974], made an extended analysis of the ground state energy obtained variationally as a function of continuous charge $Z$ and predicted the existence of the second critical charge $Z_B$ - the charge for which the ground state eigenfunction looses its normalizability, see [@Stillinger:1966]. The trial function that was used for the analysis, was the celebrated Hylleraas-Eckart-Chandrasekhar function with a property of clusterization - in average one electron is closer to the charge center than the other, they are called innermost and outermost electrons, $$\label{GS} \Psi_{trial} (r_1, r_2, r_{12})\ =\ \Psi_0 (r_1, r_2) + \Psi_0 (r_2, r_1)\ ,\ \Psi_0 (r_1, r_2)=e^{- {\alpha}_1 r_1 - {\alpha}_2 r_2}\ ,\ {\alpha}_1 \neq {\alpha}_2\ ,$$ where ${\alpha}_{1,2}(Z)$ are two variational parameters, both real, found in the minimization of the energy functional for a fixed $Z$. The $Z$ dependence of the parameters ${\alpha}_{1,2}$ was quite non-trivial. It must be noted the important technical moment which made the analysis feasible: all integrals involved are evaluated analytically and are rational functions in ${\alpha}_{1,2}$. Hence, the variational energy is a simple analytic function of parameters ${\alpha}_{1,2}$ being the ratio of two polynomials in variables ${\alpha}_{1,2}$. It allows analytic considerations of the variational energy[^1]. Stillinger and Stillinger showed the existence of two distinct critical charges: (i) $Z_{c}(=0.9538)$ for which the ionization energy vanishes, hence, the ground state energy is equal to $-Z_{c}^2/2$ and coincides (degenerates) with the lower bound of continuum (the ground state energy of the $Z_c$-Hydrogen atom, $(Z_c,e)$), and (ii) $Z_{B}(=0.9276)$ for which the trial function (\[GS\]) becomes non-normalizable: minimal ${\alpha}_{1}$ (or ${\alpha}_{2}$) vanishes, for illustration see Fig. \[E-ground-state\]. It was also demonstrated that the variational ground state energy $E_{var}=E(Z)$ is a regular function at $Z=Z_{c}$, while at $Z=Z_{B}$ it has a square-root branch point with exponent 3/2 with the branch cut going along real axis to the negative direction, $Z \in (-\infty, Z_{B}]$. Recently their analysis was extended: it was checked that, assuming the parameters ${\alpha}_{1,2}$ in (\[GS\]) can be complex, the minimum of energy functional is always reached for real ${\alpha}_{1,2}$ for $Z>Z_B$. The ground state variational energy can be constructed for a more general trial function with the electronic correlation explicitly included [@Bruno:2014], $$\label{GSB} \Psi_0 (r_1, r_2, r_{12})\ =\ (1 + c r_{12}) e^{- {\alpha}_1 r_1 - {\alpha}_2 r_2 - {\gamma}r_{12}}\ , \ {\alpha}_1 \neq {\alpha}_2\ ,$$ (c.f. (\[GS\])), where ${\alpha}_{1,2}(Z), {\gamma}(Z), c(Z)$ are four variational parameters. Similarly to (\[GS\]), the ground state was characterized by two distinct critical charges, $Z_{c} \neq Z_{B}$, where $Z_{c}(=0.9195)$ is regular point while $Z_{B}(=0.8684)$ is the square-root branch point with exponent 3/2. Thus, the introduction of two extra parameters $c, {\gamma}$ to the trial function (\[GS\]) shifts significantly the position of each critical point without changing the analytic properties. Note that both trial functions (\[GS\]), (\[GSB\]) lead to sufficiently accurate ground energy for $Z=1$ (which is a weakly-bound state): $-0.5133$ and $-0.5260$, respectively, while $E_{exact}=-0.52775$a.u. and for $Z=2$ : $-2.8757$ and $-2.9019$, respectively, while $E_{exact}=-2.9037$a.u. In fact, (\[GS\]) and (\[GSB\]) are the most accurate trial functions among two and four-parametric trial functions for any integer $Z \in [1,\infty)$: they give 1-2 and 2-3 correct decimal digits in ground state energy, respectively. Thus, both variational energies provide highly accurate, uniform(!) approximation of the exact energy in $Z \in [1,\infty)$. Striking common property of both approximate variational energies as analytic functions of $Z$ is the absence of singularity at $Z=Z_c$. Note that the existence of two critical charges was also demonstrated in the Born-Oppenheimer approximation $m_p {\rightarrow}0$ (two-center, H$_2^+$-type case) [@TM:2012]. The analysis of both variational energies $E_{var}(Z)$ obtained with (\[GS\]), (\[GSB\]) did not indicate the existence of other square-root branch points at finite $Z$ on the complex $Z$-plane [@Bruno:2014]. Hence, one can draw the conclusion, in accordance with the Landau-Zener theory of the level crossings (for discussion see e.g. [@BW; @BW2]), that the point $Z_{B}$ seems to be a natural candidate for the point of the level crossing of the two lowest spin-singlet, $1^1S$ and $2^1S$ states. However, the analytic continuation of the energy to the real positive axis of the second sheet of the Riemann surface in $Z$ in the vicinity of $Z_{B}$ is not fully supported by the analytic continuation of trial functions (\[GS\]), (\[GSB\]). Although these functions remain normalizable after analytic continuation, they do not develop a nodal surface, thus, they are NON-orthogonal to the ones corresponding to the ground state at given $Z$ and need to be orthogonalized. Extending the analysis of [@Stillinger:1966; @Bruno:2014], one can derive that for both cases (\[GS\]) and (\[GSB\]) the variational energy in $Z$ has two essentially different functional expansions near $Z_{c}$ and $Z_{B}$, namely: $$\label{Zcr} E(Z) \ =\ -\frac{Z_{c}^2}{2} + a_1 (Z - Z_{c})+ a_2 (Z - Z_{c})^2 + a_3 (Z - Z_{c})^3 + \ldots\ ,$$ which is the Taylor expansion: it indicates the absence of a singularity at $Z=Z_{c}$, and $$\label{ZB} E(Z) \ =\ b_0 + b_1 (Z - Z_{B}) + c_1 (Z - Z_{B})^{\frac{3}{2}} + b_2 (Z - Z_{B})^2 + c_2 (Z - Z_{B})^{\frac{5}{2}} + \ldots\ ,$$ for $Z \geq Z_{B}$, which is the so-called Puiseux expansion (the expansion in fractional degrees). The latter indicates the existence of the square-root branch point with exponent 3/2. In absence of other singularities in the vicinity of $Z_{c}$ other than at $Z=Z_B$, the radius of convergence of the expansion (\[Zcr\]) has to be equal to $(Z_{c}-Z_B)$. Since the continuum starts at $E_{th} = -Z_{c}^2/2$, the presence of the bound state at $E \geq E_{th}$ indicates the phenomenon of the existence of a bound state [*embedded*]{} to continuum [@Stillinger:1975] [^2]. It seems true inside the accuracy of a given variational procedure. It must be noted that even highly accurate, normalizable trial function found at $Z > Z_c$ does not guarantee the existence of the solution of the Schrödinger equation in the Hilbert space for $Z < Z_c$ after analytic continuation. Thus, this state is not necessarily bound. It is worth emphasizing that the expansions (\[Zcr\]) and (\[ZB\]) are exact functionally since they are derived from variational energies emerging from (\[GS\]) and (\[GSB\]) after the exact minimization: minimizing-the-energy variational parameters are found in a form of expansions near $Z_{c}$ and near $Z_{B}+$, respectively. [|l| l | l | l |]{}  $Z$   &$E$ (a.u.)  & &\  2.00   &-2.903 724 4  &-2.903 719  &-2.201 927\  1.00   &-0.527 751 0  &-0.527 749  &-0.515 540\  0.94   &-0.449 669 0  &-0.449 668  &-0.447 012\  0.92   &-0.425 485 3  &-0.425 485  &-0.424 741\  $Z^{EBMD}_{c}$ (see (\[zcrit\]))  &-0.414 986 2  &-0.414 986  &-0.414 793\  0.91   &-0.413 799 2  &-0.413 799  &-0.413 652\ It seems natural to check validity of the expansions (\[Zcr\]) and (\[ZB\]) for more accurate energies than ones found using (\[GS\]), (\[GSB\]). In order to do it we calculate $E(Z)$ in 12 points at $Z \in [0.91 , 1]$ and $Z=2$ in the Lagrange mesh method accurate to not less than 9 s.d. (see Table \[table1S\] for illustration). Then taking into account the points from the domain $Z \in [Z^{EBMD}_{c} , 1]$ [**only**]{}, we make a fit using a general terminated Puiseux expansion, $$E\ =\ \sum_{n=0}^{N} a_n (Z - Z_B)^{n/2 + {\epsilon}_n} \ ,$$ for different $N$ assuming that exponents are growing with $n$, searching for optimal values of the parameters $a_n, {\epsilon}_n$. As for initial values of parameters, those we use, they are ones taken from the expansion (\[ZB\]) for (\[GS\]), (\[GSB\]). Finally, we arrive at $$E_{1^{1}S}^{(fit)}(Z)\ =\ -\,0.407924\ -\ 1.12347\, (Z - Z_{B})\ -\ 0.197785\, (Z - Z_{B})^{3/2}\ -\ 0.752842\, (Z - Z_{B})^2$$ $$-\ 0.108259\, (Z - Z_{B})^{5/2}\ -\ 0.014135\, (Z - Z_{B})^3 +\ 0.00854 \, (Z - Z_{B})^{7/2}$$ $$\label{e-fit_1} +\ 0.00483\, (Z - Z_{B})^4\ -\ 0.000056\, (Z - Z_{B})^{9/2}\ ,$$ with the critical charge $$\label{ZB-1} Z_{B}\ =\ 0.904854\ ,$$ where as the result of the fit ${\epsilon}_n$ turned to be of order $10^{-6} - 10^{-7}$ as well as the parameter $a_1$. It turns out that is artifact of fitting procedure: all these parameters ${\epsilon}_n, a_1$ can be placed equal to zero without reducing the quality of the fit! The expression (\[e-fit\_1\]) reproduces up to a portion of $10^{-6} - 10^{-7}$ in energies at 12 points in $Z$ more or less equally distributed in $Z \in [Z_c, 1.]$ and also the point $Z=0.91 < Z_c$ (!) , for illustration see Table I (and Table II, see below). The quality of the fit drops with the increase of $Z$, see Table I. The numerical coefficients in the expansion decrease with the increase of the order of terms. It may be considered as the indication to the finite radius of convergence of corresponding numerical series of coefficients and thus the Puiseux expansion (\[ZB\]) itself. We must emphasize that the critical charge $Z_B$ is unphysical: it corresponds to analytically-continued ground state energy to the domain of quasi-stationary states! In the recent analysis of $1/Z$ expansion [@TL:2016] it was shown that the first ten decimal digits in all the first 401 coefficients $e_n$ are found correctly in [@Baker:1990]: their weighted sums $E_P(Z)\ =\ \sum^{401} e_n Z^n$ reproduce subsequently the first ten decimal digits in ground state energy of two-electron sequence at $Z=1,2,\ldots 10$, found in [@Nakashima:2007]. Checking convergence of $1/Z$ expansion we calculate weighted sums $E_P$ for $Z<1$, see Table \[E\], and compare them with the result of fit (\[e-fit\_1\]) and with energies obtained in Lagrange mesh calculations on the lattice $92 \times 92 \times 20$. Comparison of $E_P$ and $E_{1^{1}S}^{(fit)}$ indicates to a striking agreement between them: they differ in $\sim 2 \cdot 10^{-7}$ for $Z=0.95$, in $\sim 6 \cdot 10^{-6}$ for $Z=Z^{\infty}_{c} \approx 0.911$ and $\sim 4 \cdot 10^{-5}$ for $Z=0.905$. Hence, the fit (\[e-fit\_1\]) is in agreement with perturbation theory in $1/Z$. In a similar way $E_P$ agrees with accurate calculations of the ground state energy $E$ in the Lagrange mesh method even for $Z < Z_c$. [|l|l|l|l|]{} $Z$ &$E_P$ &$E_{1^{1}S}^{(fit)}$ (\[e-fit\_1\])  &$E$ (Lagr.mesh)\ 0.905   &  -0.408 045   & -0.408 089  & -0.408 1${}^\dagger$\ 0.908   &  -0.411 478   & -0.411 501  & -0.411 502${}^\dagger$\ 0.909   &  -0.412 629   & -0.412 648  & -0.412 648 5${}^\dagger$\ 0.91   &  -0.413 783   & -0.413 799  & -0.413 799 2\ $Z^{\infty}_{c}$ &  -0.414 972   & -0.414 986  & -0.414 986 2\ 0.92   &  -0.425 482   & -0.425 485  & -0.425 485 3\ 0.93   &  -0.437 450   & -0.437 451  & -0.437 451 3\ 0.94   &  -0.449 669   & -0.449 669  & -0.449 669 0\ 0.95   &  -0.462 125   & -0.462 125  & -0.462 124 7\ Moreover, it has to be mentioned a remarkable fact that expanding (\[e-fit\_1\]) at $Z=\infty$ the established coefficients in $1/Z$- expansion $e_{10-150}$ are reproduced with relative accuracy $10^{-2}$ (equivalently, the first two significant digits) [@TLO:2016]. [|l| l | l | l |]{}  $Z$ &$E$ (a.u.)  &  &\  2.00   &-2.903 304 6   & -2.903 178  & -2.196 303\  1.00   &-0.527 445 9  &-0.527 444  &-0.515 312\  0.96   &-0.474 536 5  &-0.474 537  &-0.469 305\  0.92   &-0.425 242 4  &-0.425 242  &-0.424 519\  0.911069725 (see (\[zcrit-mp-us\]))  &-0.414 798 1  &-0.414 798  &-0.414 617\  0.91   &-0.413 564 0  &-0.413 564   &-0.413 430\ Similar analysis to (\[e-fit\_1\]) can be made for the finite-mass case $m_p$. Making a fit for the energies at $Z \in [Z^{m_p}_{c} , 1]$ (see Table \[table1S-mp\]) we arrive at $$E_{1^{1}S}^{(fit\,, m_p)}(Z)\ =\ -0.408019 - 1.123511\, (Z - Z_{B}) - 0.198005\, (Z - Z_{B})^{3/2} - 0.751842\, (Z - Z_{B})^2\,$$ $$-\, 0.101848\, (Z - Z_{B})^{5/2}\, -\, 0.02171\, (Z - Z_{B})^3 + 0.0039\, (Z - Z_{B})^{7/2}\, +\, 0.0126\, (Z - Z_{B})^4$$ $$\label{e-fit_2} -\ 0.0028\, (Z - Z_{B})^{9/2}\ ,$$ with the critical charge $$\label{ZB-mp} Z_{B}\ =\ 0.90514\ ,$$ (cf (\[ZB\])). Note that the parameters of the fit (\[e-fit\_1\]) obtained in the static limit change very little in going from infinite to finite proton mass, see (\[e-fit\_2\]). Both critical charges $Z_{c}, Z_{B}$ increase slightly for the case of the finite proton mass. It can be immediately seen from Tables I, \[table1S-mp\] that based on the analytic continuation in $Z$ of the fit (\[e-fit\_1\]), (\[e-fit\_2\]) (it corresponds to the change of sign in front of the terms of half-integer degrees), the existence of the second bound state $2^{1}S$ of negative hydrogen ion is predicted, $$\label{2S} E_{2^{1}S}^{\infty}\ =\ -0.515 54\,a.u.\quad , \quad E_{2^{1}S}^{m_p}\ =\ -0.515 31\,a.u.\ ,$$ for the both infinite and finite proton mass, respectively. Transition energies are $$\label{1S-2S} {\Delta}E_{1^{1}S {\rightarrow}2^{1}S}^{\infty}\ =\ -0.0122\,a.u.\quad , \quad E_{1^{1}S {\rightarrow}2^{1}S}^{m_p}\ =\ -0.0121 \,a.u.\ ,$$ respectively. H${}^-$ is a “maximally” strongly-correlated Coulomb system, thus, it is an example of a general 3-body Coulomb problem, it should be treated in full generality. So far, we are unable to find accurately the wavefunction for the $2^{1}S$ state. It is evident it should [*not*]{} be of the type $(1s2s)$ as for He, Li$^+$ and other two-electron positive ions. Likely, relativistic effects (spin-orbit, spin-spin interactions, radiative effects) to energy of the lowest states (\[2S\]),(\[1S-2S\]) should not be more than $\sim 10^{-4}$a.u. similarly to ones for Helium [@Alexander:2010] ![ \[E-ground-state\] $Z$-complex plane of the ground state energy $E=E_{>}$, see text, the branch cut shown by bold wide black line; vertical line passing through $Z_c$ indicates the line of discontinuity (non-analyticity) between $E_{>}(Z)$ and $E_{<}(Z)$ (see text). ](ZbZcrZ1.eps){width="3.2in"} Note that conceptually the prediction (\[2S\]) is in agreement with D.R. Yafaev’s rigorous mathematical statement [@Yafaev:1972] about the finite number of bound states at $Z=1$ but in contradiction to the widely-known theoretical statement by R.N. Hill [@Hill; @Hill2] about the existence of a single, spin-singlet bound state of H${}^-$ only. Mathematical justification of the latter result is absent (to the best of the present author’s knowledge). It seems it is a challenge to check the prediction (\[2S\]) experimentally since the transition $1^{1}S {\rightarrow}2^{1}S$ is the forbidden type transition, hence, it should be a type of multi-photon transition and/or bound-free-bound transitions, which is difficult to observe. In principle, the spectra of H${}^-$ contains one more, but metastable spin-triplet $2^{3}P$ state [@Drake:1970]. Note that for helium atom, the main modes are (permitted) electric-dipole transitions $2^{1}S {\rightarrow}2^{1}P$ and $2^{1}P {\rightarrow}1^{1}S$, see [@Dalgarno:1966; @VanDyck:1971]. Experimental study of the (forbidden) transition $1^{1}S {\rightarrow}2^{1}S$ for helium atom was done only recently using the Doppler-free two-photon spectroscopy, see [@Eyler]. In present paper we studied the domain of bound states of three-body Coulomb problem $Z > Z_c$ and the analytic continuation to the complex $Z$-plane of the ground state $1^{1}S$, in particular, to $Z < Z_c$. We showed that the second critical charge predicted in [@Stillinger:1966] is unphysical: it appears in analytic continuation of the ground state energy to $Z < Z_c$, thus, it can not be measured. It seems important to carry out a similar study for the domain of quasi-stationary states $Z < Z_c$, in particular, checking the analytic discontinuity $Z = Z_c$. AVT thanks G. W. F. Drake (Windsor), N. Berrah, E. Eyler, W. Smith and W. Stwalley (Storrs), J.-P. Karr (Paris) for the interest to the work and useful discussions. One of us (AVT) is grateful to Stony Brook University for the kind hospitality during his sabbatical stay, where a part of the research was done, he was supported in part by PASPA program (Mexico). The final stage of the present work was carried out at Simons Center for Geometry and Physics (Stony Brook). The research by JCLV and AVT is supported in part by DGAPA grant IN108815 and CONACyT grant 166189 (Mexico). [99]{} T. Kato, [*Perturbation Theory for Linear Operators*]{},\ 2nd edition, Springer-Verlag: Berlin-Heidelberg-New York p.410-413 (1980) C.S. Estienne, M. Busuttil, A. Moini, G.W.F. Drake, [*Phys. Rev. Lett. **112***]{}, 173001 (2014) H. Olivares-Pilón, A.V. Turbiner, [*Phys. Lett. **A 379***]{} (2015) 688-690 B. Simon, [*J. Func. Anal. **25***]{}, 338-344 (1977) S. Kais and Q. Shi, [*Phys. Rev. A **62***]{} (2000) 060502(R) A. Moini, *Critical nuclear charge of quantum mechanical three-body problem*,\ Master Thesis, G.W.F. Drake (Advisor), University of Windsor, 2014 A.W. King et al, [*Phys. Rev. A **91***]{} (2015) 042512 D. Baye, *The Lagrange-mesh method*, [*Phys. Repts. **565***]{}, 1-107 (2015) A.V. Turbiner, J.C. Lopez Vieyra, H. Olivares-Pilón,\ [*From Helium-like to H${}_2^+$ like systems: Multiple nuclear critical charges*]{}, 2016 (in preparation) A.I. Baz’, Ya.B. Zel’dovich and A.M. Perelomov,\ [*Scattering, reactions and decay in nonrelativistic quantum mechanics*]{}, 2-e izd. (M.: Nauka, 1971), (in Russian), Ch.VII; [*Reactions And Decay In Nonrelativistic Quantum Mechanics*]{}, 1st ed. (Jerusalem: Israel Program for Scientific Translations, 1969) A.M. Perelomov and Y.B. Zel´dovich, *Quantum Mechanics, Selected Topics*,\ World Scientific 1998;  Ch. 8: [*Quasi-Stationary States*]{} F.H. Stillinger, [*J. Chem. Phys. **45***]{}, 3623 - 3631 (1966) F.H. Stillinger, D.K. Stillinger, [*Phys. Rev. **10***]{}, 1109-1121 (1974) B. Carballo, *About Critical Nuclear Charges for Two - Electron Atoms*,\ Talk given at *5to Taller sobre Dinamica Molecular*, Cuernavaca Mor, Mexico, July 8-10, 2014 A.V. Turbiner, H. Medel Cobaxin, [*Intern.J.Quantum Chemistry **112***]{}, 2411-2421 (2012) C.M. Bender, T.T. Wu, [*Phys. Rev. **184***]{}, 1231-1260 (1969) C.M. Bender, T.T. Wu, [*Phys. Rev. **D 7***]{}, 1620-1636 (1973) F.H. Stillinger and D.R. Herrick, [*Phys. Rev. **A11***]{}, 446 - 454 (1975) A.V. Turbiner and J.C. Lopez Vieyra, [*Can. Journal of Physics **94***]{}, 249-253 (2016) J.D. Baker, D.E. Freund, R.N. Hill, J.D. Morgan III, [*Phys. Rev. **A41***]{}, 1247-1273 (1990) H. Nakashima, H. Nakatsuji, [*J. Chem. Phys. **127***]{}, 224104 (2007) S.A. Alexander, S. Datta, R.L. Coldwell, [*Phys. Rev. A **81***]{} (2010) 032519 D.R. Yafaev, [*Funct.Anal. Appl. **6***]{}, 349-353 (1972) R.N. Hill, [*Phys. Rev. Lett. **38***]{}, 643-646 (1977) R.N. Hill, [*J. Math. Phys. **18***]{}, 2316-2330 (1977) G.W.F. Drake, [*Phys. Rev. Lett. **24***]{}, 126-127 (1970) A. Dalgarno, [*Mon. Not. R. astr. Soc. **131***]{}, 311-313 (1966) R.S. Van Dyck, C.E. Johnson, H.A. Shugart, [*Phys. Rev. **A4***]{}, 1327 - 1336 (1971) S.D. Bergeson et al, [*Phys. Rev. Lett. **80***]{}, 3475-3478 (1998) [^1]: The same is correct for the function (\[GSB\]), see below [^2]: Recently, this unexpected statement was supported in highly accurate variational study [@Drake:2014]: for $Z$ slightly smaller that $Z_c$, $Z < Z_c$ normalizable trial functions guarantee convergence in energy up to 14 decimal figures without loosing normalizability!

ArXiv

--- abstract: 'In [@PKVK] we announced the complete integrability of geodesic motion in the general higher-dimensional rotating black-hole spacetimes. In the present paper we prove all the necessary steps leading to this conclusion. In particular, we demonstrate the independence of the constants of motion and the fact that they Poisson commute. The relation to a different set of constants of motion constructed in [@KKPF] is also briefly discussed.' author: - Pavel Krtouš - David Kubizňák - 'Don N. Page' - Muraari Vasudevan date: 'June 28, 2007' title: 'Constants of Geodesic Motion in Higher-Dimensional Black-Hole Spacetimes' --- Introduction {#sc:intro} ============ Spacetimes of higher dimensions (${D>4}$) have become much studied as a result of their appearance in theories of unification, such as string/M theory. Of such spacetimes, one important class is a sequence of black-hole metrics of greater and greater generality in higher dimensions that have been discovered over the years. The first such higher-dimensional black-hole spacetime was the metric for a nonrotating black hole in ${D>4}$ (the generalization of the 1916 Schwarzschild metric in four dimensions [@Schw]), found in 1963 by Tangherlini [@Tang]. Next was the metric for a rotating black hole in higher dimensions (the generalization of the 1963 Kerr metric in four dimensions [@Kerr]), discovered in 1986 by Myers and Perry [@MP] in the case with zero cosmological constant. Then in 1998 Hawking, Hunter, and Taylor-Robinson [@HHT] found the general ${D\!=\!5}$ version of the ${D\!=\!4}$ rotating black hole with a cosmological constant (often called the Kerr–de Sitter metric) that had been found in 1968 by Carter [@Carter1; @Carter2]. In 2004 Gibbons, Lü, Page, and Pope [@GLPP1; @GLPP2] discovered the general Kerr–de Sitter metrics in all higher dimensions, and in 2006 Chen, Lü, and Pope [@CLP] put these into a simple form similar to Carter’s and were able to add a NUT [@NUT] parameter (though not charge) to get the general Kerr–NUT–(A)dS metrics for all $D$. It is important to study the properties of these higher-dimensional black-hole spacetimes, and one key property is the nature of geodesic motion in them. In [@PKVK; @KKPF] we exhibited $D$ constants of geodesic motion and announced that they are all independent (making the geodesic motion integrable) and that the Poisson brackets of any pair of them vanish (making the integrable geodesic motion completely integrable). In this paper we shall prove these assertions. After introducing the metric and its basic symmetries, we recapitulate our construction of constants of geodesic motion and show how these constants can be generated from a generating function. The two main proofs demonstrating the independence and the Poisson commutativity of these constants follow. The canonical formalism used in the text is reviewed in the Appendix. We type tensors in boldface with components in normal letters. The spacetime indices are denoted by Latin letters from the beginning of the alphabet, ${a,b,c=1,\dots,D}$, and we use the Einstein summation convention for them. For a rank-2 tensor ${{{\boldsymbol{B}}}}$ the symbol ${B}$ stands for the matrix of its components ${B^{a}_{\ b}}$. Where it cannot lead to a confusion a dot indicates contraction, i.e., ${{{\boldsymbol{a}}}\cdot{{\boldsymbol{b}}}=a^e b_e}$. We assume automatic lowering and raising of indices using the metric. ${{{{{\boldsymbol{\partial}}}}}_{x^a}}$ stands for the coordinate vector associated with the coordinate ${x^a}$. Higher-dimensional black-hole spacetimes {#sc:HDBH} ======================================== The general Kerr–NUT–(anti-)de Sitter spacetime discovered by Chen, Lü, and Pope may, after a suitable Wick rotation of the radial coordinate, be written [@CLP] $$\label{metric_coordinates} {{\boldsymbol{g}}}\!\!=\!\!\sum_{\mu=1}^n\Bigl[\frac{{{{{\boldsymbol{d}}}}}x_{\mu}^2}{Q_{\mu}} +Q_{\mu}\!\Bigl(\sum_{k=0}^{n-1} A_{\mu}^{(k)}{{{{\boldsymbol{d}}}}}\psi_k\!\Bigr)^{\!2}\Bigr] \!-\!\frac{{\varepsilon}c}{A^{(n)}}\Bigl(\sum_{k=0}^n A^{(k)}{{{{\boldsymbol{d}}}}}\psi_k\!\Bigr)^{\!2} \!,$$ with ${n=\lfloor D/2\rfloor}$ and ${\varepsilon=D-2n}$. Here, $Q_{\mu}={X_{\mu}/}{U_{\mu}}\,,$ $$\begin{gathered} U_{\mu}=\prod_{\substack{\nu=1\\\nu\ne\mu}}^{n}(x_{\nu}^2-x_{\mu}^2)\;,\;\;\; X_{\mu}=\sum\limits_{k=\varepsilon}^{n}c_kx_{\mu}^{2k}-2b_{\mu}x_{\mu}^{1-\varepsilon}+\frac{\varepsilon c}{x_{\mu}^2}\;, \nonumber\\ A_{\mu}^{(k)}=\!\!\!\!\!\sum_{\substack{\nu_1<\dots<\nu_k\\\nu_i\ne\mu}}\!\!\!\!\!x^2_{\nu_1}\dots x^2_{\nu_k}\;,\quad A^{(k)}=\!\!\!\!\!\sum_{\nu_1<\dots<\nu_k}\!\!\!\!\!x^2_{\nu_1}\dots x^2_{\nu_k}\;\label{co}.\end{gathered}$$ The coordinates ${x_\mu}$ (${\mu=1,\dots,n}$) correspond to radial and latitude directions, ${\psi_k}$ ( ${k=0,\dots,n-1+{\varepsilon}}$) to temporal and azimuthal directions. The parameter $c_n$ is proportional to the cosmological constant, and the remaining constants $c_k$, $c$ and $b_{\mu}$ are related to the rotation parameters, the mass and the NUT parameters. Hamamoto, Houri, Oota and Yasui [@KNAcurv] derived explicit formulas for the curvature and demonstrated that in all dimensions this metric obeys the Einstein equations $${R_{ab}=(-1)^{n}(D-1)c_n\, g_{ab}}\;.$$ Besides the obvious spacetime isometries generated by the $D-n$ Killing vectors ${{\boldsymbol{\partial}}}_{\psi_k}$, the spacetime possesses a whole set of hidden symmetries [@PKVK; @KKPF], which can be generated from the principal (rank-2 closed) conformal Killing–Yano tensor discovered by Kubizň' ak and Frolov [@KF]. These hidden symmetries play the crucial role for the integrability of the geodesic motion. The metric (\[metric\_coordinates\]) can be diagonalized. Let us introduce the orthonormal basis one-forms $$\begin{aligned} \label{one-forms} {{{{\boldsymbol{e}}}}}^\mu &=& Q_{\mu}^{-1/2} dx_{\mu}\;, \nonumber\\ {{{{\boldsymbol{e}}}}}^{\hat \mu}&=&{{{{\boldsymbol{e}}}}}^{n+\mu} = Q_{\mu}^{1/2} \sum_{k=0}^{n-1}A_{\mu}^{(k)}d\psi_k\;, \nonumber\\ {{{{\boldsymbol{e}}}}}^{2n+1} &=& (-c/A^{(n)})^{1/2} \sum_{k=0}^nA^{(k)}d\psi_k\;.\end{aligned}$$ Then we have $$\label{metric} {{\boldsymbol{g}}}=\sum_{a=1}^D{{{{\boldsymbol{e}}}}}^a {{{{\boldsymbol{e}}}}}^a=\sum_{\mu=1}^n \bigl({{{{\boldsymbol{e}}}}}^\mu{{{{\boldsymbol{e}}}}}^\mu + {{{{\boldsymbol{e}}}}}^{\hat\mu}{{{{\boldsymbol{e}}}}}^{\hat\mu}\bigr) + \varepsilon\, {{{{\boldsymbol{e}}}}}^{2n+1}{{{{\boldsymbol{e}}}}}^{2n+1}\;,$$ and the principal conformal Killing–Yano tensor ${{\boldsymbol{h}}}$ which obeys the equations $$\label{cKYprop} (D-1)\nabla_{\!a} {h}_{bc}=g_{ab}\xi_c-g_{ac}\xi_b\;,\ \ \xi_a = \nabla_{\!c}{h}^c{}_a\;,$$ takes the extremely simple form $$\label{cKYdef} {{\boldsymbol{{h}}}}=\sum_{\mu=1}^n x_\mu {{{{\boldsymbol{e}}}}}^\mu \wedge {{{{\boldsymbol{e}}}}}^{\hat\mu}\;.$$ In what follows we shall also use the conformal Killing tensor $$\label{Qdef} Q=-{h}{h}\;,\quad\text{i.e.,}\quad Q_{ab}={h}_{ac}{h}_{bd}\,g^{cd}\;,$$ which takes the explicit form $$\label{Qexpl} {{\boldsymbol{Q}}}=\sum_{\mu=1}^nx_{\mu}^2({{{{\boldsymbol{e}}}}}^\mu{{{{\boldsymbol{e}}}}}^\mu + {{{{\boldsymbol{e}}}}}^{\hat\mu}{{{{\boldsymbol{e}}}}}^{\hat\mu})\;,$$ and satisfies $\nabla_{\!(a} Q_{bc)}=g_{(ab}Q_{c)}\,,$ where $$Q_a =\frac{1}{D+2}(2\nabla_{\!c}\,Q^c_{\ a}+\nabla_{\!a}Q^c_{\ c}).$$ Constants of motion {#sc:cm} =================== In [@PKVK] we have claimed that in the spacetime there are ${D}$ independent constants of geodesic motion, given by the following quantities: (a) $n-1$ observables ${C}_j$, ${j=1,\dots,n-1}$, given by traces of powers of the projection ${{\boldsymbol{F}}}$ of the principal conformal Killing-Yano tensors ${{\boldsymbol{{h}}}}$ (cf. Eqs.  and below) $$\label{cpdef} {C}_j = \operatorname{\mathrm{tr}}\bigl[(-w^{-1}F^2)^j\bigr]\;,$$ (b) ${D-n}$ observables ${{p}_j}$, ${j=0,\dots,D-n-1}$, given by symmetries of the spacetime $$\label{cldef} {p}_j={{\boldsymbol{u}}}\cdot{{{{\boldsymbol{\partial}}}}}_{\psi_j}\;,$$ and (c) the square ${w}$ of the (unnormalized) velocity ${{{\boldsymbol{u}}}}$ $$\label{wdef} w={{\boldsymbol{u}}}\cdot{{\boldsymbol{u}}}=u^a u_a\;.$$ Moreover, these quantities commute in the sense of Poisson brackets on the phase space. Here we want to elucidate and prove these properties in more detail. We understand all mentioned quantities as observables (i.e., functions) on the phase space ${{{\boldsymbol{\Gamma}}}={{\mathbf{T}^{*}M}}}$. It is well known that the cotangent space ${{{\mathbf{T}^{*}M}}}$ has a natural phase space structure (cf. the Appendix or [@Arnold:book]). Since we investigate the relativistic theory and ${M}$ is a spacetime manifold describing also the physical temporal direction, the phase space ${{{\boldsymbol{\Gamma}}}={{\mathbf{T}^{*}M}}}$ is an unphysical phase space which is, however, well suited for an investigation of the geodesic motion. Doing canonical mechanics on it allows us to solve the geodesic motion in an external time which can be identified at the end with the affine parameter of the studied geodesic. We denote the momentum variable on the cotangent space as ${{{\boldsymbol{u}}}}$. Indeed, since the geodesic motion is governed by the Lagrange function ${L=\frac12\,{{\boldsymbol{u}}}\cdot{{\boldsymbol{u}}}=\frac12\, u^a u^b\, g_{ab}}$, the canonical momentum can be (up to a position of the tensor index) identified with the (unnormalized) velocity ${{{\boldsymbol{u}}}}$. The Hamiltonian then is $$\label{Ham} H=\frac12\,w=\frac12\,{{\boldsymbol{u}}}\cdot{{\boldsymbol{u}}}=\frac12\,u_a u_b\, g^{ab}\;.$$ We easily realize that ${{p}_j}$ defined in are the special components of momentum and that they are constants of motion since ${{{{\boldsymbol{\partial}}}}}_{\psi_j}$ are Killing vectors. The quantities ${C}_j$, eq. , are constants of motion because the tensor ${{\boldsymbol{F}}}$, defined as $$\label{Fdef} F_{ab}=({h}_{ab}u_c+{h}_{bc}u_a+{h}_{ca}u_b)\;u^c ,$$ and the square ${w}$ of the velocity ${{{\boldsymbol{u}}}}$, are covariantly conserved along the geodesic. Indeed, thanks to , for ${u^c\nabla_{\!c} u^a=0}$ we have $u^c\nabla_{\!c} F_{ab}=0$. Next we express the constants of motion ${{C}_j}$ in terms of the quantities related to the principal conformal Killing-Yano tensor ${{{\boldsymbol{{h}}}}}$. The components of the tensor can be rewritten as $$\label{F=PKP} w^{-1}F=P{h}P\;,$$ where ${{{\boldsymbol{P}}}}$ is the projector orthogonal to the velocity ${{{\boldsymbol{u}}}}$, ${P=I-p}$, i.e., ${P^a_{b}=\delta^a_{b}-p^a_{b}}$. Here we also introduced the projector ${{{\boldsymbol{p}}}}$, $$\label{Pdef} p^a_{b}=w^{-1} u^a u_b\;,$$ onto the direction ${{{\boldsymbol{u}}}}$. Using the cyclic property of the trace we thus have $$\label{Cexpr} {C}_j=(-1)^j w^j \operatorname{\mathrm{tr}}\bigl[({h}P)^{2j}\bigr]\;.$$ The trace of the matrix product could be viewed diagrammatically as a loop formed by joined vertices (each with two ‘legs’) corresponding to matrices in the product. In our case the loop is formed by alternating ${{h}}$ and ${P}$ vertices. Substituting ${P=I-p}$ we get a sum over all possible loops in which ${P}$ is replaced either by ${I}$ or by ${-p}$. In the case of the identity ${I}$ the corresponding vertex is effectively eliminated, and in the case of the one-dimensional projector ${{{\boldsymbol{p}}}=w^{-1}{{\boldsymbol{u}}}{{\boldsymbol{u}}}}$ the loop splits into disconnected pieces. Namely, we can use the identity $$\operatorname{\mathrm{tr}}\bigl({h}^{k_1}\!p\,{h}^{k_2}\!p\,\cdots{h}^{k_c}\!p\bigr)= \operatorname{\mathrm{tr}}\bigl({h}^{k_1}\! p\bigr)\operatorname{\mathrm{tr}}\bigl({h}^{k_2}\! p\bigr)\cdots\operatorname{\mathrm{tr}}\bigl({h}^{k_c}\! p\bigr)\;.$$ The trace in thus leads to $$\label{trhPj} \begin{split} &\operatorname{\mathrm{tr}}\bigl[({h}P)^{2j}\bigr]=\operatorname{\mathrm{tr}}\bigl({h}^{2j}\bigr)\\ &\qquad+\sum_{c=1}^{2j}\sum_{\substack{k_1\le \dots\le k_c\\k_1+\dots+k_c=2j}} \!\!\!(-1)^c N^{2j}_{k_1\dots k_j} \prod_{i=1}^c\operatorname{\mathrm{tr}}\bigl({h}^{k_i}\!p\bigr)\;. \end{split}$$ The sum over ${c}$ is the sum over the number of ‘splits’ of the loop, the indices ${k_i}$ are the ‘lengths’ of the split pieces, and the combinatorial factor ${N^{2j}_{k_1\dots k_c}}$ gives the number of ways in which the loop of the length ${2j}$ can be split to ${c}$ pieces of lengths ${k_1,\dots,k_c}$. From the fact that the tensor ${{{\boldsymbol{{h}}}}}$ is antisymmetric, it follows that traces of odd powers of ${{h}}$ (optionally multiplied by a projector) are zero. Setting ${k_i=2l_i}$ and introducing the rank-2 conformal Killing tensor ${{{\boldsymbol{Q}}}}$ from , eq.  thus reduces to $$\label{trhPjinQ} \begin{split} &\operatorname{\mathrm{tr}}\bigl[(-{h}P{h}P)^{j}\bigr]=\operatorname{\mathrm{tr}}\bigl(Q^j\bigr)\\ &\qquad+\sum_{c=1}^j\sum_{\substack{l_1\le \dots\le l_c\\l_1+\dots+l_c=j}} \!\!\!(-1)^c\, 2\,N^j_{l_1\dots l_j} \prod_{i=1}^c\operatorname{\mathrm{tr}}\bigl(Q^{l_i}\!p\bigr)\;, \end{split}$$ where we used ${N^{2j}_{2l_1\dots2l_c}=2N^j_{l_1\dots l_c}}$ which follows from the definition of the ${N}$’s. If we define the quantities $$\label{wjdef} w_j = w \operatorname{\mathrm{tr}}(Q^j p) = u_{a_0} Q^{a_0}_{a_1}Q^{a_1}_{a_2}\cdots Q^{a_{j-1}}_{a_j}u^{a_j},$$ we finally obtain $$\label{trhPjinw} \begin{split} {C}_j&=w^j\operatorname{\mathrm{tr}}\bigl[(-{h}P{h}P)^{j}\bigr]\\ &=w^j\operatorname{\mathrm{tr}}\bigl(Q^j\bigr) +2\sum_{c=1}^j\sum_{\substack{l_1\le \dots\le l_c\\l_1+\dots+l_c=j}} \!\!\!(-1)^c \,N^j_{l_1\dots l_j} w^{j-c}\prod_{i=1}^c w_{l_i}\;, \end{split}$$ which is eq. (17) of [@PKVK]. Let us note that by the same argument as that leading to eq. , we can derive the relation for the trace of a power of ${QP}$, $$\label{trQPj} \begin{split} &\operatorname{\mathrm{tr}}\bigl[(Q P)^j\bigr]=\operatorname{\mathrm{tr}}\bigl(Q^j\bigr)\\ &\qquad+\sum_{c=1}^j\sum_{\substack{l_1\le \dots\le l_c\\l_1+\dots+l_c=j}} \!\!\!(-1)^c N^j_{l_1\dots l_j} \prod_{i=1}^c\operatorname{\mathrm{tr}}\bigl(Q^{l_i}\!p\bigr)\;, \end{split}$$ Comparing with eq. , we see that we have proved the relation (16) of [@PKVK], $$\label{trtrtr} \operatorname{\mathrm{tr}}\bigl[(-{h}P{h}P)^{j}\bigr]+\operatorname{\mathrm{tr}}\bigl[Q^j\bigr]=2\operatorname{\mathrm{tr}}\bigl[(Q P)^j\bigr]\;.$$ The relation and an algorithm for computing the coefficients ${N^j_{l_1\dots l_c}}$ can be derived also in a different way. It was mentioned in [@KKPF] that the constants ${{C}_j}$ can be generated from the generating function ${Z(\beta)=\log W(\beta)}$: $$\begin{split} Z(\beta)&=\sum_{j=1}^\infty\frac{(-1)^{j+1}}{2j}\frac{\beta^j}{w^j}{C}_j =-\sum_{j=1}^\infty\frac{1}{2j}\beta^j\operatorname{\mathrm{tr}}\bigl[({h}P)^{2j}\bigr]\\ &=\operatorname{\mathrm{tr}}\log\bigl(I\!-\!\sqrt\beta\,{h}P\bigr) =\log\det{}\bigl(I\!-\!\sqrt\beta\,{h}P\bigr)\;. \end{split}$$ The third equality follows from the antisymmetry of ${{h}}$. Using properties of the determinant, the antisymmetry of ${{h}}$, ${I=P+p}$, and the fact that the projector ${p}$ is one dimensional, we can split ${Z(\beta)}$ into two pieces (cf. eq. (2.7) and (2.8) of [@KKPF]): $$\begin{aligned} Z(\beta) &= \log W_0(\beta) + \log\Sigma(\beta)\;,\\ W_0(\beta)&=\det\bigl(I\!-\!\sqrt\beta\,{h}\bigr) =\det{}^{\!\frac12\!}\bigl(I\!+\!\beta\,Q\bigr)\;,\\ \Sigma(\beta)&=\det\bigl(P+(I-\sqrt\beta\,{h})^{\!-\!1}p\bigr)\\ &=\operatorname{\mathrm{tr}}\bigl((I-\sqrt\beta\,{h})^{\!-\!1}p\bigr) =\operatorname{\mathrm{tr}}\bigl((I+\beta\,Q)^{\!-\!1}p\bigr)\;. \end{aligned}$$ Equation then corresponds to the term proportional to ${\beta^j}$ in the power expansion of ${Z(\beta)}$. The first term of is obtained from ${\log W_0(\beta)}$, and the sum over all possible splittings of the loop corresponds to the ${\beta^j}$ term of ${\log\Sigma(\beta)}$. Clearly, the ${j}$-th derivative of ${\log\Sigma(\beta)}$ (evaluated at ${\beta=0}$) contains the sum over all possible products of ${l}$-th derivatives ${\Sigma^{(l)}(0)}$ which are proportional to ${w_l}$ defined in . The factors ${N^j_{l_1\dots l_2}}$ can thus be obtained by the explicit computation of the derivatives of the generating function ${\log\Sigma(\beta)}$: $$\label{Ck} \begin{split} {C}_j&=w^j\operatorname{\mathrm{tr}}\bigl(Q^j\bigr)\\ &-\frac{2(-w)^{j}}{(j-1)!}\frac{d^j}{d\beta^j}\log \Bigr(1+\sum_{k=1}^j(-1)^k\frac{w_k}{w}\beta^k\Bigl)\Big|_{\beta=0}\;. \end{split}$$ Using software for algebraic manipulation we easily get the first five constants (sufficient for the integrability of geodesic motion up through $D=13$): $$\begin{aligned} {C}_1 &= w \operatorname{\mathrm{tr}}Q - 2 w_1\;,\notag\\ {C}_2 &= w^2 \operatorname{\mathrm{tr}}Q^2 - 4 w\, w_2 + 2 w_1^2\;,\notag\\ {C}_3 &= w^3 \operatorname{\mathrm{tr}}Q^3 - 6 w^2 w_3 + 6 w\, w_1 w_2 - 2 w_1^3\;,\notag\\ {C}_4 &= w^4 \operatorname{\mathrm{tr}}Q^4 - 8 w^3 w_4 + w^2 (4w_2^2+8w_1w_3)\label{c1-5}\\ &\qquad-8 w\, w_1^2w_2+2w_1^4\;,\notag\\ {C}_5 &= w^5 \operatorname{\mathrm{tr}}Q^5 - 10 w^4 w_5 + w^3 (10w_2w_3+10w_1w_4)\notag\\ &\qquad-w^2(10w_1w_2^2+10w_1^2w_3)+10w\,w_1^3w_2-2w_1^5\;.\notag\end{aligned}$$ Taking into account the facts that the eigenvalues of the principal conformal Killing-Yano tensor ${{{\boldsymbol{{h}}}}}$ are given by the coordinates ${x_\mu}$, cf. eq. , respectively, that the eigenvalues of ${{{\boldsymbol{Q}}}}$ are ${x_\mu^2}$, see eq. , we can write down an explicit form for ${\operatorname{\mathrm{tr}}Q^j}$ and ${w_j}$: $$\label{trQjinx} \operatorname{\mathrm{tr}}Q^j = 2 \sum_{\mu=1}^n x_\mu^{2j}\;,$$ $$\label{wjinx} w_j = \sum_{\mu=1}^n x_\mu^{2j} (u_\mu^2+u_{\hat\mu}^2)\;.$$ Let us also point out that on the level of the generating functions the relation corresponds to $$\label{detdetdet} \det\bigl(I-\beta\,{h}P{h}P\bigr)\,\det\bigl(I+\beta\,Q\bigr) =\det{}^{\!2}\bigl(I+\beta\, QP\bigr)\;.$$ It was realized in [@KKPF] that the generating function ${W(\beta)=\exp Z(\beta)}= W_0(\beta)\Sigma(\beta)$ actually generates another set of conserved quantities ${{c}_j}$ by $$\label{quadrc} W(\beta) = \frac1w\sum_{j=0}^\infty {c}_j\beta^j\;,$$ which are quadratic in the velocity ${{{\boldsymbol{u}}}}$. (That they are quadratic can be seen from the fact that ${W_0}$ does not depend on the velocity, from eq. , and from ${w \Sigma(\beta) = \sum_{j=0}^\infty (-1)^j w_j \beta^j}$.) The relation between ${W(\beta)}$ and ${Z(\beta)}$ implies that $$\label{Ckck} {C}_j=-\frac{2(-w)^{j}}{(j-1)!}\frac{d^j}{d\beta^j} \log\Bigl(w+\sum_{k=1}^jc_k\beta^k\Bigr)\Big|_{\beta=0}\;,$$ and in particular: $$\begin{aligned} {C}_1&=\;\;\,2{c}_1\;,\notag\\ {C}_2&=-4w{c}_2+2{c}_1^2\;,\notag\\ {C}_3&=\;\;\,6w^2{c}_3-6w{c}_1{c}_2+2{c}_1^3\;,\label{cCrel}\\ {C}_4&=-8w^3{c}_4+8w^2{c}_1{c}_3+4w^2{c}_2^2-8w{c}_1^2{c}_2+2{c}_1^4\;,\notag\\ {C}_5&=\;10w^4{c}_5-10w^3{c}_1{c}_4-10w^3{c}_2{c}_3\notag\\ &\qquad+10w^2{c}_1^2{c}_3+10w^2{c}_1{c}_2^2-10w{c}_1^3{c}_2+2{c}_1^5\;,\notag\end{aligned}$$ which are the inverse of the relations (3.19) of [@KKPF]. Independence of constants of motion =================================== Now we can demonstrate that the quantities ${w}$, ${{p}_j}$ and ${{C}_j}$ are independent at a generic point of the phase space ${{{\boldsymbol{\Gamma}}}={{\mathbf{T}^{*}M}}}$. This means that their gradients on the phase space are linearly independent. To prove this, it is sufficient to show that these gradients are independent in the vertical direction of the cotangent bundle ${{{\mathbf{T}^{*}M}}}$, i.e., that the derivatives of these quantities with respect to the momentum ${{{\boldsymbol{u}}}}$ are linearly independent. To achieve this we will study the wedge product of the ‘vertical’ derivatives. Let us, instead of ${w}$ and ${{C}_j}$, consider the equivalent set of observables ($j=1,\dots,n-1$) $$\label{cscal1} \begin{aligned} 2\tilde{{C}}_j&=-\frac1{2j}\,w^{1-j}{C}_j=-\frac{1}{2j}\,w\operatorname{\mathrm{tr}}{{Q}^j}+w_j+\dots\;,\\ 2\tilde{{C}}_0&=w\;, \end{aligned}$$ where dots in the first expression denote terms which contain $w_k$ with $k<j$, cf. eqs. , . We are interested in the quantity[^1] $$\label{jac} {{\boldsymbol{J}}} = {{{{\boldsymbol{\partial}}}}}\tilde{{C}}_0\wedge\dots\wedge{{{{\boldsymbol{\partial}}}}}\tilde{{C}}_{n\!-\!1}\wedge {{{{\boldsymbol{\partial}}}}}{p}_0\wedge\dots\wedge{{{{\boldsymbol{\partial}}}}}{p}_{D\!-\!n\!-\!1}\;.$$ Due to and , we have $ {{{{\boldsymbol{\partial}}}}}{p}_j={{{{\boldsymbol{\partial}}}}}_{\psi_j}\,,$ and $$\label{pderofcj} {{{{\boldsymbol{\partial}}}}}\tilde{C}_j=-\frac{1}{2j}\,\bigl(\operatorname{\mathrm{tr}}{Q^j}\bigr)\;{{\boldsymbol{u}}}+{{\boldsymbol{Q}}}^j\cdot{{\boldsymbol{u}}}+ \dots\;,$$ where dots denote linear combinations of ${{{\boldsymbol{Q}}}^k\cdot{{\boldsymbol{u}}}}$ with ${k<j}$; ${{{\boldsymbol{Q}}}^l\cdot{{\boldsymbol{u}}}}$ represents the vector with components ${Q^a_{a_1}Q^{a_1}_{a_2}\cdots Q^{a_{l\!-\!1}}_{a_j}u^{a_l}}$. From the antisymmetry of the wedge product it follows that $$\label{jacinu} {{\boldsymbol{J}}}={{\boldsymbol{u}}}\wedge({{\boldsymbol{Q}}}\cdot{{\boldsymbol{u}}})\wedge\dots\wedge({{\boldsymbol{Q}}}^{n\!-\!1}\cdot{{\boldsymbol{u}}})\wedge {{{{\boldsymbol{\partial}}}}}_{\psi_0}\wedge\dots\wedge{{{{\boldsymbol{\partial}}}}}_{\psi_{D\!-\!n\!-\!1}}\;.$$ (Matrix) powers ${{{\boldsymbol{Q}}}^j}$ of the conformal Killing tensor can be written as $$\label{powQ} {{\boldsymbol{Q}}}^j = \sum_{\mu=1}^n x_\mu^{2j}{{{{\boldsymbol{e}}}}}_\mu{{{{\boldsymbol{e}}}}}^\mu+\sum_{\mu=1}^n x_\mu^{2j}{{{{\boldsymbol{e}}}}}_{\hat\mu}{{{{\boldsymbol{e}}}}}^{\hat\mu}\;.$$ The second term acts on the subspace of the vectors spanned on ${{{{{\boldsymbol{\partial}}}}}_{\psi_j}}$. Thus, thanks to the ${{{{{\boldsymbol{\partial}}}}}_{\psi_0}\wedge\dots\wedge{{{{\boldsymbol{\partial}}}}}_{\psi_{D\!-\!n\!-\!1}}}$ term in the wedge product, this part can be ignored in . Taking into account that ${{{{{\boldsymbol{e}}}}}_\mu{{{{\boldsymbol{e}}}}}^\mu = {{{{\boldsymbol{\partial}}}}}_{x_\mu}{{{{\boldsymbol{d}}}}}x_\mu}$ and ${u^\mu={{{{\boldsymbol{d}}}}}x_\mu\cdot{{\boldsymbol{u}}}}$, the substitution of into leads to $$\label{jacfin} {{\boldsymbol{J}}} = u^1 \dots u^n\, U\, {{{{\boldsymbol{\partial}}}}}_{x_1}\wedge\dots\wedge{{{{\boldsymbol{\partial}}}}}_{x_n}\wedge{{{{\boldsymbol{\partial}}}}}_{\psi_0}\wedge\dots\wedge{{{{\boldsymbol{\partial}}}}}_{\psi_{D\!-\!n\!-\!1}}\;,$$ where $$\label{Udef} U = \mspace{-20mu}\sum_{\substack{\text{permutations ${\sigma}$}\\\text{of ${[0,\dots, n\!-\!1]}$}}}\mspace{-20mu} \operatorname{\mathrm{sign}}\sigma\; x_1^{2\sigma_1}\dots x_n^{2\sigma_n} = \mspace{-10mu}\prod_{\substack{\mu,\nu=1\dots n\\\nu<\mu}}\mspace{-8mu} (x_\mu^2-x_\nu^2)\;.$$ In a generic point of the phase space we have ${u^j\ne0}$ and ${x_\mu^2\ne x_\nu^2}$ (for ${\mu\ne\nu}$) and therefore ${{{\boldsymbol{J}}}\ne0}$ there, thus showing that the constants of motion are independent. Poisson brackets {#sc:pb} ================ Finally we show that the observables ${w}$, ${{C}_j}$, and ${{p}_j}$ Poisson commute on the phase space. The Poisson bracket of two functions on the phase space ${{{\boldsymbol{\Gamma}}}={{\mathbf{T}^{*}M}}}$ can be written as $$\label{PBdef} \{A,B\}={{{{\boldsymbol{\nabla}}}}}A \cdot {{{{\boldsymbol{\partial}}}}}B - {{{{\boldsymbol{\partial}}}}}A \cdot {{{{\boldsymbol{\nabla}}}}}B\;,$$ where ${{{{{\boldsymbol{\nabla}}}}}F}$ represents an arbitrary (torsion-free) covariant derivative which ignores the dependence of ${F}$ on the momentum ${{{\boldsymbol{u}}}}$, and ${{{{{\boldsymbol{\partial}}}}}B}$ is the derivative of ${B}$ with respect to the momentum ${{{\boldsymbol{u}}}}$, cf. the Appendix. ${{{{{\boldsymbol{\nabla}}}}}F}$ is a 1-form and ${{{{{\boldsymbol{\partial}}}}}F}$ a vector field on the spacetime ${M}$, and the dot indicates the contraction in spacetime tensor indices. We use naturally the covariant derivative ${{{{{\boldsymbol{\nabla}}}}}}$ generated by the metric on ${M}$. Clearly, the commutation of any observable with the Hamiltonian ${\frac12w}$ of the geodesic motion is equivalent to the conservation of the observable, cf. eq. , so we have $$\label{PBw} \{w,{p}_j\}=0\;,\quad\{w,{C}_j\}=0\;.$$ The Poisson bracket between observables ${{p}_j={{\boldsymbol{u}}}\cdot{{{{\boldsymbol{\partial}}}}}_{\psi_j}}$ reduces to Lie brackets of the Killing vector fields ${{{{{\boldsymbol{\partial}}}}}_{\psi_j}}$, which vanish because ${{{{{\boldsymbol{\partial}}}}}_{\psi_j}}$ are coordinate vector fields: $$\label{PBclcl} \begin{split} \{{p}_i,{p}_j\} &= {{{{\boldsymbol{\partial}}}}}_{\psi_j}\cdot({{{{\boldsymbol{\nabla}}}}}{{{{\boldsymbol{\partial}}}}}_{\psi_i})\cdot{{\boldsymbol{u}}} -{{{{\boldsymbol{\partial}}}}}_{\psi_i}\cdot({{{{\boldsymbol{\nabla}}}}}{{{{\boldsymbol{\partial}}}}}_{\psi_j})\cdot{{\boldsymbol{u}}}\\ &= [{{{{\boldsymbol{\partial}}}}}_{\psi_j},{{{{\boldsymbol{\partial}}}}}_{\psi_i}]\cdot{{\boldsymbol{u}}}=0\; \end{split}$$ The Poisson bracket of any observable with the observable ${p={{\boldsymbol{l}}}\cdot{{\boldsymbol{u}}}}$ linear in momentum leads to the Lie derivative along the vector field ${{{\boldsymbol{l}}}}$, see eq. : $$\label{PBclcp} \{{C}_i,{p}_j\} ={\pounds}_{{{{{\boldsymbol{\partial}}}}}_{\psi_j}}{C}_i=0\;.$$ Here, the Lie derivative ${{\pounds}_{{{{{\boldsymbol{\partial}}}}}_{\psi_j}}{C}_i}$ ignores the dependence of ${{C}_i}$ on the momentum ${{{\boldsymbol{u}}}}$, cf. the Appendix. It vanishes because ${{{{{\boldsymbol{\partial}}}}}_{\psi_j}}$ is a Killing vector and the definition of ${{C}_i}$ respects the symmetry of the spacetime (it does not depend explicitly on ${\psi_j}$). Finally, it remains to evaluate the brackets ${\{{C}_i,{C}_j\}}$. To simplify the following computation, we will study rescaled observables[^2] $$\label{cscal2} \tilde{C}_j=(-1)^jw^{j}{C}_j=\operatorname{\mathrm{tr}}\bigl[({h}\, \tilde{P})^{2j}\bigr]\;,$$ cf. eq. , and we denote $$\label{tPdef} \tilde{{{\boldsymbol{P}}}}=w{{\boldsymbol{P}}}=w{{\boldsymbol{I}}}-{{\boldsymbol{u}}}{{\boldsymbol{u}}}\;.$$ Using the cyclic property of the trace, the derivative of ${\tilde{C}_j}$ in the spacetime direction is $$\nabla_a\tilde{C}_j=2j\operatorname{\mathrm{tr}}\bigl[(\nabla_a{h}) \tilde{P} ({h}\tilde{P})^{2j-1}\bigr]\;.$$ Here ${\nabla_a{h}}$ is the matrix of components ${\nabla_a{h}^b{}_c}$ of the covariant derivative ${{{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{{h}}}}}$. Substituting for ${\nabla_a{h}^b{}_c}$ from eq.  and using the antisymmetry of ${{h}}$, we obtain $$\label{covdcp} \begin{split} &{\textstyle\frac{D-1}{2j}}\nabla_e\tilde{C}_j\\ &\quad=\xi_{a_0}\tilde{P}^{a_0}_{\,b_1}{h}^{b_1}{}_{\!a_1}\tilde{P}^{a_1}_{\,b_2} \dots{h}^{b_{2j-2}}{}_{\!\!\!a_{2j-1}}\tilde{P}^{a_{2j-1}}_{\,e}\\ &\qquad-g_{ea_{2j}}\tilde{P}^{a_{2j}}_{\,b_{2j-1}}{h}^{b_{2j-1}}{}_{\!\!\!a_{2j-1}} \dots{h}^{b_1}{}_{\!a_1}\tilde{P}^{a_1}_{b_0}\xi^{b_0}\\ &\quad=2\,\xi_{a_0}\tilde{P}^{a_0}_{\,b_1}{h}^{b_1}{}_{\!a_1}\tilde{P}^{a_1}_{\,b_2} \dots{h}^{b_{2j-2}}{}_{\!\!\!a_{2j-1}}\tilde{P}^{a_{2j-1}}_{\,e}\;. \end{split}$$ For the derivative with respect of the momentum ${{{\boldsymbol{u}}}}$ we get $$\label{pdercp} \begin{split} {\textstyle\frac{1}{4j}}\partial^e \tilde{C}_j &=u^e\; \bigl({h}^{d_1}{}_{\!c_1}\tilde{P}^{c_1}_{\,d_2}{h}^{d_2}{}_{\!c_2}\tilde{P}^{c_2}_{\,d_3}\dots \tilde{P}^{c_{2j\!-\!1}}_{\,d_{2j}}{h}^{d_{2j}}{}_{\!d_1}\bigr)\\ &\quad+ {h}^{e}{}_{\!c_1}\tilde{P}^{c_1}_{\,d_2}{h}^{d_2}{}_{\!c_2}\tilde{P}^{c_2}_{\,d_3}\dots \tilde{P}^{c_{2j\!-\!1}}_{\,d_{2j}}{h}^{d_{2j}}{}_{\!\!c_{2j}}u^{c_{2j}}\;. \end{split}$$ Substituting and into expression for ${\{\tilde{C}_i,\tilde{C}_j\}}$ and using ${\tilde{P}^a_{\,b}u^b=0}$, we find $$\label{PBcpcp} \begin{split} &{\textstyle\frac{D-1}{16ij}}\{\tilde{C}_i,\tilde{C}_j\}=\\ &\quad=\xi_{a_0}\tilde{P}^{a_0}_{\,b_1}{h}^{b_1}{}_{\!a_1} \dots\tilde{P}^{a_{2i-1}}_{\,b_{2i-1}} {h}^{b_{2i-1}}{}_{\!c_1}\dots \tilde{P}^{c_{2j-1}}_{\,d_{2j}}{h}^{d_{2j}}{}_{\!\!c_{2j}}u^{c_{2j}}\\ &\quad\,-\xi_{a_0}\tilde{P}^{a_0}_{\,b_1}{h}^{b_1}{}_{\!a_1} \dots\tilde{P}^{a_{2j-1}}_{\,b_{2j-1}} {h}^{b_{2j-1}}{}_{\!c_1}\dots \tilde{P}^{c_{2i-1}}_{\,d_{2i}}{h}^{d_{2i}}{}_{\!\!c_{2i}}u^{c_{2i}}\\ &\quad =0\;. \end{split}\raisetag{3ex}$$ We thus proved that the conserved quantities ${w}$, ${{p}_j}$, and ${{C}_j}$ Poisson commute with each other. Since the generating function ${Z(\beta)}$ is given by power series in ${\beta}$ with coefficients given (up to constant factors) by the constants ${{C}_j}$, then also this function (and similarly ${W(\beta)=\exp Z(\beta)}$) Poisson commute with ${w}$ and ${{p}_j}$, as well as with itself for different choices of ${\beta}$: $$\label{PBZZ} \{Z(\beta_1),Z(\beta_2)\}=0\;,\quad\{W(\beta_1),W(\beta_2)\}=0\;.$$ The same is true also for the quantities ${{c}_j}$ generated from ${W(\beta)}$ introduced in [@KKPF]. Therefore the constants of motion all Poisson commute (are in involution), so the geodesic motion is completely integrable [@Arnold:book; @Kozlov]. Summary ======= We have explicitly proved the complete integrability of geodesic motion in the general higher-dimensional rotating black-hole spacetimes [@CLP]. The ‘nontrivial’ constants of motion are associated with the Killing tensors which we generated from the principal conformal Killing–Yano tensor. Observables $c_j$ are quadratic in momenta and correspond to rank-2 Killing tensors, whereas constants $C_j$ are of higher order in momenta and correspond to Killing tensors of increasing rank. The complete integrability of the geodesic motion is related to the issue of separability of the Hamilton-Jacobi equation recently accomplished by Frolov, Krtouš, and Kubizň' ak [@FKK]. The relation between integrability and separability on a general level has been studied in the series of papers by Benenti and Francaviglia (see, e.g., [@Benenti]) where it was demonstrated that the separability is possible only if all the constants of motion, corresponding to Killing vectors and rank-2 Killing tensors, Poisson commute. Acknowledgments {#acknowledgments .unnumbered} =============== P.K. is supported by the grant GAČR 202/06/0041 and appreciates the hospitality of the University of Alberta. D.K. is grateful to the Golden Bell Jar Graduate Scholarship in Physics at the University of Alberta. D.P. thanks the Natural Sciences and Engineering Research Council of Canada for financial support. Covariant canonical formalism on the cotangent bundle {#sc:apx} ===================================================== It is textbook knowledge [@Arnold:book] that the cotangent bundle ${{{\mathbf{T}^{*}M}}}$ has the natural structure of a phase space, i.e., it is possesses a symplectic form ${{{\boldsymbol{\Omega}}}}$ which defines the Poisson bracket ${\{\;,\;\}}$. For a base manifold ${M}$ which is equipped with an additional geometric structure, it can be useful to express phase-space quantities and operations with the help of quantities and operations on the base manifold. In this Appendix we shortly review such a procedure.[^3] We call functions on the phase space ${{{\boldsymbol{\Gamma}}}={{\mathbf{T}^{*}M}}}$ observables, and we write ${F(x,{{\boldsymbol{u}}})}$ to emphasize the dependence of ${F}$ on the configuration variable ${x\in M}$ and on the momentum ${{{\boldsymbol{u}}}\in\mathbf{T}_x^{*}M}$. For the base manifold $M$ with a (torsion-free) covariant derivative ${{{{{\boldsymbol{\nabla}}}}}}$ (in our case the spacetime manifold with the metric connection), it is possible to introduce the covariant derivative of an observable ${F(x,{{\boldsymbol{u}}})}$ *in the horizontal (configurational) direction* of the phase space ${{{\boldsymbol{\Gamma}}}}$. For any base manifold vector ${{{\boldsymbol{l}}}\in\mathbf{T}M}$ we define $$\label{pscovddef} {{\boldsymbol{l}}}^{e}{{{{\boldsymbol{\nabla}}}}}_{\!e}F(x,{{\boldsymbol{u}}})=\frac{d}{d\alpha}F(x(\alpha),{{\boldsymbol{u}}}(\alpha))\Big|_{\alpha=0}\;,$$ where ${x(\alpha)}$ is a curve starting from ${x}$ with tangent vector ${{{\boldsymbol{l}}}}$, and ${{{\boldsymbol{u}}}(\alpha)}$ is parallel transport of ${{{\boldsymbol{u}}}}$ along ${x(\alpha)}$. For a 1-form ${{{\boldsymbol{p}}}\in\mathbf{T}_x^{*}M}$ we can also introduce the derivative of an observable ${F(x,{{\boldsymbol{u}}})}$ *in the vertical (momentum) direction* $${{\boldsymbol{p}}}_e {{{{\boldsymbol{\partial}}}}}^e F(x,{{\boldsymbol{u}}}) = \frac{d}{d\alpha}F(x,{{\boldsymbol{u}}}+\alpha{{\boldsymbol{p}}})\Big|_{\alpha=0}\;.$$ Thanks to the linearity of ${\mathbf{T}_x^{*}M}$, this derivative is independent of any additional geometrical structure. Derivatives ${{{\boldsymbol{l}}}^e{{{{\boldsymbol{\nabla}}}}}_{\!e}F}$ and ${{{\boldsymbol{p}}}_e{{{{\boldsymbol{\partial}}}}}^{e}F}$ are derivative operators on ${{{\boldsymbol{\Gamma}}}}$ and as such they define vector fields on ${{{\boldsymbol{\Gamma}}}}$, which we denote[^4] ${{{\boldsymbol{l}}}^e{{{{\boldsymbol{\nabla}}}}}_{\!e}}$ and ${{{\boldsymbol{p}}}_e{{{{\boldsymbol{\partial}}}}}^{e}}$. These derivatives and vector fields depend ultralocally on the base manifold quantities ${{{\boldsymbol{l}}}}$ and ${{{\boldsymbol{p}}}}$ respectively, and we can thus introduce differentials ${{{{{\boldsymbol{\nabla}}}}}_{\!e}F}$ and ${{{{{\boldsymbol{\partial}}}}}^e F}$ and mixed tensor quantities ${\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}}x}}$ and ${\frac{{{{{\boldsymbol{\partial}}}}}^e}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}}$ by ‘tearing off’ ${{{\boldsymbol{l}}}}$ and ${{{\boldsymbol{p}}}}$, respectively, and by ‘tearing off’ the function ${F}$. Clearly, ${{{{{\boldsymbol{\nabla}}}}}_{\!e}F}$ is the covariant derivative of the observable ${F(x,{{\boldsymbol{u}}})}$ which ‘ignores’ the momentum ${{{\boldsymbol{u}}}}$ leaving it covariantly constant. On the other side, the derivative ${{{{{\boldsymbol{\partial}}}}}^e F}$ ‘ignores’ the configuration variable ${x}$. For an observable ${F(x,{{\boldsymbol{u}}})}$ given by a contraction of a spacetime tensor field ${{{\boldsymbol{f}}}(x)}$ with several momenta ${{{\boldsymbol{u}}}}$, $$\label{fisfuu} F(x,{{\boldsymbol{u}}}) = {{\boldsymbol{f}}}^{abc\dots}(x)\,{{\boldsymbol{u}}}_a{{\boldsymbol{u}}}_b{{\boldsymbol{u}}}_c\dots\;,$$ the covariant derivative reduces to the standard base manifold covariant derivative $$\label{covdfuu} {{{{\boldsymbol{\nabla}}}}}_{\!e}F(x,{{\boldsymbol{u}}}) = {{{{\boldsymbol{\nabla}}}}}_{\!e}{{\boldsymbol{f}}}^{abc\dots}(x)\; {{\boldsymbol{u}}}_a{{\boldsymbol{u}}}_b{{\boldsymbol{u}}}_c\dots\;.$$ The momentum derivative leaves ${{{\boldsymbol{f}}}}$ intact $$\label{pderfuu} \begin{split} {{{{\boldsymbol{\partial}}}}}^{e}F(x,{{\boldsymbol{u}}}) &={{\boldsymbol{f}}}^{ebc\dots}(x)\,{{\boldsymbol{u}}}_b{{\boldsymbol{u}}}_c\cdots +{{\boldsymbol{f}}}^{aec\dots}(x)\,{{\boldsymbol{u}}}_a{{\boldsymbol{u}}}_c\cdots\\ &\qquad+{{\boldsymbol{f}}}^{abe\dots}(x)\,{{\boldsymbol{u}}}_a{{\boldsymbol{u}}}_b\cdots+\dots\;. \end{split}\raisetag{3ex}$$ A general phase space observable can then be written as a (infinite) sum of terms of this type. The mixed tensor ${\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}}x}}$ is a vector field on the phase space (from ${\mathbf{T}{{\boldsymbol{\Gamma}}}}$) and a 1-form on the base manifold (from ${{{\mathbf{T}^{*}M}}}$). It is actually the horizontal lift from ${\mathbf{T}M}$ to ${\mathbf{T}{{\boldsymbol{\Gamma}}}}$ corresponding to the covariant derivative ${{{{{\boldsymbol{\nabla}}}}}}$. The mixed tensor ${\frac{{{{{\boldsymbol{\partial}}}}}^e}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}}$ is a vector field on the phase space (from ${\mathbf{T}{{\boldsymbol{\Gamma}}}}$) and a vector field on the base manifold (from ${\mathbf{T}M}$). It gives a natural identification of the cotangent fiber ${\mathbf{T}_x^{*}M}$ with its vertical tangent space ${\mathbf{T}\mathbf{T}_x^{*}M}$. The inverse symplectic form ${{{\boldsymbol{\Omega}}}^{-1}}$ and the Poisson bracket can be written as $$\label{invSymplStr} {{\boldsymbol{\Omega}}}^{-1} = \frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}}x} \frac{{{{{\boldsymbol{\partial}}}}}^e}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}} -\frac{{{{{\boldsymbol{\partial}}}}}^e}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}} \frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}}x}$$ and $$\label{PB} \{A,B\} = {{{{\boldsymbol{\nabla}}}}}_{\!e}A\, {{{{\boldsymbol{\partial}}}}}^e B - {{{{\boldsymbol{\partial}}}}}^e A\, {{{{\boldsymbol{\nabla}}}}}_{\!e} B\;.$$ They do not depend on a choice of the covariant derivative. Indeed, if we choose another torsion-free covariant derivative ${\tilde{{{{{\boldsymbol{\nabla}}}}}}}$ on ${M}$, which can be done by specifying the ‘difference’ tensor ${{{\boldsymbol{\Gamma}}}^{\,b}_{ac}}$, $$\tilde{{{{{\boldsymbol{\nabla}}}}}}_{\!a}{{\boldsymbol{a}}}^b = {{{{\boldsymbol{\nabla}}}}}_{\!a}{{\boldsymbol{a}}}^b + {{\boldsymbol{\Gamma}}}^{\,b}_{ac}\,{{\boldsymbol{a}}}^c\;,$$ the induced covariant derivative of the phase space observables transforms as $$\tilde{{{{{\boldsymbol{\nabla}}}}}}_{\!a}F(x,{{\boldsymbol{u}}})={{{{\boldsymbol{\nabla}}}}}_{\!a}F(x,{{\boldsymbol{u}}}) +{{\boldsymbol{u}}}_{e}\,{{\boldsymbol{\Gamma}}}^{\,e}_{ac}(x)\,{{{{\boldsymbol{\partial}}}}}^{c}F(x,{{\boldsymbol{u}}})\;.$$ Substituting this into and using the symmetry ${{{\boldsymbol{\Gamma}}}^{\,b}_{ac}={{\boldsymbol{\Gamma}}}^{\,b}_{ca}}$ we find that $$\{A,B\} = \tilde{{{{\boldsymbol{\nabla}}}}}_{\!e}A\, {{{{\boldsymbol{\partial}}}}}^e B - {{{{\boldsymbol{\partial}}}}}^e A\, \tilde{{{{\boldsymbol{\nabla}}}}}_{\!e} B\;,$$ i.e., the Poisson bracket is independent of the choice of the covariant derivative. The argument for the symplectic structure is similar. The Poisson bracket of an observable of type with an observable ${p}$ linear in momenta ${{{\boldsymbol{u}}}}$, $$p(x,{{\boldsymbol{u}}})={{\boldsymbol{l}}}^c(x)\,{{\boldsymbol{u}}}_c\;,$$ leads, with help of and , to the Lie derivative: $$\label{PBlinobs} \begin{split} &\{F,p\} = {{\boldsymbol{l}}}^e{{{{\boldsymbol{\nabla}}}}}_{\!e}F - {{{{\boldsymbol{\partial}}}}}^e F\, ({{{{\boldsymbol{\nabla}}}}}_{\!e}{{\boldsymbol{l}}}^c)\,{{\boldsymbol{u}}}_c\\ &\;=\bigl({{\boldsymbol{l}}}^e{{{{\boldsymbol{\nabla}}}}}_{\!e}{{\boldsymbol{f}}}^{ab\dots} - {{\boldsymbol{f}}}^{eb\dots} {{{{\boldsymbol{\nabla}}}}}_{\!e}{{\boldsymbol{l}}}^a - {{\boldsymbol{f}}}^{ae\dots} {{{{\boldsymbol{\nabla}}}}}_{\!e}{{\boldsymbol{l}}}^b - \dots\bigr) \,{{\boldsymbol{u}}}_a{{\boldsymbol{u}}}_b\cdots\\ &\;=\bigl({\pounds}_{{{\boldsymbol{l}}}}{{\boldsymbol{f}}}^{ab\dots}\bigr)\,{{\boldsymbol{u}}}_a{{\boldsymbol{u}}}_b\cdots\; \equiv{\pounds}_{{{\boldsymbol{l}}}}F\;. \end{split}\raisetag{3ex}$$ Here ${{\pounds}_{{{\boldsymbol{l}}}}{{\boldsymbol{f}}}}$ is the standard Lie derivative on ${M}$ along the vector field ${{{\boldsymbol{l}}}}$. The last equality then defines the generalized Lie derivative ${{\pounds}_{{{\boldsymbol{l}}}}F}$ of the phase space observable ${F}$ along the base manifold vector field ${{{\boldsymbol{l}}}}$ which effectively ‘ignores’ the dependence of ${F}$ on the momentum ${{{\boldsymbol{u}}}}$. It can be extended to general phase space observables by linearity. It can be also defined similarly to with ${{{\boldsymbol{u}}}(\alpha)=\phi_\alpha{{\boldsymbol{u}}}}$ given by a flow ${\phi_\alpha}$ induced by the vector field ${{{\boldsymbol{l}}}}$ acting on ${{{\boldsymbol{u}}}}$. ${{\pounds}_{{{\boldsymbol{l}}}}F}$ can be also viewed as the derivative of the observable ${F}$ along the vector field ${{\pounds}_{{{\boldsymbol{l}}}}}$ on ${{{\mathbf{T}^{*}M}}}$ which is called the complete lift of the vector field ${{{\boldsymbol{l}}}}$ on ${M}$. Clearly, the Poisson bracket with the Hamiltonian leads to the covariant derivative along the ${{{\boldsymbol{u}}}}$ direction: $$\label{PBHam} \begin{split} &\{F,H\} = {{\boldsymbol{u}}}^e{{{{\boldsymbol{\nabla}}}}}_{\!e}F\;. \end{split}$$ Despite the fact that we do not need them in the main text, let us introduce for completeness the mixed tensor fields ${{{\boldsymbol{D}}}^{e}x}$ and ${{{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{e}}$ dual to ${\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}}x}}$ and ${\frac{{{{{\boldsymbol{\partial}}}}}^e}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}}$ defined by $$\begin{gathered} \frac{{{{{\boldsymbol{\nabla}}}}}_{\!b}}{{{\boldsymbol{\partial}}}x} \cdot {{\boldsymbol{D}}}^{a}x = {{\boldsymbol{\delta}}}^{a}_{b}\;,\quad \frac{{{{{\boldsymbol{\partial}}}}}^a}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}} \cdot {{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{b}= {{\boldsymbol{\delta}}}^{a}_{b}\;,\\ \frac{{{{{\boldsymbol{\nabla}}}}}_{\!a}}{{{\boldsymbol{\partial}}}x} \cdot {{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{b}= 0\;,\quad \frac{{{{{\boldsymbol{\partial}}}}}^a}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}} \cdot {{\boldsymbol{D}}}^{b}x =0\;. \end{gathered}$$\ Here the dot ‘$\cdot$’ indicates the contraction of the phase space tensor indices. ${{{\boldsymbol{D}}}^{e}x}$ is a vector field on the base manifold ${M}$ and a on the phase manifold ${{{\boldsymbol{\Gamma}}}}$. It is actually the differential of the bundle projection ${x:{{\mathbf{T}^{*}M}}\to M}$. ${{{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{e}}$ is a 1-form both on the base manifold ${M}$ and phase space ${{{\boldsymbol{\Gamma}}}}$. These phase space ‘forms’ satisfy the completeness relation $$\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}}x}\, {{\boldsymbol{D}}}^{e}x +\frac{{{{{\boldsymbol{\partial}}}}}^e}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}\, {{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{e}= {{\boldsymbol{\delta}}}\;,$$ with ${{{\boldsymbol{\delta}}}}$ being the identity tensor on ${\mathbf{T}{{\boldsymbol{\Gamma}}}}$. The symplectic structure ${{{\boldsymbol{\Omega}}}}$ can be written as $$\label{SymplStr} {{\boldsymbol{\Omega}}} = {{\boldsymbol{D}}}^{e}x\,{{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{e}-{{{{\boldsymbol{\nabla}}}}}{{\boldsymbol{u}}}_{e}\,{{\boldsymbol{D}}}^{e}x\;.$$ Finally, if we choose the coordinate derivative ${{{{{\boldsymbol{\eth}}}}}}$ associated with a coordinate system ${x^a}$ on ${M}$, $${{{{\boldsymbol{\eth}}}}}{{{{\boldsymbol{\partial}}}}}_{x^a}=0\;,\quad{{{{\boldsymbol{\eth}}}}}{{{{\boldsymbol{d}}}}}x^a=0\;,$$ instead of the covariant derivative ${{{{{\boldsymbol{\nabla}}}}}}$, the relations , , and reduce to the standard relations in terms of the canonical coordinates ${x^a,u_b}$ on ${{{\boldsymbol{\Gamma}}}}$, namely $$\label{SymplStrincoor} \begin{gathered} {{\boldsymbol{\Omega}}} ={{{{\boldsymbol{d}}}}}{x^e}\, {{{{\boldsymbol{d}}}}}{u_e}-{{{{\boldsymbol{d}}}}}{u_e}\, {{{{\boldsymbol{d}}}}}{x^e} \;,\\[1ex] {{\boldsymbol{\Omega}}}^{-1}= {{{{\boldsymbol{\partial}}}}}_{x^e}\, {{{{\boldsymbol{\partial}}}}}_{u_e} - {{{{\boldsymbol{\partial}}}}}_{u_e}\,{{{{\boldsymbol{\partial}}}}}_{x^e} \;, \end{gathered}$$ and $$\{A,B\} =\frac{\partial A}{\partial x^e}\,\frac{\partial B}{\partial u_e} -\frac{\partial A}{\partial u_e}\,\frac{\partial B}{\partial x^e}\;.$$ All coordinate vectors and 1-forms in live on the phase space ${{{\boldsymbol{\Gamma}}}}$. D.N. Page, D. Kubizňák, M. Vasudevan, and P. Krtouš, Phys. Rev. Lett. [**98**]{}, 061102 (2007), arXiv:hep-th/0611083. P. Krtouš, D. Kubizňák, D.N. Page, and V.P. Frolov, J. High Energy Phys. 02 (2007) 004, arXiv:hep-th/0612029. K. Schwarzschild, Sitzungsber. deutsch. Akad. Wiss. Berlin, Kl. Math. Phys. Technik, 189 (1916). F.R. Tangherlini, Nuovo Cimento [**27**]{}, 636 (1963). R.P. Kerr, Phys. Rev. Lett. [**11**]{}, 237 (1963). R.C. Myers and M.J. Perry, Ann. Phys. (N.Y.) [**172**]{}, 304 (1986). S.W. Hawking, C.J. Hunter, and M.M. Taylor-Robinson, Phys. Rev. D  [**59**]{}, 064005 (1999), arXiv:hep-th/9811056. B. Carter, Phys. Lett. [**26A**]{}, 399 (1968). B. Carter, Commun. Math. Phys. [**10**]{}, 280 (1968). G.W. Gibbons, H. Lü, D.N. Page and C.N. Pope, J.Geom. Phys. [**53**]{}, 49 (2005), arXiv:hep-th/0404008. G.W. Gibbons, H. Lü, D.N. Page and C.N. Pope, Phys.Rev. Lett.  [**93**]{}, 171102 (2004), arXiv:hep-th/0409155. W. Chen, H. Lü, and C.N. Pope, Class. Quant. Grav., 5323 (2006), arXiv:hep-th/0614125. E. Newman, L. Tamburino, and T. Unti, J. Math. Phys.(N.Y.) [**4**]{}, 915 (1963). N. Hamamoto, T. Houri, T. Oota, and Y. Yasui, J. Phys. [**A40**]{}, F177 (2007), arXiv:hep-th/0611285. D. Kubizňák and V.P. Frolov, Class. Quant. Grav., F1 (2007), arXiv:gr-qc/0610144. V.I. Arnol’d, [*Mathematical Methods of Classical Mechanics*]{} (Springer-Verlag, Berlin, 1978). V.V. Kozlov, Usp. Mat. Nauk [**38:1**]{}, 3 (1983) \[Russ. Math. Surv. [**38:1**]{}, 1 (1983)\]. V.P. Frolov, P. Krtouš, and D. Kubizňák, J.High Energy Phys. [**02**]{}, 005 (2007), arXiv:hep-th/0611245. S. Benenti and M. Francaviglia, Gen. Rel. Grav., 79 (1979). R. Penrose and W. Rindler, [*Spinors and Space-Time*]{} (Cambridge University Press, Cambridge, England, 1984, 1986). R. M. Wald, [*General Relativity*]{} (The University of Chicago Press, Chicago and London, 1984). [^1]: The derivative ${{{{{\boldsymbol{\partial}}}}}f}$ is the *vector* field on spacetime ${M}$ with components ${\partial{f}/\partial{u_a}}$, cf. the Appendix \[sc:apx\] (it could be written more explicitly as ${{{{{\boldsymbol{\partial}}}}}f/{{{{\boldsymbol{\partial}}}}}{{\boldsymbol{u}}}}$). The wedge product is, strictly speaking, defined for (antisymmetric) *forms*. However, we can easily define the wedge product also for the vectors or lower the vector indices with the help of the metric to get 1-forms. [^2]: The scaling differs from used in the previous section. [^3]: Similarly to the main text we type tensors in bold. Optionally, we write here the tensors with *abstract indices* [@PenroseRindler:book; @Wald:book1984] which help to indicate tensorial operations as, for example, contraction. However, the abstract indices do not refer to any particular choice of coordinates. We use small latin letters for base manifold indices (for tensors from ${\mathbf{T} M}$), but we do not introduce indices for the phase space tensors (tensors from ${\mathbf{T}{{\boldsymbol{\Gamma}}}}$). We assume implicitly the tensor product, i.e., ${{{\boldsymbol{a}}}{{\boldsymbol{b}}}={{\boldsymbol{a}}}\otimes{{\boldsymbol{b}}}}$. [^4]: We could be more explicit and write them as ${{{\boldsymbol{l}}}^e\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}} x}}$ and ${{{\boldsymbol{p}}}_e\frac{{{{{\boldsymbol{\partial}}}}}^{e}}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}}$. Similarly we could write ${\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}F}{{{\boldsymbol{\partial}}} x}}$ and ${\frac{{{{{\boldsymbol{\partial}}}}}^{e}F}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}}$ for quantities ${{{{{\boldsymbol{\nabla}}}}}_{\!e}F}$ and ${{{{{\boldsymbol{\partial}}}}}^e F}$ introduced below. However, we use such an explicit notation only for the mixed tensor fields ${\frac{{{{{\boldsymbol{\nabla}}}}}_{\!e}}{{{\boldsymbol{\partial}}} x}}$ and ${\frac{{{{{\boldsymbol{\partial}}}}}^{e}}{{{\boldsymbol{\partial}}}{{\boldsymbol{u}}}}}$ (see below) where the notation ${{{{{\boldsymbol{\nabla}}}}}_{\!e}}$ and ${{{{{\boldsymbol{\partial}}}}}^{e}}$ would be too brief.

ArXiv

--- abstract: 'In distributed systems, communication is a major concern due to issues such as its vulnerability or efficiency. In this paper, we are interested in estimating sparse inverse covariance matrices when samples are distributed into different machines. We address communication efficiency by proposing a method where, in a single round of communication, each machine transfers a small subset of the entries of the inverse covariance matrix. We show that, with this efficient distributed method, the error rates can be comparable with estimation in a non-distributed setting, and correct model selection is still possible. Practical performance is shown through simulations.' address: 'Department of Statistics, University of Michigan' bibliography: - 'strings.bib' - 'refs.bib' title: Efficient Distributed Estimation of Inverse Covariance Matrices --- Distributed Estimation, Debiased Estimators, Efficient Communication, Gaussian Graphical Models, Inverse Covariance Estimation Introduction {#sec:intro} ============ The collection of copious and meticulous amounts of information has led to the modern phenomena of datasets being both high-dimensional and very large in sample size. These massive datasets are distributed over multiple machines due to size limitations or because data is collected and stored independently. Even in the case when a single machine is large enough, there are efficiency, security, and privacy concerns in aggregating all the data onto one machine. Bandwidth restrictions can make impossible or inefficient to send large amounts of data and the more communication in the system, the more vulnerable it is to attacks. In addition, the raw dataset may contain sensitive information in the individual samples such as in medical or financial records. Thus, it is advantageous, if not necessary, for each machine to be able to calculate compact estimates that can be efficiently communicated and preserve the confidentiality of the samples. Estimation of the inverse covariance matrix is used in the analysis of different types of data, such as gene expression or brain imaging. In particular, when samples are Gaussian, the non-zero entries of the inverse covariance correspond to the edges of a Gaussian Markov random field. Thus, accurately estimating the non-zero pattern of the inverse covariance matrix is crucial. In this paper, we develop a method for estimating a sparse inverse covariance matrix when the samples of the data are distributed among machines. Our method is efficient in communication, in the sense that only a single round of communication between each machine and a central hub is sufficient, and the bandwidth required is small compared to the size of the matrix estimated. Related Work ------------ High-dimensional estimation of inverse covariance matrices is usually addressed by $\ell_1$ penalized convex optimization [@banerjee2008model; @friedman2008sparse]. In a distributed setting, general frameworks for convex optimization involve multiple rounds of communication between all machines [@Boyd:2011:DOS:2185815.2185816; @duchi2012dual], which can be expensive. For some $\ell_1$ penalization problems, the communication cost can be reduced to a single round of communication [@Zhang:2013:CAS:2567709.2567769; @lee2015] by reducing the bias with bootstrapped or debiased estimators [@2013Javanmard]. Our work follows a similar approach as [@lee2015] for penalized linear regression; however, we introduce efficiency not only in the amount of communication, but also in the size of the communication channel. Related work on inverse covariance estimation has focused on the case where variables are distributed and the structure of the associated graphical model is known [@meng2014marginal]. Here, our samples are distributed, and we must estimate the structure of the matrix and the value of its entries. Outline ------- The rest of this paper is organized as follows: Section 2 presents a communication efficient method for inverse covariance estimation in a distributed setting. Section 3 studies the error rates of the estimator and shows that model selection consistency is possible. Section 4 studies the practical performance and compares our method with other distributed and non-distributed approaches. Proofs of the theoretical results are included in the appendix. Distributed Inverse covariance estimation by debiasing and thresholding ======================================================================= We work in a setting where samples are distributed among $M$ different machines. Each observation is a vector of size $p$ coming from a distribution with covariance $\Sigma$ and inverse covariance $\Theta$. Let $ s $ be the number of non-zero entries in $\Theta$ and $ d $ the maximum number of non-zeros per row. Let $\mathcal{S}$ denote the set of non-zero entries of $\Theta$ and $\mathcal{S}^c$ the set of zeros. We assume that the data is split equally over all machines, where each machine has $n$ observations, so we denote $X_m \in \mathbb{R}^{n \times p }$ as the data matrix on machine $m$. We are interested in estimating $ \Theta$. In a distributed setting, it is desirable to have a single round of communication between each machine and the central hub. Moreover, bandwidth or storage capacity often limits the amount of data that can be shared with the central hub, forcing the data to be distributed. Thus, our approach is based on constructing sparse estimators on each machine that when aggregated in the central hub, provide a good estimation for $\Theta$. In the high dimensional setting, where $n \ll p$, a common approach to obtain a sparse estimator of $\Theta$ is by minimizing the $\ell_1$-penalized log-determinant Bregman divergence. Thus on each machine, this estimator, known as *graphical lasso*, is defined as $$\label{l1-opt} \hat{\Theta}_m := \underset{\theta \in \mathbb{S}^p_{++}}{\arg\min} \left\{ \,\text{tr}(\theta \hat{\Sigma}_m) - \log\det\theta + \lambda ||\theta||_{1, \text{off}}\right\},$$ where $\|\cdot\|_{1,\text{off}}$ denotes the $\ell_1$ norm of the off-diagonal entries of a matrix, and $\mathbb{S}^p_{++}$ is the cone of positive definite matrices of size $p$. For a single machine, this estimator has been studied and shown to be asymptotically consistent in mean squared error with rate ${\mathcal{O}_{\Bbb{P}}\left(s\log p /n\right)}$ [@rothman2008]. Moreover, the set of non-zero entries of $\hat{\Theta}$ coincides with $\mathcal{S}$ when $\lambda\asymp\sqrt{\log p/n}$ and under certain conditions [@ravikumar2011]. A naive approach for distributed estimation would be to average the estimated $\hat{\Theta}_m$ from each machine. However, this estimator is biased due to the $\ell_1$ penalty, and averaging only improves the variance, not the bias. We adopt a similar approach as [@lee2015] did for lasso regression by trading-off the bias for variance. The *debiased graphical lasso* estimator was proposed in order to construct confidence intervals for the entries of $\Theta$ [@jankova2015]. The idea of this estimator is to invert the Karush-Kuhn-Tucker (KKT) conditions of the optimization problem in equation in order to get a debiased estimator defined as $$\label{debiased est} \hat{\Theta}_m^d := \hat{\Theta}_m + \hat{\Theta}_m((\hat{\Theta}_m)^{-1}-\hat{\Sigma}_m)\hat{\Theta}_m.$$ The *debiased graphical lasso* has the appealing property that each entry of the matrix is asymptotically normal distributed. It is shown that $\hat{\Theta}_m^d$ can also be written as $$\label{loss} \hat{\Theta}_m^d = \Theta-\Theta(\hat{\Sigma} - \Sigma)\Theta + \Delta_m,$$ where $\|\Delta_m\|_\infty={\mathcal{O}_{\Bbb{P}}\left(d\log p/n\right)}$ accounts for the bias, and the second term in the equation is asymptotically normal [@jankova2015]. The bias of these estimators is of smaller order than the bias of the graphical lasso, and the variance is reduced when we average these estimators in the central hub to get an overall estimator $\frac{1}{M}\sum_{m=1}^M \hat{\Theta}^{d}_{m } $. When the data is not distributed into too many machines, the averaged estimator can get similar error rates in $\ell_\infty$ norm as the graphical lasso performed on all the data (see Lemma \[lemma:avgglasso\]). The debiased graphical lasso estimator is not sparse. This fact is problematic in a distributed setting, since it will require the transfer of $p^2$ entries, which might be larger than the data on each machine. Under the sparsity assumption on $\Theta$, we are actually only interested in the value of $s+p$ entries. Hence, on each machine we select the most significant coefficients of $\hat{\Theta}_m$ and send them to the central hub. The sparse estimator is defined as $$\hat{\Theta}^{d,\rho}_{ij} := \hat{\Theta}^{d}_{ij} \, \mathbb{I}\left(|\hat{\Theta}^d_{ij}| > \rho \, \hat{\sigma}_{ij}\right),$$ where $\hat{\sigma}_{ij}$ is an estimator for $\sigma^2_{ij} = \text{Var}(\Theta_{i\cdot} X_{1\cdot} X_{1\cdot}^T \Theta_{\cdot j})$ and $\mathbb{I}(E)$ is the indicator of the event $E$. For gaussian random vectors, a good estimator for this quantity is $\hat{\sigma}^2_{ij}=\hat{\Theta}_{ii}\hat{\Theta}_{jj}+\hat{\Theta}_{ij}^2$ (Lemma 2 of [@jankova2015]). In order to achieve correct estimation (in the central hub) of the support of $\Theta$, it is optimal to send as many entries as possible. So we let the threshold parameter $\rho$ be a function of $B$, the bandwidth of the communication channel from each machine to the central hub, and we set $\rho$ as the smallest threshold that still fits in the channel. In Algorithm \[alg machine\], we summarize the estimation procedure on each machine. \[alg machine\] Do on each machine $m$ Data matrix $X_{m}$, penalty $\lambda$, threshold $\rho$. 1) Define $\hat{\Theta}$ as the solution of with penalty $\lambda$. 2) Define $ \hat{\Theta}^d = \hat{\Theta} + \hat{\Theta}(\hat{\Theta}^{-1}-\hat{\Sigma})\hat{\Theta} $. 3) Calculate an estimate $\hat{\sigma}_{ij}^2$ for $\sigma_{ij}^2$. 4) Threshold $\hat{\Theta}^d$, so $ \hat{\Theta}^{d, \rho}_{ij} = \hat{\Theta}^{d}_{ij} \, \mathbb{I}\left(|\hat{\Theta}^d_{ij}| > \rho \, \hat{\sigma}_{ij} \right) $.\ Sparse matrix $\hat{\Theta}^{d, \rho}$. \[alg machine\] The average debiased graphical lasso estimator is also not sparse, which might be unpractical, in particular because estimating the set of non-zeros can be more important than the values of the entries themselves. However, if the averaged estimator is thresholded at a certain level $\tau\asymp \sqrt{\log p / (nM)}$, correct model selection on the non-zero entries of the matrix is possible (see Theorem \[theorem:rates\]). Each entry again requires an estimator of $\sigma^2_{ij}$. For normally distributed data, we use $\bar{\Theta}_{ii}\bar{\Theta}_{jj} +\bar{\Theta}_{ij}^2$, where $\bar{\Theta}=\frac{1}{M}\sum_{m=1}^M \hat{\Theta}^{d, \rho}_m$. Algorithm \[alg hub\] shows the complete procedure in the central hub. \[alg hub\] Estimators from each machine $\hat{\Theta}^{d, \rho}_m$, threshold $\tau$. 1) Define $ \hat{\Theta}^D :=\hat{\Theta}^D(\rho)= \frac{1}{M}\sum_{m=1}^M \hat{\Theta}^{d, \rho}_{m} $. 2) Calculate an estimate $\hat{\sigma}_{M, ij}^2$ for $\sigma_{ij}^2$. 3) Threshold $\hat{\Theta}^D$, so $ \hat{\Theta}_{ij}^{D, \tau} = \hat{\Theta}_{ij}^D \, \mathbb{I} \left(|\hat{\Theta}^D_{ij}| > \tau \, \hat{\sigma}_{M, ij} \right) $.\ Overall distributed estimator $\hat{\Theta}^{D, \tau}$. \[alg hub\] Theoretical results =================== In this section, we derive bounds for the estimation error of our method. We show that the error of our distributed estimator has the same rate as the non-distributed graphical lasso when the number of machines increases slower than $M \lesssim n/(d^2\log p)$. Moreover, we show that a bandwidth of size $O(p^{2-c})$, with $c$ an absolute constant, is enough to correctly identify the set of non-zero entries of $\Theta$. The following assumptions are necessary in order for the graphical lasso and debiased graphical lasso to have a good estimation performance [@ravikumar2011; @jankova2015], and we require them for our theoretical results. 1. \[a1\] There exists some $ \alpha \in (0, 1] $ such that $ \underset{e \in \mathcal{S}^c}{\max} \, || \Gamma_{e\mathcal{S}} (\Gamma_{\mathcal{S}\mathcal{S}})^{-1} ||_1\le (1-\alpha) $ where $ \Gamma := (\Theta)^{-1} \otimes (\Theta)^{-1} $ is the Hessian of (\[l1-opt\]). 2. \[a2\] There exists a constant $ L < \infty $ such that $ 1/L \leq \Lambda_{min}(\Sigma) \leq \Lambda_{max}(\Sigma) \leq L $ where $ \Lambda_{min}(\Sigma) $ and $ \Lambda_{max}(\Sigma) $ are the minimum and maximum eigenvalues of $ \Sigma $. 3. \[a3\] The samples of the data are subgaussian random variables $X\in\Bbb{R}^p$, with $\mathbb{E}[X]=0 $, $ \text{Cov}(X)=\Sigma$ and subgaussian norm $\|X\|_{\phi_2}\leq K$. 4. \[a4\] The quantities ${{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \Sigma \right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_\infty$ and ${{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert (\Gamma_{\mathcal{S}\mathcal{S}})^{-1} \right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_\infty$ are bounded, where ${{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \cdot \right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}$ is the opertor norm of a matrix. In the literature, it is common to allow the error rates to depend on the bounding constants from the previous assumptions. Here, in order to keep the results simple, we state the error rates as a function of the dimensionality ($n,p,M$), sparsity ($s$, $d$), and smallest entries of $\Theta$, while keeping the other quantities bounded by absolute constants. The work from [@jankova2015] studies the error rate of debiased graphical lasso. This result can be extended to the average of multiple debiased estimators as follows. \[lemma:avgglasso\] Suppose that assumptions \[a1\] through \[a4\] hold, $M<p$ and define $\lambda_m \asymp\sqrt{\log p/n}$ in equation . Then, the averaged debiased graphical lasso estimator satisfies $$\left\|\frac{1}{M}\sum_{m=1}^M\Theta_m^{d, \rho} - \Theta\right\|_\infty = {\mathcal{O}_{\Bbb{P}}\left(\max\left\{ \sqrt{\frac{\log p}{Mn}},d\frac{\log p}{n} \right\}\right)}.$$ The previous lemma splits the error rate of the distributed estimator into two parts, which correspond to the variance and the bias. The variance vanishes as the number of machines increases, but the bias remains. However, when $M\lesssim n/(d^2\log p)$, the variance becomes dominant and the error is ${\mathcal{O}_{\Bbb{P}}\left(\sqrt{\log p/(Mn)}\right)}$. This is the same error rate as the graphical lasso when the data is not distributed [@rothman2008]. Previous work shows similar results for distributed $\ell_1$ penalized regression [@lee2015]. Our method expands the framework to the graphical lasso estimator. Additionally, in the next theorem we show that a bandwith of size $B\asymp p^{2-c}$ is enough to select the correct set of entries and a similar rate for the mean squared error as the graphical lasso on the full data. \[theorem:rates\] Suppose that \[a1\] through \[a4\] hold and $M<p$. Define the tuning parameters of the algorithms as $\lambda\asymp\sqrt{\log(p)/n}$ and $\tau\asymp\sqrt{\log(p)/(Mn)}$. If $ \eta_\text{min} := \underset{(i,j)\in\mathcal{S}}{\min} \frac{|\Theta_{ij}|}{\sigma_{ij}} = \Omega(\sqrt{\log(p)/n})$, then there exists a constant $c\in(0,1]$ such that the following results hold for a bandwidth $B=\Omega\left(p^{2-c}\right)$. 1. Algorithm 2 recovers the correct set of non-zero entries of $\Theta$ with high probability, that is, $${\mathbb{P}\left(\mathcal{S}(\hat{\Theta}_m^{D, \tau}) = \mathcal{S}(\Theta)\right)}\geq 1-{\mathcal{O}\left(1/p\right)}.$$ 2. The mean squared error of the estimator given by Algorithm 2 satisfies $$\label{mse} \left\|\hat{\Theta}_m^{D, \tau} - \Theta\right\|_F^2 = {\mathcal{O}_{\Bbb{P}}\left((s+p)\max\left\{\frac{\log p}{Mn},\frac{d^2\log^2 p}{n^2}\right\}\right)}.$$ Simulation results ================== To evaluate the performance of our distributed estimator, we conduct a simulation study. We study the effect of varying the total sample size by changing the number of machines in the distributed system. We fix the number of variables to $p=1000$ and the sample size on each machine to $n=100$. Samples were generated from a normal distribution $N(0_{p}, \Sigma)$ such that $\Theta=(\Sigma)^{-1}$ has the form $\Theta_{i,i}=1$ and $\Theta_{i,i+1}=\Theta_{i,i-1}=0.4$ for $i=1,\ldots,p$. Thus, the associated Gaussian graph is a chain. All of our simulation results are calculated over 100 different trials. To solve the problem , we use GLasso [@friedman2008sparse] with the R package `huge` [@zhao2012huge]. We set the tuning parameters $\lambda$ and $\tau$ according to the rates in Theorem \[theorem:rates\], so for each machine $\lambda=\sqrt{\log(p)/n}$ and $\tau=\sqrt{\log(p)/(Mn)}$. The bandwidth is set to $B=10p$, so only 1% of the entries of $\hat{\Theta}^d_m$ are sent to the central hub. We compare the performance of our distributed estimator (Distributed) with the following estimators. 1. (Naive) An estimator based on averaging the graphical lasso estimators from each machine $ \hat{\Theta}^{\text{naive}} = \frac{1}{M}\sum_{m=1}^M \hat{\Theta}_{m} $. 2. (Full) An estimator based on the full non-distributed data given by the graphical lasso $ \hat{\Theta}^{\text{full}}$. 3. (Full Debiased) Since debiasing decreases the bias of the estimation, we also compare with a thresholded debiased graphical lasso estimator on the full data $\hat{\Theta}^{D,\text{full}}$. ![Estimation error of the different estimators based on MSE and $\ell_\infty$ norm. Note that Debiased Full has the same error in $\ell_\infty$ as the Full estimator. For a small number of machines, the performance of our distributed estimator is similar to that of the estimator on the full data, and superior to the naive estimator.[]{data-label="fig:errorplots"}](L2plot){width="\linewidth"} ![Estimation error of the different estimators based on MSE and $\ell_\infty$ norm. Note that Debiased Full has the same error in $\ell_\infty$ as the Full estimator. For a small number of machines, the performance of our distributed estimator is similar to that of the estimator on the full data, and superior to the naive estimator.[]{data-label="fig:errorplots"}](LINF){width="\linewidth"} ![FPR and FNR for recovering the correct set of non-zero entries of $\Theta$ using our distributed estimator. The number of machines is set to $10$ and we vary the tuning parameters of the algorithm. We note the robustness of the estimators in recovering the correct support using different values of $\lambda$ and $\tau$. []{data-label="fig:modelselection"}](FPR){width="\linewidth"} ![FPR and FNR for recovering the correct set of non-zero entries of $\Theta$ using our distributed estimator. The number of machines is set to $10$ and we vary the tuning parameters of the algorithm. We note the robustness of the estimators in recovering the correct support using different values of $\lambda$ and $\tau$. []{data-label="fig:modelselection"}](FNR){width="\linewidth"} In Figure \[fig:errorplots\], the errors of the different estimators are shown. In general, a better estimator can be obtained using the full data, which is expected. However, when the number of machines is small, the performance of our distributed estimator is comparable to the estimators on the full data, both in mean squared error and $\ell_\infty$ norm. In mean squared error, when $M$ is small, the debiased estimators perform better because they trade off bias for variance. Thus, Debiased Full performs the best and Distributed has similar performance. Naive performs poorly under this norm for all $M$. In $\ell_\infty$ norm, the performance of Full and Full Debiased is exactly the same because the largest error is from the entries in the diagonal, which are equal since the diagonal is not penalized. The error of Distributed is approximately the same for a wide range of values of $M$. However, in $M=2$, the performance of Distributed is affected by the bandwidth since missing edges in any machine have a larger impact on the error. But as the number of machines increases, the estimation of Distributed becomes insensitive to the bandwidth. To evaluate the robustness of the method against the selection of the tuning parameters $\lambda$ and $\tau$, we measure false positive rate (defined as percentage of zeros of $\Theta$ identified as edges by the estimator) and false negative rate (defined as percentage of edges missed by the estimator). We use the same settings as the previous scenario, but the number of machines is fixed to $M=10$. The values of the tuning parameters for our method vary proportionally to $\lambda = \beta\sqrt{\log(p)/n}$ and $\tau = \beta\sqrt{\log(p)/(Mn)}$ with $\beta$ between $0.2$ and $2$. The bandwidth is still fixed at 1% of the entries. We observe that for a wide range of parameters we recover the correct set of non-zeros. Discussion ========== We have proposed a method for estimating sparse inverse covariance matrices when the samples of a dataset are distributed over different machines. Our method agrees with other results for efficient distributed estimation in high-dimensional settings, and we also introduced efficiency in the bandwidth size. Asymptotically, the performance of our estimator is analogous to estimators with samples that are not distributed. Our simulation results are consistent with the theoretical rates and also show that we perform significantly better than a naive approach to estimation in a distributed setting. Appendix ======== Using the decomposition given in equation , we have \_[m=1]{}\^M -\_& (\_[m=1]{}\^M-)\_+ \_[m=1]{}\^M\_\ & = (- )\_+ \_[m=1]{}\^M\_m\_, where $\hat{\Sigma}$ is the covariance matrix using the full data. Thus, the first term can be bounded as $ \left\|\Theta(\hat{\Sigma}- \Sigma)\Theta\right\|_\infty = {\mathcal{O}_{\Bbb{P}}\left( \sqrt{\log p/(Mn) } \right)} $ (equation (11) of [@jankova2015]) and the second term can be bounded as & (\_[m=1]{}\^M\_m\_) ( \_m\_)\^M\ & = (1 - 1/p\^2)\^M &gt; (1 - 1/p\^2)\^p &gt; 1 - 1/p when $ M < p $ and $\delta \asymp \sqrt{\log p / n}$ [@ravikumar2011]. \[lemma:MSprob\] Suppose assumptions \[a1\] through \[a4\] hold. Let $\gamma$ be a constant in $[0,1]$ and define the events $$\begin{aligned} \mathcal{E}_\gamma(\hat{\Theta}^{D}(0)) & = & \left\{|\hat{\Theta}^{D}_{ij}(0)|>\gamma \sigma_{ij}\eta_\text{min}\text{ for }(i,j)\in\mathcal{S} \right\},\\ \mathcal{N}_{\gamma,b}(\hat{\Theta}^{D}(0)) & =& \left\{\sum_{(i,j)\in\mathcal{S}^C} \mathbb{I} \left(|\hat{\Theta}^D_{ij}(0)|>\gamma \sigma_{ij}\eta_\text{min}\right)\leq b\right\}.\end{aligned}$$ Thus, $\mathcal{E}_\gamma$ denotes the event that all non-zero entries of $\Theta$ in $\hat{\Theta}^D(0)$ are above a threshold $\gamma \sigma_{ij}\eta_\text{min}$, and $\mathcal{N}_{\gamma,b}$ the event when at most $b$ entries of $\hat{\Theta}^D(0)$ corresponding to the zeros of $\Theta$ are above the threshold. Then, $$\begin{aligned} \label{probInt} {\mathbb{P}\left(\mathcal{E}_\gamma\right)} &= &1 - {\mathcal{O}\left(s e^{-0.5 (1-\gamma)^2 n M ( \eta_\text{min} )^2 }\right)}\\ {\mathbb{P}\left(\mathcal{N}_{\gamma,b}\right)} & = & 1 - {\mathcal{O}\left((p^2-s-b) e^{-0.5 \gamma^2 n M \left( \eta_\text{min} \right)^2 }\right)}\\\end{aligned}$$ By the asymptotic normality of the debiased graphical lasso entries [@jankova2015], $\hat{\Theta}^D_{ij}(0)$ is asymptotically distributed as $N(\Theta_{ij}, \sigma_{ij}^2/(nM))$. Therefore, $$\begin{aligned} \label{probE} {\mathbb{P}\left(\mathcal{E}_\gamma^c\right)} & \leq & \notag \sum_{(i,j)\in\mathcal{S}}{\mathbb{P}\left(|\hat{\Theta}^D_{ij}(0)|<t_{ij}\right)}+o(1) \notag \\ & \leq & s\max_{(i,j)\in\mathcal{S}}{\mathbb{P}\left(|Z+\sqrt{nM}\frac{\Theta_{ij}}{\sigma_{ij}}|< \gamma \sqrt{nM} \eta_{\text{min}}\right)} +o(1)\notag \\ &\leq & s{\mathbb{P}\left(Z< -(1-\gamma)\sqrt{nM}\eta_{\text{min}}\right)}+o(1),\notag\end{aligned}$$ where $Z$ is a standard normal random variable. Using a similar argument, the second event can be bounded as $$\begin{aligned} \label{probN} {\mathbb{P}\left(\mathcal{N}_{\gamma,b}^c\right)} & = & {\mathbb{P}\left(\bigcup_{\substack{A\subset S^c \\ |A|\leq b}}\bigcap_{(i,j)\in A^c}(|\hat{\Theta}_{ij}^D(0)|\geq t_{ij})\right)} +o(1)\notag \\ & \leq & \min_{\substack{A\subset S^c \\ |A|\leq b}} \sum_{(i,j)\in A^c}{\mathbb{P}\left(|Z|\geq\gamma\sqrt{nM}\eta_{\text{min}}\right)}+o(1) \notag \\ &\leq & 2(p^2-s-b){\mathbb{P}\left(Z\leq-\gamma\sqrt{nM}\eta_{\text{min}}\right)}+o(1)\notag\end{aligned}$$ Using tails of a normal distribution, the results follow. *Part 1*. In order to recover the correct set of non-zero entries, the event $\mathcal{E}_{\tau}(\hat{\Theta}^{D}(\rho))\cap\mathcal{N}_{\tau,B-s}(\hat{\Theta}^{D}(\rho))$ must hold. Note that if in Algorithm 1, the threshold given by the bandwidth contains all non-zero entries, then $\hat{\Theta}^{D}_{ij}(\rho)=\hat{\Theta}^D_{ij}(0)$, for $(i,j)\in\mathcal{S}$. Moreover, since ${\mathbb{P}\left(|\hat{\Theta}^{D}_{ij}(\rho)|>h\right)} \leq {\mathbb{P}\left(|\hat{\Theta}^{D}_{ij}(0)|>h\right)}$ for $(i,j)\in\mathcal{S}^C$ and all $h>0$, then, by conditioning on the event $\cap_{m=1}^M\left(\mathcal{E}_{\rho}(\hat{\Theta}^d_m))\cap\mathcal{N}_{\rho,b}(\hat{\Theta}^d_m)\right)$, it holds that &[(\_(\^[D]{}())\_[,0]{}(\^[D]{}()))]{}\ &[(\_(\^[D]{}(0))\_[,0]{}(\^[D]{}(0)))]{}[(\_[m=1]{}\^M(\_(\^d\_m)\_[,b]{}(\^d\_m)))]{}. Using the conditions of Theorem \[theorem:rates\], in particular the size of $\eta_{\text{min}}$, and by Lemma \[lemma:MSprob\], it holds that the first term of the product is $1-{\mathcal{O}\left(p^{-1}\right)}$. Similarly, using Lemma \[lemma:MSprob\] again, we calculate the probability of correct recovery in all machines $$\begin{aligned} {\mathbb{P}\left(\cap_{m=1}^M\left(\mathcal{E}_{\rho}(\hat{\Theta}^d_m)\cap\mathcal{N}_{\rho,B-s}(\hat{\Theta}^d_m)\right)\right)} & = & \left(1-{\mathcal{O}\left(\frac{1}{p^{1+c_1}}\right)}\right)^M,\end{aligned}$$ which is $1-{\mathcal{O}\left(\frac{1}{p}\right)}$ when $M<p$, and $c_1>0$ is an absolute constant. Thus, correct model selection holds with high probability. *Part 2*. Using the equivalence between $\ell_\infty$ and Frobenius norm together with the correct recovery result of part 1, we obtain equation \[mse\] as a consequence of multiplying the result of Lemma \[lemma:avgglasso\] by $p+s$. Acknowledgments =============== The research in this paper was partially supported by the Consortium for Verification Technology under Department of Energy National Nuclear Security Administration award number DE-NA0002534.

ArXiv

--- author: - 'Chaoyu Chen$^{1}$, Shaolong He$^{1}$, Hongming Weng$^{1}$, Wentao Zhang$^{1}$, Lin Zhao$^{1}$, Haiyun Liu$^{1}$, Xiaowen Jia$^{1}$, Daixiang Mou$^{1}$, Shanyu Liu$^{1}$, Junfeng He$^{1}$, Yingying Peng$^{1}$, Ya Feng$^{1}$, Zhuojin Xie$^{1}$, Guodong Liu$^{1}$, Xiaoli Dong$^{1}$, Jun Zhang$^{1}$, Xiaoyang Wang$^{2}$, Qinjun Peng$^{2}$, Zhimin Wang$^{2}$, Shenjin Zhang$^{2}$, Feng Yang$^{2}$, Chuangtian Chen$^{2}$, Zuyan Xu$^{2}$, Xi Dai$^{1}$, Zhong Fang$^{1}$, X. J. Zhou$^{1,*}$' date: 'July 28, 2011' title: 'Persistence of Topological Order and Formation of Quantum Well States in Topological Insulators Bi$_2$(Se,Te)$_3$ under Ambient Conditions' --- [**The topological insulators represent a unique state of matter where the bulk is insulating with an energy gap while the surface is metallic with a Dirac cone protected by the time reversal symmetry[@Tin3D; @ZhangQiPT; @HasanKaneRev; @ZhangQiRMP; @JMoorePer]. These characteristics provide a venue to explore novel quantum phenomena in fundamental physics[@XLQiNP; @RDLiAxions; @XLQiMomo; @LFuMajorana; @RYuSience] and show potential applications in spintronics and quantum computing[@Revnextgeneration; @ZhangQiPT; @JMoorePer]. One critical issue directly related with the applications as well as the fundamental studies is how the topological surface state will behave under ambient conditions (1 atmosphere air and room temperature). In this paper, we report high resolution angle-resolved photoemission measurements on the surface state of the prototypical topological insulators, Bi$_2$Se$_3$, Bi$_2$Te$_3$ and Bi$_2$(Se$_{0.4}$Te$_{2.6}$), upon exposing to ambient conditions. We find that the topological order persists even when the surface is exposed to air at room temperature. However, the surface state is strongly modified after such an exposure. Particularly, we have observed the formation of two-dimensional quantum well states near the surface of the topological insulators after the exposure which depends sensitively on the original composition, [*x*]{}, in Bi$_2$(Se$_{3-x}$Te$_x$). These rich information are crucial in utilizing the surface state and in probing its physical properties under ambient conditions.** ]{} The angle-resolved photoemission spectroscopy (ARPES) is a powerful experimental tool to directly identify and characterize topological insulators[@HasanARPES]. A number of three-dimensional topological insulators have been theoretically predicted and experimentally identified by ARPES[@BiSbARPES; @BiSeTheo; @Bi2Se3ARPES; @Bi2Te3ARPES; @TlBiSe2Theo; @TlBiSe2ARPES]; some of their peculiar properties have been revealed by scanning tunneling microscopy (STM)[@TZhangSTM; @YazdaniSTM; @KaptulnikSTM; @PCheng; @Hanaguri]. The application of the topological surface states depends on the surface engineering that can be manipulated by incorporation of non-magnetic[@LAWraypn; @mnmi; @ccvb; @crs] or magnetic[@YLChenFe; @ZHasanPertur; @LAWraypn; @mnmi] impurities or gas adsorptions[@ZHasanPertur; @LPlucinski; @LAWraypn; @HMBenia]. While the ARPES and STM measurements usually involve the fresh surface obtained by cleaving samples [*in situ*]{} under ultra-high vacuum, for the transport measurements which are widely used to investigate the intrinsic quantum behaviors of the topological surface state[@DXQu; @JGAnalytis; @ZRen; @JChen], and particularly the ultimate utilizations of the topological insulators, the surface may be exposed to ambient conditions (1 atmosphere air and room temperature) or some gas protection environment. It is therefore crucial to investigate whether the topological order can survive under the ambient conditions, and furthermore, whether and how the surface state may be modified after such exposures. We start by first looking at the electronic structure of the prototypical topological insulators Bi$_2$(Se,Te)$_3$ under ultra-high vacuum. The Fermi surface and the band structure of the Bi$_2$(Se$_{3-x}$Te$_x$) topological insulators depend sensitively on the composition, x, as shown in Fig. 1. The single crystal samples here were all cleaved [*in situ*]{} and measured at 30 K in an ultra-high vacuum (UHV) chamber with a base pressure better than 5 $\times$ 10$^{-11}$ Torr. For Bi$_2$Se$_3$, a clear Dirac cone appears near -0.36 eV (Figs. 1d and 1e); the corresponding Fermi surface (Fig. 1a) is nearly circular but with a clear hexagon-shape in the measured data[@HexBi2Se3]. It is apparently of [*n*]{}-type because the Fermi level intersects with the bulk conduction band. On the other hand, the Dirac cone of the Bi$_2$Te$_3$ sample lies near -0.08 eV (Figs. 1h and 1i), much closer to the Fermi level than that reported before (-0.34 eV in [@Bi2Te3ARPES]). The corresponding Fermi surface (Fig. 1c) becomes rather small, accompanied by the appearance of six petal-like bulk Fermi surface sheets. These results indicate that our Bi$_2$Te$_3$ sample is of [*p*]{}-type because the Fermi level intersects the bulk valence band along the $\overline{\Gamma}$-$\overline{M}$ direction. This is also consistent with the positive Hall coefficient measured on the same Bi$_2$Te$_3$ sample[@CZhang]. This difference of the Fermi surface topology and the location of the Dirac cone from others[@Bi2Te3ARPES] may be attributed to the different carrier concentration in Bi$_2$Te$_3$ due to different sample preparation conditions. In our Bi$_2$(Se$_{3-x}$Te$_x$) samples, we have seen a crossover from n-type Bi$_2$Se$_3$ to p-type Bi$_2$Te$_3$. In order to eliminate the interference of the bulk bands on the surface state near the Fermi level, we fine tuned the composition [*x*]{} in Bi$_2$(Se$_{3-x}$Te$_x$) and found that, for [*x*]{}=2.6, nearly no spectral weight can be discerned from the bulk conduction band, as seen from both the Fermi surface (Fig. 1b) and the band structure (Figs. 1f and 1g). A slight substitution of Te by Se in Bi$_2$(Se$_{0.4}$Te$_{2.6}$) causes a dramatic drop of the Dirac point to -0.31 eV (Figs. 1f and 1g) and an obvious hexagon-shaped Fermi surface (Fig. 1b). It is interesting to note that the hexagon-shape of Bi$_2$(Se$_{0.4}$Te$_{2.6}$) (Fig. 1b) is rather pronounced, although its Fermi surface size is smaller than that of Bi$_2$Se$_3$ (Fig. 1a). The hexagonally-shaped Fermi surface observed in the topological surface states reflects the hybridization of surface electronic states with the bulk states and can be theoretically explained by considering the higher order terms in the $\emph{k}$ $\cdot$ $\emph{p}$ Hamiltonian[@HexWarp]. In order to directly examine how the topological surface state behaves under ambient conditions in the topological insulators, we carried out our ARPES measurements in different ways. (1). We first cleaved the sample *in situ* and performed ARPES measurement in the ultra-high vacuum (UHV) chamber. The sample was then pulled out to another chamber filled with 1 atmosphere N$_2$ gas, exposed for about 5 minutes, before transferring back to the UHV chamber to do ARPES measurements; (2). We cleaved and measured the sample in the UHV chamber, and then pulled the sample out to air for 5 minutes before transferring back to the UHV chamber for the ARPES measurements; (3). We cleaved the sample in air and then transferred it to the UHV chamber to do the ARPES measurement. Our measurements show that the above procedures of exposure to air or N$_2$ produce similar and reproducible results for a given sample. The surface exposure of the topological insulators to air or N$_2$ gives rise to a dramatic alteration of the surface state, as shown in Figs. 2, 3 and 4, for Bi$_2$Se$_3$, Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$), and Bi$_2$Te$_3$, respectively, when compared with those for the fresh surface (Fig. 1). The first obvious change is the shifting of the Dirac cone position relative to the Fermi level. For Bi$_2$Se$_3$, Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$) and Bi$_2$Te$_3$, it shifts from the original -0.36 eV (Figs. 1d and 1e), -0.31 eV (Figs. 1f and 1g), -0.08 eV (Figs. 1h and 1i) for the fresh surface to -0.48 eV (Fig. 2b), -0.40 eV (Figs. 3a and 3b), and -0.28 eV \[Figs. 4(c-f)\] at 30 K for the exposed surface, respectively. In all these cases, the shift of the Dirac cone to a larger binding energy indicates an additional doping of electrons into the surface state. The exposure also gives rise to a dramatic change of the surface Fermi surface. For Bi$_2$Se$_3$, in addition to a slight Fermi surface size increase, an obvious change occurs in the Fermi surface shape that the hexagon-shape becomes much more pronounced in the exposed surface (Fig. 2d) than that in the fresh sample (Fig. 1a). For Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$), one clearly observes the much-enhanced warping effect in the exposed surface (Fig. 3c) when compared with the nearly standard hexagon in the fresh surface (Fig. 1b). The most dramatic change occurs for Bi$_2$Te$_3$ where not only the Fermi surface size increases significantly, but also the warping effect in the exposed surface (Fig. 4i) becomes much stronger. Overall, the exposure causes the lowering of the Dirac cone position, an increase of the surface Fermi surface size, and an obvious enhancement of the Fermi surface warping effect in the Bi$_2$(Se$_{3-x}$Te$_x$) system. The topological order in the Bi$_2$(Se$_{3-x}$Te$_x$) topological insulators is robust even when the surface is exposed to ambient conditions, in spite of all the alterations mentioned above. One clearly observes the persistence of the Dirac cone in the exposed surface as in Bi$_2$Se$_3$ (Figs. 2b and 2c), in Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$) (Figs. 3a and 3b), and Bi$_2$Te$_3$ (Figs. 4(c-h)). Particularly, this is the case for the surface exposed to air and measured at room temperature (Fig. 2c for Bi$_2$Se$_3$, and Figs. 4g and 4h for Bi$_2$Te$_3$). On the other hand, after the exposure, although the signal of the surface state gets weaker for Bi$_2$Se$_3$ (Fig. 2), it remains rather strong for Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$) (Fig. 3) and Bi$_2$Te$_3$ (Fig. 4). This is in a stark contrast to the conventional trivial surface state where minor surface contamination will cause the extinction of the surface state[@Huefner]. The robustness of the topological order to Coulomb, magnetic and disorder perturbations has been reported before[@ZHasanPertur; @LPlucinski]. Our present observations directly demonstrate the robustness of the topological order against absorption and thermal process under ambient conditions due to the protection of the time-reversal symmetry[@HasanKaneRev; @ZhangQiRMP]. The surface exposure to air or N$_2$ in the Bi$_2$(Se$_{3-x}$Te$_x$) topological insulators produces two-dimensional electronic states near the surface. In Bi$_2$Se$_3$, the exposure gives rise to additional parabolic bands, as schematically marked by the dashed line in Figs. 2b and 2c. Correspondingly, this leads to additional Fermi surface sheet(s) inside the regular topological surface state (Figs. 2d and 2e). In Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$), this effect gets more pronounced and the newly emerged bulk conduction band splits into several discrete bands, as marked by the dashed lines in Fig. 3b. While the band quantization effect occurs in the bulk conduction band in Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$), it shows up in the valence band in the exposed Bi$_2$Te$_3$ surface, as shown in Figs. 4(c-h), where one can see a couple of discrete “M"-shaped bands. The quantized bands are obvious at low temperature and get slightly smeared out when the temperature rises to room temperature (Figs. 4g and 4h). The formation of the split bands in the exposed surface of the topological insulators is reminiscent of the quantum well states observed in the quantum confined systems[@SiAg] and in some topological insulators[@tdeg; @mnmi; @crs; @ccvb; @HMBenia; @KingReshba]. There are a couple of possibilities that the quantum well states may be formed. One usual way is due to band bending effect. As mentioned before, the surface exposure to air or N$_2$ causes an electron transfer to the surface of the topological insulators. The accumulation of these additional electrons near the surface would lead to a downward bending of the bulk bands near the surface region, as schematically shown in Fig. 3d, resulting in a “V"-shaped potential well where the bulk conduction band of electrons can be confined. This picture, as proposed before[@tdeg; @ccvb], seems to be able to explain the two-dimensional quantum well states in the conduction bands in Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$) (Figs. 3a and 3b). However, it becomes questionable to explain the quantum well states observed in the bulk valence band of Bi$_2$Te$_3$ (Figs. 4c-j). In this case, the downward band bending no longer acts as a quantum well potential for the valence band top because the charge carriers are hole-like. The band-bending is therefore not a general picture that can explain the formation of the two-dimensional quantum well states in Bi$_{2}$(Se$_{0.4}$Te$_{2.6}$) (Figs. 3a and 3b) and Bi$_2$Te$_3$ (Fig. 4) topological insulators on the same footing. An alternative scenario is the expansion of van der Waals spacings in between the quintuple layers (QLs) caused by the intercalation of gases[@SVEremeev]. The observation of multiple split bands with different spacings would ask for multiple van der Waals gaps with different expansions. Whether and how these can be realized in the exposed surface remains to be investigated. We note that our observations of multiple split bands are similar to those seen in the ultra-thin films of Bi$_2$Se$_3$[@YZhang] and Bi$_2$Te$_3$[@YYLi]. From our first principle band structure calculations on Bi$_2$Te$_3$ with different number of quintuple layers, we also find that a detached slab with a thickness of 7 quintuple layers can give a rather consistent description (Fig. 4l) of our observed results in terms of the quantitative spacings between the 3 resolved bands (VB0, VB1, and VB2 bands as marked in Figs. 4c and 4l). In addition, the distance between the conduction band bottom (CB0 band in Figs. 4i and 4l) and the first valence sub-band bottom (VB0 band in Figs. 4i and 4l) is rather consistent between the measured and calculated results. These seem to suggest that a “confined surface slab" with nearly 7 quintuple layers may be formed after the exposure that acts more or less independently from the bulk. More work needs to be done to further investigate whether such a confined surface slab can be thermodynamically stable. Overall, the formation of the two-dimensional quantum well states is a general phenomenon for the exposed surface of the Bi$_2$(Se$_{3-x}$Te$_x$) topological insulators; the effect depends sensitively on the composition [*x*]{} of the samples which may facilitate manipulation of these quantum well states. The present work has significant implications on the fundamental study and ultimate applications of the topological insulators. Many experimental measurements, such as some transport measurements, involve samples exposed to ambient conditions. The practical applications may involve sample surface either exposed to ambient condition, or be in contact with other magnetic or superconducting materials. On the one hand, the robustness of the topological order under ambient conditions sends a good signal for these experimental characterization and practical utilizations. The formation of the quantum well states may give rise to new phenomena to be studied and utilized. The sensitivity of the surface state to the Bi$_2$(Se$_{3-x}$Te$_x$) composition provides a handle to manipulate these quantum states. On the other hand, the strong modification of the electronic structure and the formation of additional quantum well states in the exposed surface have to be considered seriously in interpreting experimental data and in surface engineering. It is critical to realize before hand that the surface under study or to be utilized may exhibit totally different behaviors as those from the fresh surface cleaved in ultra-high vacuum. In addition to the alteration of electronic states upon exposure, the transport properties of the topological surface state may be further complicated by the formation of quantum well states. We thank Prof. X. H. Chen for providing us samples at the initial stage of the project, and Prof. Liling Sun and Prof. Zhong-xian Zhao for their help in the characterization of the samples. This work is supported by the NSFC (91021006) and the MOST of China (973 program No: 2011CB921703). [ ****METHODS**** ]{} [**Crystal growth methods** ]{} Single crystals of $Bi_{2}(Se_{3-x}Te_{x})$ (x=0, 2.6 and 3) were grown by the self-flux method. Bismuth, selenium and tellurium powders were weighed according to the stoichiometric $Bi_{2}(Se_{3-x}Te_{x})$ (x=0, 2.6 and 3) composition. After mixing thoroughly, the powder was placed in alumina crucibles and sealed in a quartz tube under vacuum. The materials were heated to 1000 $^\circ$C, held for 12 hours to obtain a high degree of mixing, and then slowly cooled down to 500 $^\circ$C over 100 hours before cooling to room temperature. Single crystals of several millimeters in size were obtained. The crystal structure of the resulting crystals was examined by use of a rotating anode x-ray diffractometer with Cu *K$_{\alpha}$* radiation ($\lambda$ = 1.5418 [Å]{}). The chemical composition of the crystals was analyzed by the energy dispersive X-ray spectroscopy (EDAX) and the induction-coupled plasma atomic emission spectroscopy (ICP-AES). The resistivity of the crystals was measured by the standard four-probe method. [**Laser-ARPES methods.** ]{} The angle-resolved photoemission measurements were carried out on our vacuum ultra-violet (VUV) laser-based angle-resolved photoemission system[@LiuARPES]. The photon energy of the laser is 6.994 eV with a bandwidth of 0.26 meV. The energy resolution of the electron energy analyzer (Scienta R4000) is set at 1 meV, giving rise to an overall energy resolution of $\sim$1 meV which is significantly improved from 10$\sim$15 meV from regular synchrotron radiation systems[@Bi2Se3ARPES; @Bi2Te3ARPES]. The angular resolution is $\sim$0.3$^\circ$, corresponding to a momentum resolution $\sim$0.004 $\AA$$^{-1}$ at the photon energy of 6.994 eV, more than twice improved from 0.009 $\AA$$^{-1}$ at a regular photon energy of 21.2 eV for the same angular resolution. Our superior instrumental resolution of laser ARPES has made the measured features of topological insulators in this work much sharper. The Fermi level is referenced by measuring on a clean polycrystalline gold that is electrically connected to the sample. The samples were all measured in vacuum with a base pressure better than 5$\times$10$^{-11}$ Torr. $^{*}$Corresponding author: [email protected] [99]{} Fu, L., Kane, C. L. $\&$ Mele, E. J. Topological insulators in three dimensions. Phys Rev Lett [**98**]{}, 106803 (2007). Qi, X. L. $\&$ Zhang, S. C. The quantum spin Hall effect and topological insulators. Physics Today [**63**]{}, 33-38 (2010). Hasan, M. Z. $\&$ Kane, C. L. Colloquium: Topological insulators. Rev Mod Phys [**82**]{}, 3045 (2010). Qi, X. L. $\&$ Zhang, S. C. Topological insulators and superconductors. Preprint at $<$ http://arxiv.org/abs/1008.2026 $>$ (2011). Moore, J. E. The birth of topological insulators. Nature [**464**]{}, 194-198 (2010). Qi, X.-L., Hughes, T. L. $\&$ Zhang, S.-C. Fractional charge and quantized current in the quantum spin Hall state. Nat Phys [**4**]{}, 273-276 (2008). Li, R., Wang, J., Qi, X.-L. $\&$ Zhang, S.-C. Dynamical axion field in topological magnetic insulators. Nat Phys [**6**]{}, 284-288 (2010). Qi, X.-L., Li, R., Zang, J. $\&$ Zhang, S.-C. Inducing a magnetic monopole with topological surface states. Science [**323**]{}, 1184-1187 (2009). Fu, L. $\&$ Kane, C. L. Superconducting proximity effect and Majorana fermions at the surface of a topological insulator. Phys Rev Lett [**100**]{}, 096407 (2008). Yu, R. et al. Quantized anomalous Hall effect in magnetic topological insulators. Science [**329**]{}, 61-64 (2010). Moore, J.E. Topological insulators the next generation. Nat Phys [**5**]{}, 378-380 (2009). Hasan, M. Z. A new experimental approach for the exploration of topological quantum phenomena. Preprint at $<$ http://arxiv.org/abs/1105.0396 $>$ (2011). Hsieh, D. et al. A topological Dirac insulator in a quantum spin Hall phase. Nature [**452**]{}, 970 (2008). Zhang, H. J. et al. Topological insulators in Bi$_2$Se$_3$, Bi$_2$Te$_3$ and Sb$_2$Te$_3$ with a single Dirac cone on the surface. Nat Phys [**5**]{}, 438 (2009). Xia, Y. et al. Observation of a large-gap topological insulator class with a single Dirac cone on the surface. Nat Phys [**5**]{}, 398 (2009). Chen, Y. L. et al. Experimental realization of a three-dimensional topological insulator, Bi$_2$Te$_3$. Science [**325**]{}, 178 (2009). Yan, B. H. et al. Theoretical prediction of topological insulators in thallium-based III-V-VI2 ternary chalcogenides. EPL [**90**]{}, 37002,(2010); Lin, H. et al. Single-Dirac-cone topological surface states in the TlBiSe$_2$ class of topological semiconductors. Phys Rev Lett [**105**]{}, 036404 (2010). Kuroda, K. et al. Experimental realization of a three-dimensional topological insulator phase in ternary chalcogenide TlBiSe$_2$. Phys Rev Lett [**105**]{}, 146801 (2010); Sato, T. et al. Direct evidence for the Dirac-cone topological surface states in the ternary chalcogenide TlBiSe$_2$. Phys Rev Lett [**105**]{}, 136802 (2010); Chen, Y. L. et al. Single Dirac cone topological surface state and unusual thermoelectric property of compounds from a new topological insulator family. Phys Rev Lett [**105**]{}, 266401 (2010). Zhang, T. et al. Experimental demonstration of topological surface states protected by time-reversal symmetry. Phys Rev Lett [**103**]{}, 266803 (2009). Roushan, P. et al. Topological surface states protected from backscattering by chiral spin texture. Nature [**460**]{}, 1106-1109 (2009). Alpichshev, Z. et al. STM imaging of electronic waves on the surface of Bi$_2$Te$_3$: topologically protected surface states and hexagonal warping effects. Phys Rev Lett [**104**]{}, 016401 (2010). Cheng, P. et al. Landau quantization of topological surface states in Bi$_2$Se$_3$. Phys Rev Lett [**105**]{}, 076801 (2010). Hanaguri, T. et al. Momentum-resolved Landau-level spectroscopy of Dirac surface state in Bi$_2$Se$_3$. Phys Rev B [**82**]{}, 081305(R) (2010). Wray, L. A. et al. Electron dynamics in topological insulator semiconductor-metal interfaces (topological p-n interface). Preprint at $<$ http://arxiv.org/abs/1102.3935 $>$ (2011). Bianchi, M. et al. Simultaneous quantization of bulk conduction and valence states through adsorption of nonmagnetic impurities on Bi$_2$Se$_3$. Preprint at $<$ http://arxiv.org/abs/1105.3826 $>$ (2011). Zhu, Z.-H. et al. Rashba spin-splitting control at the surface of the topological insulator Bi$_2$Se$_3$. Preprint at $<$ http://arxiv.org/abs/1106.0552 $>$ (2011). Pan, Z.-H. et al. Scattering on Magnetic and Non-magnetic Impurities on a Surface of a Topological Insulator. Preprint at $<$ http://arxiv.org/abs/1104.0966 $>$ (2011). Chen, Y. L. et al. Massive Dirac Fermion on the surface of a magnetically doped topological insulator. Science [**329**]{}, 659 (2010). Wray, L. A. et al., A topologocal insulator surface under strong Coulomb, magnetic and disorder perturbations, Nature Phys [**7**]{}, 32 (2011). Plucinski, L. et al., Robust surface electronic properties of topological insulators: Bi$_2$Te$_3$ films grown by molecular beam epitaxy. App Phys Lett [**98**]{}, 222503 (2011). Benia, H. M. et al. Reactive chemical doping of the Bi$_2$Se$_3$ topological insulator. Preprint at $<$ http://arxiv.org/abs/1105.2664 $>$ (2011). Qu, D.-X., Hor, Y. S., Xiong, J., Cava, R. J. $\&$ Ong, N. P. Quantum oscillations and Hall anomaly of surface states in the topological insulator Bi$_2$Te$_3$. Science 329, 821-824 (2010). Analytis, J. G. et al. Two-dimensional surface state in the quantum limit of a topological insulator. Nat Phys [**6**]{}, 960-964, (2010). Ren, Z., Taskin, A. A., Sasaki, S., Segawa, K. $\&$ Ando, Y. Large bulk resistivity and surface quantum oscillations in the topological insulator Bi$_2$Te$_2$Se. Phys Rev B [**82**]{}, 241306 (2010). Chen, J. et al. Gate-voltage control of chemical potential and weak antilocalization in Bi$_2$Se$_3$. Phys Rev Lett [**105**]{}, 176602 (2010). Kuroda, K. et al. Hexagonally deformed Fermi surface of the 3D topological insulator Bi$_2$Se$_3$. Phys Rev Lett [**105**]{}, 076802 (2010). Zhang, C. et al. Phase diagram of a pressure-induced superconducting state and its relation to the Hall coefficient of Bi$_2$Te$_3$ single crystals. Phys Rev B [**83**]{}, 140504 (2011). Fu, L. Hexagonal warping effects in the surface states of the topological insulator Bi$_2$Te$_3$. Phys Rev Lett [**103**]{}, 266801 (2009). Reinert, F. et al. Direct measurements of the L-gap surface states on the (111) face of noble metals by photoelectron spectroscopy. Phys Rev B [**63**]{}, 115415 (2001). Speer, N. J., Tang, S.-J., Miller, T. $\&$ Chiang, T.-C. Coherent electronic fringe structure in incommensurate silver-silicon quantum wells. Science [**314**]{}, 1132941 (2006). Bianchi, M. et al. Coexistence of the topological state and a two-dimensional electron gas on the surface of Bi$_2$Se$_3$. Nat Commun [**1**]{}, 128 (2010). King, P. D. C. Large tuneable Rashba spin splitting of a two-dimensional electron gas in Bi$_2$Se$_3$. Preprint at $<$ http://arxiv.org/abs/1103.3220 $>$ (2011). Eremeev, S. V. et al. New interpretation of the origin of 2DEG states at the surface of layered topological insulators, Preprint at $<$ http://arxiv.org/abs/1107.3208 $>$ (2011). Zhang, Y. et al., Crossover of the three-dimensional topological insulator Bi$_2$Se$_3$ to the two-dimensional limit, Nat Phys [**6**]{}, 584 (2010). Li, Y. Y. et al., Intrinsic topological insulator Bi$_2$Te$_3$ thin films on Si and their thickness limit. Adv Mater[**22**]{}, 4002 (2010). Liu, G. D. et al. Development of a vacuum ultraviolet laser-based angle-resolved photoemission system with a superhigh energy resolution better than 1 meV. Rev Sci Instrum [**79**]{}, 023105 (2008). {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}

ArXiv

--- abstract: 'Objective probabilistic forecasts of future climate that include parameter uncertainty can be made by using the Bayesian prediction integral with the prior set to Jeffreys’ Prior. The calculations involved in determining the prior can then be simplified by making parametric assumptions about the distribution of the output from the climate model. The most obvious assumption to make is that the climate model output is normally distributed, in which case evaluating the prior becomes a question of evaluating gradients in the parameters of the normal distribution. In previous work we have considered the special cases of diagonal (but not constant) covariance matrix, and constant (but not diagonal) covariance matrix. We now derive expressions for the general multivariate normal distribution, with non-constant non-diagonal covariance matrix. The algebraic manipulation required is more complex than for the special cases, and involves some slightly esoteric matrix operations including taking the expectation of a vector quadratic form and differentiating the determinants, traces and inverses of matrices.' author: - 'Stephen Jewson[^1], Dan Rowlands, Myles Allen\' bibliography: - 'arxiv.bib' title: ' Objective Climate Model Predictions Using Jeffreys’ Prior: the General Multivariate Normal Case ' --- Introduction ============ We are interested in using climate models to produce objective probabilistic forecasts of future climate that include parameter uncertainty. The word ‘objective’ is used here in a technical statistical sense that means that the prior distribution for the parameters is determined by a rule, rather than from intuition. In statistics the most widely discussed rule is the Jeffreys’ rule [@jeffreys]. Jeffreys’ rule could, in principal, be applied to a climate model directly from the definition. This would translate into using numerical methods to differentiate the predicted probabilities from initial condition ensembles and to take expectations over all simulated climate states. That could, however, be computationally demanding. As an alternative one could fit distributions to the output from the model, and differentiate the estimated parameters of the distributions instead, which is likely to be considerably easier. The most obvious distribution to fit is then the multivariate normal distribution. In previous work we have considered two special cases. In @jp1 we considered the case where the predicted variables are independent, but both the mean and the variance of initial condition ensembles from the model are allowed to vary as a function of the model parameters. In @jp2 we considered the complementary case where the predicted variables can be correlated, but with a covariance matrix that is constant as a function of the model parameters. We now consider the general case, with correlated predicted variables *and* a covariance matrix that can vary as a function of the model parameters. This general case contains the two special cases. The algebra in this case is not quite as elementary as the two special cases, in that we have to take expectations of a vector quadratic form, and differentiate the determinants, traces and inverses of matrices. Pedagogically, therefore, we consider the derivations used in the two simpler cases as still being useful, especially as climate modelling practice is perhaps unlikely to reach the stage where it would be necessary to consider the covariance matrix varying as a function of parameters for quite some time (there are many other more important challenges to be dealt with first). Detailed explanations of Jeffreys’ Prior, and the motivation for its use, are given in @jp1 and will not be repeated here. In section \[s2\] below we give the expressions for Jeffreys’ Prior and the multivariate normal density. In section \[s3\] we then derive the expression for Jeffreys’ Prior for the multivariate normal distribution where the climate model has just a single parameter. In section \[s4\] we derive the same expression, but considering multiple parameters. In both section \[s3\] and section \[s4\] we also consider four special cases: independence (taking us back to the results in @jp1), constant covariance (taking us back to the results in @jp2), constant correlation and constant variance. Jeffreys’ Prior and the Multivariate Normal Density {#s2} =================================================== We start with definitions of the Jeffreys’ Prior and the multivariate normal density. Jeffreys’ Prior --------------- Jeffreys’ Prior is given by: $$\label{eq1} p(\theta)=\sqrt{\mbox{det}\left[-\mbox{E}\left(\frac{\partial^2 \log p(x|\theta)}{\partial \theta_j \partial \theta_k}\right)\right]}\\ \label{jp_def}$$ Detailed explanations of this equation, which can at first be rather difficult to understand, are given in @jp1 and @jp2, as well as many statistics textbooks such as @peterlee. The Multivariate Normal Density ------------------------------- Probability densities from the multivariate normal distribution are given by: $$p(x|\theta)=\frac{1}{(2\pi)^\frac{n}{2}} \frac{1}{D^{\frac{1}{2}}} \mbox{exp}\left(-\frac{1}{2}(x-\mu)^T S(x-\mu)\right)$$ where, in our application: - $x$ is a vector of those variables predicted by the climate model that are to be compared with observations - $\theta$ is a vector for the underlying parameters in the climate model - $\mu$=$\mu(\theta)$ is a vector of the mean response of the model (in other words, the ensemble mean of $x$ for an infinite-sized initial condition ensemble for fixed parameters $\theta$) - $\Sigma=\Sigma(\theta)$ is the covariance matrix of the response of the model (in other words, the ensemble covariance matrix of $x$ for an infinite-sized initial condition ensemble for fixed parameters $\theta$) - $S=\Sigma^{-1}=S(\theta)$ is the inverse of the covariance matrix - $D=\mbox{det}(\Sigma)=D(\theta)$ is the determinant of the covariance matrix This gives: $$\begin{aligned} \ln p(x|\theta) &=&-\frac{n}{2}\ln 2\pi -\frac{1}{2} \ln D -\frac{1}{2}(x-\mu)^T S(x-\mu)\\ &=&-\frac{n}{2}\ln 2\pi -\frac{1}{2} \ln D -\frac{1}{2}(x^T S x-x^TS \mu-\mu^T S x+\mu^T S \mu)\end{aligned}$$ Since $S$ is symmetric we have $x^TS\mu=\mu^T S x$ which means that the above expression for $\ln p(x|\theta)$ simplifies a little to: $$\begin{aligned} \ln p(x|\theta)\label{eq1} &=&-\frac{n}{2}\ln 2\pi -\frac{1}{2} \ln D -\frac{1}{2}(x^T S x-2x^TS \mu+\mu^T S \mu)\end{aligned}$$ We now consider two cases: one parameter and multiple parameters. One parameter {#s3} ============= The first case we consider is where there is just a single parameter in the climate model. We mainly consider this case as a warm-up for the multiple parameter case, although it would also be relevant if one only wanted to model the uncertainty due to a single parameter, which might be a good approximation to the overall parameter uncertainty if that single parameter dominates the uncertainty. If we consider $\theta$ to be this single (scalar) parameter, then differentiating equation \[eq1\] by $\theta$ gives: $$\begin{aligned} \frac{\partial \ln p(x|\theta)}{\partial \theta} &=&-\frac{1}{2} \frac{\partial \ln D}{\partial \theta} -\frac{1}{2}x^T \frac{\partial S}{\partial \theta} x +x^T\frac{\partial (S \mu)}{\partial \theta} -\frac{1}{2}\frac{\partial (\mu^T S \mu)}{\partial \theta}\end{aligned}$$ Expanding the third and fourth terms: $$\begin{aligned} \frac{\partial \ln p(x|\theta)}{\partial \theta} &=&-\frac{1}{2} \frac{\partial \ln D}{\partial \theta} -\frac{1}{2}x^T \frac{\partial S}{\partial \theta} x +x^T \left(\frac{\partial S}{\partial \theta}\mu+S\frac{\partial \mu}{\partial \theta}\right) -\frac{1}{2}\left(\frac{\partial \mu^T }{\partial \theta}S \mu+\mu^T\frac{\partial S }{\partial \theta}\mu+\mu^T S \frac{\partial \mu}{\partial \theta}\right)\\ &=&-\frac{1}{2} \frac{\partial \ln D}{\partial \theta} -\frac{1}{2}x^T \frac{\partial S}{\partial \theta} x +x^T \frac{\partial S}{\partial \theta}\mu+x^T S\frac{\partial \mu}{\partial \theta} -\frac{1}{2}\frac{\partial \mu^T }{\partial \theta}S \mu-\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\mu-\frac{1}{2}\mu^T S \frac{\partial \mu}{\partial \theta}\end{aligned}$$ Since $S$ is symmetric $\mu^T S \frac{\partial \mu}{\partial \theta}=\frac{\partial \mu^T }{\partial \theta}S \mu$ and so the 5th and 7th terms combine, giving: $$\begin{aligned} \label{eq10} \frac{\partial \ln p(x|\theta)}{\partial \theta} &=&-\frac{1}{2} \frac{\partial \ln D}{\partial \theta} -\frac{1}{2}x^T \frac{\partial S}{\partial \theta} x +x^T \frac{\partial S}{\partial \theta}\mu +x^T S\frac{\partial \mu}{\partial \theta} -\frac{\partial \mu^T }{\partial \theta}S \mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\mu\end{aligned}$$ Differentiating again wrt $\theta$ gives: $$\begin{aligned} \frac{\partial^2 \ln p(x|\theta)}{\partial \theta^2} &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta^2} x +x^T \frac{\partial}{\partial \theta}\left(\frac{\partial S}{\partial \theta}\mu\right) +x^T \frac{\partial}{\partial \theta}\left(S\frac{\partial \mu}{\partial \theta}\right)\\ && -\frac{\partial}{\partial \theta}\left(\frac{\partial \mu^T }{\partial \theta}S \mu\right) -\frac{1}{2}\frac{\partial}{\partial \theta}\left(\mu^T\frac{\partial S }{\partial \theta}\mu\right)\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta^2} x +x^T \left(\frac{\partial^2 S}{\partial \theta^2}\mu+\frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta}\right) +x^T \left(\frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta}+S\frac{\partial^2 \mu}{\partial \theta^2}\right)\\ && -\left(\frac{\partial^2 \mu^T }{\partial \theta^2}S \mu +\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \theta} \mu +\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}\right) -\frac{1}{2}\left(\frac{\partial \mu^T}{\partial \theta}\frac{\partial S }{\partial \theta}\mu +\mu^T\frac{\partial^2 S }{\partial \theta^2}\mu +\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \theta}\right)\\ &&\mbox{(where we have expanded the derivatives of products)}\nonumber\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta^2} x +x^T \frac{\partial^2 S}{\partial \theta^2}\mu +x^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta} +x^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta} +x^T S\frac{\partial^2 \mu}{\partial \theta^2}\\ && -\frac{\partial^2 \mu^T }{\partial \theta^2}S \mu -\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \theta} \mu -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\frac{\partial \mu^T}{\partial \theta}\frac{\partial S }{\partial \theta}\mu -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta^2}\mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \theta}\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta^2} x +x^T \frac{\partial^2 S}{\partial \theta^2}\mu +2x^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta} +x^T S\frac{\partial^2 \mu}{\partial \theta^2}\\ && -\frac{\partial^2 \mu^T }{\partial \theta^2}S \mu -2\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \theta} \mu -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta^2}\mu\\ &&\mbox{(where we have combined the 4th and 5th, and 8th, 10th and 12th terms)}\nonumber\end{aligned}$$ Taking expectations: $$\begin{aligned} E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta^2}\right) &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2}E\left(x^T \frac{\partial^2 S}{\partial \theta^2} x\right) +E(x^T) \frac{\partial^2 S}{\partial \theta^2}\mu +2E(x^T) \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta}\\ && +E(x^T) S\frac{\partial^2 \mu}{\partial \theta^2} -\frac{\partial^2 \mu^T }{\partial \theta^2}S \mu -2\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \theta} \mu -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta^2}\mu\end{aligned}$$ But $E(x)=\mu$, and, from appendix 2, $$E\left(x^T \frac{\partial^2 S}{\partial \theta^2} x\right)=\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta^2}\Sigma\right)+\mu^T \frac{\partial^2 S}{\partial \theta^2} \mu$$ and so: $$\begin{aligned} E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta^2}\right) &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2} \left(\mbox{tr}\left(\frac{\partial^2 S}{\partial \theta^2}\Sigma \right)+\mu^T \frac{\partial^2 S}{\partial \theta^2} \mu\right) +\mu^T \frac{\partial^2 S}{\partial \theta^2}\mu +2\mu^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta}\\ && +\mu^T S\frac{\partial^2 \mu}{\partial \theta^2} -\frac{\partial^2 \mu^T }{\partial \theta^2}S \mu -2\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \theta} \mu -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta^2}\mu\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta^2} -\frac{1}{2} \mbox{tr}\left( \frac{\partial^2 S}{\partial \theta^2}\Sigma\right) -\frac{1}{2} \mu^T \frac{\partial^2 S}{\partial \theta^2} \mu +\mu^T \frac{\partial^2 S}{\partial \theta^2}\mu +2\mu^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \theta}\\ && +\mu^T S\frac{\partial^2 \mu}{\partial \theta^2} -\frac{\partial^2 \mu^T }{\partial \theta^2}S \mu -2\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \theta} \mu -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta^2}\mu\\\label{eq2} &=&-\frac{1}{2} \left[\frac{\partial^2 \ln D}{\partial \theta^2}+\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta^2}\Sigma\right)\right] -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}\\ &&\mbox{(where we have cancelled various terms)}\nonumber\end{aligned}$$ But $$\begin{aligned} \frac{\partial^2 \ln D}{\partial \theta^2} &=&\frac{\partial}{\partial \theta}\left(\frac{\partial \log D}{\partial \theta}\right)\\ &=&\frac{\partial}{\partial \theta}\left(\frac{1}{D}\frac{\partial D}{\partial \theta}\right)\\ &=&\frac{\partial}{\partial \theta}\left(\frac{1}{D}\left(D\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \theta}\right)\right)\right)\\ &&\mbox{(using a standard result for the derivative of a determinant known as Jacobi's formula)}\nonumber\\ &=&\frac{\partial}{\partial \theta}\left(\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \theta}\right)\right)\\ &=&\mbox{tr}\left(\frac{\partial}{\partial \theta}\left(S\frac{\partial \Sigma}{\partial \theta}\right)\right)\\ &&\mbox{(using a standard result for the derivative of a trace)}\nonumber\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}+S\frac{\partial^2 \Sigma}{\partial \theta^2}\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)+\mbox{tr}\left( S \frac{\partial^2 \Sigma}{\partial \theta^2}\right)\\ &&\mbox{(using the linearity of the trace operator)}\nonumber\end{aligned}$$ Now note that $$\begin{aligned} \frac{\partial}{\partial \theta}(S \Sigma)&=&S \frac{\partial \Sigma}{\partial \theta}+\frac{\partial S}{\partial \theta}\Sigma\end{aligned}$$ and $$\begin{aligned} \frac{\partial^2}{\partial \theta^2}( S \Sigma) &=&S \frac{\partial^2 \Sigma}{\partial \theta^2}+\frac{\partial^2 S}{\partial \theta^2}\Sigma+2\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\end{aligned}$$ But $S\Sigma=I$ and so $\frac{\partial^2}{\partial \theta^2}( S \Sigma)=0$, implying that $$S\frac{\partial^2 \Sigma}{\partial \theta^2}=-\frac{\partial^2 S}{\partial \theta^2}\Sigma -2\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}$$ and so $$\begin{aligned} \frac{\partial^2 \ln D}{\partial \theta^2} &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)+\mbox{tr}\left( S \frac{\partial^2 \Sigma}{\partial \theta^2}\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)+\mbox{tr}\left(- \frac{\partial^2 S}{\partial \theta^2}-2\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\Sigma\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)-\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta^2}\Sigma\right)-\mbox{tr}\left(2\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)\\ &=&-\mbox{tr}\left(\frac{\partial^2 S}{\partial \theta^2}\Sigma \right)-\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)\end{aligned}$$ Giving: $$\begin{aligned} \frac{\partial^2 \ln D}{\partial \theta^2}+\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta^2}\Sigma\right) &=&\mbox{tr}\left(- \frac{\partial^2 S}{\partial \theta^2}\Sigma\right)-\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)+\mbox{tr}\left(S \frac{\partial^2 \Sigma}{\partial \theta^2}\right)\\\label{eq3} &=&-\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)\end{aligned}$$ Returning to equation \[eq2\] and substituting in the expression given in equation \[eq3\] gives: $$\begin{aligned} E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta^2}\right) &=&-\frac{1}{2} \left[\frac{\partial^2 \ln D}{\partial \theta^2}+\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta^2}\Sigma\right)\right] -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}\\ &=&-\frac{1}{2} \left[-\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)\right] -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}\\ &=&\frac{1}{2}\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right) -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}\\ &=&-\frac{1}{2}\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \theta}S\frac{\partial \Sigma}{\partial \theta}\right) -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}\\ &&\mbox{(using the standard result that $\frac{\partial A^{-1}}{\partial \theta}=-A^{-1}\frac{\partial A}{\partial \theta}A^{-1}$)}\end{aligned}$$ This gives the prior: $$\begin{aligned} p(\theta) &=&\sqrt{-E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta^2}\right)}\\ &=&\sqrt{\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)}\label{eq100}\\ &=&\sqrt{\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} +\frac{1}{2}\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \theta}S\frac{\partial \Sigma}{\partial \theta}\right)}\label{eq100b}\end{aligned}$$ If there are $n$ observations then: - $\mu$ is an $n$ x $1$ vector - $\frac{\partial \mu}{\partial \theta}$ is an $n$ by $1$ vector - $\mu^T$ is a $1$ x $n$ vector - $\frac{\partial \mu^T}{\partial \theta}$ is a $1$ by $n$ vector - $S$ is an $n$ by $n$ matrix - $\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}$ is a scalar - $\frac{\partial S}{\partial \theta}$ is an $n$ by $n$ matrix - $\frac{\partial \Sigma}{\partial \theta}$ is an $n$ by $n$ matrix - $\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}$ is an $n$ by $n$ matrix - tr$\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right)$ is a scalar We now consider various special cases of this general formula, starting with the two cases described in @jp1 and @jp2. Independence ------------ If the observations are modelled as independent then: $$\begin{aligned} \Sigma &=&\mbox{diag}\left(\sigma_1^2,\sigma_2^2,...,\sigma_n^2\right)\\ S &=&\mbox{diag}\left(\sigma_1^{-2},\sigma_2^{-2},...,\sigma_n^{-2}\right)\\ \frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} &=&\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \mu_i}{\partial \theta}\right)^2\\ \frac{\partial \Sigma}{\partial \theta} &=&\mbox{diag}\left(2\sigma_1 \frac{\partial \sigma_1}{\partial \theta},2\sigma_2\frac{\partial \sigma_2}{\partial \theta},...,2\sigma_n\frac{\partial \sigma_n}{\partial \theta}\right)\\ \frac{\partial S}{\partial \theta} &=&\mbox{diag}\left(-2\sigma_1^{-3} \frac{\partial \sigma_1}{\partial \theta},-2\sigma_2^{-3}\frac{\partial \sigma_2}{\partial \theta},...,-2\sigma_n^{-3}\frac{\partial \sigma_n}{\partial \theta}\right)\\ \frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta} &=&\mbox{diag}\left(-4\sigma_1^{-2} \frac{\partial \sigma_1}{\partial \theta},-4\sigma_2^{-2}\frac{\partial \sigma_2}{\partial \theta},...,-4\sigma_n^{-2}\frac{\partial \sigma_n}{\partial \theta}\right)\\ &=&-4\mbox{diag}\left(\frac{1}{\sigma_1^2} \frac{\partial \sigma_1}{\partial \theta},\frac{1}{\sigma_2^2}\frac{\partial \sigma_2}{\partial \theta},...,\frac{1}{\sigma_n^2}\frac{\partial \sigma_n}{\partial \theta}\right)\\ -\frac{1}{2}\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta}\right) &=&2\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \sigma_i}{\partial \theta}\right)^2\end{aligned}$$ and so the prior is $$\begin{aligned} p(\theta)&=&\sqrt{\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \mu_i}{\partial \theta}\right)^2 +2\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \sigma_i}{\partial \theta}\right)^2}\end{aligned}$$ which agrees with equation 19 in @jp1. Constant Covariance ------------------- If the covariance $\Sigma$ is constant then equation \[eq100\] reduces immediately to $$\begin{aligned} p(\theta)&=&\sqrt{\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta}}\end{aligned}$$ which agrees with equation 19 in @jp2. Constant Correlation -------------------- If the correlations are modelled as constant (but the variances are allowed to vary) then: $$\begin{aligned} \Sigma&=&VCV\\ \mbox{where }V&=&\mbox{diag}(\sigma_1^2,\sigma_2^2,...,\sigma_n^2)\\ \mbox{and }C&=&\mbox{the correlation matrix}\\ \frac{\partial \Sigma}{\partial \theta} &=& \frac{\partial V}{\partial \theta}CV+VC\frac{\partial V}{\partial \theta}\\ &=& 2\frac{\partial V}{\partial \theta}CV\\ \mbox{where } \frac{\partial V}{\partial \theta} &=&2\mbox{diag}(\sigma_1\frac{\partial \sigma_1}{\partial \theta},\sigma_2\frac{\partial \sigma_2}{\partial \theta},...,\sigma_n\frac{\partial \sigma_n}{\partial \theta})\\ S&=&V^{-1} C^{-1} V^{-1}\\ \mbox{where }V^{-1}&=&\mbox{diag}(\sigma_1^{-2},\sigma_2^{-2},...,\sigma_n^{-2})\\ \mbox{and }C^{-1}&=&\mbox{the inverse correlation matrix}\\ \frac{\partial \Sigma}{\partial \theta} &=& \frac{\partial V^{-1}}{\partial \theta}C^{-1}V^{-1}+V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta}\\ &=& 2V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta}\\ \mbox{where } \frac{\partial V^{-1}}{\partial \theta} &=&-2\mbox{diag}\left(\frac{1}{\sigma_1^2}\frac{\partial \sigma_1}{\partial \theta},\frac{1}{\sigma_2^2}\frac{\partial \sigma_2}{\partial \theta},...,\frac{1}{\sigma_n^2}\frac{\partial \sigma_n}{\partial \theta}\right)\\ \frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta} &=&4 V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta}\frac{\partial V}{\partial \theta}CV\\ \frac{\partial V^{-1}}{\partial \theta}\frac{\partial V}{\partial \theta} &=&-4\mbox{diag}\left(\frac{1}{\sigma_1}\left(\frac{\partial \sigma_1}{\partial \theta}\right)^2,\frac{1}{\sigma_2}\left(\frac{\partial \sigma_2}{\partial \theta}\right)^2,...,\frac{1}{\sigma_n}\left(\frac{\partial \sigma_n}{\partial \theta}\right)^2\right)\end{aligned}$$ and the prior is: $$\begin{aligned} p(\theta) &=&\sqrt{\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mbox{tr}\left(4 V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta}\frac{\partial V}{\partial \theta}CV\right)}\end{aligned}$$ Constant Variance ----------------- If the variances are modelled as constant (but the correlations are allowed to vary) then: $$\begin{aligned} \Sigma&=&VCV\\ \mbox{where }V&=&\mbox{diag}(\sigma_1^2,\sigma_2^2,...,\sigma_n^2)\\ \mbox{and }C&=&\mbox{the correlation matrix}\\ \frac{\partial \Sigma}{\partial \theta}&=& V \frac{\partial C}{\partial \theta}V\\ S&=&V^{-1} C^{-1} V^{-1}\\ \mbox{where }V^{-1}&=&\mbox{diag}(\sigma_1^{-2},\sigma_2^{-2},...,\sigma_n^{-2})\\ \mbox{and }C^{-1}&=&\mbox{the inverse correlation matrix}\\ \frac{\partial S}{\partial \theta}&=&V^{-1} \frac{\partial C^{-1}}{\partial \theta}V^{-1}\\ \frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \theta} &=&V^{-1}\frac{\partial C^{-1}}{\partial \theta}V^{-1}V\frac{\partial C}{\partial \theta}V\\ &=&V^{-1}\frac{\partial C^{-1}}{\partial \theta}\frac{\partial C}{\partial \theta}V\\ &=&-V^{-1}C^{-1}\frac{\partial C}{\partial \theta}C^{-1}\frac{\partial C}{\partial \theta}V\end{aligned}$$ and so the prior is: $$\begin{aligned} p(\theta) &=&\sqrt{\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} -\frac{1}{2}\mbox{tr}\left(V^{-1}\frac{\partial C^{-1}}{\partial \theta}\frac{\partial C}{\partial \theta}V\right)}\label{eq101}\\ &=&\sqrt{\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \theta} +\frac{1}{2}\mbox{tr}\left(V^{-1}C^{-1}\frac{\partial C}{\partial \theta}C^{-1}\frac{\partial C}{\partial \theta}V\right)}\label{eq101b}\end{aligned}$$ Multiple Parameters {#s4} =================== We now consider the multiparameter case. We start with two parameters and generalise to multiple parameters later. The derivations are only slightly more complex than those for the single parameter case. Starting from equation \[eq10\], which was: $$\begin{aligned} \frac{\partial \ln p(x|\theta)}{\partial \theta} &=&-\frac{1}{2} \frac{\partial \ln D}{\partial \theta} -\frac{1}{2}x^T \frac{\partial S}{\partial \theta} x +x^T \frac{\partial S}{\partial \theta}\mu +x^T S\frac{\partial \mu}{\partial \theta} -\frac{\partial \mu^T }{\partial \theta}S \mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\mu\end{aligned}$$ we now take the derivative wrt a second parameter $\phi$: $$\begin{aligned} \frac{\partial^2 \ln p(x|\theta)}{\partial \theta \partial \phi} &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta \partial \phi} x +x^T \frac{\partial}{\partial \phi}\left(\frac{\partial S}{\partial \theta}\mu\right) +x^T \frac{\partial}{\partial \phi}\left(S\frac{\partial \mu}{\partial \theta}\right)\\ && -\frac{\partial}{\partial \phi}\left(\frac{\partial \mu^T }{\partial \theta}S \mu\right) -\frac{1}{2}\frac{\partial}{\partial \phi}\left(\mu^T\frac{\partial S }{\partial \theta}\mu\right)\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta \partial \phi} x +x^T \left(\frac{\partial^2 S}{\partial \theta \partial \phi}\mu+\frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \phi}\right) +x^T \left(\frac{\partial S}{\partial \phi}\frac{\partial \mu}{\partial \theta}+S\frac{\partial^2 \mu}{\partial \theta \partial \phi}\right)\\ && -\left(\frac{\partial^2 \mu^T }{\partial \theta \partial \phi}S \mu +\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \phi} \mu +\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi}\right) -\frac{1}{2}\left(\frac{\partial \mu^T}{\partial \phi}\frac{\partial S }{\partial \theta}\mu +\mu^T\frac{\partial^2 S }{\partial \theta \partial \phi}\mu +\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \phi}\right)\nonumber\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} -\frac{1}{2}x^T \frac{\partial^2 S}{\partial \theta \partial \phi} x +x^T \frac{\partial^2 S}{\partial \theta \partial \phi}\mu +x^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \phi} +x^T \frac{\partial S}{\partial \phi}\frac{\partial \mu}{\partial \theta} +x^T S\frac{\partial^2 \mu}{\partial \theta \partial \phi}\\ && -\frac{\partial^2 \mu^T }{\partial \theta \partial \phi}S \mu -\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \phi} \mu -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi} -\frac{1}{2}\frac{\partial \mu^T}{\partial \phi}\frac{\partial S }{\partial \theta}\mu -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta \partial \phi}\mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \phi}\end{aligned}$$ Taking expectations: $$\begin{aligned} E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta^2}\right) &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} -\frac{1}{2}E\left(x^T \frac{\partial^2 S}{\partial \theta \partial \phi} x\right) +E(x^T) \frac{\partial^2 S}{\partial \theta \partial \phi}\mu +E(x^T) \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \phi}\\ && +E(x^T) \frac{\partial S}{\partial \phi}\frac{\partial \mu}{\partial \theta} +E(x^T) S\frac{\partial^2 \mu}{\partial \theta \partial \phi} -\frac{\partial^2 \mu^T }{\partial \theta \partial \phi}S \mu -\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \phi} \mu\\ && -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi} -\frac{1}{2}\frac{\partial \mu^T}{\partial \phi}\frac{\partial S }{\partial \theta}\mu -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta \partial \phi}\mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \phi}\end{aligned}$$ But $E(x)=\mu$, and, from appendix 2, $$E\left(x^T \frac{\partial^2 S}{\partial \theta \partial \phi} x\right) =\mbox{tr}\left(\frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma \right)+\mu^T \frac{\partial^2 S}{\partial \theta \partial \phi} \mu$$ and so: $$\begin{aligned} E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta \partial \phi}\right) &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} -\frac{1}{2}\left(\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma\right)+\mu^T \frac{\partial^2 S}{\partial \theta \partial \phi} \mu\right) +\mu^T \frac{\partial^2 S}{\partial \theta \partial \phi}\mu +\mu^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \phi}\\ && +\mu^T \frac{\partial S}{\partial \phi}\frac{\partial \mu}{\partial \theta} +\mu^T S\frac{\partial^2 \mu}{\partial \theta \partial \phi} -\frac{\partial^2 \mu^T }{\partial \theta \partial \phi}S \mu -\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \phi} \mu\\ && -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi} -\frac{1}{2}\frac{\partial \mu^T}{\partial \phi}\frac{\partial S }{\partial \theta}\mu -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta \partial \phi}\mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \phi}\\ &=&-\frac{1}{2} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} -\frac{1}{2}\mbox{tr}\left(\frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma \right) -\frac{1}{2}\mu^T \frac{\partial^2 S}{\partial \theta \partial \phi} \mu +\mu^T \frac{\partial^2 S}{\partial \theta \partial \phi}\mu +\mu^T \frac{\partial S}{\partial \theta}\frac{\partial \mu}{\partial \phi}\\ && +\mu^T \frac{\partial S}{\partial \phi}\frac{\partial \mu}{\partial \theta} +\mu^T S\frac{\partial^2 \mu}{\partial \theta \partial \phi} -\frac{\partial^2 \mu^T }{\partial \theta \partial \phi}S \mu -\frac{\partial \mu^T }{\partial \theta}\frac{\partial S}{\partial \phi} \mu\\ && -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi} -\frac{1}{2}\frac{\partial \mu^T}{\partial \phi}\frac{\partial S }{\partial \theta}\mu -\frac{1}{2}\mu^T\frac{\partial^2 S }{\partial \theta \partial \phi}\mu -\frac{1}{2}\mu^T\frac{\partial S }{\partial \theta}\frac{\partial \mu}{\partial \phi}\\ &=&-\frac{1}{2} \left[\frac{\partial^2 \ln D}{\partial \theta \partial \phi} +\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma\right)\right] -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi}\label{eq4}\end{aligned}$$ But $$\begin{aligned} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} &=&\frac{\partial}{\partial \theta}\left(\frac{\partial \log D}{\partial \phi}\right)\\ &=&\frac{\partial}{\partial \theta}\left(\frac{1}{D}\frac{\partial D}{\partial \phi}\right)\\ &=&\frac{\partial}{\partial \theta}\left(\frac{1}{D}\left(D\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \phi}\right)\right)\right)\\ &=&\frac{\partial}{\partial \theta}\left(\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \phi}\right)\right)\\ &=&\mbox{tr}\left(\frac{\partial}{\partial \theta}\left(S\frac{\partial \Sigma}{\partial \phi}\right)\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}+S\frac{\partial^2 \Sigma}{\partial \theta \partial \phi}\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}\right)+\mbox{tr}\left( S \frac{\partial^2 \Sigma}{\partial \theta \partial \phi}\right)\end{aligned}$$ However $$\begin{aligned} \frac{\partial}{\partial \theta}\left(S \Sigma\right)&=&S \frac{\partial \Sigma}{\partial \theta}+\frac{\partial S}{\partial \theta}\Sigma\\ \frac{\partial^2}{\partial \theta \partial \phi}\left( S \Sigma\right) &=&S \frac{\partial^2 \Sigma}{\partial \theta \partial \phi} +\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta} +\frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma +\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}\end{aligned}$$ But $S\Sigma=I$ and so $\frac{\partial^2}{\partial \theta \partial \phi}( S \Sigma)=0$, implying that $$S\frac{\partial^2 \Sigma}{\partial \theta \partial \phi}=- \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma -\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi} -\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}$$ and so $$\begin{aligned} \frac{\partial^2 \ln D}{\partial \theta \partial \phi} &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}\right)+\mbox{tr}\left( S \frac{\partial^2 \Sigma}{\partial \theta \partial \phi}\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}\right) +\mbox{tr}\left(- \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma -\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi} -\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}\right)\\ &=&\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}\right) +\mbox{tr}\left(- \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma\right) -\mbox{tr}\left(\frac{\partial S}{\partial \theta}\frac{\partial \Sigma}{\partial \phi}\right) -\mbox{tr}\left(\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}\right)\\ &=&\mbox{tr}\left(- \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma\right) -\mbox{tr}\left(\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}\right)\end{aligned}$$ Giving: $$\begin{aligned} \frac{\partial^2 \ln D}{\partial \theta \partial \phi}+\mbox{tr}\left(\frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma \right) &=&\mbox{tr}\left(- \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma\right) -\mbox{tr}\left(\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}\right) +\mbox{tr}\left(\frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma \right)\\ &=&-\mbox{tr}\left(\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}\right)\label{eq5}\end{aligned}$$ Substituting expression \[eq5\] into equation \[eq4\] gives: $$\begin{aligned} E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta \partial \phi}\right) &=&-\frac{1}{2} \left[\frac{\partial^2 \ln D}{\partial \theta \partial \phi} +\mbox{tr}\left( \frac{\partial^2 S}{\partial \theta \partial \phi}\Sigma\right)\right] -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi}\\ &=&\frac{1}{2}\mbox{tr}\left(\frac{\partial S}{\partial \phi}\frac{\partial \Sigma}{\partial \theta}\right) -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi}\\ &=&-\frac{1}{2}\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \phi}S\frac{\partial \Sigma}{\partial \theta}\right) -\frac{\partial \mu^T }{\partial \theta}S \frac{\partial \mu}{\partial \phi}\end{aligned}$$ We now generalize from two to multiple parameters. We change the notation so that $\theta$ is the vector of all parameters. The prior is then: $$\begin{aligned} p(\theta) &=&\sqrt{-\mbox{det}E\left(\frac{\partial^2 \ln p(x|\theta)}{\partial \theta_j \partial \theta_k}\right)}\\ &=&\sqrt{\mbox{det}\left(\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k} -\frac{1}{2}\mbox{tr}\left(\frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k}\right)\right)}\label{eq101}\\ &=&\sqrt{\mbox{det}\left(\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k} +\frac{1}{2}\mbox{tr}\left(S\frac{\partial \Sigma}{\partial \theta_j}S\frac{\partial \Sigma}{\partial \theta_k}\right)\right)}\label{eq101b}\end{aligned}$$ If there are $n$ observations and $m$ parameters then: - $\mu$ is an $n$ x $1$ vector - $\frac{\partial \mu}{\partial \theta_j}$ is an $n$ by $1$ vector - $\mu^T$ is a $1$ x $n$ vector - $\frac{\partial \mu^T}{\partial \theta_k}$ is a $1$ by $n$ vector - $S$ is an $n$ by $n$ matrix - $\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k}$ is a scalar (which is the $(j,k)$th element of a matrix) - $\frac{\partial S}{\partial \theta_j}$ is an $n$ by $n$ matrix - $\frac{\partial \Sigma}{\partial \theta_k}$ is an $n$ by $n$ matrix - $\frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k}$ is an $n$ by $n$ matrix - tr$\left(\frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k}\right)$ is a scalar (which is the $(j,k)$th element of a matrix) We now again consider various special cases of this general formula, starting with the two cases described in @jp1 and @jp2. Independence ------------ If the observations are modelled as independent then: $$\begin{aligned} \Sigma &=&\mbox{diag}\left(\sigma_1^2,\sigma_2^2,...,\sigma_n^2\right)\\ S &=&\mbox{diag}\left(\sigma_1^{-2},\sigma_2^{-2},...,\sigma_n^{-2}\right)\\ \frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k} &=&\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \mu_i}{\partial \theta_j}\right)\left(\frac{\partial \mu_i}{\partial \theta_k}\right)\\ \frac{\partial \Sigma}{\partial \theta_k} &=&\mbox{diag}\left(2\sigma_1 \frac{\partial \sigma_1}{\partial \theta_k},2\sigma_2\frac{\partial \sigma_2}{\partial \theta_k},...,2\sigma_n\frac{\partial \sigma_n}{\partial \theta_k}\right)\\ \frac{\partial S}{\partial \theta_j} &=&\mbox{diag}\left(-2\sigma_1^{-3} \frac{\partial \sigma_1}{\partial \theta_j},-2\sigma_2^{-3}\frac{\partial \sigma_2}{\partial \theta_j},...,-2\sigma_n^{-3}\frac{\partial \sigma_n}{\partial \theta_j}\right)\\ \frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k} &=&\mbox{diag}\left(-4\sigma_1^{-2} \frac{\partial \sigma_1}{\partial \theta_j}\frac{\partial \sigma_1}{\partial \theta_k},-4\sigma_2^{-2}\frac{\partial \sigma_2}{\partial \theta_j}\frac{\partial \sigma_2}{\partial \theta_k},...,-4\sigma_n^{-2}\frac{\partial \sigma_n}{\partial \theta_j}\frac{\partial \sigma_n}{\partial \theta_k}\right)\\ &=&-4\mbox{diag}\left(\frac{1}{\sigma_1^2} \frac{\partial \sigma_1}{\partial \theta_j}\frac{\partial \sigma_1}{\partial \theta_k},\frac{1}{\sigma_2^2}\frac{\partial \sigma_2}{\partial \theta_j}\frac{\partial \sigma_2}{\partial \theta_k},...,\frac{1}{\sigma_n^2}\frac{\partial \sigma_n}{\partial \theta_j}\frac{\partial \sigma_n}{\partial \theta_k}\right)\\ -\frac{1}{2}\mbox{tr}\left(\frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k}\right) &=&2\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\frac{\partial \sigma_i}{\partial \theta_j}\frac{\partial \sigma_i}{\partial \theta_k}\end{aligned}$$ and so $$\begin{aligned} p(\theta)&=&\sqrt{\mbox{det}\left(\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \mu_i}{\partial \theta_j}\right)\left(\frac{\partial \mu_i}{\partial \theta_k}\right) +2\sum_{i=1}^{n}\frac{1}{\sigma_i^2}\left(\frac{\partial \sigma_i}{\partial \theta_j}\right)\left(\frac{\partial \sigma_i}{\partial \theta_k}\right)\right)}\end{aligned}$$ which agrees with equation 36 in @jp1. Constant Covariance ------------------- If the covariance $\Sigma$ is constant then equation \[eq101\] reduces immediately to $$\begin{aligned} p(\theta)&=&\sqrt{\mbox{det}\left(\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k}\right)}\end{aligned}$$ which agrees with equation 28 in @jp2. Constant Correlation -------------------- If the correlations are modelled as constant (but the variances are allowed to vary) then: $$\begin{aligned} \Sigma&=&VCV\\ \mbox{where }V&=&\mbox{diag}(\sigma_1^2,\sigma_2^2,...,\sigma_n^2)\\ \mbox{and }C&=&\mbox{the correlation matrix}\\ \frac{\partial \Sigma}{\partial \theta_k} &=& \frac{\partial V}{\partial \theta_k}CV+VC\frac{\partial V}{\partial \theta_k}\\ &=& 2\frac{\partial V}{\partial \theta_k}CV\\ \mbox{where } \frac{\partial V}{\partial \theta_k} &=&2\mbox{diag}(\sigma_1\frac{\partial \sigma_1}{\partial \theta_k},\sigma_2\frac{\partial \sigma_2}{\partial \theta_k},...,\sigma_n\frac{\partial \sigma_n}{\partial \theta_k})\\ S&=&V^{-1} C^{-1} V^{-1}\\ \mbox{where }V^{-1}&=&\mbox{diag}(\sigma_1^{-2},\sigma_2^{-2},...,\sigma_n^{-2})\\ \mbox{and }C^{-1}&=&\mbox{the inverse correlation matrix}\\ \frac{\partial S}{\partial \theta_j} &=& \frac{\partial V^{-1}}{\partial \theta_j}C^{-1}V^{-1}+V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta_j}\\ &=& 2V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta_j}\\ \mbox{where } \frac{\partial V^{-1}}{\partial \theta_j} &=&-2\mbox{diag}\left(\frac{1}{\sigma_1^2}\frac{\partial \sigma_1}{\partial \theta_j},\frac{1}{\sigma_2^2}\frac{\partial \sigma_2}{\partial \theta_j},...,\frac{1}{\sigma_n^2}\frac{\partial \sigma_n}{\partial \theta_j}\right)\\ \frac{\partial V^{-1}}{\partial \theta_j}\frac{\partial V}{\partial \theta_k} &=&-4\mbox{diag}\left(\frac{1}{\sigma_1}\frac{\partial \sigma_1}{\partial \theta_j}\frac{\partial \sigma_1}{\partial \theta_k},\frac{1}{\sigma_2}\frac{\partial \sigma_2}{\partial \theta_j}\frac{\partial \sigma_2}{\partial \theta_k},...,\frac{1}{\sigma_n}\frac{\partial \sigma_n}{\partial \theta_j}\frac{\partial \sigma_n}{\partial \theta_k}\right)\\ \frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k} &=&4 V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta_j}\frac{\partial V}{\partial \theta_k}CV\end{aligned}$$ and the prior is: $$\begin{aligned} p(\theta) &=&\sqrt{\mbox{det}\left(\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k} -\frac{1}{2}\mbox{tr}\left(4 V^{-1}C^{-1}\frac{\partial V^{-1}}{\partial \theta_j}\frac{\partial V}{\partial \theta_k}CV\right)\right)}\end{aligned}$$ Constant Variance ----------------- If the variances are modelled as constant (but the correlations are allowed to vary) then: $$\begin{aligned} \Sigma&=&VCV\\ \mbox{where }V&=&\mbox{diag}(\sigma_1^2,\sigma_2^2,...,\sigma_n^2)\\ \mbox{and }C&=&\mbox{the correlation matrix}\\ \frac{\partial \Sigma}{\partial \theta_k}&=& V \frac{\partial C}{\partial \theta}V\\ S&=&V^{-1} C^{-1} V^{-1}\\ \mbox{where }V&=&\mbox{diag}(\sigma_1^{-2},\sigma_2^{-2},...,\sigma_n^{-2})\\ \mbox{and }C{-1}&=&\mbox{the inverse correlation matrix}\\ \frac{\partial S}{\partial \theta_j}&=&V^{-1} \frac{\partial C^{-1}}{\partial \theta_j}V^{-1}\\ \frac{\partial S}{\partial \theta_j}\frac{\partial \Sigma}{\partial \theta_k} &=&V^{-1}\frac{\partial C^{-1}}{\partial \theta_j}VV^{-1}\frac{\partial C}{\partial \theta_k}V\\ &=&V^{-1}\frac{\partial C^{-1}}{\partial \theta_j}\frac{\partial C}{\partial \theta_k}V\\ &=&-V^{-1}C^{-1}\frac{\partial C}{\partial \theta_j}C^{-1}\frac{\partial C}{\partial \theta_k}V\end{aligned}$$ and the prior is: $$\begin{aligned} p(y|x) &=&\sqrt{\mbox{det}\left(\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k} -\frac{1}{2}\mbox{tr}\left(V^{-1}\frac{\partial C^{-1}}{\partial \theta_j}\frac{\partial C}{\partial \theta_k}V\right)\right)}\\ &=&\sqrt{\mbox{det}\left(\frac{\partial \mu^T }{\partial \theta_j}S \frac{\partial \mu}{\partial \theta_k} +\frac{1}{2}\mbox{tr}\left(V^{-1}C^{-1}\frac{\partial C}{\partial \theta_j}C^{-1}\frac{\partial C}{\partial \theta_k}V\right)\right)}\end{aligned}$$ Summary {#s5} ======= Climate models are statistical models, in that they produce probabilistic predictions (when run as initial condition ensembles) and have certain parameters that can only be determined by comparison of model results with observations. As a result, the Bayesian framework, in which probabilistic predictions made from models with different parameter values are combined together to make a single best probabilistic prediction, can be applied. Within that framework one has to specify a prior, and one must choose between a prior based on intuition or a prior based on a rule. The former is known as subjective Bayesian statistics, and the latter, objective Bayesian statistics. The authors are pursuing a research programme that is exploring methods by which objective Bayesian statistics can be applied in climate modelling. In this article we have discussed the application of the most standard objective prior, known as Jeffreys’ Prior. Climate models are complex and the relationship between the parameters and the predicted distributions is also complex. However, the form of the predicted distributions themselves can often be rather simple, and for many groups of variables a multivariate normal may be a good approximation. We have shown that, by making this approximation, the calculation of Jeffreys’ Prior can be reduced to differentiating the parameters of the multivariate normal by the parameters of the underlying climate model. We derive expressions for Jeffreys’ Prior in this situation. The results from our two previous articles on this topic (@jp1 and @jp2) are special cases of the general results shown here. In all this work we have expressed Jeffreys’ Prior in terms of the true, rather than the estimated, parameters of the distributions from climate model predictions. In other words, we have assumed infinite rather than finite size initial condition ensembles. A further challenge is to rederive the expressions given above but incorporating estimation uncertainty. Proof that if $S$ is symmetric then $a^TSb=b^TSa$ ================================================= Since $a^TSb$ is a scalar it is equal to its tranpose, as so: $$a^TSb=(a^TSb)^T=b^TS^Ta$$ but if $S$ is symmetric then $$b^TS^Ta=b^TSa$$ Putting these two together gives: $$a^TSb=b^TSa$$ Proof that $E(x^TAx)=\mbox{tr}\left(\Sigma A\right)+\mu^TA\mu$ ============================================================== $$\begin{aligned} E(x^TAx) &=&E\left(\sum_{ij} x_i A_{ij} x_j\right)\\ &=&E\left(\sum_{ij} x_i x_j A_{ij} \right)\\ &=&\sum_{ij} E(x_i x_j) A_{ij}\end{aligned}$$ But, by definition: $$\begin{aligned} \Sigma_{ij} &=&E((x_i-\mu_i)(x_j-\mu_j))\\ &=&E(x_i x_j-x_i \mu_j-\mu_i x_j+\mu_i \mu_j)\\ &=&E(x_i x_j)-E(x_i) \mu_j-\mu_i E(x_j)+\mu_i \mu_j)\\ &=&E(x_i x_j)-\mu_i \mu_j-\mu_i \mu_j+\mu_i \mu_j)\\ &=&E(x_i x_j)-\mu_i \mu_j\end{aligned}$$ So $$\begin{aligned} E(x_i x_j)&=&\Sigma_{ij}+\mu_i \mu_j\end{aligned}$$ and so $$\begin{aligned} E(x^TAx) &=&\sum_{ij} (\Sigma_{ij}+\mu_i \mu_j) A_{ij}\\ &=&\sum_{ij} \Sigma_{ij}A_{ij}+\sum_{ij}\mu_i\mu_j a_{ij}\\ &=&\mbox{tr}\left(\Sigma A\right)+\mu^T A \mu\\ &=&\mbox{tr}\left(A\Sigma \right)+\mu^T A \mu\end{aligned}$$ [^1]: *Correspondence email*: `[email protected]`

ArXiv

--- abstract: 'In the current era of data-intensive science, it is increasingly important for researchers to be able to have access to published results, the supporting data, and the processes used to produce them. Six years ago, recognizing this need, the American Astronomical Society and the Astrophysics Data Centers Executive Committee (ADEC) sponsored an effort to facilitate the annotation and linking of datasets during the publishing process, with limited success. I will review the status of this effort and describe a new, more general one now being considered in the context of the Virtual Astronomical Observatory.' author: - Alberto Accomazzi title: 'Linking Literature and Data: Status Report and Future Efforts' --- Introduction ============ Links between papers in ADS and data products hosted by astronomy archives have existed since 1995. These links have been created and curated by librarians and archivists as part of the data center’s effort to collect information about the scientific use of the data being hosted by the archive. The links provide more than just useful connections between bibliographic records and observations that allow users to access related material. They represent part of the scientific artifacts created during the research lifecycle of an astronomer, and as such are needed to fully document and describe the research activity itself [@LISAVI]. One obvious benefit which comes from maintaining such links is the ease with which one can generate metrics about the scientific impact of the observations in the form of published papers or citations. Thus, in a highly competitive scientific discipline such as astronomy, maintaining the complete record of paper-data connections has become an accepted way to evaluate a project, mission, or even an entire research field ([@Uta], [@2010AN....331..338T]). The metadata collected during the creation of links maintained by ADS and its collaborators have so far been limited to some very basic information about the location of the resources linked together. Typically these are simple mappings of ADS’s bibliographic identifiers ([*bibcodes*]{}) and URLs pointing to one or more particular data product(s) hosted by an archive. The ADS record will of course also have a link to the published paper itself, acting as a “bridge” between the manuscript and the data described therein. Figure 1 shows the connection between a paper and data products available from the Multimission Archive at Space Telescope (MAST) and the Chandra X-Ray Archive. ![Links between an ADS record, the full-text manuscript hosted by a publisher, and data products available from MAST and Chandra[]{data-label="fig1"}](accomazzi_fig1.pdf){width="100.00000%"} Development of Dataset Linking Infrastructure ============================================= In 2002, it became apparent that the methodology adopted to create, maintain and share these linkages could be improved. Thus, in 2003, the NASA Astrophysics Data Centers Executive Council (ADEC) and the AAS journals issued guidelines aimed at improving the situation [@2004AAS...204.7502E]. This new effort was aimed at addressing four separate issues in the management of these links: their curation, naming, resolution and persistence. The creation of links to data products has been a time-consuming activity usually carried out by a librarian or archivist. [@2004ASPC..314..605R] describe the effort required to perform this activity, which typically consists of scanning the literature to identify which papers mention one or more data products from a particular archive, and then link those papers with the relevant datasets. In 2004, in order to facilitate this activity, and in coordination with the ADEC proposal, the Astrophysical Journal introduced the capability for authors to properly tag the datasets analyzed in the paper. This introduced a mechanism to formally “cite” data in a way similar to how scientists cite other papers. According to this plan, citations to data products would be vetted by both editors and referees during the manuscript editorial process, and links would be created to the corresponding data products as part of the process which generates the online HTML version of the paper. The correlation between a paper and the datasets referenced therein would then be propagated back to the ADS and the participating data centers via metadata exchange. The implementation of this linking proposal would not only benefit end-users, but would potentially provide significant savings in the curation efforts of archivists and librarians, who could now harvest these linkages directly from ADS, thus reducing the need for the manual scanning of the literature. In order to properly cite the datasets in the literature, the ADEC and AAS adopted a standard way to uniquely identify data resources based on the IVOA Identifier standard [@IVOAIDS]. The proposed system of nomenclature [@2007ASPC..376..467A] provided a standard for dataset identifiers which featured some important properties. Among them: uniqueness (one resource corresponds to a single identifier), and persistence (identifiers do not change even when data products are migrated to a different archive). The identifiers were designed to support the naming of resources with a broad range of granularity and included a “public” prefix identifying the archive or mission that generated the dataset as well as a “private” key identifying the data item within a specific collection. In order to ensure the proper use and persistence of links to datasets, the ADEC charged the ADS with the task of setting up a verification and resolution service for dataset identifiers. In this role, the ADS would act as the registration authority on behalf of the community, creating the infrastructure necessary to enable the dataset linking. During copy-editing of a paper, the editors would use an automated tool provided by ADS to verify that a particular dataset identifier is known and can be resolved to an online resource. Upon successful verification, the identifier would be incorporated into the paper with a link to a resolution service provided by ADS (rather than a simple link to the current URL for the resource). This model provides a level of redirection which can be used to properly track a dataset if and when it moves from one archive to another, and allows the resolver to provide options should multiple versions of a data product be available. A complete description of this implementation can be found in [@2007ASPC..376..467A]. Elements of this architecture are similar to the Digital Object Identifier standard used for the persistent linking of scholarly publications, which is discussed in section 4. However, this system was designed to be fully managed by the astronomical community requiring a minimal level of effort for institutional buy-in. Current Status ============== Six years have passed since the introduction of the dataset linking infrastructure, so now is a good time to take stock of this effort. From a system design point of view, some of the features that made the implementation of this system attractive have, in retrospect, proven to be obstacles to its long-term success. Chief among all problems with the registration of dataset identifiers has been enforcing their persistence. Since data products ultimately reside within archives that participate in the ADEC but which are run independently of each other, the implementation and maintenance of services that provide access to the data is left to the archives themselves. Given that the thrust of this effort is completely voluntary, there is no contract or reward system which can be leveraged to enforce the long-term resolution of and access to a particular dataset. Experience shows that unless requirements for the preservation of these linking services become part of the archive operations, a simple system upgrade is enough to break valuable links to dataset resources. As an example, over 200 dataset identifiers which were published in a 2004 ApJ Supplement special issue on the Spitzer Space Telescope are no longer resolvable due to a change in the Spitzer Science Center interface. ![The number of links to data entered in AJ and ApJ papers as part of the ADEC/AAS dataset linking effort[]{data-label="fig2"}](accomazzi_fig2.png){width="100.00000%"} Unfortunately the adoption and use of dataset identifiers in the literature has not been a success story. Citations to datasets began appearing in 2005 and increased in the following two years, peaking in 2007, before decreasing in 2008 and finally going down to zero in 2009 (see fig. 2). The reasons for this reversal are not entirely clear, but can probably be attributed to a variety of factors. First and foremost, even though the ADEC approved a policy encouraging archives and users to make an effort to more widely publish dataset identifiers, anecdotal evidence shows a low level of awareness from scientists of this possibility. Researchers tend to be busy and, unless properly coached by editors and archivists, will easily overlook new demands or stipulations requiring additional work on their part. In addition, data archives don’t always make it obvious how a particular dataset (or data file) should be cited in the literature by the scientists who make use of the data in their research. Since astronomers have been accustomed to referring to the data they have used in terms of specific observations or regions of the sky, this general practice is still used. Rather than unambiguously identifying the data using dataset identifiers, astronomers describe how the data can be obtained. While this is a reasonable way for an author to convey the necessary information about the data being studied, it obviously defeats the goal of creating persistent, unambiguous, machine-readable links to the data products. Finally, it was hoped that after the initial adoption of the standard for dataset identifiers within the ADEC more data centers and more journals would follow suit, but this did not materialize. The critical mass and community awareness necessary to make this common practice was never reached. A Way Forward ============= Despite the lack of adoption from the community of the proposal described above, every scientist, librarian and archivist agrees that preserving data products and publishing links to data in the literature is a worthwhile effort. We believe that at this point in time we should even be more ambitious, and recognize that in fact for our discipline to flourish in the digital era all artifacts related to the research lifecycle need to be available online, and properly interlinked [@LISAVI]. Thus, the issue of creating links from the literature to data products can be recast in a wider scope – the preservation and interlinking of digital assets in astronomy. The use of the term “digital assets” in this context refers to artifacts used and generated during the research activity of an astronomer. This includes observing proposals, observations, archival data from surveys and catalogs, observing logs, tables and plots that are published in manuscripts. In short, we advocate capturing all the data and knowledge that has gone into the research activity itself, with the aim of providing a digital environment that can support the repeatability of the research described in a publication. While there are many possible implementations of a digital environment for data preservation, it is clear that any such effort must satisfy a set of principles. Below we identify some of the basic requirements that we believe will need to be addressed in the near future. Management of Digital Assets ---------------------------- First, one should consider the issue of nomenclature and persistence of digital assets. Since we can expect that the data referred to in a paper will be hosted on a distributed set of digital repositories, naming and linking standards need to be clearly defined and adopted in order to create persistent links to such resources . The solution proposed by the ADEC was primarily designed to satisfy the requirements of uniqueness and persistence for data already available in well-established archives using community-developed standards. While this approach is technically sound and seemed at the outset to provide the best solution to the problem, new technologies and standards since developed by the digital library community now offer attractive alternatives that should be considered. For the creation and management of unique identifiers, the Handle System is a general purpose distributed information system for the minting and resolution of unique identifiers on the internet. The Digital Object Identifier (DOI) system is an application built upon the Handle system and is widely used by the digital publishing industry. Organizations making use of the DOI system agree to a business model that requires the deposit and active curation of metadata for the digital assets registered in the system, and are subject to fines if found to be in breach of the DOI Foundation rules. This elevated level of commitment provides a certain level of assurance that the digital assets registered in the system will be properly maintained. In addition, the DOI foundation explicitly imposes requirements on its members to ensure the long-term survival of the system. For instance, should one of its members cease operations, the DOI resolution of its content would be transferred to other members of the foundation. When it comes to the preservation of data products, this type of long-term commitment has never been formalized or made explicit by most of the publishers, societies or even astronomical data centers (except for the case of the active NASA missions, whose digital assets are transferred to archival centers at the end of a mission). However, it is exactly the type of commitment we believe is essential for our community to make at this time. Archival and Preservation ------------------------- In order to enable the publication and broader re-use of scientific high-level data products, researchers should be required to upload such data to one or more trusted, community-curated, digital repositories. Not only does this requirement allow repeatability of experiment and analysis, but it promotes a level of transparency and trust that is an important component of the scientific discourse. While much of the tabular data now published in scientific articles ends up being stored in services such as NED and Vizier, a significant amount of supplementary material does not make it into such archives. To be sure, authors are often encouraged to submit machine-readable versions of the data (or even computer code) as supplementary material submitted to the journal, but the uniformity, longevity, re-usability and discoverability of such material are at this point highly inconsistent and questionable. In addition, no explicit or common migration plan has so far been defined by publishers or learned societies, so future access and curation of these assets is not assured. We believe that it is essential to encourage the deposit of digital assets in a wide range of trusted repositories curated by the community in collaboration with the journals. The kind of material deposited in such repositories will supplement the products which are currently curated by projects such as NED and Vizier, and consist of anything which does not fit in the typical description of a data table or catalog. This may include published images, plots, observing notes, workflows, software, intermediate results, and large data collections. In order for these data products to be useful, it is essential for the user to deposit and for the repository to expose not just raw data but also its related metadata. This should include, at a minimum, a description of the datasets in the sample, a set of applicable keywords, and some notes relating the data in question to the published paper(s) in which they were used. The need to create a digital infrastructure in support of these activities is increasingly being recognized both in the US and in Europe by funding agencies such as NASA, the NSF, JISC and digital preservation programs are now being defined [@Sayeed]. In particular, the NSF DataNet Data Conservancy program has established as one of its goals the support of scientific inquiry through the adoption of a comprehensive data curation strategy. Today there are a number of open-source digital repository systems available, some of which have already been deployed by universities and projects involved in preserving digital institutional assets (e.g. Fedora or DSpace). Other initiatives, such as the Dataverse Network [@King], provide a scientist-centered framework for storing data products associated with publications and encourage their citation through the use of unique, permanent dataset identifiers based on the Handle system. Discussion ========== Even though there is general agreement that publishing and citing data is a noble goal and worthwhile effort, the experience of the ADEC data linking effort has shown that it takes sustained community engagement to turn a proposal into a successful activity. When many people and organizations are involved in providing crucial components of such a distributed system, the risk of multiple points of failure becomes significant and can ultimately spell the demise of even the best thought-out technical scenario. Rather than giving up on this worthwhile idea, we should take the failure of adoption as a learning opportunity to devise a more robust system that can not only provide links to existing datasets available from well-established archives, but also provide the capability of storing author-supplied data and metadata related to publications. In the long run, it is likely that there will be an ecosystem of different repositories and technologies used for the preservation of research products. Some of them will be more focused on actively capturing and curating datasets, as is today done by projects such as Vizier and NED, and others which will provide an infrastructure which can be used by scientists and publishers to self-manage data products associated with published papers. The overarching goal of such systems should be to provide useful services to the astronomy community and guarantee that the data deposited in such repositories will be properly curated and preserved for the foreseeable future. This includes providing a migration path for obsolete data formats, curating and exposing metadata of digital assets in the repository, and providing discovery services to its content. This commitment comes at a cost, which should be shared by the community as part of the effort which funds the infrastructure supporting astronomical research. In addition, it is essential that we promote and encourage policies that foster and facilitate the growth of the digital scholarly environment that the Virtual Observatory has been envisioning. The recent funding of digital preservation frameworks such as the Data Conservancy project suggests that the time has come for the VO to play a major role in the capture and preservation of the astronomy research lifecycle. We look forward for the members of the International Virtual Observatory Alliance to take a pro-active role over the next decade in order to make this vision a reality. Accomazzi, A., & Eichhorn, G. 2004, Astronomical Data Analysis Software and Systems (ADASS) XIII, 314, 181 Accomazzi, A., Eichhorn, G., & Rots, A. 2007, Astronomical Data Analysis Software and Systems XVI, 376, 467 Accomazzi, A. 2010, in: proceedings of Library and Information Systems VI, in press Choudhury, S. et al. 2007, The International Journal of Digital Curation, 2, 20 Eichhorn, G., & Astrophysics Datacenter Executive Committee (ADEC) 2004, Bulletin of the American Astronomical Society, 36, 805 Grothkopf, U. & Lagerstrom, J. 2010, these proceedings King, G. 2007, Sociological Methods & Research, 36, 173 Plante, R., Linde, T., Williams, R., & Noddle, K. 2006, http://www.ivoa.net/Documents/latest/IDs.html Rots, A. H., Winkelman, S. L., Paltani, S., Blecksmith, S. E., & Bright, J. D. 2004, Astronomical Data Analysis Software and Systems (ADASS) XIII, 314, 605 Trimble, V., & Ceja, J. A. 2010, Astronomische Nachrichten, 331, 338

ArXiv

--- bibliography: - 'Hourglass\_Weyl\_bib.bib' title: Supplemental materials --- The full band structures and methods of calculation =================================================== Our first-principles calculations have been performed on BiPd, ZnTe under high pressure, and the high temperature phase of Tl$_3$PbBr$_5$, which possess a single $N$-fold screw axis with $N=2,3,4$, respectively, as well as TRS. The SOC is expected to be sizable in these materials. BiPd is a monoclinic crystal with 16 atoms in the unit cell. It becomes superconducting below 3.8 K[@PhysRevB.84.064518] and a Dirac cone surface state exists on the (010) surface[@BiPd1]. The band structure shows metallic behavior, with twelve bands crossing the Fermi energy. The SOC causes band splitting except at the eight TRIM where Kramers degeneracy appears. According to our symmetry analysis, the bands along the screw-invariant lines must show 2-hourglass-like dispersion. In Fig.\[fig:BiPd\], we show the full band structure of BiPd, and that along CE, a twofold screw-invariant line. It should be mentioned that all the bands, from the bottom valence bands to the top conduction bands show similar behavior – every four bands form a 2-hourglass if the SOC is large enough. ZnTe is a B- semiconductor in zinc-blende (ZB) structure at ambient pressure, and turns into ZnTe- phase at 9.6 GPa and further into ZnTe- phase at 12.1 GPa[@ZnTe1]. The ZnTe- phase is in the SG $Cmcm$ in the orthorhombic system, while controversy exists in identifying the ZnTe- phase. Both cinnabar structure (SG $P3_121$)[@ZnTe3] and $P3_1$[@ZnTe2](Fig.\[fig:ZnTe\] (a)) have been claimed. Here we presume the $P3_1$ ZnTe exists and consider its band structures. Similar to BiPd, 3-hourglass-like dispersions are observed for the bands along the TRIM-ended screw-invariant line ${\Gamma}$A, as shown in Fig.\[fig:ZnTe\]. Note that here the 3-hourglass is not standard due to the permutation of the vertices, thus more than two crossings occur. What is physically interesting here, as discussed in the main text, is that in half of the BZ, a minimal number of two Weyl points (the $1+1$ case) appears as the crossings of the eighth and the ninth band below the Fermi energy, and a minimal number of six Weyl points (the $3+3$ case) exists as the crossings of the second and the third band. Tl$_3$PbBr$_5$ can achieve effective doping of rare-earth element and can be applied as non-linear optical devices[@Tl3PbBr5A]. Two phases are identified, the low-temperature orthorhombic phase (SG $P2_12_12_1$) and the high-temperature tetragonal phase (SG $P4_1$)[@Tl3PbBr5B], with the transition temperature about 239 $^\circ$C. For the high-temperature phase there are 36 atoms per unit cell. 4-hourglass-like dispersions along the fourfold screw-invariant lines ${\Gamma}$Z and MA, and 2-hourglass-like dispersions along the twofold screw-invariant line XR, are observed. A minimal number of two Weyl points (the $1+1$ case) might also exist for the fourth and the fifth band above the Fermi energy, one at ${\Gamma}$Z and the other at MA. Band structure calculations were performed by density functional theory using the Vienna Ab-initio Simulation Package (VASP)[@PhysRevB.54.11169]. The projector augmented-wave (PAW)[@PhysRevB.50.17953] pseudopotentials were used and the generalized gradient approximation of the PBE type functional[@PhysRevLett.77.3865] was used describing the exchange-correlation energy. A cutoff energy of 700 eV (500 eV for Tl$_3$PbBr$_5$) was used for the plane wave expansion to ensure the convergence of total energy to be less than 1 meV. The Monkhorst-Pack grid[@PhysRevB.13.5188] for BZ sampling was set to $7 \times 7 \times 7$, $15\times15\times9$, $7\times7\times 5$ for BiPd, ZnTe and Tl$_3$PbBr$_5$, respectively. The structures were fully optimized with experimental lattice constants used, until the forces were smaller than 0.01 eV/. ![(a) Crystal structure of BiPd in SG $P2_1$. (b) The first BZ and the high symmetry points and lines. The screw-invariant lines are highlighted by glowing orange. (c) Band structure of BiPd (with SOC). A 2-hourglass along CE is enlarged and shown on the right.[]{data-label="fig:BiPd"}](figA1.jpg){width="8cm"} ![(a) Crystal structure of ZnTe in SG $P3_1$. (b) The first BZ and the high symmetry points and lines. The screw-invariant line ended with two TRIM is highlighted. (c) Band structure of ZnTe (with SOC). The valence bands near Fermi energy showing 3-hourglass band crossings are enlarged and shown on the right.[]{data-label="fig:ZnTe"}](figA2.jpg){width="8cm"} ![(a) Crystal structure of Tl$_3$PbBr$_5$ in SG $P4_1$. (b) The first BZ and high symmetry points and lines. The fourfold and twofold screw-invariant lines are highlighted by orange and blue, respectively. (c) Band structure of Tl$_3$PbBr$_5$ (with SOC). The conduction and valence bands near the Fermi energy showing $4$- and 2-hourglass band crossings are enlarged and shown on the right.[]{data-label="fig:Tl3PbBr5"}](figA3.jpg){width="8cm"} Compatibility relations {#sec:compat} ======================= The [*n*]{}-hourglass-like band structures in the nonsymmorphic SGs $P2_1$,$P3_1$ and $P4_1$ can also be derived by considering the compatibility relatons[@PhysRevMaterials.2.074201] between irreducible representations at different high symmetry points and lines. We derive the compatibility relations for $P4_1$ and $P3_1$ below. We will see that the $n$-hourglasses are automatically generated as a rigorous result of the symmetry requirement. Consider SG $P4_1$ first. For nonsymmorphic SGs, the SG representation $D(\{\alpha|\tau+t\})$ can be obtained from the corresponding point group representations $\Gamma(\{\alpha|0\})$ by $D(\{\alpha|\tau+t\})=e^{ik\tau}e^{ikt}\Gamma(\{\alpha|0\})$. For ${\Gamma}(0,0,0)$, $\Lambda (0,0,\omega)$, and Z$(0,0,0.5\pi)$ in the BZ, the corresponding point groups – or the group of the [*k*]{} vectors are the same, being $\{E,\{C_{2z}|\frac{1}{2}t_3\},\{C_{4z}|\frac{1}{4}t_3\},\{C_{4z}^{-1}|\frac{3}{4}t_3\}\}$. Since SOC is considered, the symmetry of bands are described by the double-valued representation[@Elcoro:ks5574] of the double SG. The corresponding character table for the double-valued representations of the group of wave vectors at $\Gamma$,$\Lambda$, and $Z$ are shown in Table.\[tab:P4\_1\]. From the character table it is obvious that the representations at $\Gamma$ and Z are both doubly degenerate. This results in the double degeneracy of the bands at $\Gamma$ and Z, which are both TRIM. Along $\Lambda$, the representations are non-degenerate. By comparing the characters of each representation, it is easy to get the following compatibility relations between $\Gamma$ and $\Lambda$, $$\begin{split} \overline \Gamma_5 \overline \Gamma_7 \to \overline \Lambda_5 + \overline \Lambda_7, \\ \overline \Gamma_6 \overline \Gamma_8 \to \overline \Lambda_6 + \overline \Lambda_8, \end{split}$$ and the compatibility relations between Z and $\Lambda$, $$\begin{split} \overline Z_5 \overline Z_5 \to \overline \Lambda_5 + \overline \Lambda_5, \\ \overline Z_6 \overline Z_6 \to \overline \Lambda_6 + \overline \Lambda_6, \\ \overline Z_7 \overline Z_8 \to \overline \Lambda_7 + \overline \Lambda_8. \end{split}$$ Repres. $E$ $\{C_{2z}|\frac{1}{2}t_3\}$ $\{C_{4z}|\frac{1}{4}t_3\}$ $\{C_{4z}^{-1}|\frac{3}{4}t_3\}$ T. I. ------------------------- ----- ----------------------------- ---------------------------------------------------- ---------------------------------------------------- ------- ${\overline\Gamma_5}$ $1$ $-i$ $e^{i\frac{3}{4}\pi}$ $e^{-i\frac{3}{4}\pi}$ $b$ ${\overline\Gamma_6}$ $1$ $-i$ $e^{-i\frac{1}{4}\pi}$ $e^{i\frac{1}{4}\pi}$ $b$ ${\overline\Gamma_7}$ $1$ $i$ $e^{-i\frac{3}{4}\pi}$ $e^{i\frac{3}{4}\pi}$ $b$ ${\overline\Gamma_8}$ $1$ $i$ $e^{i\frac{1}{4}\pi}$ $e^{-i\frac{1}{4}\pi}$ $b$ ${\overline Z_5}$ $1$ $1$ $-1$ $1$ $a$ ${\overline Z_6}$ $1$ $1$ $1$ $-1$ $a$ ${\overline Z_7}$ $1$ $-1$ $-i$ $-i$ $b$ ${\overline Z_8}$ $1$ $-1$ $i$ $i$ $b$ ${\overline \Lambda_5}$ $1$ ${-ie^{i\pi \omega}}$ ${e^{i\frac{3}{4}\pi}e^{i\frac{1}{2}\pi \omega}}$ ${e^{-i\frac{3}{4}\pi}e^{i\frac{3}{2}\pi \omega}}$ $c$ ${\overline \Lambda_6}$ $1$ ${-ie^{i\pi \omega}}$ ${e^{-i\frac{1}{4}\pi}e^{i\frac{1}{2}\pi \omega}}$ ${e^{i\frac{1}{4}\pi}e^{i\frac{3}{2}\pi \omega}}$ $c$ ${\overline \Lambda_7}$ $1$ ${ie^{i\pi \omega}}$ ${e^{-i\frac{3}{4}\pi}e^{i\frac{1}{2}\pi \omega}}$ ${e^{i\frac{3}{4}\pi}e^{i\frac{3}{2}\pi \omega}}$ $c$ ${\overline \Lambda_8}$ $1$ ${ie^{i\pi \omega}}$ ${e^{i\frac{1}{4}\pi}e^{i\frac{1}{2}\pi \omega}}$ ${e^{-i\frac{1}{4}\pi}e^{i\frac{3}{2}\pi \omega}}$ $c$ : The double-valued irreducible representations of SG $P4_1$, for the group of wave vectors at $\Gamma$, Z and $\Lambda$. The degeneracy due to TRS for each representation is added at the end of each row, with $a$ and $b$ meaning doubly degenerate and $c$ non-degenerate[@PhysRev.52.361]. Here $0<\omega<0.5$. For $\omega=0$, $\Lambda$ becomes $\Gamma$ and for $\omega=0.5$, $\Lambda$ becomes Z. \[tab:P4\_1\] ![(a) The compatibility relations between bands at $\Gamma$, $\Lambda$ and Z of crystals with $P4_1$ SG symmetry. (b) The compatibility relations between bands at $\Gamma$, $\Delta$ and A of crystals with $P3_1$ SG symmetry.[]{data-label="fig:compat"}](figA4.jpg){width="8cm"} The compatibility relations can be schematically shown in Fig.\[fig:compat\](a), which result in the band structure similar to that obtained in the main text by symmetry analysis. We see that each representation has to appear twice to make the set of bands closed, i.e. separated from another set. In Fig.\[fig:compat\](a), they are arranged in such a way to result in a standard 4-hourglass band structure with three crossings along $\Lambda$. By tuning the sequence of the representations at $\Gamma$ or Z we can get a variant of the standard 4-hourglass with more crossings. Similarly, the double-valued irreducible representations of SG $P3_1$ for the group of wave vectors at $\Gamma$, $\Delta$ and A are shown in Table.\[tab:P3\_1\], from which we have the compatibility relations between $\Gamma$ and $\Delta$, $$\begin{split} \overline \Gamma_4 \overline \Gamma_4 \to \overline \Delta_4 + \overline \Delta_4, \\ \overline \Gamma_5 \overline \Gamma_6 \to \overline \Delta_5 + \overline \Delta_6, \end{split}$$ and the compatibility relation between A and $\Delta$ $$\begin{split} \overline A_4 \overline A_6 \to \overline \Delta_4 + \overline \Delta_6, \\ \overline A_5 \overline A_5 \to \overline \Delta_5 + \overline \Delta_5. \end{split}$$ The compatibility relations are shown schematically in Fig.\[fig:compat\](b), with minimally two crossings along $\Delta$. Repres. $E$ $\{C_{3z}|\frac{1}{3}t_3\}$ $\{C_{3z}^{-1}|\frac{2}{3}t_3\}$ T. I. ---------------------- ----- -------------------------------------------------- -------------------------------------------------- ------- $\overline \Gamma_4$ $1$ $-1$ $-1$ $a$ $\overline \Gamma_5$ $1$ $e^{-i\frac{1}{3}\pi}$ $e^{i\frac{1}{3}\pi}$ $b$ $\overline \Gamma_6$ $1$ $e^{i\frac{1}{3}\pi}$ $e^{-i\frac{1}{3}\pi}$ $b$ $\overline A_4$ $1$ $e^{-i\frac{2}{3}\pi}$ $e^{-i\frac{1}{3}\pi}$ $b$ $\overline A_5$ $1$ $1$ $-1$ $a$ $\overline A_6$ $1$ $e^{i\frac{2}{3}\pi}$ $e^{i\frac{1}{3}\pi}$ $b$ $\overline \Delta_4$ $1$ $-e^{i\frac{2}{3}\pi \omega}$ $-e^{i\frac{4}{3}\pi \omega}$ $c$ $\overline \Delta_5$ $1$ $e^{-i\frac{1}{3}\pi}e^{i\frac{2}{3}\pi \omega}$ $e^{i\frac{1}{3}\pi}e^{i\frac{4}{3}\pi \omega}$ $c$ $\overline \Delta_6$ $1$ $e^{i\frac{1}{3}\pi}e^{i\frac{2}{3}\pi \omega}$ $e^{-i\frac{1}{3}\pi}e^{i\frac{4}{3}\pi \omega}$ $c$ : Double-valued irreducible representations of SG $P3_1$ for the group of wave vectors at $\Gamma (0,0,0)$, A$(0,0,0.5)$ and $\Delta(0,0,\omega)$ with $0<\omega<0.5$. The degeneracy due to TRS for each representation is added at the end of each row, with $a$ and $b$ meaning doubly degenerate and $c$ non-degenerate. \[tab:P3\_1\] Band structures without SOC =========================== The $n$-hourglass-like band structures appear in the presence of SOC. In the main text, we have pointed out that when SOC is ignored, the $n$-hourglass-like structures turn into V-like, N-like and W-like structures for $n=2,3,4$, respectively, using our minimal model. Here we give the band structures without SOC and show that the V-, N-, and W-like band structures can also be understood from the compatibility relations between the single-valued irreducible representations of the group of wave vectors. Repres. $E$ $\{C_{2z}|\frac{1}{2}t_3\}$ $\{C_{4z}|\frac{1}{4}t_3\}$ $\{C_{4z}^{-1}|\frac{3}{4}t_3\}$ T. I. --------------- ----- ----------------------------- ---------------------------------- ---------------------------------- ------- ${\Gamma_1}$ $1$ $1$ $1$ $1$ $a$ ${\Gamma_2}$ $1$ $1$ $-1$ $-1$ $a$ ${\Gamma_3}$ $1$ $-1$ $i$ $-i$ $b$ ${\Gamma_4}$ $1$ $-1$ $-i$ $i$ $b$ ${Z_1}$ $1$ $i$ $e^{i\frac{1}{4}\pi}$ $e^{i\frac{3}{4}\pi}$ $b$ ${Z_2}$ $1$ $i$ $e^{-i\frac{3}{4}\pi}$ $e^{-i\frac{1}{4}\pi}$ $b$ ${Z_3}$ $1$ $-i$ $e^{i\frac{3}{4}\pi}$ $e^{i\frac{1}{4}\pi}$ $b$ ${Z_4}$ $1$ $-i$ $e^{-i\frac{1}{4}\pi}$ $e^{-i\frac{3}{4}\pi}$ $b$ ${\Lambda_1}$ $1$ ${e^{i\pi \omega}}$ ${e^{i\frac{1}{2}\pi \omega}}$ ${e^{i\frac{3}{2}\pi \omega}}$ $a$ ${\Lambda_2}$ $1$ ${e^{i\pi \omega}}$ ${-e^{i\frac{1}{2}\pi \omega}}$ ${-e^{i\frac{3}{2}\pi \omega}}$ $a$ ${\Lambda_3}$ $1$ ${-e^{i\pi \omega}}$ ${ie^{i\frac{1}{2}\pi \omega}}$ ${-ie^{i\frac{3}{2}\pi \omega}}$ $a$ ${\Lambda_4}$ $1$ ${-e^{i\pi \omega}}$ ${-ie^{i\frac{1}{2}\pi \omega}}$ ${ie^{i\frac{3}{2}\pi \omega}}$ $a$ : The single-valued irreducible representations of SG $P4_1$, for the group of wave vectors at $\Gamma$, Z and $\Lambda$. The degeneracy due to TRS for each representation is added at the end of each row, with $a$ meaning non-degenerate and $b$,$c$ doubly degenerate. Here $0<\omega<0.5$. For $\omega=0$, $\Lambda$ becomes $\Gamma$ and for $\omega=0.5$, $\Lambda$ becomes Z. \[tab:P4\_1\_single\_valued\] ![Band structure of ZnTe without SOC. The valence bands near Fermi energy showing N-like dispersions are enlarged and shown on the right.[]{data-label="fig:ZnTe_woSOC"}](figA6.jpg){width="8cm"} ![Band structure of Tl$_3$PbBr$_5$ without SOC. The conduction and valence bands near Fermi energy showing W-like dispersions are enlarged and shown on the right of each screw-invariant path.[]{data-label="fig:Tl3PbBr5_woSOC"}](figA7.jpg){width="8cm"} For SG $P4_1$, the groups of wave vectors at $\Gamma$, Z and $\Lambda$ are the same with that in Sec.\[sec:compat\]. The only difference is that now we are concerned with the single-valued irreducible representations[@Tinkham], which describe the symmetry of bands when the Hamiltonian contains no spin-dependent terms. The results are shown in Tab.\[tab:P4\_1\_single\_valued\], from which we can derive the compatibility relations between irreducible representations of the group of wave vectors at $\Gamma$ and $\Lambda$, $$\begin{split} \Gamma_1 \to \Lambda_1, \\ \Gamma_2 \to \Lambda_2, \\ \Gamma_3 \Gamma_4 \to \Lambda_3 + \Lambda_4, \end{split}$$ and the compatibility relations between the irreducible representation at Z and $\Lambda$, $$\begin{split} Z_1 Z_4 \to \Lambda_1 + \Lambda_4, \\ Z_2 Z_3 \to \Lambda_2 + \Lambda_3. \end{split}$$ From the compatibility relations, the W-like band structure along ${\Gamma}$Z can be deduced and is shown schematically in Fig.\[fig:compat\_woSOC\](a). ![(a) The compatibility relations between bands at $\Gamma$, $\Lambda$ and $Z$ of crystals with $P4_1$ SG, ignoring SOC. (b) Compatibility relation for bands at $\Gamma$, $\Delta$ and $A$ of crystals with $P3_1$ SG, also ignoring SOC.[]{data-label="fig:compat_woSOC"}](figA5.jpg){width="8cm"} Repres. $E$ $\{C_{3z}|\frac{1}{3}t_3\}$ $\{C_{3z}^{-1}|\frac{2}{3}t_3\}$ T. I. ------------ ----- -------------------------------------------------- -------------------------------------------------- ------- $\Gamma_1$ $1$ $1$ $1$ $a$ $\Gamma_2$ $1$ $e^{i\frac{2}{3}\pi}$ $e^{-i\frac{2}{3}\pi}$ $b$ $\Gamma_3$ $1$ $e^{-i\frac{2}{3}\pi}$ $e^{i\frac{2}{3}\pi}$ $b$ $A_1$ $1$ $e^{i\frac{2}{3}\pi}$ $e^{-i\frac{2}{3}\pi}$ $b$ $A_2$ $1$ $e^{-i\frac{2}{3}\pi}$ $e^{i\frac{2}{3}\pi}$ $b$ $A_3$ $1$ $1$ $1$ $a$ $\Delta_1$ $1$ $e^{i\frac{2}{3}\pi \omega}$ $e^{i\frac{4}{3}\pi \omega}$ $a$ $\Delta_2$ $1$ $e^{i\frac{2}{3}\pi}e^{i\frac{2}{3}\pi \omega}$ $e^{-i\frac{2}{3}\pi}e^{i\frac{4}{3}\pi \omega}$ $a$ $\Delta_3$ $1$ $e^{-i\frac{2}{3}\pi}e^{i\frac{2}{3}\pi \omega}$ $e^{i\frac{2}{3}\pi}e^{i\frac{4}{3}\pi \omega}$ $a$ : Single-valued irreducible representations of SG $P3_1$ for the group of wave vectors at $\Gamma (0,0,0)$, A$(0,0,0.5)$ and $\Delta(0,0,\omega)$, with $0<\omega<0.5$. The degeneracy due to TRS for each representation is added at the end of each row, with $a$ meaning non-degenerate and $b$ and $c$ doubly degenerate. \[tab:P3\_1\_single\_valued\] In a similar way we can obtain the single-valued irreducible representation of crystals with $P3_1$ SG symmetry, as shown in Table.\[tab:P3\_1\_single\_valued\], from which the compatibility relations between $\Gamma$ and $\Delta$, $$\begin{split} \Gamma_1 \to \Delta_1, \\ \Gamma_2 \Gamma_3 \to \Delta_2 + \Delta_3, \end{split}$$ and the compatibility relations between A and $\Delta$, $$\begin{split} A_1 A_2 \to \Delta_1 + \Delta_2, \\ A_3 \to \Delta_3, \end{split}$$ are obtained. The N-like band structures along $\Gamma$A can be deduced, as shown in Fig.\[fig:compat\_woSOC\](b). The band structures of ZnTe and Tl$_3$PbBr$_5$ without SOC are shown in Fig.\[fig:ZnTe\_woSOC\] and Fig.\[fig:Tl3PbBr5\_woSOC\], respectively. For ZnTe, along the screw-invariant line ($\Gamma$A), the band structure shows N-like dispersion. The shape can vary due to permutations of the vertices at $\Gamma$ and $A$ as well as curving. Similarly, for Tl$_3$PbBr$_5$, along $\Gamma$Z and MA, the band structure show W-like dispersions. Along XR, which is a twofold screw-invariant line, the bands show V-like dispersion. In the whole plane ZRA the bands are doubly degenerate, due to the presence of a twofold screw axis[@Wang2017PRB].

ArXiv

--- abstract: 'We have used data from the TeV $\gamma$-ray flare associated with the active galaxy Markarian 421 observed on 15 May 1996 to place bounds on the possible energy-dependence of the speed of light in the context of an effective quantum gravitational energy scale. The possibility of an observable time dispersion in high energy radiation has recently received attention in the literature, with some suggestions that the relevant energy scale could be less than the Planck mass and perhaps as low as $10^{16}$GeV. The limits derived here indicate this energy scale to be in excess of $4\times10^{16}$GeV at the 95% confidence level. To the best of our knowledge, this constitutes the first convincing limit on such phenomena in this energy regime.' --- =8.0 in =5.5 in Limits to Quantum Gravity Effects from\ Observations of TeV Flares in Active Galaxies .2in S.D. Biller,[$^{1}$ ]{}A.C. Breslin,[$^{2}$ ]{}J. Buckley,[$^{3}$ ]{} M. Catanese,[$^{4}$ ]{}M. Carson,[$^{2}$ ]{}D.A. Carter-Lewis,[$^{4}$ ]{} M.F. Cawley,[$^{5}$ ]{}D.J. Fegan,[$^{2}$ ]{}J. Finley,[$^{6}$ ]{}J.A. Gaidos,[$^{6}$ ]{}A.M. Hillas,[$^{7}$ ]{}F. Krennrich,[$^{4}$ ]{}R.C. Lamb,[$^{8}$ ]{}R. Lessard,[$^{6}$ ]{}C. Masterson,[$^{2}$ ]{}J.E. McEnery,[$^{9}$ ]{}B. McKernan,[$^{2}$ ]{}P. Moriarty,[$^{10}$ ]{}J. Quinn,[$^{11}$ ]{}H.J. Rose,[$^{7}$ ]{}F. Samuelson,[$^{4}$ ]{} G. Sembroski,[$^{6}$ ]{}P. Skelton,[$^{7}$ ]{}T.C. Weekes[$^{11}$ ]{}\ [*[$^{1}$ ]{}Oxford University, Oxford, United Kingdom\ [$^{2}$ ]{}University College, Dublin, Ireland\ [$^{3}$ ]{}Washington University, St. Louis, Missouri 63130\ [$^{4}$ ]{}Iowa State University, Ames, Iowa 50011\ [$^{5}$ ]{}Maynooth University, Maynooth, Ireland\ [$^{6}$ ]{}Purdue University, West Lafayette, Indiana 47907\ [$^{7}$ ]{}University of Leeds, Leeds, United Kingdom\ [$^{8}$ ]{}California Institute of Technology, California 91125\ [$^{9}$ ]{}University of Utah, Salt Lake City, UT 84112\ [$^{10}$ ]{}Galway-Mayo Instutute of Technology, Galway, Ireland\ [$^{11}$ ]{}Whipple Observatory, Amado, Arizona 85645\ *]{} .2in 0.3in [*Submitted to Physical Review Letters*]{} It has recently been pointed out that many quantum gravity scenarios may result in an observable time dispersion for high energy radiation originating at large distances from the Earth [@Amelino] [@Garay] [@Gambini]. This would result from an effective energy-dependence to the velocity of light in vaccum owing to propagation through a gravitational medium containing quantum fluctuations on distance scales near the Planck length, $L_P \simeq 10^{-33}$cm, with timescales on the order of $1/E_P$, where $E_P$ is the Planck Mass ($\simeq10^{19}$GeV). In particular, it has been indicated [@Amelino] that different approaches to quantum gravity lead to a similar description of the first-order effects of such a time dispersion: $$\Delta t \simeq \xi \frac{E}{E_{QG}} \frac{L}{c}$$ where $\Delta t$ is the time delay relative to the standard energy-independent speed of light, $c$; $\xi$ is a model-dependent factor of order 1; $E$ is the energy of the observed radiation; $E_{QG}$ is the assumed energy scale for quantum gravitational effects which can couple to electromagnetic radiation; and $L$ is the distance over which the radiation has propagated. While $E_{QG}$ is generally assumed to be on the order of $E_P$, recent work within the context of string theory suggests that the onset of noticeable quantum gravitational effects may correspond to a characteristic energy scale smaller than the Planck mass and perhaps as low as $10^{16}$GeV [@Witten]. Thus, any experimental probe of such scales or higher would be of great interest. In a recent paper [@Amelino] it was suggested that $\gamma$-ray bursts (GRBs) could provide a natural way to test such predictions owing to the short duration, high energies and the apparent cosmological origin of at least some of these bursts. Based on current data, these authors indicate that if [**1)**]{} time structure on the scale of 0.01 seconds or smaller can be established for energies $\sim200$keV and [**2)**]{} an association of such a burst can be made with an object possessing a redshift of order 1, energy scales of $E_{QG} \sim 10^{16}$GeV could be probed. Unfortunately, establishing the distance of any particular GRB from earth has proven to be non-trivial, with only a handful positively associated with optical counterparts. Also, some of the highest energies seen from GRBs are associated with an “afterglow” which seems to occur over much longer timescales than the initial burst. However, more stringent and robust limits to $E_{QG}$ can already be set based instead on the rapidly rising TeV flares seen to occur in active galaxies. The Whipple Observatory $\gamma$-ray telescope, located in Arizona, detects the Čerenkov light generated by electromagnetic cascades resulting from the interaction of high-energy $\gamma$-rays in the atmosphere. Images taken of such cascades are used to discriminate backgrounds and derive energies of the primary $\gamma$-rays in the regime above $\sim$250 GeV. To date, three extragalactic sources, all active galaxies of the blazar class, have been identified as emitters of TeV radiation [@Punch] [@Quinn] [@Catanese1]. Two of these, Markarian 421 and Markarian 501, produce particularly strong emission with energy spectra approximated by an $\sim E^{-2.5}$ power law (although Markarian 501 shows evidence for additional curvature) between energies of 300 GeV and 10 TeV [@Krennrich]. These same sources have also exhibited dramatic changes in flux level on timescales ranging from minutes to days. On several occasions, such variations have been simultaneously studied and correlated with x-ray, UV and optical measurements [@Buckley] [@Catanese2]. The most rapid flare observed thus far was seen from Markarian 421 on 15 May 1996 [@Gaidos]. This data is shown in figure 1, where the excess rate of $\gamma$-ray selected events above a threshold of 350 GeV is binned in intervals of 280 seconds duration, as it appeared in the original publication of this observation. To avoid confusion (and potential bias), we will retain this same binning for the current analysis. The doubling time of the flare is less than 15 minutes, although variability is apparent on the scale of the binning at the 99% confidence level. Because of the rapidly falling energy spectrum, the $\gamma$-ray data is dominated by events near the triggering threshold. Thus, the peak of the flare is almost entirely defined by events with $\gamma$-ray energies less than 1 TeV, as shown at the top of figure 2. In the 280 second interval corresponding to this peak, 4 events with $\gamma$-ray energies in excess of 2 TeV have been identified, whereas no such events could be identified in either of the 2 adjacent intervals (figure 2, bottom). The probability for this to occur by chance is less than 5%. Hence, at the 95% confidence level, emission above 2 TeV appears to keep in step with emission below 1 TeV for variability timescales less than 280 seconds. The redshift of Markarian 421 is 0.031, which translates to $1.1\times10^{16}$ light-seconds for an assumed Hubble constant of 85 km/s/Mpc. From equation 1, this leads to lower bound on $E_{QG}$/$\xi$ of $4\times10^{16}$GeV. We note that an earlier limit on the energy-dependence of the speed of light, which would be more restrictive than that given here, had been derived from the possible ultra-high-energy detection of anomalous pulsed emission from Her X-1 in 1986 [@Haines]. However, more recent analyses and the lack of further such detections suggests that the interpretation of that observation as a statistical fluctuation is not an unreasonable one [@Biller]. We therefore believe that the limit presented in this paper represents the most credible and stringent bound thus far obtained. The next generation of proposed ground-based instruments, such as VERITAS and HESS, will feature multi-telescope systems with much improved sensitivity, energy coverage and resolution, along with the ability to track candidate sources of flares more continuously using dedicated telescopes. This will allow for both a more detailed study of the time structure of currently known TeV sources and the prospect of discovering and studying more distant objects. It is therefore reasonable to expect to probe $E_{QG}$ to even higher energies in the near future from further studies of TeV flares. As has already been pointed out [@Amelino], the distinctive dependence of the shortest observable variability timescale on both energy and source distance for quantum gravitational dispersion should allow source-specific effects to be distinguished. Thus, future TeV studies could conceivably provide convincing evidence for quantum gravity, particularly if the resulting time-dispersion effects are associated with characteristic energy scales less than the Planck mass. We hope that this prospect, in addition to the bounds derived here, will encourage more detailed predictions of such phenomena to be calculated in the context of specific quantum gravity frameworks. We would like to thank Subir Sarkar at Oxford for several useful discussions on this topic. This work has been supported in part by PPARC, Forbairt, the US Department of Energy, NASA and the Smithsonian Institution. [99]{} Amelino-Camelia, G. [*et al.*]{}, Nature, [**393**]{}, 319, 25 June 1998 Garay, Luis J., Phys. Rev. Lett, [**80**]{}, 23 March 1998 Gambini, Rodolfo and Pullin, George, preprint gr-qc/9809038 Witten, E., Nucl. Phys. B, [**471**]{}, 135, 1996 Punch, M. [*et al.*]{}, Nature, [**358**]{}, 477, 1992 Quinn, J. [*et al.*]{}, Ap. J., [**458**]{}, L83, 1996 Catanese, M. [*et al.*]{}, Ap. J. (to appear) Krennrich, F. [*et al.*]{}, Ap. J. (to appear) Buckley, J. H. [*et al.*]{}, Ap. J., [**472**]{}, L9, 1996 Catanese, M. [*et al.*]{}, Ap. J., [**487**]{}, L143, 1997 Gaidos, J. [*et al.*]{}, Nature, [**383**]{}, 319, 1996 Haines, T. J. [*et al.*]{}, Phys. Rev. D, [**41**]{}, 692, 1990 Biller, S. D., PhD thesis, University of California, 1992

ArXiv

--- abstract: 'The dc transport properties of long arrays of small Al Josephson junctions, biased through on-chip Cr resistors, are studied. The $IV$-characteristics show a large Coulomb threshold for current as well as negative-slope regions indicating the regime of autonomous Bloch oscillations up to rather high frequencies of $f = I/2e \sim \unit[1]{GHz}$, comparable to those reported by other groups for single junctions. On the other hand, a small depth of the back-bending implies a low duty cycle and a broad spectrum of the oscillations, which we attribute to the insufficiently high impedance of the bias resistors. A self-sustained switching process at a small bias current is used to study the statistics of the switching voltages and to determine the effective Bloch capacitance which was found to considerably exceed the geometric junction capacitance.' author: - 'Sergey V. Lotkhov, Vladimir A. Krupenin, and Alexander B. Zorin' title: 'Cooper Pair Transport in a Resistor-Biased Josephson Junction Array' --- [^1][^2] [^3] Charge transfer, current, Josephson arrays, SQUIDs, superconductor-insulator-superconductor devices, thin-film devices, stripline components. Introduction ============ The application of single-charge tunneling effects in electrical metrology has been considered in a view of capacitance and current standards based on charge quantization phenomena (see, e.g., review [@RevKell]). The proper charge quantization in circuits with small tunnel junctions has been achieved for low junction transparencies: $R_{\rm T} \gg R_{\rm Q} \equiv h/e^2 \approx \unit[25.8]{k\Omega}$ for the single-electron tunnelling resistance, and $E_{\rm J} \ll E_{\rm C}$ for Cooper pair devices, with $E_{\rm J} \propto R_{\rm T}^{-1}$ and $E_{\rm C}$ being the Josephson coupling and charging energies, respectively. For example, remarkable advances have been made in development towards the single-electron standard of capacitance [@SciKell]. The accurate operation of these circuits, based on pumping of single electrons, has been, however, limited to small currents, $I \sim 1$ to $\unit[10]{pA}$, because of the stochastic nature of single electron tunneling with a relatively long time constant. Here we address the Cooper pair transport in the case of substantial Josephson coupling $E_{\rm J} \ge E_{\rm C}$, i.e., when both the Josephson phase and the charge exhibit significant quantum uncertainties. In such rather transparent, and therefore large-current-capable junctions, high accuracy could still be achieved due to the exact $2e$-periodicity of their energy bands as a function of the quasicharge [@LiZo]. Basic idea and sample ===================== The related transport phenomenon is known as Bloch oscillations in a current-biased junction with a fundamental current-frequency relation $I = 2ef$ [@LiZo]. Successful attempts to phase-lock the Bloch oscillation by an external signal were made in the early 1990s (see, e.g., [@Kuzm1; @Havi; @Kuzm2]) in the single junctions biased through high-ohmic on-chip microresistors for frequencies up to $f \sim \unit[10]{GHz}$. Unfortunately, the linewidth of oscillations was found to be rather large, $\delta f \sim \unit[1]{GHz}$, which was supposedly due to significant thermal fluctuations which smeared the current plateaus. Moreover, a possible size of these plateaus was expected to be rather small, scaling as a small threshold voltage in a single junction. Our approach is based on 1D arrays of small Josephson junctions. As was shown theoretically for current-biased arrays [@AL] and proved experimentally for very long voltage-biased Al arrays ($N \sim 200$) [@AAH], these arrays are analogous to long Josephson junctions: Similar to the fluxons formed in the long junctions, the charge profile in the array is governed by the sine-Gordon type of equation, describing the Cooper pair solitons of size (in number of junctions) $M \sim \sqrt {\tilde C/C_{\rm 0}} < N$ [@AAH]. Here we denote as $\tilde C$ the effective junction capacitance, related to the curvature of the ground state energy $E(q)$ and, hence, the shape of this 2$e$-soliton; $C_0 \ll \tilde C$ is the self-capacitance of the intermediate islands. The coherent motion of a train of 2$e$-solitons along an array biased via high-ohmic resistors can be described by a set of non-stationary equations: $$\label{SineGordon} L_B (q_i)\frac{\partial ^2 q_i }{\partial t^2} + \frac{\partial E(q_i )}{\partial q_i } = \frac{1}{C_0}(q_{i - 1} - 2q_i + q_{i + 1} ), \quad i = 1,2,...$$ which includes the quasicharge-dependent Bloch inductances of the junctions $L_{\rm B}(q_{\rm i})$ which were recently introduced in [@Z], with corresponding initial and boundary conditions. Here we denote the set of junction quasicharges as $\{q_{i}(t)\}$. In the limit of small Josephson coupling, $\lambda \equiv E_{\rm J}/E_{\rm C} \ll 1$, the periodic term $\partial E(q_i )/\partial q_i $ takes the form of the sawtooth function with the amplitude $e/C$, where $C$ is the capacitance of the individual junctions in the array. In the opposite case, $\lambda \gg 1$, there is an analytic expression [@LiZo]: $$\label{LargeLambda} \frac{\partial E(q_i)}{\partial q_i } =a(\lambda)\frac{e}{C}{\rm sin}(\pi q_i/e),$$ where $a(\lambda) = 2^{11/4} \pi ^{1/2} \lambda ^{3/4} \exp \left[ { - \left( {8\lambda } \right)^{1/2} } \right]$ is a numerical factor describing exponential suppression of the Bloch band width for $\lambda \to \infty$. In this paper, we report the dc properties of autonomous arrays of $N =60$ SQUIDs (see the SEM photo in Fig. \[SEM\]), fabricated using the Al/oxide/Al-Cr shadow evaporation process (see, e.g., [@Napoli]) in a 4-point layout with Cr microstrips of $\unit[50]{\mu m}$-length with $R_{\rm Cr} \approx \unit[550]{k\Omega}$ each. The tunnel resistance was $R_{\rm T} \approx \unit[8]{k\Omega}$ and an estimated capacitance was $C_{\rm T} \sim \unit[0.36]{fF}$ per junction, which corresponds to the maximum Josephson coupling energy of each SQUID, $E_{\rm J} \approx \unit[160]{\mu eV}$ and the charging energy $E_{\rm C} \equiv e^2/4C_{\rm T} \approx \unit[110]{\mu eV}$ per link, yielding their ratio $\lambda \approx 1.5$. We roughly estimate $C_{\rm 0} \sim \unit[50]{aF}$. The samples were measured in the current-bias mode with an external high-ohmic bias resistor, $R_{\rm B} = \unit[100]{M\Omega}$. The capacitance of each connection line to ground was $C_{\rm L} \sim \unit[1]{nF}$, resulting in an $RC$-constant of the bias circuitry of $\tau_{\rm D} \sim R_{\rm B}C_{\rm L} = \unit[0.1]{s}$. Our microwave-tight holder was supplied with Thermocoax$^{\textregistered}$ filters (jacket $\O~\unit[0.35]{mm}$, 110 cm long), anchored to the mixing chamber. Results ======= Several important peculiarities can be observed in the $IV$-characteristics of the arrays shown in Fig. \[IVC\]. In our opinion, the most remarkable transport property of the system relates to the bias current range $I_{\rm bias} \le 300$ to $\unit[350]{pA}$, with zero to negative slope of the $IV$-curve (see the blow-up of the corresponding region in the right inset in Fig. \[IVC\]), which we assume to be an indication of the Bloch oscillation regime [@LiZo]. Due to the moderate current values, the ramping of quasicharge in this transport regime occurs sufficiently slowly that the system always stays in the zero Bloch band and experiences 2$e$-periodic oscillations of the voltage with the frequencies up to $f = I_{\rm bias}/2e \sim \unit[1]{GHz}$. At larger currents, the average voltage across the array gradually increases, presumably due to excitation of the upper energy bands [@LiZo], thus setting a high-frequency limit for possible application of these particular Josephson arrays as sources of a quantized current. Note that, despite a large number of junctions in our arrays, this limitation, being at the level of GHz, is virtually of the same scale with the frequencies of phase-locking experiments on single junctions [@Kuzm1], whereas the typical voltage scale is much larger, which is promising for a better observability of the Bloch oscillations. On the other hand, small depth of the back-bending, typically, few percent, points out the low duty cycle of Bloch oscillations with the short pulse duration of $\tau_{\rm arr} \le R_{\rm Cr}C_{\rm arr} \sim (\unit[5]{GHz})^{-1} \ll f^{-1}$, where we use a rough estimate for the “input” capacitance of the array, $C_{\rm arr} \approx (2\tilde C C_{\rm 0})^{1/2} \approx \unit[0.4]{fF}$ (see below for our estimation of the effective value $\tilde C$). We attribute the present form of oscillations to insufficiently high impedance of the biasing resistors. Simple estimations show that the observed upper frequency limit for a single-band behaviour of about $\unit[1]{GHz}$ is achieved by the system biased only slightly above its Coulomb blockade threshold, therefore resulting in a typical decay time profile of oscillations. A broad spectrum of such oscillations, with considerable contribution of higher harmonics, makes it difficult to ensure their effective phase-locking to an external high-frequency signal. In particular, the preliminary experiments on microwave irradiation of our resistor-biased arrays up to frequencies of about $\unit[1]{GHz}$ did not result in observable features in the $IV$-curves. On the other hand, it was shown in [@Sawdrive] for the Josephson oscillations that even such oscillations may be effectively phase-locked using a driving signal of an appropriate waveform. As an indication of a hysteretical behaviour due to the inductance term in the equation of motion (\[SineGordon\]) [@AAH; @Z], the biasing point at small currents $I_{\rm bias} \sim \unit[10]{pA}$ was found to be unstable, exhibiting relatively slow irregular voltage oscillations, see left inset in Fig. \[IVC\]. As shown schematically in Fig. \[hystere\], random switching of the array from the blockade to the finite-current state was followed by a relatively slow recharging of the line capacitance $C_{\rm L}$, which made possible the real-time observation of the switching cycles. In the blockade state, A$\to$B, with a life-time $\Delta t \sim \tau_{\rm D}$, the current source charged the input terminal of the array until its switching, B$\to$C, to the finite-current state, $I_{\rm bias} \sim \unit[1]{nA}$, followed by a rapid discharging, C$\to$D, down to voltages $V_{\rm r} \approx \unit[500]{\mu V}$ and retrapping, D$\to$A, facilitated by the presence of fluctuations (cf. the thermally enhanced retrapping process in a shunted Josephson junction [@Ben-Jacob]). Whilst at $T < \unit[100]{mK}$ the average escape time, $t_{\rm mean} \equiv \left\langle {\Delta t} \right\rangle$, was nearly temperature-independent (cf. [@Escape]), as shown in Fig. \[Times\], the switchings at $T > \unit[100]{mK}$ were due to thermal escape over the barrier whose height we have evaluated to be $\Delta U \approx \unit[70]{\mu eV}$, as estimated from the slope of the Ahrrenius plot, see inset to Fig. \[Times\]. Using the approach [@Kazacha] which is valid, strictly speaking, only in the weak-coupling case $\lambda \ll 1$, we obtained a rough estimate of the effective junction capacitance and the length of a soliton in the array. It is possible to express the barrier height $\Delta U(V)$ through the threshold voltage $V_{\rm T}$ of soliton motion: $$\label{DeltaU} \Delta U(V) \approx \frac{eV_{\rm T}}{2}\left( {1 - \frac{V}{V_{\rm T}}} \right)^2.$$ Taking advantage of the narrow range of switching voltages at $T > \unit[100]{mK}$, $V \approx V_{\rm r}$, we estimated $V_{\rm T} \approx \unit[830]{\mu V}$, a soliton energy $E \equiv \Delta U(0) = eV_{\rm T}/2 = e^2/(2\tilde C C_{\rm 0})^{1/2} \approx \unit[420]{\mu eV} > E_{\rm C}$, the effective Bloch capacitance $\tilde C \sim \unit[1.6]{fF} \gg C_{\rm T}$, and the size of a soliton $M \sim 8$. One can see that the effective Bloch capacitance greatly exceeds the geometric junction capacitance which is obviously due to the strong suppression of the Bloch band width, $\Delta E = E^{\rm max}(q) - E^{\rm min}(q)$, for the values of $\lambda >1$ \[see, e.g., the limiting case, $\lambda \gg 1$, described by equation (\[LargeLambda\])\]. The shape of a static 2$e$-soliton in the infinitely long array is shown in Fig. \[Soliton\] as calculated by numerically solving the equation (\[SineGordon\]) in the stationary case $\partial q_i /\partial t \equiv 0$, for several representative values of the ratio $E_{\rm J}/E_{\rm C}$. Conclusions and outlook ======================= Our dc measurements pointed out the presence of autonomous Bloch oscillations in a resistively biased array of small Josephson junctions. Compared to the experiments with single Bloch junctions reported by other groups, similar Bloch frequencies are expected, whereas the Coulomb voltage threshold and, as a consequence, the amplitudes of oscillations are considerably larger, making a potential advantage of using the arrays for phase-locking experiments. However, to enable an external synchronization of these oscillations one should realize at least several times higher bias impedance, possibly in form of very resistive microstrips. A certain increase in the Bloch frequency could be expected for the junctions with a larger Josephson energy. But since the voltage scale is also affected by the values of $E_{\rm J}$ and $E_{\rm C}$, a detailed optimization is still important. Using the statistics of the spontaneous state switchings, we evaluated the effective junction capacitance and the length of a Cooper pair soliton. These data can be helpful for future device development. For practical applications, the influence of the background charges in the islands of the array on the soliton dynamics should also be investigated. Acknowledgements ================ The authors would like to thank D. V. Averin and A. V. Ustinov for helpful discussions. Technical assistance from T. Weimann (e-beam lithography), S. A. Bogoslovsky (measurement setup), F.-J. Ahlers and V. A. Rogalya (software) is gratefully acknowledged. The work was partially supported by the EU through the projects RSFQubit and EuroSQIP. [14]{} M. W. Keller, “Standards of current and capacitance based on single-electron tunneling devices,” in: *Recent Advances in Metrology and Fundamental Constants*, T. J. Quinn, S. Leschiutta, and P. Tavella Eds., Amsterdam: IOS Press, 2001, pp. 291-316. M. W. Keller, A. L. Eichenberger, J. M. Martinis, and  N. M. Zimmerman, “A capacitance standard based on counting electrons,” *Science*, vol. 285, pp. 1706-1709, 1999. K. K. Likharev and A. B. Zorin, “Theory of the Bloch-wave oscillations in small Josephson junctions,” *J. Low Temp. Phys.*, vol. 59, pp. 347-382, 1985. L. S. Kuzmin and D. B. Haviland, “Observation of the Bloch oscillations in an ultrasmall Josephson junction,” *Phys. Rev. Lett.*, vol. 67, pp. 2890-2893, 1991. D. B. Haviland, L. S. Kuzmin, P. Delsing, K. K. Likharev, and T. Claeson, “Experimental evidence for the Coulomb blockade of Cooper pair tunneling and Bloch oscillations in single Josephson junctions,” *Z. Phys. B - Condensed Matter*, vol. 85, pp. 339-347, 1991. L. S. Kuzmin, “Experimental evidence for the autonomous Bloch oscillations in single Josephson junctions,” *IEEE Trans. Appl. Supercond.*, vol. 3, pp. 1983-1986, 1993. D. V. Averin and K. K. Likharev, “Single-electronics: correlated transfer of single electrons and Cooper pairs in small tunnel junctions,” in *Mesoscopic Phenomena in Solids*, B. L. Altshuler, P. A. Lee, and R. A. Webb Eds., Amsterdam: Elsevier, 1991, pp. 175-271. P. $\rm {\AA}$gren, K. Andersson, and D. B. Haviland, “Kinetic inductance and Coulomb blockade in one dimensional Josephson junction arrays,” *J. Low Temp. Phys.*, vol. 124, pp. 291-304, 2001. A. B. Zorin, “Bloch inductance in small-capacitance Josephson juncitons,” *Phys. Rev. Lett.*, vol. 96, pp. 167001, 2006. S. V. Lotkhov, S. A. Bogoslovsky, A. B. Zorin, and J. Niemeyer, “Frequency-locked current of Cooper pairs in superconducting single electron transistor with ohmic resistor,” in *International Workshop on Superconducting Nano-Electronics Devices*, J. Pekola et al. Eds., New York: Kluwer Academic/Plenum Publishers, 2002, pp. 105–114. R. Monaco, “Enhanced ac Josephson effect,” *Appl. Phys. Lett.*, vol. 68, pp. 679-687, 1990. E. Ben-Jacob, D. J. Bergman, B. J. Matkowsky, and Z. Schuss, “Lifetime of oscillatory steady states,” *Phys. Rev. A*, vol. 26, pp. 2805-2816, 1982. K. Andersson and D. B. Haviland, “Escape from a zero-current state in a one-dimensional array of Josephson junctions,” *Phys. Rev. B*, vol. 67, pp. 092507, 2003. N. S. Bakhvalov, G. S. Kazacha, K. K. Likharev, and S. I. Serdykova, “Single-electron solitons in one-dimensional tunnel structures,” *Sov. Phys. JETP*, vol. 68, pp. 581-586, 1989. FIGURES {#figures .unnumbered} ======= Fig.1. SEM picture of a fragment of the resistor-biased array of two-junction SQUIDs fabricated with the three-angle evaporation technique. Fig.2. The $IV$-curves of two slightly different samples. Left inset: a typical time trace of the self-sustaining switching cycles observed at low bias currents. Right inset: a close-up look of the negative-slope segments in the $IV$-curves (shown by horizontal arrows) corresponding to single-band Bloch oscillations. Fig.3. An equivalent electrical circuit (a) and a cycle diagram in the $IV$-plane (b), both explaining the switching/retrapping process. For the sake of simplicity, the on-chip Cr resistors and the four-point connection layout of the experiment are not shown. The Josephson array is schematically represented by the hatched box. In the presence of fluctuations (external noise, thermal fluctuations, etc.) there is no stable bias point along the load line, which results in the self-sustaining voltage relaxation oscillations observed in experiment. Fig.4. Mean lifetimes of the blockade state. Inset: Ahrrenius plot at high temperatures. Fig.5. Calculated shape of a 2e-soliton presented in a form of the island charges $Q_i = q_{i+1}-q_i = C_{\rm 0}V_i$, where $q_i$ is a quasicharge of an $i$-th junction and $V_i$ is a voltage on the $i$-th island. ![[]{data-label="SEM"}](fig1cm){width="\columnwidth"} S. V. Lotkhov $et~al.$ “Cooper pair transport in a resistor-biased Josephson junction array” ![ []{data-label="IVC"}](fig2cm){width="\columnwidth"} S. V. Lotkhov $et~al.$ “Cooper pair transport in a resistor-biased Josephson junction array” ![ []{data-label="hystere"}](fig3cm){width="\columnwidth"} S. V. Lotkhov $et~al.$ “Cooper pair transport in a resistor-biased Josephson junction array” ![[]{data-label="Times"}](fig4cm){width="\columnwidth"} S. V. Lotkhov $et~al.$ “Cooper pair transport in a resistor-biased Josephson junction array” ![[]{data-label="Soliton"}](fig5cm){width="\columnwidth"} S. V. Lotkhov $et~al.$ “Cooper pair transport in a resistor-biased Josephson junction array” [^1]: Manuscript received July 10, 2006. This work was partially supported by the European Commission within the project EuroSQIP. [^2]: S. V. Lotkhov and A. B. Zorin are with Physikalisch-Technische Bundesanstalt, D-38116 Braunschweig, Germany. [^3]: V. A. Krupenin is with Laboratory of Cryoelectronics, Moscow State University, 119899 Moscow, Russian Federation.

ArXiv

--- abstract: 'We present an [*ab-initio*]{} study of the spin-transfer torque in a Fe/MgO/FePt/Fe magnetic tunnel junctions. We consider a FePt film with a thickness up to six unit cells, either in direct contact with the MgO spacer or with an intercalated ultra-thin Fe seed layer. We find that in the FePt layer the torque is not attenuated as strongly as in the case of pure Fe. Moreover, in FePt the torque alternates sign at the Fe and Pt atomic planes throughout the stack for all FePt thicknesses considered. Finally, when Fe is intercalated between MgO and L$1_0$-FePt, the torque is sharply attenuated and it is transferred to FePt only for a Fe seed layer that is less than two-atomic-planes thick. We attribute these features to the different spatial profiles of the exchange and correlation field and the induced non-equilibrium spin accumulation. The calculated tunnelling magneto-resistance of the Fe/MgO/FePt/Fe junctions studied is enhanced with respect to the one of Fe/MgO/Fe, while it is reduced with Fe intercalation. Our work shows that L$1_0$-FePt junctions can be promising candidates for current-operated magnetic devices and that the magnetic texture at the atomic scale has an important effect on the spin transfer torque.' author: - Mario Galante - 'Matthew O. A. Ellis' - Stefano Sanvito title: 'Non-trivial spatial dependence of the spin torques in L1$_0$ FePt-based tunnelling junctions' --- Introduction ============ Magnetic random access memories are believed to be among the most promising candidates to deliver the future of scalable, non-volatile, rapidly accessible data storage. At the heart of these devices are magnetic tunnel junctions (MTJs), which store data on the relative orientation of the magnetisation vectors of two magnetic layers separated by an insulating barrier. Reading and writing such junctions can be efficiently performed by applying an electric current through the device; exploiting the tunnelling magneto-resistance (TMR) [\[[@TMR]\]]{} effect for reading and using spin-transfer torque (STT) [\[[@STT]\]]{} to write. STT arises when a current passes across two ferromagnets having different magnetisation directions and it is caused by the transfer of angular momentum between the two mediated by the current. The conduction electrons become spin polarised by passing through the first magnetic layer and their angular momentum is then transferred to the second. The ideal insulating barrier acts as a spin-filter maximising the spin-polarisation of the current and hence the torque. Optimising the device structure to achieve low write currents is an important challenge in realising the potential of these devices. Whilst early demonstrations of MTJs focused on devices with in-plane layers magnetisation, the write current can be reduced significantly by adopting an out-of-plane geometry, where the magnetisation direction of both layers is oriented normally to the barrier interface. In junctions with this configuration, known as perpendicular MTJs (pMTJs), a large perpendicular magnetic anisotropy (PMA) is required to overcome the shape anisotropy of the thin film and enforce thermal stability in scalable devices. State-of-the-art devices are based upon CoFeB/MgO thin films, which can reach a TMR of up to 604% at room temperature and 1144% at low temperature [\[[@Ikeda2010]\]]{}. Furthermore, a large PMA has been observed at the CoFeB/MgO interface which is sufficient to achieve a perpendicular geometry in ultra-thin layers [\[[@Worledge2011a]\]]{}. Alternatively, L$1_0$ FePt is a popular material choice for high-density magnetic recording, since it has a large magneto-crystalline PMA, $K_u = \SI{7e6}{Jm^{-3}}$, allowing stable grain sizes down to a few nanometres [\[[@Weller2000]\]]{}. Despite the large uniaxial anisotropy, switching has been observed in FePt/Au giant magneto-resistance pillars with the aid of an applied magnetic field [\[[@Seki2006]\]]{}. Theoretical calculations of a FePt/MgO MTJ predicts a TMR of 340% for a Fe terminated interface [\[[@Taniguchi2008]\]]{}. Unfortunately, growing FePt/MgO devices can be challenging since the lattice mismatch between L1$_0$ FePt and MgO is large, $\sim 8.5\%$ [\[[@Cuadrado2014a]\]]{}. This may cause issues during the growth process, such as the inability of preserving the epitaxy across uneven layers. Strain can also cause a significant change in the magnetic properties of the FePt layer. In particular calculations have shown that a strain of 4% can reduce the PMA to about 10% of its original value [\[[@Seki2006]\]]{}. Practically, such strain can be reduced by inserting a seed-layer with a more amenable lattice constant at the MgO/FePt interface. In this work we investigate a series of FePt/MgO-based pMTJs in order to establish their potential for future device applications. We utilise *ab-initio* models to calculate the spin-transfer torque and the TMR for a range of FePt-based MTJ structures. We begin by detailing our computational method, before presenting results on the atom-resolved STT in the zero-bias limit for an Fe/MgO/Fe junction. This has an electronic structure analogous to that of CoFeB-based MTJs and hence provides a useful starting point for the discussion. We then continue with the analysis of the torque acting on the MTJs with FePt/Fe free layers and with a thin Fe seed layer intercalated at the MgO interface. In this case we vary the thickness of both the FePt layer and the seed layer (including the case where there is no seed layer). We find that a MgO/FePt interface yields a STT that decays more slowly in the free layer than in the MgO/Fe case, while the insertion of a Fe seed layer produces results similar to the FePt-free case. We then present the outcome of our TMR calculations and the STT acting on the Fe reference layer for some representative cases. Finally we replace the Fe atoms in the seed layer with Ni. This provides a comparison and helps us to formulate an argument about the origin of the spatial dependence of the STT. Computational Method ==================== Our approach for calculating the spin-transfer torque follows the prescription provided by Haney et al. in reference [\[[@Haney2007]\]]{} and is based on isolating the transport (non-equilibrium) contribution to the density matrix from the equilibrium part. The influence that an electric current has on the system can be estimated from first principles by combining density functional theory and the non-equilibrium Green’s functions method for transport (DFT+NEGF). All calculations have been performed with the [Smeagol]{} code [\[[@Rocha2005; @Rocha2006; @Rungger2008a; @Rungger2009]\]]{}, which implements the DFT+NEGF scheme within the numerical atomic orbital framework of the [Siesta]{} package [\[[@Soler2002]\]]{}. ![(Colour Online) Set up for a quantum transport calculation of a Fe/MgO/Fe/FePt/Fe junction. The dashed rectangle delimits the scattering region from the leads. The green arrows indicate the different direction of the magnetisations of the magnetic layers at the left- and right-hand side of the insulating barrier. The coloured spheres represent atoms of different species: Fe atoms are in red, Pt in grey, O in light blue, Mg are small red spheres.[]{data-label="setup"}](./Fig1.pdf){width="\columnwidth"} The system set up for the quantum transport calculation sandwiches the magnetic tunnel junction between two semi-infinite leads (see Fig. \[setup\]). These are assumed to be made of bulk material and at equilibrium. Note that a certain portion of the electrodes has to be included in the scattering region in order to ensure the continuity of the electrostatic potential. Here the magnetisation of the reference or fixed layer, ${\bf M}_\text{ref}$, is considered to be magnetised along $z$ (the transport direction) and the one of the free layer, ${\bf M}_\text{free}$, along $x$, so that the two form a $\pi/2$ angle. A voltage is applied in such a way that the electron flux is flowing along the stacking direction, $z$, in our convention from the reference layer to the free one. The component of the torque vector, ${\bf T}$, which is responsible for the switching between the parallel and the anti-parallel magnetisation configurations is the one that lies in the plane defined by ${\bf M}_\text{free}$ and ${\bf M}_\text{ref}$, namely the $x-z$ plane. In the free layer this component coincides with $T_z$, which is the main focus of our study. In order to reduce the computational costs, we limit our analysis to the torque response to a small bias, the *torkance*, meaning that all calculations are performed in the linear response approximation. At an atom $a$ in the free layer the torkance is defined as $$\label{torque:def} \tau_z^a \equiv \frac{dT^a_z}{dV}\Big|_{V=0} = \frac{1}{2}\text{Re}\sum_{i\in a}\sum_j \left(\boldsymbol\Delta_{ij}\times\frac{d\boldsymbol m_{ji}}{dV}\Big|_{V=0}\right)_z\:,$$ and this can be estimated with a zero-bias calculation. Here $\boldsymbol\Delta$ denotes the exchange and correlation field, namely the derivative of the exchange and correlation energy, $E_\mathrm{XC}$, with respect to the magnetisation density, $\boldsymbol m$, $\boldsymbol\Delta={\delta E_\mathrm{XC}}/{\delta {\boldsymbol m}}$. Thus, the derivative of $\boldsymbol m$ with respect to voltage embodies the spin contribution due to the rearrangement of the electronic population under non-equilibrium conditions. Henceforth this will be referred to as the non-equilibrium spin density or the spin accumulation. As such, the torque is the result of the interaction between the internal static field $\boldsymbol\Delta$ and the non-equilibrium spin density generated by the current flow. Further details on the calculation of the spin-transfer torque and the torkance can be found in Refs. [\[[@Stamenova2016; @Ellis2017]\]]{}. A series of junctions are constructed, all having a barrier of 6 MgO layers sandwiched between two semi-infinite leads of bulk [*bcc*]{} Fe oriented along the (001) direction. Periodic boundary conditions are applied in the plane perpendicular to the transport, as a result of the perfect epitaxy of the junction. The in-plane lattice constant is taken to be $a_\text{Fe}=\SI{2.866}{\angstrom}$ throughout the system. The out-of-plane lattice constants of the remaining materials were chosen according to information provided in reference [\[[@Kohn2013; @Cuadrado2014a]\]]{}, in particular $c_\text{MgO}=4.05/\sqrt{2}$ $\SI{}{\angstrom}$, $c_\text{FePt} = \SI{1.737}{\angstrom}$. The same studies assess that the most stable interfacial configuration is made of a Fe-terminated FePt surface on top of O (Fe) for the FePt/ MgO (FePt/Fe) interface, with an inter-plane distance of $\SI{2.2}{\angstrom}$ ($\SI{1.585}{\angstrom}$). The accuracy of such estimates was found satisfactory by relaxation of the different structures. The local spin density approximation (LSDA) for the exchange correlation potential was adopted. A real-space mesh cut-off of  [Ry]{} along with a 15$\times$15 $k$-point mesh in the plane orthogonal to transport were found to yield converged results. We adopted double $\zeta$ polarised orbitals for each atomic species and the convergence of the radial cut-offs was verified by comparing the band structure of bulk materials with the result of all-electron calculations. Since the introduction of spin-orbit coupling effects did not yield a sizeable change to our calculated torques we have omitted relativistic corrections. Results ======= ![(Color Online) Real space profiles of the relevant components of (a) the exchange and correlation field, $\boldsymbol\Delta$, (b) the non-equilibrium spin density, $\mathrm{d}\boldsymbol{m}/\mathrm{d}V$, and (c) the torkance, $\boldsymbol\tau$, per unit $\mu_\mathrm{B}/e$ and area acting on the [*bcc*]{} Fe free layer. The coloured background indicates the atomic species in the stack: red for Fe, blue for O, green for Mg.[]{data-label="FeMgO"}](./Fig2.pdf){width=".5\textwidth"} We begin by examining the properties of a Fe/MgO/Fe MTJ to later discuss their modification upon the introduction of a FePt layer. As shown in Eq. (\[torque:def\]), the torkance is given by the vector product of the exchange and correlation field and the non-equilibrium spin density. Since the free layer is magnetised in the $x$-direction and within the LSDA the exchange and correlation field is proportional and locally parallel to the magnetisation, the only relevant components to the torkance are $\Delta_x$ and $\mathrm{d}m_y/\mathrm{d}V$. These two components and the resulting torkance, $\tau_z$, are shown in Fig. \[FeMgO\]. In general, $\Delta_x$ peaks at the Fe/MgO interface and then presents small oscillations with the period of the interlayer Fe separation, $a_\text{Fe}$. Such profile does correlate with the real space profile of the equilibrium magnetic moment (not displayed), which is also enhanced at the Fe/MgO interface. In contrast, the non-equilibrium spin density \[panel (b)\] has an appreciable magnitude only in the region around the Fe/MgO interface. This decays in the Fe layer and is almost fully attenuated a few monolayers from the interface. Such behaviour will later be compared with that in FePt and in Ni. Finally note that there is an appreciable non-equilibrium spin density also in the MgO, although it does not contribute to the torkance since the exchange and correlation field vanishes in absence of a local magnetization \[see panel (a)\]. If we now consider the torkance we note that this is sharply peaked at the Fe/MgO interface and is attenuated in the Fe layer at the same speed of the non-equilibrium spin density. In fact, for this Fe/MgO/Fe case the spatial dependance of the torkance resembles closely that of the spin accumulation, given the fact that the exchange and correlation field has little spatial dependence in Fe. Let us remark, however, that the point-by-point vector product of the quantities in panels (a) and (b) does not give the torkance in panel (c), since the sum of the products of the matrix elements does not equal the product of the sums \[namely, $\sum_{ij} \left(\boldsymbol\Delta_{ij}\times\frac{d\boldsymbol m_{ji}}{dV}\right)\ne \left(\sum_{i}\boldsymbol\Delta_{ii}\right)\times\left(\sum_{i}\frac{d\boldsymbol m_{ii}}{dV}\right)$ - see formula (\[torque:def\])\]. {width="17.78cm"} We now explore the effects of inserting a layer of FePt at the MgO/free layer interface. Figures \[FePt:noSL\](a)-(c) show, as with the case of the Fe/MgO/Fe MTJ, the relevant components of $\boldsymbol\Delta$ and $\mathrm{d}\boldsymbol{m}/\mathrm{d}V$ contributing to the total torkance along $z$ for a FePt layer 4-unit-cell thick. From panel (a) it is clear that the exchange and correlation field is enhanced at the Fe sites, and also finite at the Pt ones. This is because in L$1_0$ FePt there is an induced magnetic moment on the Pt ions (this is about $0.4~\mu_\mathrm{B}$ as calculated from the Mülliken population analysis), in agreement with previous *ab-inito* calculations [\[[@Cuadrado2014a]\]]{}. The oscillations in the $\Delta_x$ profile remain constant in the FePt layer without any sign of decay, and then in the Fe layer the $\Delta_x$ profile returns to resemble the one observed before in Fig. \[FeMgO\]. Note that $\boldsymbol\Delta$ is an equilibrium property, which essentially depends on the presence of an exchange splitting in a given material. As such one does not expect a decay of $\boldsymbol\Delta$ unless there is a decay in the magnetisation. {width="17.78cm"} In contrast to the pure Fe case, the non-equilibrium spin density has lower intensity in FePt than in Fe but a significantly less attenuated decay \[panel (b)\]. The total non-equilibrium spin density shows regular oscillations within the FePt layer, whilst it is enhanced at both the FePt/Fe and the MgO/FePt interfaces, and then vanishes within a few unit cells of the Fe lead. Furthermore we observe that $\mathrm{d}{m_y}/\mathrm{d}V$ in Pt has opposite sign with respect to that of the first Fe layer in contact to MgO. Finally, the torkance \[panel (c)\] is again peaked at the interface with MgO but its strength is reduced in comparison to that computed for the Fe/MgO/Fe MTJ with the same MgO thickness. Within the FePt layer the torkance does not attenuate as in Fe but persists to reach the Fe-only side of the free layer. Most interestingly the torkance has an oscillatory behaviour in FePt, presenting small negative values at the Pt layers and positive at the Fe ones. Such oscillations are common in antiferromagnets \[[@Stamenova2016]\] and here are observed also in a ferromagnet with non-trivial magnetic texture. It is also interesting to note that, despite the larger spin accumulation at Pt sites, the resulting torque is smaller than that at the Fe ones. This is due to the fact that the exchange and correlation field in Pt is significant weaker than in Fe (because the magnetisation is smaller). The persistence of the torkance in the FePt layer remains as we change the FePt thickness, $n_\text{FePt}$ (number of unit cells). This can be seen in the panels (d)-(f) of Fig. \[FePt:noSL\]. For a thin layer \[panel (d)\] the torque is enhanced at the FePt/Fe interface, while it is attenuated for all the other cases \[e.g. see $n_\text{FePt}=6$ in panel (f)\]. Furthermore, for all the thicknesses considered the torkance remains strikingly positive at all the Fe atomic planes of FePt, while it is small and negative at the Pt ones. Moreover, the intensity of the peak at the MgO/FePt is not modified by the increase in thickness. ![(Colour Online) Panel (a): the total torkance per unit $\mu_\mathrm{B}/e$ and area acting on the free layer of Fe/MgO/Fe/FePt/Fe junctions with 2, 4 and 6 FePt monolayers and a Fe seed layer of $n_\mathrm{SL}=0, 2, 4$ atomic planes. Panel (b): the calculated TMR in Fe/MgO/FePt/Fe and Fe/MgO/Fe/FePt/Fe junctions with $n_\mathrm{SL}=2, 4, 8$ and 4 FePt unit cells. In both graphs the black dashed line represents the same quantity calculated for the Fe/MgO/Fe junction.[]{data-label="TotT+TMR"}](./Fig5.pdf){width="\columnwidth"} Although Fe/MgO/FePt/Fe junctions provide an interesting case of study, the significant lattice mismatch between MgO and L1$_0$ FePt ($\sim$ 8.5%) makes their experimental realisation troublesome. This problem may be overcome by inserting a compatible seed layer at the MgO/FePt interface. Hence, we have analysed the influence of incorporating a thin Fe seed layer (SL) between the MgO and the FePt, keeping the thickness of the FePt layer constant at 4 unit cells. The Fe SL has different effects depending on its thickness (see Fig. \[FePt:FeSL\]). We notice from panel (a) that the exchange and correlation field profile in FePt is analogous to the previous case (since the equilibrium magnetisation profile is also unchanged), while $\Delta_x$ is almost constant in the seed layer. The non-equilibrium spin density still oscillates in FePt, although the amplitude of such oscillations is much smaller than that obtained in absence of the SL. Consequently, the torkance \[panel (c)\] is peaked at the MgO/Fe interface with the SL and its intensity is comparable to that observed for the Fe/MgO/Fe case (see figure \[FePt:noSL\]). The torkance, however, is not exactly zero away from the SL, in particular on the Fe atoms of FePt and at the FePt/Fe interface. This does not happen for thicker Fe SLs \[panels (e) and (f)\], for which the total torkance decays before reaching the interface with FePt. In general, however, the main effect of the seed layer is to suppress the persistence of the torkance in FePt, so that all the angular momentum transfer takes place in the seed layer. ![(Colour Online) Torkance per unit $\mu_\mathrm{B}/e$ and area acting on the reference layer of Fe/MgO/Fe (a), Fe/MgO/FePt(4)/Fe (b) and Fe/MgO/Fe(2)/FePt(4)/Fe (c) MTJs. The coloured background indicates the atomic species to which each point corresponds to: red for Fe, blue for O, green for Mg. []{data-label="RefL"}](./Fig6.pdf){width="\columnwidth"} We now move to analyse the total torkance and the TMR of each junction. Figure \[TotT+TMR\](a) shows the total torkance integrated over the free layer, $\tau_z^\mathrm{tot}=\sum_a^{\alpha\in\mathrm{FL}}\tau_z^a$, for different thicknesses of the FePt layer. We present results for the situation where there is no SL (red squares), and for a Fe SL of respectively 2 (green squares) and 4 atomic planes (blue squares). For each SL thickness, the torkance shows little dependence on the thickness of the FePt layer. When there is no SL this is attributed to the oscillatory behaviour without attenuation of the torkance profile as observed in Fig. \[FePt:noSL\]. In contrast, when a SL is present most of the torque resides at the first MgO/Fe interface so that the thickness of the FePt becomes irrelevant (see Fig. \[FePt:FeSL\]). Interestingly, when a SL is present the total torkance transferred into the Fe/MgO/Fe/FePt MTJ is larger than that of a simpler Fe/MgO/Fe MTJ with identical barrier (dashed black line). This is no longer true when the SL is absent. Such finding means that the introduction of a Fe seed layer not only helps in achieving a better epitaxy during the growth but also facilitates a larger spin transfer torque. Figure \[TotT+TMR\](b) shows the calculated TMR for each junction (in all cases the FePt layer comprises 4 layers) and a comparison with that of a Fe/MgO/Fe MTJ with an identical barrier. We observe that the junction with no Fe SL presents the largest TMR, despite having the lowest torkance. This is unexpected, since in FePt bands with $\Delta_1$ symmetry, namely those with the largest transmission across MgO, are present for both spin channels [\[[@Taniguchi2008]\]]{}. Such feature returns a predicted TMR for MTJs with FePt leads not exceeding 340% [\[[@Taniguchi2008]\]]{}. However, here the situation is different since in all our MTJs the leads are made of Fe, so that spin filtering is always in place. As such, in our case the addition of a FePt layer (or a complex Fe/FePt layer) changes the details of the spin-dependent scattering potential, but does not alter the main spin-filtering mechanism at play in Fe/MgO/Fe junctions. Interestingly, as the thickness of the Fe SL gets larger, the value of the TMR is reduced. {width="17.78cm"} So far the left electrode has been considered to be the fixed layer, namely the one producing the spin-polarised current. It is now interesting to look at the opposite case, namely the one where the electrons flux flows from the right-hand side to the left-hand side electrode. This is the situation where the FePt/Fe composite electrode acts as the fixed, current polarising, layer. Since in the right electrode the magnetisation is along the $z$ direction, the relevant torque in this case is $\tau_x$. This is presented in Fig. \[RefL\] for three representative junctions: (a) Fe/MgO/Fe, (b) Fe/MgO/FePt(4)/Fe, and (c) Fe/MgO/Fe(2)/FePt(4)/Fe, where the numbers in parentheses indicate the number of unit cells. Since in this geometry the current flows in the opposite direction than previously, we have plotted $-\tau_x$, namely the torque component that will lead to an alignment of the magnetisations of the fixed and free layers. The trend of $-\tau_x$ is in all cases analogous to that of $\tau_z$ for the Fe/MgO/Fe MTJ \[see Fig. \[FeMgO\] (c)\], namely the STT is peaked at the magnet/insulator interface and is negligible elsewhere. The only significant difference between the three MTJs is the reduction of approximately a factor two of the peak intensity for the Fe/MgO/FePt(4)/Fe stack \[panel (b)\]. Discussion ========== The results presented so far indicate that the STT (the torkance) varies strongly with the distance from the MgO interface, and that the details depend subtly on the specific layer structure. In general, Fe seems capable of absorbing a significant amount of angular momentum, so that only a few Fe monolayers are enough to make the STT decay sharply from the MgO interface. The main cause of such effect has to be found in the intense Fe exchange field. In fact, the strong exchange interaction in Fe relaxes the non-equilibrium spin density (the spin accumulation) toward the local direction of the magnetisation within a few atomic layers from the interface, so that there is little $\mathrm{d}\boldsymbol{m}/\mathrm{d}V$ away from the interface itself. In addition the exchange and correlation field remains almost constant within the Fe layer, resulting in a torque that persists little away from the interface with MgO. In L$1_0$ FePt the alternating planes of Fe and Pt lead to a magnetisation texture that is non-uniform at the atomic scale. In particular ${\bf \Delta}$ is small at the Pt sites so that the average exchange and correlation field is reduced with respect to that of the pure Fe case. As a consequence the spin accumulation can penetrate longer into the free layer so that the STT decays less sharply. In order to further investigate the effects of the exchange field on the spatial decay of the torque we now consider a Ni seed layer since it has a much smaller moment, and thus exchange field, than Fe. The calculation has been simplified by maintaining the [*bcc*]{} structure and the lattice constant of Fe. As such our device stack does not correspond to a likely experimental situation but just serves the purpose of comparing the different seed layers. The atomic resolved torkance for a Fe/MgO/Ni/FePt/Fe stack with a Ni seed layer comprising 2, 4 and 6 atomic planes is shown in figure \[Ni\]. As in the case of a Fe seed layer, the torque \[panel (c)\] is strongly peaked at the Ni/MgO interface, but now it does not decay entirely and thus a non-vanishing STT with an oscillatory behaviour persists into the FePt layer. A closer look at the profile of ${\bf \Delta}$ across the junction \[panel (a)\] reveals that the exchange and correlation field in Ni is about half of that of Fe \[see Figure \[FePt:FeSL\](a)\]. As a consequence, in Ni the spin accumulation does not relax along the local direction of the magnetization as efficiently as in Fe, a fact that can be appreciated by comparing Fig. \[Ni\](b) with Fig. \[FePt:FeSL\](b). Interestingly, the attenuation of the spin accumulation and thus of the torque is not complete even for relatively thick Ni seed layers, as can be seen in panels (d) through (f). A second interesting observation concerns the phase of the oscillations of the STT in the FePt layer. In fact for a junction where FePt is in direct contact with the MgO barrier, the torque is positive at the Fe planes and negative (although rather small) at the Pt ones. The same behaviour, although with a much reduced torque is observed for Fe intercalation (in the presence of a Fe seed layer). In contrast when the seed layer is made of Ni the sign of the STT on the FePt layer changes, becoming negative at the Fe planes and positive (although small) at the Pt ones. As a result the total integrated torque over the entire free layer (seed layer plus FePt) for Ni intercalation is two thirds than that obtained with Fe intercalation. Finally, we wish to make a few general remarks on the spatial dependence of the STT. Macroscopic models combining the Landau-Lifshitz-Gilbert equation for the magnetisation dynamics with a diffusion model for the spin accumulation [\[[@Abert2015; @Wang2006]\]]{} suggest that the spin accumulation is maximised in regions where there is a large magnetisation gradient, namely at interfaces. This is confirmed here at the microscopic level. In all cases investigated we find the maximum spin accumulation, and hence torque, at the interface between the free layer and MgO regardless of the presence of a seed layer. Furthermore, we also find an enhanced spin accumulation and torque at the second interface between the free layer and the Fe lead, although this is small since the spin accumulation always decays in the free layer. The fine details of the spin accumulation profile depend on how the entire stack responds to the application of an external bias. This in turn is affected by the reorganisation in the occupation of the states around the Fermi surface, which is indeed a subtle effect. In general a large exchange splitting causes the spin accumulation to relax faster along the local magnetisation direction. As such we expect the spin accumulation to decay more severely in the free layer of stacks where there is a large torque at the first few atomic layers in contact with the MgO barrier. This in turn depends on the strength of the exchange and correlation field, which in the LSDA can be written as $${\bf\Delta}^\mathrm{LSDA}({\bf r})=\frac{\delta E^\mathrm{LSDA}_\mathrm{XC}}{\delta\mu_\mathrm{B}{\bf m}({\bf r})}= -\frac{\partial \epsilon_\mathrm{XC}}{\partial m({\bf r})}\:\frac{n({\bf r})}{\mu_\mathrm{B}}\:\frac{{\bf m}({\bf r})}{m({\bf r})}\:,$$ where ${\bf m}({\bf r})$ is the local magnetisation vector, $m({\bf r})=|{\bf m}({\bf r})|$, $E^\mathrm{LSDA}_\mathrm{XC}$ is the LSDA exchange and correlation energy, $\epsilon_\mathrm{XC}$ is the exchange and correlation energy density of the homogeneous electron gas, $n({\bf r})$ is the charge density and $\mu_\mathrm{B}$ the Bohr magneton. Crucially the LSDA ${\bf\Delta}$ is locally parallel to the magnetisation direction. As such one expects ${\bf\Delta}$ (and hence the torque) to change sign as the local magnetisation changes sign (as in the case of antiferromagnets). Furthermore one can show that $|{\bf\Delta}|\sim Im$, where $I$ is the Stoner parameter [\[[@Simoni2017]\]]{}. This means that for similar Stoner coupling the exchange and correlation field is more intense for materials presenting larger magnetization. This last feature explains the difference in ${\bf\Delta}$ and torque between the Fe and the Ni seed layer. In fact Fe and Ni have rather similar Stoner parameter but their magnetization differ by more than a factor three. Conclusions =========== In conclusion, we have calculated the STT acting upon the free ferromagnetic layer in a series of FePt-based magnetic tunnel junctions. For a simple Fe/MgO/Fe MTJ the torkance is peaked at the MgO interface and decays within 4 atomic planes. When the stack is modified to include FePt \[Fe/MgO/FePt/Fe\] the torkance decays much slower and persists into the free layer up to at least 12 atomic planes. Such retention is associated to torkance oscillations at the length scale of the Fe-Pt plane separation. Since the lattice mismatch between MgO and FePt is large we have explored the option to intercalate a Fe seed layer at the interface between MgO and FePt. Also in this case the torkance is significant only at the first MgO/Fe interface and it vanishes in FePt. This is the result of the strong reduction of the spin accumulation beyond the Fe seed layer. Such strong attenuation appears to originate from the large exchange and correlation field in Fe, which rapidly aligns the spin accumulation along the local direction of magnetization. Such hypothesis is confirmed by calculations for the STT in some hypothetical MTJs incorporating a Ni seed layer. Since Ni has an exchange and correlation field that is weaker than that of Fe, it is less effective at suppressing the spin accumulation (in absorbing angular momentum) and thus the attenuation of the torkance is weaker. All together our results suggest that the atomic and materials details of the MTJs stack play an important role in determining the total STT that a free layer can experience. This knowledge can help in designing stacks with maximal torkance, so that a reduction in the critical current for switching can be achieved. Acknowledgements ================ This work has been supported by the Science Foundation Ireland Principal Investigator award (grant no. 14/IA/2624 and 16/US-C2C/3287) and TCHPC (Research IT, Trinity College Dublin). The authors wish to acknowledge the DJEI/DES/SFI/HEA Irish Centre for High-End Computing (ICHEC) for the provision of computational facilities and support. [99]{} M. Julliere, Phys. Lett. A [**54**]{}, 225 (1975). J. Slonczewski, J. Magn. Magn. Mater. [**159**]{}, L1 (1996). S. Ikeda, K. Miura, H. Yamamoto, K. Mizunuma, H.D. Gan, M. Endo, S. Kanai, J. Hayakawa, F. Matsukura and H. Ohno, Nature Materials [**9**]{}, 721 (2010). D.C. Worledge, G. Hu, D.W. Abraham, J.Z. Sun, P.L. Trouilloud, J. Nowak, S. Brown, M.C. Gaidis, E.J. O’Sullivan and R.P. Robertazzi, Appl. Phys. Lett. [**98**]{}, 022501 (2011). D. Weller, A. Moser, L. Folks, and M. Best, IEEE Trans. Magn. [**36**]{}, 10 (2000). T. Seki, S. Mitani, K. Yakushiji and K. Takanashi, Appl. Phys. Lett. [**88**]{}, 172504 (2006). Y. Taniguchi, Y. Miura, K. Abe and M. Shirai, IEEE Trans. Magn. [**44**]{}, 2585 (2008). R. Cuadrado and R.W. Chantrell, Phys. Rev. B [**89**]{}, 094407 (2014). P.M. Haney, D. Waldron, R.A. Duine, A.S. Núñez, H. Guo and A.H. MacDonald, Phys. Rev. B [**76**]{}, 024404 (2007). A.R. Rocha, V.M. García-Suárez, S.W. Bailey, C.J. Lambert, J. Ferrer and S. Sanvito, Nature Materials [**4**]{}, 335 (2005). A.R. Rocha, V.M. García-Suárez, S.W. Bailey, C.J. Lambert, J. Ferrer and S. Sanvito, Phys. Rev. B [**73**]{}, 085414 (2006). I. Rungger and S. Sanvito, Phys. Rev. B [**78**]{}, 035407 (2008). I. Rungger, O. Mryasov and S. Sanvito, Phys. Rev. B [**79**]{}, 094414 (2009). J.M. Soler, E. Artacho, J.D. Gale, A. García, J. Junquera, P. Ordejòn, and D. Sànchez-Portal, J. Phys.: Condens. Matter [**14**]{}, 2745 (2002). M. Stamenova, R. Mohebbi, J. Seyedyazdi, I. Rungger and S. Sanvito, Phys. Rev. B [**95**]{}, 060403 (2017). M.O.A. Ellis, M. Stamenova and S. Sanvito, Phys. Rev. B [**96**]{}, 224410 (2017). A. Kohn, N. Tal, A. Elkayam, A. Kovacs, D. Li, S. Wang, S. Ghannadzadeh, T. Hesjedal and R.C.C. Ward, Appl. Phys. Lett. [**102**]{}, 062403 (2013). C. Abert, M. Ruggeri, F. Bruckner, C. Vogler, G. Hrkac, D. Praetorius and D. Suess, Scientific Reports [**5**]{}, 14855 (2015). C.J. García-Cervera and X.-P. Wang, J. Comput. Phys. [**224**]{}, 699 (2006). J. Simoni, M. Stamenova and S. Sanvito, Phys. Rev. B [**96**]{}, 054411 (2017).

ArXiv

--- abstract: 'Following the field theoretic approach of Basko *et al*., Ann. Phys. **321**, 1126 (2006), we study in detail the real-time dynamics of a system expected to exhibit many-body localization. In particular, for time scales inaccessible to exact methods, we demonstrate that within the second-Born approximation that the temporal decay of the density-density correlation function is non-exponential and is consistent with a finite value for $t\to\infty$, as expected in a non-ergodic state. This behavior persists over a wide range of disorder and interaction strengths. We discuss the implications of our findings with respect to dynamical phase boundaries based both on exact diagonalization studies and as well as those established by the methods of Ref. .' author: - Yevgeny Bar Lev - 'David R. Reichman' bibliography: - 'MBL.bib' title: 'Dynamics of Many-Body Localization' --- \#1[\_[\#1]{}]{} \#1[\_[\#1]{}\^]{} \#1[\#1|]{} \#1[|\#1]{} \#1\#2[.\#1|\#2]{} \#1\#2\#3[\#1|\#2|\#3]{} \#1\#2[..|\#1\#2|]{} It has been known for more than 50 years that *non-interacting* particles in a one-dimensional disordered system exhibit Anderson localization [@Anderson1958b], namely the exponential suppression of transport. While a localized system is non-ergodic and thus does not thermalize, coupling the system to other degrees of freedom with a continuous spectrum, such as a heat bath, allows thermalization to occur via processes such as variable-range hopping [@Mott1969]. For an isolated many-body system, only interactions between the particles may lead to thermalization. The question of whether or not localization is stable in the presence of interactions was first considered by Fleishman and Anderson [@Fleishman1980a], who concluded that short-ranged interactions cannot destabilize the insulating phase. A similar and still open question also exists for Bose-Einstein condensates, treated in the framework of the time-dependent Gross-Pitaevskii (or nonlinear Schrödinger) equation [@Fishman2009b]. In this case numerics, as well as analytical arguments, suggest a temporally sub-diffusive or even logarithmic thermalization behavior for not very strong interactions [@Fishman2009b]. Using a diagrammatic approach, Basko *et al.* argued that for a general class of isolated, disordered and interacting systems, a many-body mobility edge exists, similarly to the Anderson mobility edge in a three-dimensional non-interacting system [@Basko2006a]. Namely, a critical energy separates “insulating” and “metallic” eigenstates, which can be distinguished by evaluating the spatial correlations of any local operator. “Metallic” eigenstates will have non-vanishing or slowly decaying correlations, while “insulating” states will have exponentially decaying correlations. By changing the energy (or the micro canonical temperature) of the system across the mobility-edge, the system will undergo an insulator–metal transition. Similar to the Anderson transition, the many-body localization (MBL) transition is a dynamical and *not* a thermodynamic phenomena [@Basko2006a]. However, the MBL transition is also not a conventional quantum phase transition since the critical energy, which depends on the parameters of the system, may be very far from the ground state. In fact, for systems of bounded energy density (e.g., finite number of states per site), Oganesyan and Huse suggested that the transition will persist up to infinite temperature [@Oganesyan2007a]. Namely, nontrivial parameters of the system may be found such that essentially *all* the eigenstates are “insulating”. For a zero dimensional system mapped to the Bethe lattice, it was theoretically proposed [@Altshuler1997] and recently numerically examined [@DeLuca2013], that for some range of parameters the metallic phase can be *non-ergodic*. The existence of a non-ergodic *metallic* phase for *finite* dimensional systems had been conjectured [@Altshuler2010], but has been numerically tested only for small systems [@Luca2013]. By calculation of the DC conductivity or the properties of eigenfunctions for sufficiently small systems, the MBL transition has gained support from numerical studies that utilize either direct diagonalization [@Oganesyan2007a; @Karahalios2009a; @Pal2010a; @Berkelbach2010a; @Barisic2010a] or methods of similar numerical complexity [@Monthus2010a]. However, all of these studies suffer from the drawback that the numerically accessible system size is about $16$ sites, which does not allow for a systematic analysis of finite size effects. This is not an issue deep within the insulating phase, but due to the divergence of the interacting localization length at the transition, it introduces severe difficulties for examining the system near the putative transition and in the metallic phase. Other studies have examined the dynamical nature of the transition using time-dependent density matrix renormalization group (tDMRG) or similar methods [@Znidaric2008; @Bardarson2012]. However, these studies are restricted to the localized phase due to the growth of entanglement entropy. While the existence of an insulating phase for some range of parameters and energies appears to be quite well established even at the rigorous level [@Aizenman2009b], this is not the case for the metallic phase, where there are currently no rigorous results and the existing numerical schemes are quite limited. In this letter, we examine the dynamics of an isolated one-dimensional system across the putative MBL transition predicted by exact diagonalization studies of the same system. In particular, we study the relaxation dynamics of our system starting from a far-from-equilibrium initial condition that is a pure state of the corresponding non-interacting system. We follow the diagrammatic approach of Ref. , while relaxing several assumptions used in that work. Unlike the work of Ref. , we compute in detail the dynamics of the system from an appropriately chosen initial condition. By doing so, we are able to assess the accuracy of the approximations used in Ref.  against exact numerical results (where available) as well as to describe how the MBL transition should manifest within the very framework that first predicted its existence. Following previous studies [@Znidaric2008; @Karahalios2009a; @Barisic2010a; @Berkelbach2010a; @Monthus2010a; @Pal2010a], we investigate a one-dimensional system of spinless and interacting fermions in a disordered potential, $$\begin{aligned} H & = & -t\sum_{i}\left(\hat{c}_{i}^{\dagger}\hat{c}_{i+1}+\hat{c}_{i+1}^{\dagger}\hat{c}_{i}\right)\label{eq:t-V_model}\\ & + & V\sum_{i}\left(\hat{n}_{i}-\frac{1}{2}\right)\left(\hat{n}_{i+1}-\frac{1}{2}\right)+\sum_{i}h_{i}\left(\hat{n}_{i}-\frac{1}{2}\right),\nonumber \end{aligned}$$ where $t$ is the hopping matrix element, $V$ is the interaction strength and $h_{i}$ are random fields independently distributed on the interval $h_{i}\in\left[-W,W\right]$. Note that by using the Jordan-Wigner transformation, this model can be exactly mapped onto the XXZ model. There are only two independent parameters in the Hamiltonian and we therefore choose $t=1$. Since this model has a bounded energy density $\left|\left\langle H\right\rangle \right|/L\leq\left(2t+V/4+W/2\right)$ (where $L$ is the length of the system), critical parameters may be found such that the system will transition from a mixture of “insulating” and “metallic” states to a situation where *all* of the many-body eigenstates are “insulating.” The existence of such parameters is equivalent to the assumption used in Ref.  that the MBL transition will survive at infinite temperature. Under these conditions, the critical disorder strength has been determined in previous studies for rather strong interaction $V=2t$, to be about $W_{c}/t\approx7-8$ [@Berkelbach2010a; @Pal2010a]. Since our approach is based on many-body perturbation theory, we are limited to small interaction strengths. Therefore, to find the relevant critical parameters, we extend the calculations of Ref.  to evaluate the critical *line* in the space of $W/t$ and $V/t$ using exact diagonalization (see Fig. \[fig:params\]). We also estimate the theoretical dynamical phase boundaries by using the critical temperatures and a thermodynamic relation [@Basko2006a; @Aleiner2013] $$T_{el}=\frac{t^{2}}{\nu\xi V^{2}},\qquad T_{c}=\frac{t}{12\nu\xi V\ln t/V},\qquad\frac{E}{L}=\frac{\pi^{2}}{6}\nu T^{2}.\label{eq:BAA0}$$ Here $\xi$ is the non-interacting localization length, $T_{el}$ is the temperature which separates ergodic and non-ergodic metals (see e.g., [@Altshuler2010]), $T_{c}$ is the transition temperature and $\nu$ is one-particle density of states. Although for high temperatures and for a system with a bounded energy density the thermodynamic relation is not strictly valid, we use it to extrapolate from low to high temperatures by setting $E_{c}/L$ to be equal to half of the energy band $\Delta\thickapprox\left(2t+W/2\right)$ (the interaction contribution is negligible) [@Aleiner2013]. The critical interaction can now be obtained in terms of the non-interacting localization length, $\xi\left(W\right)$ and $W$, **$$\frac{V_{el}}{t}=\frac{1}{\sqrt{\alpha}},\qquad\frac{V_{c}}{t}=\frac{1}{12\alpha\left(\ln12\alpha\left(\ln\left(12\alpha\cdots\right)\right)\right)},\label{eq:BAA}$$** where $\alpha=\sqrt{6\nu\Delta\xi^{2}/\left(\pi^{2}t\right)}$. The theoretical lines are valid for $V_{c}/t<1$ and $\xi>1,$ and are plotted at Fig. \[fig:params\]. It is clearly seen that the numerical critical line is one order of magnitude higher than the theoretical one, lying predominately in the non-ergodic metal phase. We will show, however, that at least for small interaction strengths the numerical critical line suffers from severe finite size effects. Taking this into account suggests that the line should move towards higher values of $V$, which will drive it even further away from the phase boundary where theory predicts a stable insulator. One possible explanation of this discrepancy could be that the the insulating phase is stable up to $T_{el}$ and not $T_{c}$ as predicted by the theory. In other words, the non-ergodic metal phase is also insulating. Alternatively, the approach used to determine the numerical critical line [@Pal2010a] might be sensitive to an ergodic non-ergodic transition within the metal phase and not the true metal-insulator transition.  Dynamical phase boundaries of the system as a function of $W/t$ and $V/t$. Gray solid line indicates the numerical critical values calculated using the method of Ref.  ($N=12$, see their Fig. 6). The red star represents the transition point obtained in Ref.  (note the factor of 2 difference in $W_{c}$). The blue crosses represent the parameters used in this work. Vertical cut: $W=3.5$ and $V=0.1,\,0.2\,0.3,\,0.4$. Horizontal cut: $W=1,3,5$ and $V=0.25$. The solid black line is the solution of Eq. , and separates metal from insulator. The dotted black line is obtained from Eq. , and demarcates the boundary between ergodic and non-ergodic metals. The dashed lines are extrapolations to regions below which such equations are *not* valid.](Fig1_new.pdf){width="8.6cm"} We next outline our dynamic scheme which is similar in spirit to Ref. , but relaxes several approximations of that work. We start with the one-particle greater and lesser non-equilibrium Green’s functions, $$\begin{aligned} G_{ij}^{>}\left(t;t'\right) & = & -i\tr\left\{ \hat{\rho}_{0}\hat{c}_{i}\left(t\right)\hat{c}_{j}^{\dagger}\left(t'\right)\right\} \label{eq:gtr_less_G}\\ G_{ij}^{<}\left(t;t'\right) & = & i\tr\left\{ \hat{\rho}_{0}\hat{c}_{j}^{\dagger}\left(t'\right)\hat{c}_{i}\left(t\right)\right\} ,\nonumber \end{aligned}$$ where $\hat{\rho}_{0}$ is the initial density matrix. For a non-interacting initial density matrix, the Green’s functions obey the Kadanoff–Baym equations of motion [@Kadanoff1994], $$\begin{aligned} i\partial_{t}G^{\gtrless}\left(t,t'\right) & = & \left(\hat{h}_{0}+\Sigma^{HF}\left(t\right)\right)G^{\gtrless}\left(t,t'\right)\nonumber \\ & + & \int_{0}^{t}\Sigma^{R}\left(t,t_{2}\right)G^{\gtrless}\left(t_{2},t'\right)\mathrm{d}t_{2}\nonumber \\ & + & \int_{0}^{t'}\Sigma^{\gtrless}\left(t,t_{2}\right)G^{A}\left(t_{2},t'\right)\mathrm{d}t_{2},\label{eq:KB_eq}\end{aligned}$$ where spatial indices and summations are suppressed for clarity; $\hat{h}_{0,nm}=-t\left(\delta_{n,m+1}+\delta_{n,m-1}\right)+h_{n}\delta_{nm}$ is the one particle Hamiltonian; $\Sigma^{HF}\left(t\right)$, $\Sigma^{\gtrless}\left(t\right)$ are the Hartree-Fock greater and lesser self-energies of the problem respectively; and the superscripts ’R’ and ’A’ represent retarded and advanced Green’s functions and self-energies, which are defined as $$\begin{aligned} \Sigma^{R}\left(t,t_{2}\right) & = & \theta\left(t-t_{2}\right)\left(\Sigma^{>}\left(t,t_{2}\right)-\Sigma^{<}\left(t,t_{2}\right)\right)\\ G^{A}\left(t_{2},t'\right) & = & -\theta\left(t'-t_{2}\right)\left(G^{>}\left(t_{2},t'\right)-G^{<}\left(t_{2},t'\right)\right).\nonumber \end{aligned}$$ Note that due to the complexity of (\[eq:KB\_eq\]), what was actually considered in Ref.  is the corresponding quantum Boltzmann equation. For this purpose the authors have neglected the off-diagonal spatial elements of the Green’s functions, and performed a gradient expansion of the time variable. Additionally, the real part of the self-energy as well as the Hartree-Fock contributions have been neglected [@Basko2006a]. Although it is numerically feasible to solve (\[eq:KB\_eq\]) within the specified approximation for the self-energies (see below), this approach is turns unstable for sufficiently long times of any parameters of the Hamiltonian. To eliminate this spurious behavior we reduce (\[eq:KB\_eq\]) to a quantum master equation for the one-particle density matrix by introduction of the generalized Kadanoff-Baym anzatz, $$G^{\lessgtr}\left(t,t'\right)=i\left[G^{R}\left(t,t'\right)G^{\lessgtr}\left(t',t'\right)-G^{\lessgtr}\left(t,t\right)G^{A}\left(t,t'\right)\right],\label{eq:GKBA}$$ and by approximating the retarded and advanced Green’s functions with their Hartree-Fock (HF) values [@Spicka2005; @Spicka2005a; @Latini2013]. This approach is in the spirit of the Boltzmann approach of Ref. , and is ostensibly more precise, since it considers the full density matrix and not just its diagonal values. Additionally, the quasi-classical approximation as well as gradient expansions are not needed, which allows us to describe systems far from equilibrium. It is not our purpose to examine the validity of this approximation in this letter and the reader is referred to Refs. . We leave the discussion of the full solution of (\[eq:KB\_eq\]) to a future study. As in Ref.  we utilize both the HF and the self-consistent second-Born (2B, called SCBA in [@Basko2006a]) approximations for the self energies, $$\begin{aligned} \Sigma_{ij}^{HF}\left(t\right) & = & -i\delta_{ij}\sum_{k}V_{ik}G_{kk}^{<}\left(t;t\right)+iV_{ij}G_{ij}^{<}\left(t;t\right)\nonumber \\ \Sigma_{ij}^{>}\left(t,t'\right) & = & \sum_{k,l}V_{il}V_{jk}G_{kl}^{<}\left(t',t\right)\times\label{eq:self-energies}\\ & & \left[G_{lk}^{>}\left(t,t'\right)G_{ij}^{>}\left(t,t'\right)-G_{lj}^{>}\left(t,t'\right)G_{ik}^{>}\left(t,t'\right)\right],\nonumber \end{aligned}$$ where $V_{ij}=V\left(\delta_{i,j+1}+\delta_{i,j-1}\right)$ is the interaction potential. The fact that the insulating phase is non-ergodic means that great care must be exercised with regards to the choice of the initial conditions [@Aleiner2013; @Basko2006a]. To illustrate the issues involved, consider a disconnected interacting system, namely by setting $t=0$. This system can be solved exactly, given the fact that the Hamiltonian is already diagonal in the position basis. A similar model was studied by several authors in the context of the MBL, but for distinct purposes, see Refs. . The Green’s function is a periodic function of time with period $V$. Ignoring the fact that the system is solvable and utilizing the diagrammatic perturbation theory up to second order in $V$ (as in the 2B approximation) gives, $$\begin{aligned} \Sigma_{i}^{<}\left(t\right) & = & i\left\langle \hat{n}_{i}\right\rangle e^{-i\varepsilon_{i}t}\sum_{j}V_{i,j}\left\langle \hat{n}_{j}\right\rangle \left\langle 1-\hat{n}_{j}\right\rangle ,\label{eq:self-energy-disconn}\end{aligned}$$ where $\left\langle \hat{n}_{j}\right\rangle =\tr\hat{\rho}_{0}\hat{n}_{j}$. Notably, the imaginary part of the self-energy yields an unphysical decay of the Green’s function, and only vanishes if $\left\langle \hat{n}_{i}\right\rangle =n_{i}=0,1$ (which implies a density matrix which is one Slater determinant localized on lattice sites). It can be shown that with this form of initial density matrix the problem may be solved using only the Hartree term. Thus, if we would like to recover the proper $t\to0$ limit in the framework of 2B, any selected initial state should have the properties disclosed above in this limit. To satisfy this restriction, we use the following initial condition: $$G_{ij}^{<}\left(0,0\right)=i\sum_{k}\phi_{k}\left(i\right)\phi_{k}\left(j\right)n_{k}^{0},\label{eq:init_cond}$$ where $n_{k}^{0}\in\left(0,1\right)$, and $\phi_{k}\left(i\right)$ are the non-interacting one-particle eigenstates. We consider half-filling throughout this letter, although other fillings have been investigated and do not lead to distinct behavior. For a MBL transition at finite temperature, the mobility edge is found within the many-body spectrum, and therefore the average energy as well as the energetic width of the initial condition are of great importance, since they determine the position of the state with respect to the many-body mobility edge. For a MBL transition at infinite temperature, for certain parameters, *all* eigenstates become “insulating,” and therefore the initial condition should not be important. Nevertheless, to eliminate one control parameter, we set the location of the mean energy density of the initial condition to be in the middle of the many-body energy band, namely $\left\langle H\right\rangle =0$. This closely corresponds to the $T\to\infty$ limit invoked in Ref. , and allows one to establish the onset of the transition when the mobility edge converges to zero. The relaxation of the system is monitored by calculating the correlation function $$\delta\rho\left(t\right)=\frac{1}{L}\sum_{k}\overline{\left\langle \left(\hat{n}_{k}\left(t\right)-\frac{1}{2}\right)\left(\hat{n}_{k}\left(0\right)-\frac{1}{2}\right)\right\rangle },\label{eq:measure}$$ where $\hat{n}_{k}=\sum_{ij}\phi_{k}\left(i\right)\phi_{k}\left(j\right)\hat{c}_{i}^{\dagger}\hat{c}_{j}$, and the over-line indicates averaging over disorder realizations. Generically a two-particle Green’s function is needed to calculate this quantity, however the chosen initial condition renders $\left\langle \hat{n}_{k}\left(t\right)\hat{n}_{k}\left(0\right)\right\rangle =\left\langle \hat{n}_{k}\left(t\right)\right\rangle \left\langle \hat{n}_{k}\left(0\right)\right\rangle $. Since the total number of particles is conserved, this correlation function measures the diffusion rate of the one-particle energy, and for an initial conditions in the ergodic phase it will typically decay as $t^{-1/2}$ or faster as a function of time (see, e.g., Ref.  for studies of clean systems). For initial conditions in a non-ergodic phase temporal decay will cease after some finite time, or alternatively, a sub-diffusive relaxation is expected. Note that although a non-decaying correlation indicates that some of the particles are pinned to their initial positions, it does not preclude a finite mobility for the rest of the particles, and therefore does not rule out metallic behavior. Nevertheless, in this case the mobility of the particles is expected to be considerably impaired with comparison to the ergodic case. It is precisely this non-ergodic conducting situation which defines the non-ergodic metal region of the dynamic phase diagram of Fig. \[fig:params\]. We solve the quantum master equation (QME) numerically for the one-particle Green’s functions (\[eq:KB\_eq\]) as a function of time, with the initial conditions (\[eq:init\_cond\]) and the self-energies (\[eq:self-energies\]). For this purpose we use the numerical method developed in [@Stan2009]. The correlation function is averaged over 256 realizations of the disordered potential. In Fig. \[fig:exact-2B\]a we compare the perturbative calculation to the exact solution of a small chain, $L=12$, obtained by exact diagonalization (ED). A remarkable correspondence between the QME and the exact solutions is seen for $V=0.25$ and times $t\lesssim40$. This correspondence becomes better for larger $W$ and does not exist at the HF level, which produces only non-decaying solutions (not shown). For longer times there is only a qualitative correspondence between the QME and the exact solution. It should also be noted that for smaller values of $W$ such as $W=1$ (where the non-interacting localization length, $\xi=25$, is larger than the system size) the $L=12$ system is simply a model for testing the approximation scheme and has little bearing on the dynamics of MBL in the thermodynamic limit, although surprisingly the dynamics is not very different from that at much larger disorder strengths.  **(a)** Density-density correlation function as a function of time for averaged over $256$ disorder realizations $\left(V=0.25\right)$. The solid black lines designate the exact solution calculated using ED $\left(L=12\right)$ and red (gray) lines designate the solution using QME. **(b)** Finite size behavior for same parameters using QME. The system sizes used: $W=0,$ $N=2^{14}$ (black dashed); $W=1,$ $N=12,48,96$ (solid, dashed, dot dashed); $W=3,$ $N=12,28,36$ (solid, dashed, dot dashed) and $W=5,$ $N=12,28$ (solid, dashed).](Fig2_new.pdf){width="8.6cm"}  **(a)** Modified density-density correlation $\left(\delta\bar{\rho}\left(t\right)\equiv1-4\delta\rho\left(t\right)\right)$ as a function of $V\cdot t$ , averaged over $256$ disorder realizations, for $W=3.5,\, L=36$. **(b)** Left panel rescaled by plotting $V^{-0.78}\delta\bar{\rho}$ as a function of $V\cdot t$ (see text).](Fig3_new.pdf) From Fig. \[fig:exact-2B\]b we observe that after the initial relaxation there is a slow decay of the correlation function. The decay of the clean system is exponential with a time scale of $t_{1}\sim V^{-2}$, which indicates that the QME within the 2B approximation *overestimates* the relaxation in the clean (and presumably also in the nearly metallic) region of the phase diagram [@Fabricius1998]. To eliminate finite size effects we have increased the size of the system until no changes are observed in our measure. Note that, up to the considered time, $L=12$ suffices only for $W>3$, which suggests that the numerically obtained critical line of Fig. \[fig:params\] is prone to large finite size effects, at least for the considered interaction strengths. Thus, if that line exists in the $L\to\infty$ limit, it should move toward higher values of $V$ in the $V-W$ plane. Although at the MBL transition we expect to observe a steep change in the functional dependence of the correlation functions, the behavior of density fluctuations appear surprisingly smooth over a broad range of parameters. To demonstrate this, we fix $W$ and cross the numerically determined critical line by changing $V$ (see Fig. \[fig:params\]). The correlation function can be reasonably well rescaled by setting $\delta\bar{\rho}\left(t\right)\equiv1-4\delta\rho\left(t\right)$, while plotting $V^{-0.78}\delta\bar{\rho}\left(t\right)$ as a function of rescaled time, $V\cdot t$ (see Fig. \[fig:rescaling\]). The exponent was obtained by fitting. We could not find satisfactory rescaling for fixed $V$. The initial fast relaxation time may be inferred from Fig. \[fig:rescaling\], $t_{1}\sim6V^{-1}$. During this time the system dephases across the exact many-body states, which span the initial state. The scaling suggests that, $\delta\rho\left(t\right)=\frac{1}{4}-AV^{0.78}f\left(Vt\right)$, and therefore assuming its asymptotic validity it should decay to zero at the ergodic-non-ergodic transition $V=V_{*}$. This yields the form $\delta\rho\left(V,t\to\infty\right)=\frac{1}{4}\left(1-\left(V/V_{*}\right)^{0.78}\right)$. In particular, the smooth character of the temporal behavior suggests that $V_{*}>0.4$. *Thus our numerics suggest that all correlation functions in the region of investigation asymptotically decay to a finite value corresponding to a non-ergodic state.* To summarize, this work presents, for the first time, a detailed description of the dynamical phase diagram and of how dynamical quantities manifest in the MBL scenario. Although our approach is approximate, it is based on the same approximations that first predicted the existence of MBL [@Basko2006a], and thus should provide important qualitative guidelines to the long-time behavior that is out of reach by more rigorous methods. We find that for all values of interaction and disorder strengths studied, density fluctuations decay in a remarkably slow, non-exponential manner. Furthermore, the dynamics do not qualitatively change in a broad region of parameter space, and are consistent with a non-ergodic phase. This is rather surprising, given that our scan of parameters takes the system across the transition line of Ref.  and in the vicinity of the ergodic metal region proposed in Ref.  . Our results, however, may be viewed as consistent with both works, if the phase boundary of Ref. , which based on our finite size analysis is expected to drift upward in the thermodynamic limit, demarcates an “ergodic-non-ergodic” transition where the non-ergodic phase is actually a non-ergodic metal. To resolve these questions would require calculation of the conductivity in addition to the density fluctuations. This is much more difficult to do within the approach described here. Research along these lines will be presented in a future publication. We would like to thank I. Aleiner, O. Agam, D. Huse and M. Schiro for many enlightening and helpful discussions. This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number OCI-1053575. This work was supported by the Fulbright foundation and by grant NSF-CHE-1213247.

ArXiv

--- abstract: 'We study supersymmetry breaking perturbations of the simplest dual pair of 2+1-dimensional ${\cal N}$=2 supersymmetric field theories – the free chiral multiplet and ${\cal N} = 2$ super-QED with a single flavor. We find dual descriptions of a phase diagram containing four distinct massive phases. The equivalence of the intervening critical theories gives rise to several non-supersymmetric avatars of mirror symmetry: we find dualities relating scalar QED to a free fermion and Wilson-Fisher theories to both scalar and fermionic QED. Thus, mirror symmetry can be viewed as the multicritical parent duality from which these non-supersymmetric dualities directly descend.' author: - Shamit Kachru - Michael Mulligan - Gonzalo Torroba - Huajia Wang bibliography: - 'QHE.bib' title: Nonsupersymmetric dualities from mirror symmetry --- Introduction {#sec:intro} ============ Duality plays a central role in the modern understanding of quantum field theory. In some cases, as with S-duality of maximally supersymmetric four-dimensional (4d) Yang-Mills theory, it refers to an exact symmetry exchanging strong and weak coupling limits of the same theory. In others, as with the dualities of ${\cal N}=1$ supersymmetric field theories in four dimensions [@Seibergduality] or mirror symmetry of 3d supersymmetric gauge theories [@Intriligator:1996ex; @Aharony; @Kapustin:1999ha], it relates the low-energy physics arising from two distinct high energy theories. Dualities have found diverse applications in high energy physics, condensed matter physics, and mathematics. An important way to deepen our understanding of duality is to relate one duality to another. This simplifies the logical structure of the assumptions we must make, yielding a web of dualities from a single starting point, and can also allow us to derive new dualities. Motivated in part by the proposal [@Son2015] for a duality governing the physics of the half-filled Landau level, there have been several recent discussions for dualities relating some of the simplest non-supersymmetric 3d field theories [@Aharony2016; @Karch:2016sxi; @Seiberg:2016gmd; @2016arXiv160601912M; @HsinSeiberg2016; @awesome] (with closely related earlier work appearing in [@GMPTWY2012; @AharonyGurAriYacoby2012; @AharonyGurAriYacobysecond2012; @WangSenthilfirst2015; @MetlitskiVishwanath2016; @Kachru:2015rma; @Geraedtsetal2015; @XuYou2015selfdual; @MrossAliceaMotrunichexplicitderivation2016; @MrossAliceaMotrunichbosonicph2016]). In this note, we show that many of these dualities can be derived as a consequence of the most basic avatar of mirror symmetry of 3d ${\cal N}=2$ supersymmetric gauge theories.[^1] We build on our work [@awesome] to show that, after breaking supersymmetry, the duality between the theory of a free chiral multiplet – theory A – and supersymmetric quantum electrodynamics (SQED) with a single flavor – theory B – yields a rich phase diagram with four distinct phases. Duality relates both the phases and the intervening critical points of the dual pairs. Along the four walls separating distinct phases (see Figs. \[fig:thA\] and \[fig:thB\]), we find critical theories with dual descriptions realized in the A and B pictures. This unifies the simplest duality of 3d supersymmetric field theory with various dualities relating fermionic and scalar QED to theories of free fermions or Wilson-Fisher bosons. It provides a logical completion of [@awesome], where the duality between phases I and II (and the intervening critical point) was already derived. The dualities studied here are of interest both for their intrinsic importance in understanding the structure of 3d quantum field theory, and for potential applications to problems in condensed matter physics including the study of topological order and metallic criticality. Chiral mirror symmetry {#sec:chiral} ====================== We first review the essential properties of the chiral mirror symmetry duality [@awesome; @Tong:2000ky]. Theory A consists of a free chiral superfield $(v, \Psi)$ which contains a complex scalar $v$ and its (two-component) Dirac fermion superpartner $\Psi$. The theory enjoys two global abelian symmetries, $U(1)_{J}$ and $U(1)_R$, whose actions on $(v, \Psi)$ are given below. \[eq:chiraltab1A\] $U(1)_J$ $U(1)_R$ ----------------- ---------- ---------- \[-8pt\] $v$ 1 1 \[-8pt\] $\Psi$ 1 0 Introducing background gauge fields $\hat{A}_{J,R}$ associated to the global $U(1)_{J, R}$ symmetries, the theory A Lagrangian, $$\begin{aligned} \label{eq:LA-backgrounds} {\mathcal}L^{(A)} & = |D_{\hat{A}_J + \hat{A}_R} v|^2 - m^2_v |v|^2 \cr & + \bar \Psi i \slashed{D}_{\hat{A}_J} \Psi - m_\Psi \bar \Psi \Psi - {1 \over 8 \pi} \hat{A}_J d \hat{A}_J\,.\end{aligned}$$ For abelian gauge fields $A$ and $B$, the covariant derivatives $D_{\pm A} \equiv \partial_\mu \mp i A_\mu$ with $\mu \in \{0,1,2\}$; $\slashed{D}_B \equiv \gamma^\mu (\partial_\mu - B_\mu)$ and $\bar{\Psi} \equiv \Psi^\dagger \gamma^0$ with $\gamma$-matrices[^2] satisfying $\{\gamma^\mu, \gamma^\nu \} = 2 \eta^{\mu \nu}$. Chern-Simons (CS) terms are written as $A d B \equiv \epsilon^{\mu \nu \rho} A_\mu \partial_\nu B_\rho$ with $\epsilon^{012} = 1$.[^3] Besides the relevant bosonic and fermionic mass terms, theory A admits a relevant perturbation proportional to $|v|^4$ (which will play a key role below), and a classically marginal interaction $|v|^2\bar\Psi \Psi $. The $\hat{A}_{J}$ gauge field can be included in a background vector multiplet ${\cal{\hat{V}}}_J = (\hat{A}_{J}, \hat{\sigma}_{J}, \hat{\lambda}_{J}, \hat{D}_{J})$ with the couplings occurring in dictated by unbroken ${\cal N}=2$ supersymmetry: \[theoryAmasses\] m\^2\_v = \_J\^2 + \_J, m\_= \_J. Chiral mirror symmetry allows us to map all components in this multiplet across the duality. The scalar $\hat \sigma_J$ and D-term $\hat D_J$, in particular, play important roles in our derivation. Theory B is ${\cal N}=2$ SQED with a single chiral flavor $(u, \psi)$ and $U(1)_a$ gauge group. The corresponding charge assignments are given below. \[eq:thBchargescartanchiral\] $U(1)_J$ $U(1)_R$ $U(1)_a$ ---------------------------------- ---------- ---------- ---------- \[-8pt\] $u$ 0 0 -1 \[-8pt\] $\psi$ 0 -1 -1 \[-8pt\] $e^{2\pi i \gamma/g^2}$ 1 0 0 \[-8pt\] $\sigma$ 0 0 0 \[-8pt\] $\lambda$ 0 -1 0 $\sigma$ is the real scalar partner of the dynamical 3d gauge field $a_\mu$, $\lambda$ is the gaugino[^4], and $\gamma$ is the dual photon ($\partial^\mu \gamma = {1 \over 2 \pi} \epsilon^{\mu \nu \rho} \partial_\nu a_\rho$). Including the background $U(1)_{J,R}$ terms considered previously, the dual theory B Lagrangian: $$\begin{aligned} {\mathcal}L^{(B)}= {\mathcal}L_V + {\mathcal}L_\text{matter} + {\mathcal}L_{CS} - {\mathcal}L_{BF}\end{aligned}$$ with \[lagrangiancomponents\] L\_V&= & (- f\_\^2 + ( )\^2+ |i \_[- A\_R]{} + D\^2 ),\ L\_& = & |D\_[-a]{} u|\^2 + |i \_[-a - A\_R]{} - ( \^2 - D ) |u|\^2\ & + &|+ u\^\* | + u | ,\ L\_[CS]{}&= & (a da+2 D + |),\ L\_[BF]{}&= & [1 2 ]{} (a d A\_J+ D\_J + \_J D )+ [1 4 ]{} a d A\_R .$f_{\mu \nu} = \partial_\mu a_\nu - \partial_\nu a_\mu$ is the field strength of the $U(1)_a$ gauge field. Chiral mirror symmetry says that the IR ($g \rightarrow \infty$) limit of theory B has the free field description given by theory A. In order to understand the RG flows below, let us now discuss the map of relevant deformations. $\hat \sigma_J$ and $\hat D_J$ give rise to masses in theory A via (\[theoryAmasses\]). These parameters are external backgrounds for the $U(1)_J$ supercurrent, and are mapped exactly to theory B according to the last line in (\[lagrangiancomponents\]). If we further restrict to high energies, theory B is weakly coupled and $\sigma$ and $D$ may be integrated out to show that $\hat \sigma_J$ and $\hat D_J$ produce masses for $u$ and $\psi$. The other renormalizable deformations of theory A, $|v|^4$ and $|v|^2 \bar \Psi \Psi$, do not correspond to conserved currents, and hence their effect in theory B is more involved. Nevertheless, we may derive an approximate correspondence by noting that $v^* \Psi \sim \lambda$ from the quantum numbers in both theories, and $|v|^2 \sim \sigma$ from the couplings of $\hat D_J$ on both sides. These also agree with the Taub-NUT map of the underlying ${\mathcal}N=4$ theory [@SachdevYin; @Hook:2014dfa]. Therefore, $|v|^4$ and $|v|^2 \bar \Psi \Psi$ map to masses for $\sigma$ and the gaugino in theory B. We stress that this map is approximate and is expected to receive large quantum corrections in the IR. Dynamics and phases of theory A {#sec:thA} =============================== Theory A has four distinct phases parameterized by the signs of the effective masses $m_v^2$ and $m_\Psi$ – see Fig. \[fig:thA\]. ![Phase diagram of theory A. Massless fields occur along the red lines (second order phase transitions).[]{data-label="fig:thA"}](thA "fig:"){width=".9\linewidth"}\ These gapped phases can be invariantly characterized by their responses to the background $\hat{A}_{J,R}$ fields.[^5] Integrating out the massive degrees of freedom according to the prescription in Fig. \[fig:thA\], this response is captured by the effective Lagrangians: $$\begin{aligned} \label{effectiveresponse} {\cal L}^{\rm I} & = - {1 \over 4 \pi} \hat{A}_J d \hat{A}_J, \cr {\cal L}^{\rm II} & = 0, \cr {\cal L}^{\rm III} & = - {1 \over 2 \pi} b d (\hat{A}_J + \hat{A}_R), \cr {\cal L}^{\rm IV} & = - {1 \over 4 \pi} \hat{A}_J d \hat{A}_J - {1 \over 2 \pi} b d (\hat{A}_J + \hat{A}_R).\end{aligned}$$ For simplicity, we have truncated the effective Lagrangians at the leading quadratic order in the derivative expansion. In ${\cal L}^{{\rm III}}$ and ${\cal L}^{{\rm IV}}$, we introduced the 3d gauge field $b$ [@MaldacenaMooreSeiberg2001] whose equation of motion imposes the symmetry-breaking constraint, $$\begin{aligned} \hat{A}_J + \hat{A}_R = 0,\end{aligned}$$ that occurs when $m^2_v < 0$. Phases I and II, along with the intervening critical point, were studied in [@awesome]. To reliably study phases III and IV, where the effective mass-squared $m_v^2 < 0$, we introduce a stabilizing interaction $|v|^4$. Even though this interaction breaks supersymmetry, it does not qualitatively affect the analysis or conclusions for the phase structure when $m_v^2 > 0$. It does modify the precise location at which $m_v^2 = 0$: the classical location $\hat{\sigma}_J^2 + \hat{D}_J = 0$ – see – is modified quantum mechanically. To add the $|v|^4$ interaction in a way that can be tracked across the duality, we promote $\hat{D}_J \rightarrow D_J$ to a dynamical field and integrate it out with Gaussian weight ${\cal L}_{D_J} = {1 \over 2 h^2} (\hat{D}_J - \hat{D}_0)^2$. (This technique was used in [@Gur-AriYacoby2015] in a different context.) The resulting classical scalar potential, V\_= (\_J\^2 + D\_0) |v|\^2 + |v|\^4, features a stable vacuum at $|v|^2= (|\hat D_0| - \hat \sigma_J^2)/h^2$ when $\hat D_0<- \hat \sigma_J^2$. The phase diagram in Fig. \[fig:thA\] contains four lines of phase transitions: these transitions are described either by critical Wilson-Fisher theories with interaction strength determined by $h^2$ or by the mass sign-changing transition of a free Dirac fermion. Dynamics and phases of theory B {#sec:thB} =============================== Direct analysis of theory B for all values of the background parameters is subtle because the theory is strongly coupled. Duality and supersymmetry (when unbroken), however, may together be used to determine the theory B dynamics. We have found it possible to uniquely realize the theory A response summarized in using the effective mass parameters $m_u^2$ and $m_{f_\pm}$ for the charged degrees of freedom of theory B with the prescription given in Fig. \[fig:thB\]. ![Phase diagram of theory B.[]{data-label="fig:thB"}](thB "fig:"){width=".9\linewidth"}\ We have introduced mass parameters for Dirac fermions $f_{\pm}$ that diagonalize the effective fermion mass matrix in $$\begin{aligned} \label{massmatrix} {\cal L}^{(B)}_{\rm mass} = - m_\psi \bar{\psi} \psi - m_\lambda \bar{\lambda} \lambda + \delta m^\ast\, \bar{\lambda} \psi + \delta m \,\bar{\psi} \lambda.\end{aligned}$$ The third and fourth terms – see the second line of ${\cal L}_{\rm matter}$ – are parameterized by the fermion mass mixing $\delta m$ which is only non-zero in phases when $\langle |u^2| \rangle \neq 0$, i.e., $m_u^2 < 0$. The mass matrix in has the two eigenvalues, $$\begin{aligned} \label{masseigenvalues} m_{f_{\pm}} = {1 \over 2} \Big(m_\psi + m_\lambda \pm \sqrt{(m_\psi - m_\lambda)^2 + 4 |\delta m|^2} \Big).\end{aligned}$$ $f_\pm$ are composed of linear combinations of $\psi$ and $\lambda$ with relative weights that vary as the background parameters are tuned: Fig. \[fig:thB\] implies that upon transitioning from phase II to phase I, $f_+$ becomes $\psi$ and $f_-$ becomes $\lambda$ and vice versa in the transition from phase II to phase III. The matching of the theory B responses to those of theory A in follows upon integrating out the massive degrees of freedom, $$\begin{aligned} \label{theoryBeffresponse} {\cal L}_{\rm eff}^{(B)} & = {1 \over 8 \pi} {\rm sgn}(m_{f_+}) (a + \hat{A}_R) d (a + \hat{A}_R) + {1 \over 8 \pi} a d a \cr & + {1 \over 8 \pi} {\rm sgn}(m_{f_-}) \hat{A}_R d \hat{A}_R - {1 \over 4 \pi} a d \hat{A}_R - {1 \over 2 \pi} \hat{A}_J d a.\cr\end{aligned}$$ This formula should be read with the replacement $f_+ = \psi$ and $f_- = \lambda$ in phases where $m^2_u>0$. In phases I, III, and IV where $m_u^2 > 0$, the response in is seen to directly match that in for the values of the effective mass parameters given in Fig. \[fig:thB\]. In phase I, the CS $a da$ coefficient is nonzero, and integrating out $a$ matches the result from theory A. On the other hand, in phases III and IV, this coefficient vanishes and $a$ is identified with the gauge field $b$ introduced in (\[effectiveresponse\]); this reproduces the correct symmetry breaking pattern and CS response. In phase II, the condensation $\langle |u|^2 \rangle \neq 0$ essentially fixes $a =0$ and we find agreement with ${\cal L}^{\rm II}$ in after using the effective mass values in Fig. \[fig:thB\]. The response in phase II requires $m_{f_+} m_{f_-} < 0$ or $$\begin{aligned} m_\psi m_\lambda < |\delta m^2|\,.\end{aligned}$$ Happily, this requirement is consistent with the values of the effective masses in the adjacent phases I and III. To provide additional justification for this picture, we analyze theory B near the line $\hat{D}_0 = 0$ where supersymmetry is preserved. Because supersymmetry precludes phase transitions as a function of the coupling $g$, we may transfer the qualitative information gleaned at weak coupling to the strong coupling regime of interest. Note that we are ignoring the effects of the dynamical $D_J$ field here; it provides a stabilizing potential for $\sigma$, but can otherwise be ignored near the origin of Fig. \[fig:thB\]. We first verify the effective mass parameter assignments in phases I and II along $\hat{D}_0 = 0$. In [@awesome], we demonstrated that the supersymmetry-preserving vacuum for the scalar fields lies at $$\begin{aligned} \label{susyvacua} \langle |u|^2 \rangle & = 0, & \langle \sigma \rangle & = \hat{\sigma}_J, & {\rm for}\ & \hat{\sigma}_J < 0,\cr \langle |u|^2 \rangle & = {\hat{\sigma}_J \over 2 \pi}, & \langle \sigma \rangle & = 0, & {\rm for}\ & \hat{\sigma}_J > 0. \end{aligned}$$ From this analysis we abstract the following. For $\hat{\sigma}_J < 0$, we determine the effective masses $m^2_u > 0$, $m_\psi = - \langle \sigma \rangle > 0$, and $m_\lambda < 0$ from ${\cal L}_{\rm CS}$ in . For $\hat{\sigma}_J > 0$, we find that $m_u^2 < 0$, $m_\psi = 0$, and $m_\lambda < 0$. Using the mass matrix eigenvalues in , we find agreement with the inferred values in Fig. \[fig:thB\]. Perturbation theory in $\hat{D}_0 < 0$ is likewise consistent with the assignment of effective masses. To leading order in $\hat{D}_0$, the vacua in are shifted as follows: $$\begin{aligned} \label{perturbationresults} \delta \langle u \rangle & = 0, & \delta \langle \sigma \rangle & \sim g^2 |\hat{D}_0| \hat{\sigma}_J^{-2}, & {\rm for}\ & \hat{\sigma}_J < 0,\cr \delta \langle u \rangle & \sim - |\hat{D}_0| \hat{\sigma}_J^{-3/2}, & \delta \langle \sigma \rangle & \sim |\hat{D}_0| \hat{\sigma}_J^{-1}, & {\rm for}\ & \hat{\sigma}_J > 0,\end{aligned}$$ where $\sim$ indicates equality up to multiplication by a positive constant. For $\hat{\sigma}_J < 0$, we see that $m_\psi$ is decreased to leading order, consistent with the putative fermion mass sign-changing transition at $\hat{D}_0 = - {\rm sgn}(\hat{\sigma}_J) \hat{\sigma}_J^2$. For $\hat{\sigma}_J > 0$, the perturbative decrease of $\langle u \rangle$ is consistent with symmetry restoration across the $\hat{D}_0 = - {\rm sgn}(\hat{\sigma}_J) \hat{\sigma}_J^2$ line. Likewise, the perturbative decrease of $m_\psi = - \langle \sigma \rangle$ matches the expected behavior in phase III. Integrating out $(u,\psi)$ near the origin in phases III and IV generates a 1-loop correction with the result, $m_\lambda\sim |m_\psi|-|m_u|$. This agrees with the sign assignments for $m_\lambda$ near the II/III and IV/I phase boundaries, where $u$ and $\psi$ are becoming massless. Dualities and implications {#sec:dualities} ========================== Examination of Figs. \[fig:thA\] and \[fig:thB\] shows that a single field becomes light at a given phase transition. Mirror symmetry implies the resulting critical theories are dual. We thus arrive at the following dualities (indicated by $\leftrightarrow$): $$\begin{aligned} \label{fermionsqed} & \bar{\Psi} i \slashed{D}_{\hat{A}_J} \Psi - {1 \over 8 \pi} \hat{A}_J d \hat{A}_J \leftrightarrow |D_{-a} u|^2 - |u|^4 + {1 \over 4 \pi} a d a - {1 \over 2 \pi} \hat{A}_J d a, \\ \label{wfsqed} & |D_{\hat{A}_J + \hat{A}_R}v|^2 - |v|^4 \leftrightarrow -\frac{1}{4g^2}f_{\mu\nu}^2+ |D_{-a} u|^2 - |u|^4 - {1 \over 2 \pi} a d (\hat{A}_J + \hat{A}_R), \\ \label{trivialfermionduality} & \bar{\Psi} i \slashed{D}_{\hat{A}_J} \Psi - {1 \over 8 \pi} \hat{A}_J d \hat{A}_J - {1 \over 2 \pi} b d (\hat{A}_J + \hat{A}_R) \leftrightarrow \bar{\lambda} i \slashed{D}_{- \hat{A}_R} \lambda - {1 \over 8 \pi} \hat{A}_J d \hat{A}_J - {1 \over 2 \pi} a d (\hat{A}_J + \hat{A}_R), \\ \label{wffqed} & |D_{\hat{A}_J + \hat{A}_R}v|^2 - |v|^4 - {1 \over 4 \pi} \hat{A}_J d \hat{A}_J \leftrightarrow \bar{\psi} i \slashed{D}_{- a - \hat{A}_R} \psi + {1 \over 8 \pi} a d a - {1 \over 2 \pi}a d (\hat{A}_J + {1 \over 2} \hat{A}_R) - {1 \over 8 \pi} \hat{A}_R d \hat{A}_R.\end{aligned}$$ and are two examples of the bosonization dualities discussed in [@Aharony2016; @Seiberg:2016gmd]. The supersymmetric chiral mirror duality can be viewed as the multicritical parent duality from which these non-supersymmetric dualities descend. provides dual descriptions for the integer quantum Hall plateau transition between phases I and II where the level of the $\hat{A}_J$ CS term changes by unity. is Peskin-Dasgupta-Halperin duality [@Peskin:1977kp; @DasguptaHalperin1981] and may be used to describe a superfluid transition where the diagonal component of the $U(1)_J \times U(1)_R$ symmetry is broken. is a “trivial" duality relating a free fermion of theory A to a free fermion of theory B. again describes a $U(1)_J \times U(1)_R$ symmetry-breaking transition; the distinction from lies in the presence of a level-1 CS term for $\hat{A}_J$ in the adjacent massive phases. In [@Karch:2016sxi; @Seiberg:2016gmd], it was shown how the action of the modular group on dual conformal field theories (CFTs) can generate additional dualities. Here, we discuss how this action relates the above dualities to one another. Denoting the Lagrangian of a general CFT by ${\cal L}(\Phi, \hat{A})$, where $\Phi$ collectively represents the fields of the CFT and $\hat{A}$ is a background field for the global $U(1)$ symmetry, the modular group acts as follows [@WittenSL2Z2003; @LeighPetkouSL2Z2003]: $$\begin{aligned} \label{CFTaction} & {\cal T}: {\cal L}(\Phi, \hat{A}) \mapsto {\cal L}(\Phi, \hat{A}) + {1 \over 4 \pi} \hat{A} d \hat{A}, \cr & {\cal S}: {\cal L}(\Phi, \hat{A}) \mapsto {\cal L}(\Phi, a) - {1 \over 2 \pi} \hat{B} d a. \end{aligned}$$ The action ${\cal S}$ deserves further explanation: this transformation makes the background field dynamical $\hat{A} \rightarrow a$ and adds to the Lagrangian a new background field via the BF coupling $- {1 \over 2 \pi} \hat{B} d a$. The rules in induce the action ${\cal T}: {\tilde}\sigma \mapsto \begin{pmatrix}1 & 1 \cr 0 & 1 \end{pmatrix} {\tilde}\sigma$ and ${\cal S}: {\tilde}\sigma \mapsto \begin{pmatrix}0 & 1 \cr -1 & 0 \end{pmatrix} {\tilde}\sigma$ on the complexified conductivity ${\tilde}\sigma = \sigma_{xy} + i \sigma_{xx} \in \mathbb{H}$ (extracted from the two-point functions of the $U(1)$ current). Using , we see that is the ${\cal S}$-transform of . Examining the sides of the initial duality that a given theory occurs, we may say that ${\cal S}$ exchanges theory A and B in a loose sense. Performing a ${\cal T}$ transformation on , we obtain a fermionic description for the $U(1)_J \times U(1)_R$ symmetry-breaking transition complementary to . (A closely related duality to was previously argued for in [@Mulligan2016].) Using the marvelous property of transitivity, we can relate the right-hand side of to the ${\cal T}$-transform of the right-hand side of . This research was supported in part by grant NSF PHY-1316699 (S.K.), the University of California (M.M.), Conicet PIP-11220110100752 (G.T.), and DARPA YFA contract D15AP00108 (H.W.). M.M. is grateful for the generous hospitality of the Aspen Center for Physics NSF PHY-1066293 and the Kavli Institute for Theoretical Physics NSF PHY-11-25915. [^1]: On a related direction, it is also possible to deform three-dimensional Seiberg duality [@benini11; @Dimofte2011; @Beem2012; @Intriligator:2013lca] to derive non-supersymmetric dualities in large N limits, as was done in [@JainMinwallaYokoyama2013; @Gur-AriYacoby2015]. [^2]: We choose the metric $\eta^{\mu \nu} = {\rm diag}(1, -1, -1)$ and $\gamma$-matrices that satisfy $(\gamma^0 \gamma^1 \gamma^2)_{\alpha \beta} = - i \delta_{\alpha \beta}$, e.g., $\gamma^0 = \sigma^3, \gamma^1 = i \sigma^1, \gamma^2 = i \sigma^2$ where $\sigma^j$ are the Pauli-$\sigma$ matrices. [^3]: We denote the effect of integrating out a single Dirac fermion of unit charge and mass $m$ by the level-$1/2$ CS term ${{\rm sgn}(m) \over 8 \pi} A d A$ in the Wilsonian effective Lagrangian. A gauge-invariant expression utilizes the eta-invariant – see [@AlvarezGaume:1984nf; @Wittenfermionpathintegrals2016; @Seiberg:2016gmd] for further information. [^4]: $\lambda$ is the charge-conjugate of the fermion that appears in the conventionally-defined ${\cal N}=2$ vector multiplet. [^5]: This statement is slightly imprecise as the response can be modified by regularization-dependent contact terms [@ClossetDumitrescuFestucciaKomargodskiSeiberg]; the difference of the response across a phase transition is, however, physical and this is how our expressions should be understood.

ArXiv

--- abstract: 'We develop a sequential-topological study of rational points of schemes of finite type over local rings typical in higher dimensional number theory and algebraic geometry. These rings are certain types of multidimensional complete fields and their rings of integers and include higher local fields. Our results extend the constructions of Weil over (one-dimensional) local fields. We establish the existence of an appropriate topology on the set of rational points of schemes of finite type over any of the rings considered, study the functoriality of this construction and deduce several properties.' author: - 'Alberto Cámara[^1]' title: Topology on rational points over higher local fields --- Introduction ============ The study of higher dimensional local fields was started in the 1970s by Parshin in positive characteristic and Kato in the general case. They are natural objects that arise in algebra and geometry by a process that iterates localization and completion. Let us recall the definition. \[df:hlfs\] A zero dimensional local field is a finite field. For $n \geq 1$, an $n$-dimensional local field is a complete discrete valuation field, such that the residue field for the valuation is an $(n-1)$-dimensional local field. In positive characteristic, the theory is also developed allowing zero dimensional local fields to be any perfect field of positive characteristic (cf. \[sec:otherhlfs\]). In the geometric context, these fields generalize the role played by local fields on curves over finite fields or, more generally, on global fields. The point of view is that instead of focusing on a point and a neighbourhood, we consider a maximal chain of subschemes ordered by inclusion. The development of techniques to study the arithmetic of higher local fields, generalizing what is known for (one-dimensional) local fields, would have applications to the study of the arithmetic properties of higher dimensional schemes, including Néron models of elliptic curves. There are mainly four different approaches to the study of higher local fields: through topology; analysis and measure theory (see the works of Fesenko [@fesenko-aoas1], Morrow [@morrow-integration-on-valuation-fields], and Kim-Lee [@kim-lee-spherical-hecke-algebras]); category theory (see for example works by Beilinson [@beilinson-perverse-sheaves], Drinfeld [@drinfeld-infinite-dimensional-vector-bundles], Kapranov [@kapranov-double-affine-hecke-algebras] and Previdi [@previdi-locally-compact-objects-exact-cats]) and model theory(see for example works by Hrushovski-Kazhdan [@hrushovski-kazhdan-integration-in-valued-fields]). Finding relations among these points of view and unifying them is quite a challenging problem. Regarding the topological approach, with which this work is concerned, the major achievement is local higher class field theory. This theory was established in the general case by Kato [@kato-lCFT-1], [@kato-lCFT-2], [@kato-lCFT-3]; Fesenko [@fesenko-cft-of-multidim-local-fields-char-zero], [@fesenko-multidimensional-lt-of-cf]; Spiess [@spiess-class-formations-hlcft] and others. It uses Milnor $K$-groups endowed with a certain topology and the higher reciprocity map in order to describe abelian extensions of a higher local field. For a collection of surveys on higher local class field theory, see [@ihlf]. We should remark that the valuation topology on a higher local field and the usual topology on its units do not suffice for this purpose, and the introduction of more sophisticated topologies is required. Working simultaneously with several topologies helps to study deep arithmetic properties in higher number theory. Of course, there are many other similar examples, e.g. the use of six topologies in the work on the Bloch-Kato conjecture by Voevodsky. Such topologies are what we call [*higher topologies*]{}. When taken into account, multiplication on a higher local field of dimension greater than 1 ceases to be continuous, and hence we are no longer working in a category of topological fields. However, the fact that multiplication is sequentially continuous is enough to show many interesting properties of the arithmetic of these fields: Fesenko [@fesenko-abelianlocalpclass] seems to have been the first one to exploit the sequential properties of higher topologies in the study of higher class field theory. One of the conclusions of this work is the utility of the sequential point of view on higher topology when it comes to endowing sets of rational points over higher local fields with a topology. .5cm This work is organised as follows. In \[sec:sectopgpsandrings\], we review basic facts about sequential topology and introduce a new concept: a sequential ring (Definition \[df:seqring\]). \[sec:examples\] provides the main examples of sequential rings. It consists of a small survey on higher topologies on higher local fields. Although these sections are mainly expository, several well-known facts about higher topology are stated and proved in \[sec:examples\]; despite being known to experts in the area, it is very difficult to refer to a proof. In particular, (v) in Proposition \[prop:hdtop\] and Theorem \[thm:norelationlinearseq\] are new; we have not found these results in the existing literature. In \[sec:ratpoints\], schemes of finite type over a sequential ring are considered. Under some conditions on the ring, we construct a topology on the set of rational points, and we study the properties of this construction. The key result in this direction is Theorem \[thm:topratptsgeneral\]. The main available examples of sequential rings are complete discrete valuation fields such that their residue fields are endowed with a topology that has certain properties. For such fields, we may apply a general construction that produces a topology which turns the field into a sequential ring. Following the construction given by Theorem \[thm:topratptsgeneral\], we obtain a topology on the set of rational points on a scheme of finite type over a higher local field and also on schemes over the ring of integers of a higher local field. Several properties of this topology are deduced from this general setting. Finally, we discuss further work and connections to other topics in \[sec:future\]. In particular, the possibility of applications of the results in this work to the study of higher adelic rational points of algebraic varieties and to the study of representation theory of algebraic groups over higher local fields ought to be emphasized. #### Notation. By the word *ring* we will always mean a commutative, unital ring. Throughout this text, if $F$ denotes a complete discrete valuation field, ${\mathcal{O}}_F$ will denote the ring of integers for the unique normalized discrete valuation on $F$ and ${\mathfrak{p}}_F \subset {\mathcal{O}}_F$ will denote the unique maximal ideal of ${\mathcal{O}}_F$. By $\pi_F$ we will typically denote an element of $F$ such that ${\mathfrak{p}}_F = \pi_F {\mathcal{O}}_F$. We will denote $\overline{F} = {\mathcal{O}}_F / {\mathfrak{p}}_F$ for the residue field of $F$ and $\rho_F: {\mathcal{O}}_F \rightarrow \overline{F}$ will denote the residue homomorphism. If $F$ is a higher local field of dimension $n$ as in Definition \[df:hlfs\], the rank-$n$ ring of integers of $F$, defined in \[sec:examples\], will be denoted by $O_F$. In sections \[sec:sectopgpsandrings\] and \[sec:ratpoints\], $R$ and $S$ will denote sequential rings as in Definition \[df:seqring\]. In section \[sec:ratpoints\], $X$ and $Y$ will denote schemes, which we will always consider of finite type, over a sequential ring. Throughout the text, ${\mathbf{Sets}}$ stands for the category of sets. When $R$ is a ring, ${\mathbf{Alg}}_R$ and ${\mathbf{Sch}}_R$ stand for the categories of finitely generated $R$-algebras and schemes of finite type over $R$, respectively. ${\mathbf{AffSch}}_R$ is the subcategory of ${\mathbf{Sch}}_R$ whose objects are affine schemes of finite type over $R$. Finally, ${\mathbf{Top}}$ is the category of topological spaces and ${\mathbf{Seq}}$ is the subcategory of sequential topological spaces. When dealing with the elements of a sequence, the expression *almost all* may be safely replaced by *all but finitely many*. #### Acknowledgement. I am grateful to my advisor I. B. Fesenko for suggesting this direction of work, for many instructive conversations and for his useful guidance. I am also very grateful to M. T. Morrow and O. Bräunling, who during countless conversations have given invaluable advice and suggestions, have read early versions of this work and have pointed out many improvements. Sequential topology on groups and rings {#sec:sectopgpsandrings} ======================================= We review a few aspects about the category of sequential topological spaces. Details and proofs for the statements in this section may be found in [@franklin-sequences1], [@franklin-sequences2]. Let $(X, \tau)$ be a topological space. A sequence $(x_n)_n \subset X$ is convergent to $x \in X$ if for every open neighbourhood $U$ of $x$ in $X$, almost all of the elements in the sequence $(x_n)_n$ belong to $U$. That is: there exists $n_0 \in \mathbb{N}$ such that $x_n \in U$ for $n \geq n_0$. We will use the notation $x_n \rightarrow x$ whenever $(x_n)_n \subset X$ is a sequence which converges to $x \in X$. In general, specifying the set of convergent sequences of a set $X$ does not determine a unique topology on $X$, but rather a whole family of topologies. Among them, there is one which is maximal in the sense that it is the finest: the sequential saturation. A subset $A$ of $X$ is sequentially open if for any sequence $(x_n)_n$ in $X$ convergent to $x \in A$, there is an index $n_0 \in \mathbb{N}$ such that $x_n \in A$ for $n \geq n_0$. Every open set is sequentially open. The collection of sequentially open sets of $X$ defines a topology $\tau_s$ on $X$, finer than $\tau$. The natural map $(X, \tau_s) \to (X, \tau)$, given by the identity on the set $X$, is continuous. The space $(X,\tau_s)$ is called the sequential saturation of $(X,\tau)$. Whenever $\tau = \tau_s$, we shall simply say that $X$ is sequential. A subset $C \subseteq X$ is sequentially closed if for every $x_n \to x$ in $X$, $x_n \in C$ for every $n$ implies $x \in C$. This condition is equivalent to $X \setminus C$ being sequentially open. We could have defined $\tau_{s}$ by specifying that its closed sets are all sequentially closed sets in $(X, \tau)$. Let $f: X \rightarrow Y$ be a map between topological spaces. The following are equivalent: 1. For any convergent sequence $x_n \rightarrow x$ in $X$, the sequence $(f(x_n))_n$ converges to $f(x)$ in $Y$. 2. The preimage under $f$ of any sequentially open set in $Y$ is sequentially open. 3. The map $f: (X,\tau_s) \rightarrow Y$ is continuous. When this situation holds, we say that $f$ is a sequentially continuous map. Any continuous map is sequentially continuous. Hence, sequential continuity is a weaker condition than ordinary continuity. In a sequential space, the topology is essentially controlled by sequences. Any metric space is a sequential space. More generally, any first countable space is sequential [@franklin-sequences1 1]. The Stone-Čech compactification of a topological space $X$ is the unique Hausdorff compact space $\beta X$ provided with a continuous map $X \to \beta X$ such that for any Hausdorff compact space $Y$ and any continuous map $f: X \to Y$ there is a unique map $\beta f: \beta X \to Y$ such that the diagram $$\xymatrix{ X \ar[r]^f \ar[d] & Y \\ \beta X \ar[ru]_{\beta f} }$$ commutes. The Stone-Čech compactification of the natural numbers, $\beta \mathbb{N}$, is an example of a space which is not sequential [@goreham-sequential-convergence-in-top-spaces Example 1.1]. Despite being compact, $\beta \mathbb{N}$ is not sequentially compact. In this space, the closure of a set does not consist only of the limits of all sequences in that set, and a function from $\beta \mathbb{N}$ to another topological space may be sequentially continuous but not continuous. Let ${\mathbf{Seq}}$ denote the subcategory of sequential topological spaces. Taking the sequential saturation of a topological space defines a functor ${\mathbf{Top}}\rightarrow {\mathbf{Seq}}$. Some of the usual constructions in ${\mathbf{Top}}$ are not inherited by ${\mathbf{Seq}}$. The product of topological spaces is a remarkable example of such failure. Other examples of operations which do not behave well with respect to sequential saturation are function spaces and subspaces [@franklin-sequences1 Example 1.8]. However, open and closed subspaces and open and closed images are closed constructions in ${\mathbf{Seq}}$ [@franklin-sequences1 Prop. 1.9]. These sort of problems may be addressed by performing the usual construction in ${\mathbf{Top}}$, and then taking the sequential saturation of the resulting space. In this fashion, the sequential saturation of the product topology provides a product object in ${\mathbf{Seq}}$. In the words of Steenrod [@steenrod-convenient-category-topological-spaces], ${\mathbf{Seq}}$ is a [*convenient*]{} category of topological spaces. .5cm We are interested in the compatibility between sequential topology and algebraic structures. A sequential group is a group $G$ provided with a topology, such that multiplication $G \times G \rightarrow G$ and inversion $G \rightarrow G$ are sequentially continuous ($G \times G$ is provided with the product topology). A homomorphism of sequential groups is a sequentially continuous group homomorphism. In other words: if $(G,\tau)$ is a sequential group, then $(G,\tau_s)$ is a group object in ${\mathbf{Seq}}$. $(G, \tau_s)$ is also a group object in ${\mathbf{Top}}$. If $(G,\tau)$ is a topological group, then so is $(G,\tau_s)$. When considering rings, we could be interested in topologies for which subtraction and multiplication are sequentially continuous. However, we will deal with rings and topologies on them such that their additive groups are topological groups. \[df:seqring\] A sequential ring is a commutative ring $R$ provided with a topology and such that: 1. $(R,+)$ is a topological group. 2. Multiplication $R\times R \rightarrow R$ is sequentially continuous. A homomorphism of sequential rings is a continuous ring homomorphism. We could have required homomorphisms of sequential rings to be sequentially continuous maps. However, there is a good reason to prefer this stronger condition: in this way we preserve the underlying topological structure of the additive groups. Still, a continuous homomorphism of sequential rings $R \to S$ does not furnish $S$ with the structure of a topological $R$-module. This is why we consider the notion of sequential module. Let $M$ be an $R$-module. We say that $M$ is a sequential module if $M$ is provided with a topology such that $(M,+)$ is a topological group and multiplication $$R \times M \rightarrow M$$ is sequentially continuous. A homomorphism of sequential $R$-modules is a continuous $R$-module homomorphism. Note that with this definition $R$ is a sequential $R$-module. If $R \rightarrow S$ is a homomorphism of sequential rings, $S$ is automatically endowed with the structure of a sequential $R$-module. .5cm When $R$ is a topological ring, it is always possible to provide the units with a group topology. Because we do not demand inversion to be continuous, the correct way to topologize $R^\times$ is by considering the initial topology for the map $$\label{eqn:topringembedintotwocopies} R^\times \rightarrow R \times R, \quad x \mapsto \left( x, x^{-1} \right).$$ The situation when $R$ is a sequential ring is not very different, since we do not demand inversion on $R^\times$ to be sequentially continuous. The topology on $R^\times$ given by the sequential saturation of the initial topology for the map (\[eqn:topringembedintotwocopies\]) turns $R^\times$ into a topological group. A priori, we do not require sequential groups, rings and modules to be sequential topological spaces. Examples of sequential groups and rings {#sec:examples} ======================================= Of course, any topological group (resp. ring) is an example of a sequential group (resp. ring). We will consider other objects provided with a sequential structure which is not a topological one in the usual sense. Higher local fields {#subsec:HLFS} ------------------- We have introduced higher local fields in Definition \[df:hlfs\]. An alternative to this definition: a higher local field is a sequence of fields $F = F_n, F_{n-1},\dots, F_0$ where $F_0$ is finite and for $1 \leq i \leq n$, $F_i$ is a complete discrete valuation field with residue field $F_{i-1}$. See [@yekutieli-explicit-construction-grothendieck-residue-complex Definition 2.1.1] for yet another equivalent definition. \[exa:equi2dlf\] The fields $\mathbb{F}_q {\! \left(\! \left( t \right)\!\right)}{\! \left(\! \left( u \right)\!\right)}$ and $\mathbb{Q}_p {\! \left(\! \left( t \right)\!\right)}$ are both examples of higher local fields of dimension two. The example below is slightly more difficult, but very important nonetheless. Consider the field $$\mathbb{Q}_p {\! \left\{\! \left\{ t \right\}\!\right\}} = \left\{ \sum_{i=-\infty}^{\infty} a_i t^i, a_i \in \mathbb{Q}_p, \inf_{i}v_p(a_i) > -\infty, \lim_{i\rightarrow -\infty} a_i = 0 \right\},$$ with operations given by the usual addition and multiplication of power series. Note that we need to use convergence of series in $\mathbb{Q}_p$ in order to define the product. With the discrete valuation given by $$v\left( \sum_{i=-\infty}^{\infty} a_i t^i \right) := \inf_i v_p(a_i),$$ the field $\mathbb{Q}_p {\! \left\{\! \left\{ t \right\}\!\right\}}$ turns into a 2-dimensional local field, its first residue field being $\mathbb{F}_p {\! \left(\! \left( t \right)\!\right)}$. This field may also be obtained as the completion of $\mathbb{Q}_p {\! \left[ t \right]}$ with respect to the gaussian valuation. For an introduction to the topic of higher local fields and a collection of surveys on their structure and the structure of their extensions, see [@ihlf]. For the details of the topological issues we will discuss in this section, see [@madunts-zhukov-topology-hlfs]. .5cm Besides the ring of integers ${\mathcal{O}}_F$ for the unique normalized discrete valuation of $F$, we might consider other valuation rings in $F$ which are of arithmetic interest. Fix a system of parameters $t_1, \ldots, t_n$ of $F$. That is, $t_n$ is a uniformizer of $F$, $t_{n-1}$ is a unit of ${\mathcal{O}}_F$ such that its image in $\overline{F}$ is a uniformizer, and so on. We may use our chosen system of parameters to define a valuation of rank $n$: $$\mathbf{v} = (v_1, \ldots, v_n): F^\times \rightarrow {\mathbb{Z}}^n,$$ where ${\mathbb{Z}}^n$ is ordered with the inverse of the lexicographical order. The procedure we follow is: $v_n = v_F$, $v_{n-1}(\alpha) = v_{\overline{F}}(\alpha t_n^{-v_n(\alpha)})$, and so on. Although $\mathbf{v}$ does depend on the choice of the system of parameters, the valuation ring $$O_F = \left\{ \alpha \in F;\; \mathbf{v}(\alpha) \geq 0 \right\}$$ does not. The ring $O_F$, which is a local ring with maximal ideal $\left\{ \alpha;\;\mathbf{v}(\alpha) > 0 \right\}$, is called the rank-$n$ ring of integers of $F$. It is a subring of ${\mathcal{O}}_F$. Once a system of parameters is chosen, we may consider a valuation of rank $r$ for every $1 \leq r \leq n$ by mimicking the same procedure and stopping after $r$ steps. The valuation rings we obtain are independent of any choice of parameters and are ordered by inclusion: $$\label{eqn:higherranksubrings} O_F = O_1 \subset O_2 \subset \cdots \subset O_n = {\mathcal{O}}_F.$$ .5cm A higher local field has a topology determined by its unique normalized discrete valuation. This topology turns $F$ into a Hausdorff and complete topological group which is not locally compact if $n >1$, due to the fact that the residue field is not finite. There are several reasons to consider a different topology, such as the fact that the formal series in fields such as the ones in Example \[exa:equi2dlf\] do not converge in the valuation topology. A higher topology takes into account the topology of the residue field and has some good properties, but in general it is not a ring topology. While the additive group is always a topological group with respect to this topology, multiplication is only sequentially continuous. It is for this reason that sequential topology plays an important role in the theory of higher local fields. Let $E$ be a field provided with a topology such that addition and multiplication by a fixed element are continuous. Consider the field of Laurent power series $E {\! \left(\! \left( t \right)\!\right)}$. We define a topology as follows. Let $\left\{ U_i \right\}_{i \in \mathbb{Z}}$ be a sequence of open neighbourhoods of zero of $E$, with the property that there is an $i_0 \in \mathbb{Z}$ such that $U_i = E$ for $i \geq i_0$. Define $$\label{eqn:definebasicopenhighertop} \mathcal{U} = \left\{ \sum_{i \gg -\infty} f_i t^i\in E{\! \left(\! \left( t \right)\!\right)};\; f_i \in U_i \right\}.$$ \[prop:liftoftop\] Let $E$ be a field endowed with a topology such that addition and multiplication by a fixed element are continuous. The collection of sets defined by (\[eqn:definebasicopenhighertop\]) is a basis of open neighbourhoods of zero for a group topology on $E{\! \left(\! \left( t \right)\!\right)}$. A similar procedure can be applied to higher local fields. If ${\mathrm{char}\:}F = {\mathrm{char}\:}\overline{F}$, then an isomorphism $F \simeq \overline{F}{\! \left(\! \left( t \right)\!\right)}$ is available. The mixed characteristic case, in which ${\mathrm{char}\:}F \neq {\mathrm{char}\:}\overline{F}$, is more complicated. Assume first that $F$ is a standard mixed-characteristic higher local field, i.e; that there is a local field $K$ and an isomorphism $F = K {\! \left\{\! \left\{ t_1 \right\}\!\right\}} \cdots {\! \left\{\! \left\{ t_{n-1} \right\}\!\right\}}$ (note that for this field we may choose $\pi_F = \pi_K$), and assume that the higher topology has already been constructed for $F' = K {\! \left\{\! \left\{ t_1 \right\}\!\right\}}\cdots{\! \left\{\! \left\{ t_{n-2} \right\}\!\right\}}$, which is an $(n-1)$-dimensional local field. The choice of local parameters $t_1 = \pi_F, t_2, \ldots, t_n$ determines a [*canonical*]{} lifting [^2] $$h: \overline{F} \rightarrow {\mathcal{O}}_F.$$ Assume that the topology on $\overline{F}$ has already been constructed. Let $\left\{ U_i \right\}_{i \in \mathbb{Z}}$ be a sequence of open neighbourhoods of zero in $\overline{F}$, and assume that there is an index $i_0 \in \mathbb{Z}$ such that $U_i = \overline{F}$ for all $i \geq i_0$. Define $$\label{eqn:definebasicopenhighertopmixed} \mathcal{U} = \left\{ \sum_{i \gg -\infty} h(f_i) \pi_F^i \in F;\; f_i \in U_i, \text{ for all } i \gg -\infty \right\}.$$ The sets of the form (\[eqn:definebasicopenhighertopmixed\]) define the basis of open neighbourhoods of zero for a group topology on $F$. In the general case in which $F$ is a nonstandard mixed-characteristic local field, we may find a finite subextension $M \subset F$ which is a standard mixed-characteristic local field, apply the previous construction to $M$ and provide $F$ with the product topology through the linear isomorphism $F \simeq M^{\left[ F:M \right]}$. .5cm By iterating the above constructions and using classification results for higher local fields and choices of parameters we may construct a higher topology on a general $n$-dimensional local field. Details (and proofs) may be found in [@madunts-zhukov-topology-hlfs 1.1 – 1.4]. Any topology constructed on an $n$-dimensional local field using the above construction is called a higher topology. In general, a higher topology on $F$ depends on the choice of a system of parameters. However, we should remark that whenever ${\mathrm{char}\:}\overline{F} = p$, the topology depends neither on the choice of a system of parameters or a standard subfield $M \subset F$ in the mixed-characteristic case [@madunts-zhukov-topology-hlfs Theorem 1.3]. In the case ${\mathrm{char}\:}F = {\mathrm{char}\:}\overline{F} = 0$, the topology does not depend on the choice of a parameter, but it depends on the choice of a coefficient field [@madunts-zhukov-topology-hlfs 1.4]. When an isomorphism such as $F \simeq \overline{F}{\! \left(\! \left( t \right)\!\right)}$ or $F \simeq F'{\! \left\{\! \left\{ t \right\}\!\right\}}$ has been fixed, we will drop the indefinite article and refer to [*the*]{} higher topology. We summarize the properties of higher topologies below. \[prop:hdtop\] Let $F$ be an $n$-dimensional local field. Then any higher topology on $F$ satisfies the following properties: 1. $(F,+)$ is a topological group which is complete and Hausdorff. 2. If $n > 1$, every base of neighbourhoods of zero is uncountable. 3. If $n > 1$, multiplication $F \times F \rightarrow F$ is not continuous. However, both multiplication $F \times F \to F$ and inversion $F^\times \to F^\times$ are sequentially continuous. 4. Multiplication by a fixed nonzero element $F \rightarrow F$ is a homeomorphism. 5. The residue homomorphism $\rho: {\mathcal{O}}_F \rightarrow \overline{F}$ is open. See [@ihlf 1.3.2] for parts [*(i)*]{} to [*(iv)*]{}. Regarding [*(v)*]{}, the statement is a well known fact but unfortunately seems to be unavailable in the literature; we shall provide a proof for completeness. First assume that $F \simeq \overline{F}{\! \left(\! \left( t \right)\!\right)}$ is of equal-characteristic. In such case, let $U_0$ be an open neighbourhood of zero in $\overline{F}$. Then $$\rho^{-1}\left( U_0 \right) = U_0 + \sum_{i \geq 1} \overline{F} t^i = U_0 + t {\mathcal{O}}_F,$$ which is open in $F$. Moreover, an open neighbourhood of zero $\mathcal{U} \subseteq {\mathcal{O}}_F$ is of the form $\mathcal{U} = \sum_{i\geq 0} U_i t^i$ where $U_i \subseteq \overline{F}$ are open neighbourhoods of zero, and we have $\rho\left( \mathcal{U} \right) = U_0$. Second, assume that $F$ is of mixed-characteristic. Without loss of generality, we may assume that $F \simeq F'{\! \left\{\! \left\{ t \right\}\!\right\}}$ is standard, as openness of maps is preserved on finite cartesian products. Let $U_0 \subseteq \overline{F}$ be an open neighbourhood of zero. Then $\rho^{-1}\left( U_0 \right) = h(U_0) + \pi_F {\mathcal{O}}_F$, which is open. Finally, if $\mathcal{U}$ is as in (\[eqn:definebasicopenhighertopmixed\]), then $\rho(\mathcal{U}) = U_0$, which is open. Failure of a higher topology on $F$ to be a ring topology is the original reason that led to the consideration of sequential rings in this work. A group topology on an abelian group is said to be linear if the filter of neighbourhoods of the identity element admits a collection of subgroups as a basis. A commutative ring $R$ provided with a topology $\tau$ which is linear for the additive group and for which multiplication maps $$R \to R,\quad y \mapsto xy$$ are continuous for all $x \in R$ is said to be a semitopological ring. In view of (i) and (iv) in Proposition \[prop:hdtop\], it is possible to show that a higher local field endowed with a higher topology is a semitopological ring. A theory of semitopological rings has been developed and applied to the study of higher fields by Yekutieli [@yekutieli-explicit-construction-grothendieck-residue-complex] and others. There seems to be a disagreement between the sequential and linear approaches to topologies on higher local fields. \[thm:norelationlinearseq\] Let $F$ be a higher local field. Denote $\tau$ for the topology on $F$ defined by Proposition \[prop:hdtop\], and let $\tau_s$ be its sequential saturation. The collections of open subgroups for $\tau$ and $\tau_s$ agree. We split the proof in several steps. Let $F_i = t^i {\mathcal{O}}_F$. *Step 1. A subset $A \subseteq F$ is sequentially open if and only if $A \cap F_i$ is sequentially open in $F_i$ for every $i$*. Suppose that $A \cap F_i$ is sequentially open in $F_i$ for every $i$. Let $x_n \to x \in A$. Then, there is an index $j$ such that $x_n \in F_j$ for all $n$, and since $F_j$ is sequentially closed, $x \in F_j$. As $A \cap F_j$ is sequentially open and $x \in A \cap F_j$, almost all of the $x_n$ belong to $A \cap F_j \subset A$, showing that $A$ is sequentially open. *Step 2. Let $Y \subset F_i$ be a subset such that $0 \in Y$. $Y$ is sequentially open if and only if it contains a subgroup $\mathcal{U} = \sum_{j \geq i} U_j t^j$, where the residues of elements in $U_j$ are open subgroups of $\overline{F}$ and $U_j = \overline{F}$ if $j$ is large enough*. Let $u$ be a lift of a uniformizer of $\overline{F}$. If no such subgroup $\mathcal{U}$ is contained in $Y$, then for every $n$ there is an element $y_n \in t^i\left( u^n O_F + t^n {\mathcal{O}}_F \right)$. However, $y_n \to 0 \in Y$ and therefore almost all of the $y_n$ must belong to $Y$, a contradiction. *Step 3. A subgroup of $H$ of $F$ is open for $\tau$ if and only if it is open for $\tau_s$*. Suppose that $H$ is open for $\tau_s$. By step 1, $H \cap F_i$ is sequentially open and by step 2 it contains a subset $U_i t^i$, such that the image of $U_i$ in $\overline{F}$ is open and it contains $F_n$ for some $n$. Put $\mathcal{U} = \sum U_i t^i \cap F$. If ${\mathrm{char}\:}F = {\mathrm{char}\:}\overline{F}$, $U_i$ may be chosen so that it is an open subgroup of $\overline{F}$ viewed inside $F$. Otherwise, $U_i$ may be modified to ensure that $\mathcal{U}$ is a subgroup of $H$ such that the image of $U_i$ in $\overline{F}$ is an open subgroup and $\mathcal{U}$ contains some $F_n$. In both cases, $\mathcal{U}$ is an open subgroup for $\tau$. Since $H$ is the union of $\mathcal{U}$-cosets, $H$ is also an open subgroup for $\tau$. For a general higher local field $F$ and a higher topology $\tau$, we have that $\tau_s$ is not a linear topology. A linear topology is completely determined by its collection of open subgroups. After the previous theorem, it suffices to show that $\tau \neq \tau_s$. So we recover a counterexample from [@fesenko-sequential-topologies]. Let $F = \mathbb{F}_p {\! \left(\! \left( u \right)\!\right)} {\! \left(\! \left( t \right)\!\right)}$. Let $C = \left\{ t^a u^{-c} + t^{-a} u^c,\; a, c \geq 1 \right\}$. Then $W = F \setminus C$ is open for $\tau_s$. Suppose that $U_i \subset \mathbb{F}_p {\! \left(\! \left( u \right)\!\right)}$ are open subgroups such that $U_i = \mathbb{F}_p {\! \left(\! \left( u \right)\!\right)}$ if $i$ is large enough and such that $\mathcal{U} = \sum U_i t^i \cap F$ is contained in $W$. Then, for any positive $c$ such that $u^c \in U_{-a}$ we would have $t^a u^{-c} + t^a u^{c} \in W$, a contradiction. Hence, $W$ is not open for $\tau$. Although the approach to higher topologies by using linear topologies and semitopological rings is useful for the study of the construction of higher adeles by means of ind-pro functors, the sequential approach is very important from the point of view of higher class field theory. When dealing with rational points over higher local fields, the sequential approach will allow us to say something about the continuity of polynomial maps $R^n \to R$, whereas this is not possible with a semitopological ring. The unit group of a higher local field -------------------------------------- There are several approaches to constructing topologies on $F^\times$. Once we know that $F$ is a sequential ring, and that inversion is sequentially continuous, a natural definition is the following. \[df:toponunitsHLF\] The topology we consider on $F^\times \subset F$ is the sequential saturation of the subspace topology. Compare with (\[eqn:topringembedintotwocopies\]); in the case of a higher local field, taking the embedding into two copies of $F$ is unnecessary, as we know that inversion is sequentially continuous. The proof of the result below is obvious. The topology on $F^\times$ given by Definition \[df:toponunitsHLF\] is a group topology; $F^\times$ is a Hausdorff complete group. .5cm Assume, until the end of this paragraph, that ${\mathrm{char}\:}F > 0$. In this case, Parshin suggested a different approach, which was later refined by Yekutieli [@yekutieli-explicit-construction-grothendieck-residue-complex 3]. After choosing parameters, we obtain a decomposition $$\label{eqn:decomposeunitsofhlf} F^\times \simeq \mathbb{Z} t_n \oplus \cdots \oplus \mathbb{Z} t_1 \oplus F_0^\times \oplus V_F,$$ where $F_0$ is the last residue field and $V_F$ is the group of principal units. $F^\times$ is provided with a topology by (\[eqn:decomposeunitsofhlf\]) once $\mathbb{Z} t_n \oplus \cdots \oplus \mathbb{Z} t_1 \oplus F_0^\times$ is given the discrete topology and $V_F \subset {\mathcal{O}}_F$ the subspace topology for the higher topology; this is called the Parshin topology. We summarize the properties of this topology on units below. \[prop:propsoftoponunits\] Suppose ${\mathrm{char}\:}F > 0$. The topology we have defined on $F^\times$ satisfies: 1. The group $F^\times$ is a topological group with respect to the Parshin topology only when $F$ is of dimension at most 2. 2. The Parshin topology is weaker than the valuation topology on $F^\times$. 3. The Parshin topology does not depend on the choice of a system of parameters. 4. $F^\times$ is complete. 5. Multiplication on $F^\times$ is sequentially continuous. See [@madunts-zhukov-topology-hlfs 3] for [*(i)*]{} to [*(iii)*]{}, [@ihlf 1.4] for the rest. Suppose ${\mathrm{char}\:}F > 0$. The sequential saturation of the Parshin topology on $F^\times$ agrees with the sequential saturation of the topology on $F^\times$ defined by (\[eqn:topringembedintotwocopies\]). Denote the Parshin topology by $\tau$, and the initial topology defined by (\[eqn:topringembedintotwocopies\]) by $\lambda$. Although this is implicitly included in [@fesenko-sequential-topologies], we describe the explicit argument here. A sequence $a_m=t_n^{i_{n,m}} \cdots t_1^{i_{1,m}} u_m$ of elements of $F^\times$, with $u_m$ in the unit group of the ring of integers of $F$ with respect to any of its discrete valuations of rank $n$ tends to $1$ in $\lambda$ if and only if the sequence $u_m -1$ tends to $0$ with respect to the higher dimensional topology on $F$ (described by Proposition \[prop:hdtop\]) and for every $j$ such that $1 \leq j \leq m$, the sequence $i_{j,m}$ is constant for sufficiently large $m$. But this last condition is equivalent to the sequence $a_m$ converging to $1$ with respect to $\tau$. The $n$-th Milnor $K$-group of any field $F$ is presented as the term in the right in the following exact sequence of abelian groups: $$0 \rightarrow I_n \rightarrow F^\times \otimes_{\mathbb{Z}}\stackrel{(n)}{\cdots} \otimes_{\mathbb{Z}}F^\times \rightarrow K_n(F) \rightarrow 0,$$ where $I_n = \langle a_1 \otimes \cdots\otimes a_n;\; a_i + a_j = 1 \text{ for some } i \neq j \rangle_{\mathbb{Z}}$. By definition, $K_1(F) = F^\times$ and by convention $K_0(F) = {\mathbb{Z}}$. When $F$ is an $n$-dimensional local field, $K_n$ generalises the role of $K_1 = \mathbb{G}_m$ in describing abelian extensions of the field [@ihlf]. From this point of view, it provides a correct higher dimensional generalization of the group of units. The functor $K_n$ is not representable for $n \geq 2$, meaning that in general $K_n(F)$ is not the set of $F$-rational points on any scheme. Let $F$ be an $n$-dimensional local field, and consider the topology on $F$ (resp. $F^\times$) given by Proposition \[prop:hdtop\] (resp. \[prop:propsoftoponunits\]). Consider the finest topology on $K_m(F)$ for which: 1. The symbol map $F^\times \times \cdots \times F^\times \rightarrow K_m(F)$ is sequentially continuous. 2. Subtraction $K_m(F) \times K_m(F) \rightarrow K_m(F)$ is sequentially continuous. This topology is sequentially saturated [@fesenko-sequential-topologies 4, Remark 1]. The topological $m$-th Milnor $K$-group is $$\label{eqn:deftopmilnorgps} K_m^{\mathrm{t}}(F) = K_m(F) / \Lambda_m(F),$$ where $\Lambda_m(F)$ is the intersection of all open neighbourhoods of zero. A sequentially continuous Steinberg map $F^\times \otimes_\mathbb{Z} \cdots \otimes_\mathbb{Z} F^\times \rightarrow G$ where $G$ is a Hausdorff topological group induces a continuous homomorphism $K_m^{\mathrm{t}}(F) \rightarrow G$. The Artin-Schreier-Parshin pairing, the Vostokov pairing and the tame symbol are examples of such continuous homomorphisms defined on $K_m^{\mathrm{t}}(F)$ [@ihlf I.6.4]. Other higher complete fields {#sec:otherhlfs} ---------------------------- There are other fields which can be defined in a similar way as a higher local field, and for which the construction of a higher topology is still valid. We give here two important examples of such. Let $k$ be any perfect field. A zero dimensional complete field is a perfect field. An $n$-dimensional complete field is a complete discrete valuation field such that its residue field is an $(n-1)$-dimensional complete field. If $F$ is an $n$-dimensional complete field with last residue field $k$, we say that $F$ is an $n$-dimensional complete field over $k$. Higher local fields coincide with higher dimensional complete fields over a finite field. Higher dimensional complete fields over arbitrary perfect fields may be classified using the same techniques we have discussed for higher local fields [@madunts-zhukov-topology-hlfs Theorem in 0]. For example, if ${\mathrm{char}\:}F > 0$ and the last residue field is $k$, we may choose a system of parameters $t_1, \ldots, t_n$ such that $$F \simeq k{\! \left(\! \left( t_1 \right)\!\right)} \cdots {\! \left(\! \left( t_n \right)\!\right)}.$$ The reason why we can consider topological issues in this more general setting is because the construction of a higher topology does not use at any point the fact that the last residue field is finite. In [@fesenko-abelianlocalpclass], Fesenko considered the class field theory of $n$-dimensional complete fields over a perfect field of positive characteristic. In particular, [@fesenko-abelianlocalpclass 2] contains a brief description on how to topologize such fields. The main result of interest, which completely resembles Proposition \[prop:hdtop\], is the following. An $n$-dimensional complete field $F$ over a perfect field of positive characteristic may be topologized in such a way that: 1. $(F,+)$ is a topological group which is complete and Hausdorff. 2. If $n > 1$, every base of open neighbourhoods of zero is uncountable. 3. Multiplication $F \times F \to F$ and inversion $F^\times \to F^\times$ are sequentially continuous. 4. Multiplication by a fixed non-zero element $F \to F$ is a homeomorphism. 5. The residue homomorphism $\rho: {\mathcal{O}}_F \to \overline{F}$ is open. Moreover, the topology does not depend on the choice of parameters. Consider the last residue field of $F$ as a topological field with respect to the discrete topology, and apply the inductive process described in \[subsec:HLFS\]. Hence, these are more general examples of sequential rings. .5cm A notion of higher archimedean local field exists, and these may be topologized applying the same tools. The only one-dimensional archimedean local fields are $\mathbb{R}$ and $\mathbb{C}$. Let $n > 1$. An $n$-dimensional archimedean local field is a complete discrete valuation field such that its residue field is a $(n-1)$-dimensional archimedean local field. Let $\mathbb{K}$ be either $\mathbb{R}$ or $\mathbb{C}$. The dimension of an archimedean higher local field does not agree with its dimension as a higher complete field over $\mathbb{K}$: an $n$-dimensional archimedean local field is an $(n-1)$-dimensional complete field over $\mathbb{K}$. The reason for this shift in the dimension is because, since $\mathbb{K}$ is itself already a local field, we want to put fields such as $\mathbb{R}{\! \left(\! \left( t \right)\!\right)}$ and $\mathbb{Q}_p {\! \left(\! \left( t \right)\!\right)}$ in the same box; both fields are two-dimensional with the definitions we have taken. As the field $\mathbb{K}$ is of characteristic zero, higher archimedean local fields are easily classified: they are isomorphic to Laurent power series fields over $\mathbb{K}$. Let $F$ be an $n$-dimensional archimedean local field. Then, there are parameters $t_2,\ldots,t_n \in F$ such that $$F \simeq \mathbb{K}{\! \left(\! \left( t_2 \right)\!\right)}\cdots{\! \left(\! \left( t_n \right)\!\right)}.$$ Since ${\mathrm{char}\:}\mathbb{K} = 0$, all the residue fields of $F$ have characteristic zero; the result follows by induction in the dimension and by Cohen structure theory for equidimensional complete fields. Regarding the way to topologize such fields; we will always consider the euclidean topology on $\mathbb{K}$. This topology satisfies the conditions of Proposition \[prop:liftoftop\], and it is Hausdorff and complete. We apply the construction specified in the aforementioned Proposition inductively, in order to obtain the following result. An $n$-dimensional archimedean local field $F$ may be topologized in such a way that: 1. $(F,+)$ is a topological group which is complete and Hausdorff. 2. If $n > 1$, every base of open neighbourhoods of zero is uncountable. 3. Multiplication $F \times F \to F$ and inversion $F^\times \to F^\times$ are sequentially continuous. 4. Multiplication by a fixed non-zero element $F \to F$ is a homeomorphism. 5. The residue homomorphism $\rho: {\mathcal{O}}_F \to \overline{F}$ is open. In the case of archimedean higher local fields, a higher topology does depend on the choice of a system of parameters. Higher rank valuation rings (\[eqn:higherranksubrings\]) may also be defined in the case of archimedean higher local fields and higher complete fields. In the first case, the system of parameters has $n-1$ elements and we can only construct a chain of $n-1$ subrings by this procedure. Rational points over sequential rings {#sec:ratpoints} ===================================== Affine case {#sec:ratpointsaffine} ----------- \[df:toponrationalpoints\] Let $R$ be a sequential ring, and $X \rightarrow {\mathrm{Spec}\:}R$ an affine scheme of finite type. Let $X = {\mathrm{Spec}\:}A$. By using the topology on $R$, it is possible to topologize $X(R)$ as follows. There is a natural inclusion of sets $$X(R) = {\mathrm{Hom}}_{{\mathbf{Sch}}_R}({\mathrm{Spec}\:}R, X) = {\mathrm{Hom}}_{{\mathbf{Alg}}_R}(A,R) \subset {\mathrm{Hom}}_{\mathbf{Sets}}(A,R) = R^A.$$ We endow $R^A$ with the product topology and provide $X(R) \subset R^A$ with the sequential saturation of the subspace topology. After taking a sequential saturation, the inclusion map $$X(R) \hookrightarrow R^A$$ is continuous, but not an embedding. As we will see in \[sec:ratpointsgnral\], taking a sequential saturation in Definition \[df:toponrationalpoints\] is important. Every element $a \in A$ induces a map $\varphi_a: X(R) \rightarrow R$ by evaluating $R$-algebra homomorphisms $A \rightarrow R$ at $a$. Such a map agrees with the composition $$X(R) \hookrightarrow R^A \rightarrow R$$ where the second map is given by projection to the $a$-th coordinate. By the previous remark, $\varphi_a$ is continuous. There is another way to construct a topology on $X(R)$ which is more explicit. The choice of a closed embedding into an affine space identifies $X(R)$ with a subset of $R^n$ for some $n$, and we may endow $R^n$ with the product topology and $X(R) \subset R^n$ with the saturation of the subspace topology. If this procedure is taken, it is necessary to show that this topology on $X(R)$ does not depend on the choice of embedding into affine space. We explain how this works, and show that it is essentially equivalent to Definition \[df:toponrationalpoints\]. Let us choose an $R$-algebra isomorphism $$\label{eqn:chooseisom} A \simeq R{\! \left[ t_1, \dots, t_n \right]} /I$$ and identify $X(R)$ with the set $V(I)$ of elements in $R^n$ on which all polynomial functions belonging to the ideal $I$ vanish. Consider the product topology on $R^n$ and endow $X(R)$ with the sequential saturation of the subspace topology. The choice of isomorphism corresponds to the choice of elements $a_i \in A$ that map to $t_i \pmod{I}$ under (\[eqn:chooseisom\]), for $0 \leq i \leq n$. These induce a continuous map $R^A \rightarrow R^n$ by projecting to the coordinates indexed by $(a_1,\dots,a_n)$. The inclusion $X(R) \subset R^n$ factors then into $$X(R) \hookrightarrow R^A \rightarrow R^n.$$ On one hand, the topology on $X(R)$ given by Definition \[df:toponrationalpoints\] makes all inclusions $X(R) \hookrightarrow R^n$ given by choosing an $R$-algebra isomorphism continuous and hence provides a stronger topology. On the other hand, assume that $X(R)$ is topologized according to an embedding into $R^n$ determined by (\[eqn:chooseisom\]). Every element of $A$ is an $R$-polynomial in $a_1,\ldots,a_n$, and $R$ is a sequential ring. Hence, all polynomial maps $R^n \rightarrow R$ are sequentially continuous. It follows that the inclusion map $X(R) \hookrightarrow R^A$ is sequentially continuous. But on a sequential space a sequentially continuous map is necessarily continuous. Over the affine line ${\mathbb{A}}_R^1 = {\mathrm{Spec}\:}R[t]$, the topology we have just described on ${\mathbb{A}}_R^1(R) = R$ is the sequential saturation of the topology on $R$. We summarize the construction discussed above in the statement below, along with some properties. \[thm:topratptsaffine\] Let $R$ be a sequential ring. There is a unique covariant functor $${\mathbf{AffSch}}_R \to {\mathbf{Seq}}, \quad X \mapsto X(R)$$ which carries fibred products to products, closed immersions to topological embeddings and gives ${\mathbb{A}}_R^1\left( R \right)=R$ the sequential saturation of its topology. If $R$ is Hausdorff and $X \to {\mathrm{Spec}\:}R$ is affine and of finite type, then $X(R)$ is Hausdorff and closed immersions $X \hookrightarrow X'$ induce closed embeddings $X(R) \hookrightarrow X'(R)$. Regarding uniqueness, let $X \to {\mathrm{Spec}\:}R$ be a finite type affine scheme. Choose a closed embedding $X \hookrightarrow {\mathbb{A}}_R^n$ for some $n\geq 0$. We look at the induced map on $R$-points: since we have compatibility with fibred products and we may view ${\mathbb{A}}^n$ as a product of $n$ copies of ${\mathbb{A}}^1$, and since closed immersions carry on to topological embeddings, $X(R) \subset R^n$ is an embedding into $R^n$ viewed as an object in ${\mathbf{Seq}}$. Hence, the topology is unique and $X(R)$ is Hausdorff whenever $R$ is Hausdorff, for a product of Hausdorff spaces in ${\mathbf{Seq}}$ is Hausdorff. Regarding existence, take the topology on $X(R)$ given by Definition \[df:toponrationalpoints\]; we check the rest of the claimed properties. If $X \rightarrow Y$ is a morphism of affine $R$-schemes, put $X = {\mathrm{Spec}\:}A$, $Y = {\mathrm{Spec}\:}B$. The homomorphism of $R$-algebras $B \rightarrow A$ induces a continuous map $$\label{eqn:pfaffinecase1} R^{A} \rightarrow R^{B}.$$ Then, the commutative diagram $$\xymatrix{ R^{A} \ar[r] & R^{B} \\ X(R) \ar[u] \ar[r] & Y(R) \ar[u] }$$ shows that the natural map $X(R) \rightarrow Y(R)$ is continuous, and hence it is also continuous after taking sequential saturations. If $X \hookrightarrow Y$ is a closed immersion, then the map $B \rightarrow A$ is a surjective homomorphism of $R$-algebras and the map (\[eqn:pfaffinecase1\]) is an embedding. From a topological point of view, this identifies $R^{A}$ with the subset of $R^{B}$ cut out by some equalities among components. This implies that when $R$ is Hausdorff, $X(R) \hookrightarrow Y(R)$ is a closed embedding of topological spaces. Finally, we check compatibility with fibred products. First, over the final object in the category: if $X$ and $Y$ are two affine schemes of finite type over $R$, we have an identification of sets $$\left( X \times_R Y \right)\left( R \right) = X\left( R \right) \times Y\left( R \right),$$ and both spaces are homeomorphic in ${\mathbf{Seq}}$, since we take the sequential saturation of the product topology on the left-hand side of the previous equation. In the general setting, assume that $X,Y$ and $Z$ are affine $R$-schemes of finite type and that we are given morphisms $X \rightarrow Z$, $Y \rightarrow Z$. Consider the isomorphism $$X \times_Z Y \simeq \left( X \times_R Y \right) \times_{Z \times_R Z} Z$$ and the topological homeomorphism that it induces. Since $Z$ is separated over $R$, and we already have compatibility over the final object in the category, we need only consider the case in which one of the projection maps is a closed immersion. Since closed immersions induce topological embeddings, we are done. Now we wish to study the behaviour of the topology defined in \[df:toponrationalpoints\] with respect to base change. Let $R \rightarrow S$ be a morphism of sequential rings, as in Definition \[df:seqring\], and let $X \rightarrow {\mathrm{Spec}\:}R$ be an affine scheme of finite type. We will identify $$X(S) = X_S(S),$$ where $X_S = X \times_R S$. As $X_S \rightarrow {\mathrm{Spec}\:}S$ is an affine scheme of finite type, the topology we will consider on $X_S(S)$ is that given by applying Definition \[df:toponrationalpoints\] to $X_S \rightarrow {\mathrm{Spec}\:}S$. \[prop:basechange\] Let $R \rightarrow S$ be a morphism of sequential rings and $X \rightarrow {\mathrm{Spec}\:}R$ an affine scheme of finite type. Then the natural map $X(R) \rightarrow X(S)$ is continuous. Moreover, if $R \rightarrow S$ is an open (resp. closed) embedding, then $X(R) \rightarrow X(S)$ is an open (resp. closed) embedding. Let $X = {\mathrm{Spec}\:}A$, and let $X_S = {\mathrm{Spec}\:}B$, where $B = A \otimes_R S$. We pick an isomorphism $$A \simeq R{\! \left[ t_1, \ldots, t_n \right]} / I$$ and identify $X(R) = V(I) \subset R^n$. After changing base over $S$, we have $$B \simeq S{\! \left[ t_1, \ldots, t_n \right]} / I^e,$$ where $I^e = IS{\! \left[ t_1, \ldots, t_n \right]}$ is the extension of the ideal $I$ along the ring homomorphism $$R{\! \left[ t_1, \ldots t_n \right]} \rightarrow S{\! \left[ t_1, \ldots, t_n \right]}.$$ We may identify $X(S) = X_S(S) = V(I^e) \subset S^n$. Now, the diagram $$\label{eqn:diagramincaffine} \xymatrix{ R^n \ar[r] & S^n \\ V(I) \ar[r] \ar[u] & V(I^e), \ar[u] }$$ whose vertical arrows are inclusions, shows that $X(R) \rightarrow X(S)$ is continuous when $X(R)$ and $X(S)$ are viewed as subspaces of $R^n$ and $S^n$. Therefore, we may take sequential saturations: the map $X(R) \rightarrow X(S)$ is continuous. Suppose now that $R \rightarrow S$ is a closed immersion. Then, $R^n \rightarrow S^n$ is also a closed immersion and, by restricting and taking sequential saturations, so is the map $X(R) \rightarrow X(S)$. This is also the case when we deal with an open immersion. A construction which may also be considered in this setting is the Weil restriction. Let $R \rightarrow S$ be an extension of rings, and $Y \rightarrow {\mathrm{Spec}\:}S$ be an affine scheme of finite type. Consider the functor $X : {\mathbf{Sch}}_R \rightarrow \mathbf{Sets}$ defined by $$X(T) = \mathrm{Hom}_{{\mathbf{Sch}}_S}(T \times_R S, Y), \quad T \in \mathrm{Ob}({\mathbf{Sch}}_R).$$ Whenever $X$ is representable, the associated scheme is called the Weil restriction of $Y$ along $R \rightarrow S$. Assume that $R \hookrightarrow S$ is an injective morphism of sequential rings such that $S$ is a finite type, locally free $R$-module and the topology on $S$ is the quotient topology from a presentation (equivalently, any presentation) as a quotient of a finite type free $R$-module. In particular, as $S$ is a projective $R$-module, the inclusion map $R \rightarrow S$ admits an $R$-linear splitting and $R$ may be viewed as a subspace of $S$. A finite extension of higher local fields $F \hookrightarrow L$ satisfies the above conditions, and so does the extension of valuation rings ${\mathcal{O}}_F \hookrightarrow {\mathcal{O}}_L$. The conditions stated above are enough to guarantee the existence of the Weil restriction of $Y \rightarrow {\mathrm{Spec}\:}S$ along $R \hookrightarrow S$ [@bosch-neron-models 7.6]. Let $R \hookrightarrow S$ be an injective morphism of sequential rings, such that $S$ is a locally free, finite type $R$-module. Let $Y \rightarrow {\mathrm{Spec}\:}S$ be an affine scheme of finite type . The Weil restriction $\mathcal{R}_{S|R}(Y) = X \rightarrow {\mathrm{Spec}\:}R$ is an affine scheme of finite type. By definition, $$X(R) = \mathrm{Hom}_{{\mathbf{Sch}}_S}({\mathrm{Spec}\:}R \times_R S, Y) = \mathrm{Hom}_{{\mathbf{Sch}}_S}({\mathrm{Spec}\:}S, Y) = Y(S).$$ The two topologies we have defined on this set agree. The reason why this is true is because the topologies on $X(R)$ and $Y(S)$ already coincide before taking sequential saturations. We reproduce the argument of Conrad [@conrad-weil-grothendieck-approaches-to-adelic-points Example 2.4]. As Weil restriction respects closed immersions, it is enough to prove the result when $Y$ is an affine space. Let $P$ be a free $R$-module such that $S$ is a direct summand of it. By duality, we have a surjection of $R$-modules $$P^* = \mathrm{Hom}_R(P, R) \twoheadrightarrow \mathrm{Hom}_R(S, R) = S^*.$$ Let $A$ be any $R$-algebra. The natural map $S \otimes_R A \rightarrow P \otimes_R A$ is injective and functorially defined by a system of $R$-linear equations in $A$. Now let $M = R^n$, such that $Y = {\mathrm{Spec}\:}\left( \mathrm{Sym}_S(M') \right)$, with $M' = M \otimes_R S$. Then $X$ is in a natural way a closed subscheme of ${\mathrm{Spec}\:}\left( \mathrm{Sym}_R(M \otimes_R P^*) \right)$. Then $Y(S) = \mathrm{Hom}_S(M', S) = \mathrm{Hom}_R(M,S)$ has a natural topology as a finite type free $S$-module. Because of the inclusion $S \hookrightarrow P$, we may regard $$X(R) = \mathrm{Hom}_R(M,S) = M^* \otimes_R S \hookrightarrow M^* \otimes_R P$$ as a subspace. The topologies on $X(R)$ and $Y(S)$ agree if $S$ is a subspace of $P$. But $S$ is a direct summand of $P$ and the subspace topology agrees with the quotient topology given by the surjection $P \twoheadrightarrow S$, which is the original topology on $S$ by hypothesis. Since the two topologies on $X(R) = Y(S)$ agree, they are also equal after taking sequential saturations. The general case {#sec:ratpointsgnral} ---------------- In order to adapt the construction explained in \[sec:ratpointsaffine\] to general schemes of finite type $X \rightarrow {\mathrm{Spec}\:}R$, with $R$ a sequential ring, our argument has two steps. First, we take an affine open cover of $X$ and use Theorem \[thm:topratptsaffine\] to topologize the affine open sets in the cover. Second, we need a compatibility condition in order to be able to define a topology on $X$. There are a few obstacles before we can apply this argument, which will be sorted out by assuming further properties on the sequential ring $R$. Namely, we need to show that for any open immersion of affine schemes $U \rightarrow X$, we have an open embedding of topological spaces $U(R) \hookrightarrow X(R)$. For a general ring $R$ (even for a topological ring) this property will not hold in general and $U(R) \rightarrow X(R)$ may even fail to be a topological embedding. There is one elementary example that illustrates this situation. Suppose that $R$ is a topological ring and consider $U = \mathbb{G}_m$ as the complement of the origin in the affine line $\mathbb{A}_R^1$. The map $U(R) \rightarrow {\mathbb{A}}_R^1(R)$ is the inclusion $R^\times \rightarrow R$, which is not an embedding unless inversion on $R^\times$ is continuous, since the topology on $\mathbb{G}_m$ is given by Theorem \[thm:topratptsaffine\]. From this, we gather two necessary conditions that need to be imposed on a topological ring $R$, if we want to be able to use our argument. The first of these conditions is continuity of the inversion map, so that $R^\times \rightarrow R$ is an embedding. The second condition is that $R^\times$ is an open subset of $R$, so that the embedding is open. For a sequential ring, the above example gives the inclusion $R^\times \subset R$, where both sets have the topologies described in \[sec:ratpointsaffine\], which are sequentially saturated. In order for this inclusion to be an open embedding, it is enough to require inversion on $R$ to be sequentially continuous and the subset $R^\times$ to be open in $R$. There is another aspect we need to worry about. Whenever $X = \cup_i U_i$ is an affine cover, we need $X(R)$ to be covered by the subsets $U_i(R)$. This will be the case whenever $R$ is local. The three conditions imposed on $R$ (sequential continuity of inversion, openness of the unit group and being local) are analogous to the restrictions for the similar argument to work when $R$ is a topological ring [@conrad-weil-grothendieck-approaches-to-adelic-points Proposition 3.1]. The above conditions are sufficient to extend Theorem \[thm:topratptsaffine\] to general schemes of finite type. \[thm:topratptsgeneral\] Let $R$ be a local sequential ring such that $R^\times \subset R$ is open and such that inversion $R^\times \rightarrow R^\times$ is sequentially continuous. There is a unique covariant functor $${\mathbf{Sch}}_R \to {\mathbf{Seq}}$$ which carries open (resp. closed) immersions of schemes to open (resp. closed) topological embeddings, fibred products to products, and giving $\mathbb{A}_R^1(R) = R$ the sequential saturation of its topology. This agrees with the construction in Theorem \[thm:topratptsaffine\] for affine schemes. If $R$ is Hausdorff and $X \rightarrow {\mathrm{Spec}\:}R$ is separated and of finite type, then $X(R)$ is Hausdorff. We remark that under the hypotheses on $R$, the inclusion $R^\times \hookrightarrow R$ is an open embedding in ${\mathbf{Seq}}$. The key to the argument is showing that whenever $U \rightarrow X$ is an open immersion of affine schemes of finite type over $R$, then $U(R) \hookrightarrow X(R)$ is an open embedding. It is enough to show this in the case in which $U$ is a principal open set, so $X = {\mathrm{Spec}\:}A$, $U = {\mathrm{Spec}\:}A_f$, for some $f \in A$. Then, $U(R)$ is the preimage of $R^\times$ under the map $X(R) \rightarrow R$ associated to $f$. The fibre square $$\label{eqn:topratptsgeneral} \xymatrix{ U \ar[r]^f \ar[d] & \mathbb{G}_m \ar[d] \\ X \ar[r] & {\mathbb{A}}_R^1 }$$ reduces the problem to the special case $U = \mathbb{G}_m$, $X = {\mathbb{A}}_R^1$, because all schemes involved in (\[eqn:topratptsgeneral\]) are affine and, in such case, we already have compatibility with fibred products. But, after saturating, $R^\times \rightarrow R$ is an open embedding by hypothesis. If $X = \cup_i U_i$ is an open cover, then $X(R) = \cup_i U_i\left( R \right)$, because a map ${\mathrm{Spec}\:}R \rightarrow X$ that carries the closed point into $U_i$ must land entirely inside $U_i$, since the only open subscheme of ${\mathrm{Spec}\:}R$ that contains the closed point is the entire space. Let $R$ be a sequential ring and $G \rightarrow {\mathrm{Spec}\:}R$ a finite type group scheme. The structure morphisms $$\begin{aligned} \mu: G \times_R G \longrightarrow G, \\ \varepsilon: {\mathrm{Spec}\:}R \longrightarrow G, \\ \iota: G \longrightarrow G, \end{aligned}$$ together with their properties, turn the set $G(R)$ into a group by taking $R$-points. Since the topology we have constructed is functorial, $G(R)$ becomes a topological group. Let $F$ be a higher dimensional local field equipped with a higher topology. On one hand, as any higher topology is Hausdorff, $F^\times$ is open in $F$, as it is the complement of a singleton. On the other hand, inversion is sequentially continuous. Finally, a field is a local ring. It is because of this that $F$ satisfies the hypotheses of Theorem \[thm:topratptsgeneral\]. Since the residue map $\rho: {\mathcal{O}}_F \rightarrow \overline{F}$ is continuous and the topology on $\overline{F}$ is Hausdorff, the maximal ideal ${\mathfrak{p}}_F = \rho^{-1}(\left\{ 0 \right\})$ is a closed set. Since ${\mathcal{O}}_F$ is local, ${\mathcal{O}}_F^\times = {\mathcal{O}}_F \setminus {\mathfrak{p}}_F$ is open in ${\mathcal{O}}_F$. Inversion on ${\mathcal{O}}_F^\times$ is sequentially continuous, and this means that Theorem \[thm:topratptsgeneral\] may also be applied to the ring ${\mathcal{O}}_F$. By iteration of the argument in the previous remark, Theorem \[thm:topratptsgeneral\] may also be applied to $O_F$ and to any of the higher rank valuation rings of $F$, displayed in (\[eqn:higherranksubrings\]). For the reasons exposed above, the results in this section also apply to archimedean higher local fields, higher fields over perfect fields and their higher rank rings of integers. Regarding the behaviour of this topology under base change, Proposition \[prop:basechange\] holds if we remove the affineness condition on $X$. The reason is that in order to check the conditions stated, we may restrict ourselves to an affine open subscheme, for which Proposition \[prop:basechange\] is valid. For the sake of completeness, we state this result as a proposition. \[prop:basechangegnrl\] Let $R \rightarrow S$ be a morphism of sequential rings with $R$ and $S$ satisfying the hypotheses of Theorem \[thm:topratptsgeneral\]. Let $X \rightarrow {\mathrm{Spec}\:}R$ be a scheme of finite type. Then, the map $X(R) \rightarrow X(S)$ is continuous. Moreover, if $R \rightarrow S$ is an open (resp. closed) immersion, then $X(R) \rightarrow X(S)$ is also an open (resp. closed) immersion. Future work {#sec:future} =========== #### Reduction maps. These are a particular case of base change. For a higher local field $F$ of dimension $n \geq 1$ endowed with a higher topology, $\overline{F}$ also is endowed with a higher topology (if $\overline{F}$ is finite, we consider the discrete topology), and the topologies on $F$ and $\overline{F}$ are compatible. If $X \to {\mathrm{Spec}\:}{\mathcal{O}}_F$ is of finite type, flat and irreducible, then the reduction map $$\rho: X({\mathcal{O}}_F) \rightarrow X_{\overline{F}}(\overline{F})$$ is surjective [@liu-algebraic-geometry Prop. 10.1.36]. Besides the map $\rho$ being continuous by Proposition \[prop:basechangegnrl\], one can say more in certain cases. For example, if $X = {\mathbb{A}}^1_{{\mathcal{O}}_F}$, then $\rho$ is open: we are dealing with the map $\rho: {\mathcal{O}}_F \to \overline{F}$ with respect to the sequential topologies and, since these are group topologies, it is enough to check that the image of a sequentially open neighbourhood of zero in ${\mathcal{O}}_F$ is sequentially open in $\overline{F}$. The obvious argument works because we are able to choose very particular lifts of elements of a sequence converging to zero in $\overline{F}$. The same argument could be adapted to the case $X={\mathbb{A}}^n_{{\mathcal{O}}_F}$. A general argument along these lines would not be possible: affine schemes may be viewed as closed embeddings into affine spaces and the restriction of an open map to a closed subspace is not necessarily open. It would be very interesting to know if the two topologies on either side of the map $\rho$ are related. More precisely: is $\rho$ open, perhaps under certain conditions on $X$? If not, is it at least a quotient mapping? #### Relation between topology and integration. One of the goals of the higher adelic programme is to generalize the techniques used for the study of local fields to higher local fields. In this direction, Fesenko [@fesenko-aoas1] and Morrow [@morrow-integration-on-valuation-fields], [@morrow-integration-on-product-spaces] have introduced and developed a two-dimensional measure and a theory of harmonic analysis on two-dimensional local fields, generalising the local results of [@tate-thesis]. The global counterpart to this theory has been developed by Fesenko in [@fesenko-aoas2]. The extension of these methods to sets of rational points over higher local fields and the development of a new integration theory on such spaces would be a very important achievement in the direction of a better understanding of the arithmetic of higher dimensional schemes. More precisely, harmonic analysis on rational points over higher local fields would be very helpful for the study of sets of rational points over higher adelic rings and establishment of a theory of higher dimensional Tamagawa numbers. It seems that a good understanding of the topology of these spaces, as well as the relation between topology and two-dimensional measure, could be an important contribution in this direction. #### Points over higher adelic rings. There are several higher adelic objects which may be defined as a restricted product of higher local fields. In the work of Fesenko in dimension two [@fesenko-aoas2] there are several rings that may be realised as a restricted product of higher local fields and rings. In Beilinson’s general simplicial approach to higher adeles on noetherian schemes [@huber-parshin-beilinson-adeles], some of the adelic objects described are also definable in terms of restricted products of higher local fields. It seems that an approach to endowing sets of rational points over such higher adelic rings with topologies would require the use of the results in this work as a starting point. #### Representation theory of algebraic groups over two-dimensional local fields. Such representation theory has been developed by Gaitsgory-Kazhdan [@gaitsgory-kazhdan-reps-alg-grps-over-2ldf] and Braverman-Kazhdan [@braverman-kazhdan-examples-hecke-algs-for-2ldfs] among other works. Algebraic groups over two dimensional local fields and their central extensions are a generalization of formal loop groups and are related to a generalization of a class of affine Kac-Moody groups. Hence the results in this work can find applications in the corresponding representation theory. .5cm [10]{} A. A. Be[ĭ]{}linson. How to glue perverse sheaves. In [*[$K$]{}-theory, arithmetic and geometry ([M]{}oscow, 1984–1986)*]{}, volume 1289 of [*Lecture Notes in Math.*]{}, pages 42–51. Springer, Berlin, 1987. S. Bosch, W. L[ü]{}tkebohmert, and M. Raynaud. , volume 21 of [*Ergebnisse der Mathematik und ihrer Grenzgebiete (3) \[Results in Mathematics and Related Areas (3)\]*]{}. Springer-Verlag, Berlin, 1990. A. Braverman and D. Kazhdan. Some examples of [H]{}ecke algebras for two-dimensional local fields. , 184:57–84, 2006. B. Conrad. eil and [G]{}rothendieck approaches to adelic points. Available at [ http://math.stanford.edu/$\sim$conrad/papers/adelictop.pdf]{}. V. Drinfeld. Infinite-dimensional vector bundles in algebraic geometry: an introduction. In [*The unity of mathematics*]{}, volume 244 of [*Progr. Math.*]{}, pages 263–304. Birkhäuser Boston, Boston, MA, 2006. I. Fesenko. Analysis on arithmetic schemes. [I]{}. , (Extra Vol.):261–284 (electronic), 2003. Kazuya Kato’s fiftieth birthday. I. Fesenko. Analysis on arithmetic schemes. [II]{}. , 5(3):437–557, 2010. I. Fesenko and M. Kurihara, editors. , volume 3 of [*Geometry & Topology Monographs*]{}. Geometry & Topology Publications, Coventry, 2000. Papers from the conference held in M[ü]{}nster, August 29–September 5, 1999. I. B. Fesenko. Class field theory of multidimensional local fields of characteristic [$0$]{} with residue field of positive characteristic. , 3(3):165–196, 1991. I. B. Fesenko. A multidimensional local theory of class fields. [II]{}. , 3(5):168–189, 1991. I. B. Fesenko. Abelian local [$p$]{}-class field theory. , 301(3):561–586, 1995. I. B. Fesenko. Sequential topologies and quotients of [M]{}ilnor [$K$]{}-groups of higher local fields. , 13(3):198–221, 2001. S. P. Franklin. Spaces in which sequences suffice. , 57:107–115, 1965. S. P. Franklin. Spaces in which sequences suffice. [II]{}. , 61:51–56, 1967. D. Gaitsgory and D. Kazhdan. Representations of algebraic groups over a 2-dimensional local field. , 14(3):535–574, 2004. A. Goreham. equential convergence in topological spaces. . E. Hrushovski and D. Kazhdan. Integration in valued fields. In [*Algebraic geometry and number theory*]{}, volume 253 of [ *Progr. Math.*]{}, pages 261–405. Birkhäuser Boston, Boston, MA, 2006. A. Huber. On the [P]{}arshin-[B]{}eĭlinson adèles for schemes. , 61:249–273, 1991. M. Kapranov. Double affine [H]{}ecke algebras and 2-dimensional local fields. , 14(1):239–262 (electronic), 2001. K. Kato. A generalization of local class field theory by using [$K$]{}-groups. [I]{}. , 26(2):303–376, 1979. K. Kato. A generalization of local class field theory by using [$K$]{}-groups. [II]{}. , 27(3):603–683, 1980. K. Kato. A generalization of local class field theory by using [$K$]{}-groups. [III]{}. , 29(1):31–43, 1982. H. H. Kim and K.-H. Lee. Spherical [H]{}ecke algebras of [$\rm SL_2$]{} over 2-dimensional local fields. , 126(6):1381–1399, 2004. Q. Liu. , volume 6 of [ *Oxford Graduate Texts in Mathematics*]{}. Oxford University Press, Oxford, 2002. Translated from the French by Reinie Ern[é]{}, Oxford Science Publications. A. I. Madunts and I. B. Zhukov. Multidimensional complete fields: topology and other basic constructions. In [*Proceedings of the [S]{}t. [P]{}etersburg [M]{}athematical [S]{}ociety, [V]{}ol. [III]{}*]{}, volume 166 of [*Amer. Math. Soc. Transl. Ser. 2*]{}, pages 1–34, Providence, RI, 1995. Amer. Math. Soc. M. Morrow. Integration on product spaces and [${\rm GL}_n$]{} of a valuation field over a local field. , 2(3):563–592, 2008. M. Morrow. Integration on valuation fields over local fields. , 33(1):235–281, 2010. L. [P]{}revidi. ocally compact objects in exact categories. . M. Spiess. Class formations and higher-dimensional local class field theory. , 62(2):273–283, 1997. N. E. Steenrod. A convenient category of topological spaces. , 14:133–152, 1967. J. T. Tate. Fourier analysis in number fields, and [H]{}ecke’s zeta-functions. In [*Algebraic [N]{}umber [T]{}heory ([P]{}roc. [I]{}nstructional [C]{}onf., [B]{}righton, 1965)*]{}, pages 305–347. Thompson, Washington, D.C., 1967. A. Yekutieli. An explicit construction of the [G]{}rothendieck residue complex. , (208):127, 1992. With an appendix by Pramathanath Sastry. [^1]: The author is supported by a Doctoral Training Grant at the University of Nottingham. [^2]: a set-theoretic section of the residue homomorphism with some special properties, see [@madunts-zhukov-topology-hlfs Lemma 1.2] for the definition. Indeed, obtaining this [*canonical*]{} lifting is the most difficult step in the construction of a higher topology

ArXiv

--- abstract: 'Inspired by the recently proposed Magnetic Resonance Fingerprinting (MRF) technique, we develop a principled compressed sensing framework for quantitative MRI. The three key components are: a random pulse excitation sequence following the MRF technique; a random EPI subsampling strategy and an iterative projection algorithm that imposes consistency with the Bloch equations. We show that theoretically, as long as the excitation sequence possesses an appropriate form of persistent excitation, we are able to accurately recover the proton density, $\operatorname*{T1}$, $\operatorname*{T2}$ and off-resonance maps simultaneously from a limited number of samples. These results are further supported through extensive simulations using a brain phantom.' author: - 'Mike Davies, Gilles Puy, Pierre Vandergheynst and Yves Wiaux' title: A Compressed Sensing Framework for Magnetic Resonance Fingerprinting --- Compressed sensing, MRI, Bloch equations, manifolds, Johnston-Linderstrauss embedding Introduction ============ Inspired by the recently proposed procedure of Magnetic Resonance Fingerprinting (MRF), which gives a new technique for quantitative MRI, we investigate this idea from a compressed sensing perspective. While MRF itself, was inspired by the recent growth of compressed sensing (CS) techniques in MRI [@Ma-MRF2013], the exact link to CS was not made explicit, and the paper does not consider a full CS formulation. Indeed the role of sparsity, random excitation and sampling are not clarified. The goal of this current paper is to make the links with CS explicit, shed light on the appropriate acquisition and reconstruction procedures and hence to develop a full compressed sensing strategy for quantitative MRI. In particular, we identify separate roles for the pulse excitation and the subsampling of $k$-space. We identify the Bloch response manifold as the appropriate low dimensional signal model on which the CS acquisition is performed, and interpret the “model-based” dictionary of [@Ma-MRF2013] as a natural discretization of this response manifold We also discuss what is necessary in order to have an appropriate CS-type acquisition scheme. Having identified the underlying signal model we next turn to the reconstruction process. In [@Ma-MRF2013] this was performed through pattern matching using a matched filter based on the model-based dictionary. However, this does not offer the opportunity for exact reconstruction, even if the signal is hypothesised to be 1-sparse in this dictionary due to the undersampling of $k$-space. This suggests that we should look to a model based CS framework that directly supports such manifold models [@Baraniuk-Wakin2009]. Recent algorithmic work in this direction has been presented by Iwen and Maggioni [@Iwen-Maggioni-2012], however, their approach is not practical in the present context as the computational cost of their scheme grows exponentially with the dimension of the manifold. Instead, we leverage recent results from [@TB-2011] and develop a recovery algorithm based on the Projected Landweber Algorithm (PLA). This method also has the appealing interpretation of an iterated refinement of the original MRF scheme. The remainder of the paper is set out as follows. We begin by giving a brief overview of MRI acquisition. Then we discuss the challenges of quantitative imaging in MRI and review the recently proposed MRF scheme [@Ma-MRF2013]. We next develop the detailed mathematical model associated with MRF acquisition which leads us to the voxel-wise Bloch response manifold model observed through a sequence of partially sampled $k$-space measurements. In §\[sec: CS-MRF\], using the MRF acquisition model, we set out a framework for a compressed sensing solution to the quantitative MRI problem followed by a simple extension that provides a degree of spatial regularization. In the simulation section we demonstrate the efficacy of our methods on an anatomical brain phantom [@Collins-1998], available at the BrainWeb repository [@Brainweb]. Our results show that our CS method offers substantial gains in reconstruction accuracy over the original MRF matched filter scheme [@Ma-MRF2013]. We also demonstrate the efficiency of the proposed algorithm in terms of speed of convergence and the empirical trade-off between undersampling in $k$-space and excitation sequence length. Finally, we summarize what we have learnt by placing the MRF procedure within a CS framework and highlight a number of open questions and research challenges. Magnetic Resonance Imaging Principles ===================================== MRI, with its ability to image soft tissue, provides a very powerful imaging tool for medicine. The basic principles of MRI lie in the interaction of proton spins with applied magnetic fields. While a full review of these principles is beyond the scope of this paper, following [@Wright1997], we now introduce the basics required in order to understand the motivation for the proposed acquisition scheme and the subsequent mathematical models. For a more detailed treatment of MRI from a signal processing perspective we refer the reader to one of the excellent reviews on the subject, such as [@Wright1997; @Fessler2010]. Bloch Equations --------------- The main source of the measured signal in MRI comes from the magnetic moments of the proton spins. In a single volume element (voxel) the net magnetization $\m = (m^x,m^y,m^z)^T$ is the vector sum of all the individual dipole moments within the voxel. If there is no magnetic field then at equilibrium the net magnetization is zero. If a static magnetic field, ${\bf B}_0$ (usually considered to lie in the $[0,0,1]^T$ direction), is then applied the spins align with this field and the net magnetization at equilibrium, $\m_{\mbox{eq}}$, is proportional to the proton density $\rho$ within the volume. However, equilibrium is not achieved immediately after the field is applied, but is controlled by the longitudinal relaxation time, $\operatorname*{T1}$, such that the net magnetization at time $t$ is given by: $m^z(t) = m_{\mbox{eq}} (1- \exp(-t/\operatorname*{T1}))$. If there is magnetization in the plane orthogonal to ${\bf B}_0$ then the magnetization, $\{m^x,m^y\}$, precesses about the $z$ axis at a frequency called the Lamor frequency, $\omega_L = \gamma |{\bf B}_0|$ (approximately $42.6$MHz per Tesla), where the quantity $\gamma$ is called the gyromagnetic ratio. This in turn emits an electromagnetic signal which is the signal that is measured. As the individual dipoles dephase the net transverse magnetization decays exponentially at a rate $\operatorname*{T2}$, called the transverse relaxation time. In MRI the magnetic field is composed of a static magnetic field and a dynamic component which is manipulated through a radio frequency (RF) coil aligned with the $x$ direction. When a transverse magnetic field is applied via an RF pulse the proton dipoles rotate about the applied magnetic field. The overall macroscopic dynamics of the net magnetization can be summarized by a set of linear differential equations called the Bloch equations [@Wright1997]: $$\frac{\partial \m(t)}{\partial t} = \m(t) \times \gamma {\bf B}(t) -\left(\begin{array}{c} m^x(t)/\operatorname*{T2}\\ m^y(t)/\operatorname*{T2}\\ (m^z-m_{\mbox{eq}})/\operatorname*{T1}\end{array} \right)$$ The response at a given readout time ($\operatorname*{TE}$) from an initial RF pulse can be determined by integrating these equations over time. When a specific sequence of pulses is applied (assuming pulse length $\ll \operatorname*{T1}, \operatorname*{T2}$) then the dynamics of the magnetization from pulse to pulse or readout to readout can be described simply by a three dimensional discrete time linear dynamical system [@Jaynes1955]. Spatial Encoding and Image Formation ------------------------------------ In order to produce an image it is necessary to spatially encode the magnetization in the received signal. This is done through the application of various magnetic gradients. First, a slice can be selected through the application of a magnetic gradient along the $z$ direction, while appropriately restricting the frequency band of the excitation pulses. The gradient changes the Larmor frequency as a function of $z$, and only those positions that are excited by the pulses generate a magnetization in the transverse plane. In order to encode the transverse magnetization spatially at the acquisition time (called the echo time (TE)) the magnetic field can be modified further to have gradients $G_x$ and $G_y$ in the $x$ and $y$ directions. For example, if a linear gradient is applied along the $x$ direction so that $B^z = (B_0+G_x x)$, then the spatial variation of the transverse magnetization is encoded in the Larmor frequency and hence in the frequency of the received signal (it is assumed that the duration of the signal read out time is sufficiently short such that the magnetization can be treated as stationary). The received signal therefore corresponds to a line in the spatial Fourier transform, known as $k$-space, of the transverse magnetization. By careful selection of $G_x$ and $G_y$ it is possible to sample different lines of $k$-space until it is adequately sampled. The most popular technique is to take measurements, which we denote $\y$, that sample $k$-space on a Cartesian grid, so that the image can be formed by the application of the inverse 2D Discrete Fourier transform (DFT), $F$. Thus we can generate a discrete image $\x$ (represented here in vector form) by using $\x=F^H \y$, where $H$ denotes the conjugate transpose. For simplicity, unless stated otherwise, we will work with this discrete representation and assume that samples in $k$-space have been taken on the Cartesian grid. Rapid Imaging ------------- A key challenge in MRI is acquiring the signals in a reasonably short time. Long scan times are costly, unpopular with patients and can introduce additional complications such as motion artefacts. However, the set up described so far for MRI requires the application of repeated excitation pulses and gradients to acquire the multiple lines of $k$-space. Furthermore, after each acquisition sufficient time must be left in order for the magnetization to achieve equilibrium once again. One way to accelerate the imaging is to acquire more samples from $k$-space per acquisition. By varying the transverse gradients $G_x$ and $G_y$ as a function of time it is possible to generate more sophisticated sampling patterns. For example, in echo-planar imaging (EPI) [@McKinnon:; @multi-shot; @EPI] multiple lines of $k$-space are acquired at each pulse. Another strategy is to generate spiral trajectories in $k$-space. However, in both cases as the readout time gets longer artefacts are introduced by variation in the transverse magnetization over the read out time. Furthermore, in the case of spiral and other non-Cartesian trajectories there is the added complication of requiring more complicated image formation algorithms, such as gridding techniques [@Bydder2007], that attempt to approximate the pseudo-inverse of the non-uniform Fourier transform [@Fessler2003]. A second approach to rapid imaging is to take fewer samples. Since the emergence of compressed sensing in MRI [@Lustig2008], the idea of subsampling $k$-space has become very popular. Compressed Sensing exploits the fact that the image being acquired can be approximated by a low dimensional model, e.g. sparse in the spatial or wavelet domain. Then, under certain circumstances, the image can be recovered from a subsampled $k$-space using an appropriate iterative reconstruction algorithm. Parallel imaging techniques can also be used in conjunction with the above strategies to provide further acceleration. However, these are outside the scope of the current work. Quantitative MRI ---------------- Rather than simply forming an image that measures the transverse magnetization response from a single excitation, the aim of quantitative imaging is to provide additional physiological information by estimating the spatial variation of one or more of the physical parameters that control the Bloch equations, specifically: $\operatorname*{T1}$, $\operatorname*{T2}$ and proton density. These can help in the discrimination of different tissue types and provide useful information in numerous application areas, such as diffusion and perfusion imaging. The standard approach to parameter estimation is to acquire a large sequence of images in such a way that for each voxel the sequence of values is dependent on either $\operatorname*{T1}$ and/or $\operatorname*{T2}$ as well as certain nuisance parameters. For example, the most common techniques acquire a sequence of images at different echo times from an initial excitation pulse. For $\operatorname*{T1}$ estimation, this is typically an inversion recovery pulse (full $180^{\circ}$ rotation of the magnetic field) and for $\operatorname*{T2}$ it is a spin-echo pulse ($90^{\circ}$ rotation). The image sequences encode the exponential relaxation and the parameter of interest can be estimated by fitting an exponential curve to each voxel sequence. Another approach [@Deoni-DESPOT-2003] uses a set of well tailored steady state sequences, such that each voxel sequence encodes the relevant parameter values. Such techniques require the acquisition of multiple lines for multiple images, and it is very challenging to achieve within a reasonable time and with an acceptable signal-to-noise ratio (SNR) and resolution. Recently there have been a number of papers attempting to address this problem taking a compressed sensing approach [@Block2009; @Doneva2010; @Zhao2013; @Tran-Gia2013; @Huang2013]. All these techniques accelerate the parameter acquisition using an exponential fitting model combined with only partially sampling $k$-space for each image. Model-based optimization algorithms [@Fessler2010] are then used to retrieve the parameter values. However, while such approaches take their inspiration from compressed sensing and exploit sparse signal models, these techniques mainly focus on the development of novel reconstruction algorithms, and do not tackle the fundamental issue of how to design the acquisition in such a way as to meet the compressed sensing sampling criteria. In contrast to this previous body of work, here we will set out a principled compressed sensing approach to the simultaneous determination of all parameters of interest. That is, we will develop an acquisition framework that can be shown to satisfy the compressed sensing criteria thereby enabling us to develop a model based parameter estimation algorithm with exact recovery guarantees. The basis of our acquisition scheme is the recently proposed ‘magnetic resonance fingerprinting’ technique [@Ma-MRF2013] which we describe next. Magnetic Resonance Fingerprinting ================================= In the recent paper [@Ma-MRF2013] a new type of MRI acquisition scheme is presented that enables the quantification of multiple tissue properties simultaneously through a single acquisition process. The procedure is composed of 4 key ingredients: 1. The material magnetization is excited through a sequence of random RF pulses. There is no need to wait for the signal to return to equilibrium between pulses or for the response to reach a steady state condition as in other techniques. 2. After each pulse the response is recorded through measurements in $k$-space. Due to the time constraints only a proportion of $k$-space can be sampled between each pulse. In [@Ma-MRF2013] this is achieved through Variable Density Spiral (VDS) sampling. 3. A sequence of magnetization response images is formed using gridding to approximate the least square solution. These images suffer from significant aliasing due to the high level of undersampling. 4. Parameter maps (proton density, $\rho$, $\operatorname*{T1}$, $\operatorname*{T2}$ and off-resonance,[^1] $\delta f$) are formed through a pattern matching algorithm that matches the alias-distorted magnetization response sequences per voxel to the response predicted from the Bloch equations. Below we develop the relevant mathematical models for the MRF acquisition system that will allow us to develop a full CS strategy for quantitative MRI. Pulse excitation and the Bloch response manifold ------------------------------------------------ The MRF process is based upon an Inversion Recovery Steady State Free Precession (IR-SSFP) pulse sequence.[^2] The dynamics of the magnetization for each voxel, assuming a single chemical composition, are described by the response of the Bloch equations when ’driven’ by the excitation parameters. Let $i = 1, \ldots, N$ index the voxels of the imaged slice. The MRF excitation generates a magnetization response that can be observed (or at least partially observed) at each excitation pulse. The magnetization at a given voxel at the $l$th echo time is then a function of the excitation parameters of the $l$th excitation pulse, the magnetization at the $(l-1)$th echo time, the overall magnetic field and the unknown parameters associated with the given voxel. The overall dynamics can be described by a parametrically excited linear system and are summarized in appendix \[app: bloch dynamics\]. The magnetization dynamics at voxel $i$ are parameterized by the voxel’s parameter set ${\bf \theta }_i = \{\operatorname*{T1}_i,\operatorname*{T2}_i,\delta f_i \} \in \M$, where $\M \subset \R^3$ denotes the set of feasible values for $\theta_i$, and the voxel’s proton density, $\rho_i$. The magnetization response dynamics are also characterized by the excitation parameters of the $l$th pulse, namely the flip angle, $\alpha_l$, and the repetition time, $\operatorname*{TR}_l$. Now and subsequently we will denote the magnetization image sequence by the matrix $X$, with $X_{i,l}$ denoting the magnetization for voxel $i$ at the $l$th read out time. Note we are representing the response image at a the $l$th readout by a column vector which we denote as: $X_{:,l}$, using a Matlab style notation for indexing. Similarly, we will denote the magnetization response sequence for a given voxel $i$ as $X_{i,:}$. Given the initial magnetic field, the initial magnetization of any voxel is known up to the unknown scaling by its proton density $\rho_i$. Thus the magnetization response at any voxel can be written as a parametric nonlinear mapping from $\{\rho_i, {\theta}_i\}$ to the sequence, $X_{i,:}$: $$\label{eq: bloch map} X_{i,:} = \rho_i B({\bf \theta}_i; \alpha, \operatorname*{TR}) \in \C^{1\times L}.$$ Here $ \rho_i \in \R_+$ is the proton density at voxel $i$, $L$ is the excitation sequence length and $B$ is a smooth mapping induced by the Bloch equation dynamics: $B:{ \cal{M}} \rightarrow \C^{1\times L}$, where its smoothness can be deduced by the smooth dependence of the dynamics and with respect to $\theta_i$. In order to be able to retrieve the Bloch parameters ${\bf \theta}_i$ and proton density from $X_{i,:}$ it is necessary that the excitation sequence is “sufficiently rich” such that the voxel’s magnetization response can be distinguished from a response with different parameters. Mathematically this means that there is an embedding of $\R_+ \times \M$ into $\C^L$.[^3] We will call $\B = B(\M; \alpha, \operatorname*{TR}) \subset \C^L$ the Bloch response manifold and denote its cone by $\R_+\B$. Note that this component of the MRF procedure is not compressive, as the mapping will typically need to map to a higher dimension than $\dim (\R_+\M)$ in order to induce an embedding. The primary role of the excitation sequence is therefore to ensure identifiability and this can typically be achieved through random excitation as is commonly used in system identification. We will see, however, that the excitation sequence will also need to induce a sufficiently persistent excitation in order for it to be observed in a compressive manner. The aim of a good excitation sequence should be to minimize the time taken to acquire the necessary data rather than minimizing the total number of samples. To this end, the total acquisition time for the sequences, $\sum_l \operatorname*{TR}_l$ is the relevant cost. Here, while more samples may be taken in MRF in comparison with other quantitative techniques the benefit comes in not having to wait for the magnetization to relax to its equilibrium state between samples. \[rem: proton density\] While it is clear that the proton density, $\rho_i$, will necessarily be real valued and non-negative, it is common practice in MRI to allow this quantity to absorb additional phase terms due to, for example, coil sensitivity or timing errors. Therefore $\rho_i$ is often allowed to take a complex value. In this work we will retain the idealized model, treating it as non-negative real, however, we note that the subsequent theory presented here can typically be easily modified to work with $\rho_i \in \C$ instead of $\rho_i \in \R_+$, albeit with an increase in the dimensionality of the unknown parameter set. We will highlight specific differences along the way. MRF imaging and $k$-space sampling ---------------------------------- So far we have considered the signal model for a single voxel. For a complete spatial image, assuming a discretization into $N$ voxels and treating each voxel as independent we have ${\bf \theta} \in \M^N$ and $\rho \in \R_+^N$. Similary $X \in \C^{N\times L}$. We can therefore define the full response mapping, $X = f(\rho,\theta)$, $f: \R_+^N \times \M^N \rightarrow (\R_+\B)^N \subset \C^{N\times L}$, as: $$\label{eq: full response map} X = f(\rho,\theta) = [\rho_1 B({\bf \theta}_1; \alpha, \operatorname*{TR}), \ldots, \rho_N B({\bf \theta}_N; \alpha, \operatorname*{TR})]^T.$$ Unfortunately, it is impractical to observe the full spatial magnetization (via $k$-space) at each repetition time within a sufficiently small time for the magnetization to remain approximately constant. It is therefore necessary to resort to some form of undersampling. Let us denote the observed sequence of $k$-space samples as $Y \in \C^{M \times L}$, such that the samples taken at the $l$th read out, $Y_{:,l} \in \C^M$ are given by: $$\label{eq: observation map} Y_{:,l} = P(l) F X_{:,l}$$ where $F$ again denotes the 2D discrete Fourier transform and $P(l)$ is the projection onto a subset of coefficients measured at the $l$th read out (although the original MRF scheme used a sequence of spiral trajectories, for simplicity we will assume that the Fourier samples are only taken from a Cartesian grid). We can finally define the full linear observation map from the spatial magnetization sequence to the observation sequence as $Y=h(X)$ where $h$ is given by: $$\label{eq: full observation map} Y = h(X) = [P(1)F X_{:,1}, \ldots, P(N)FX_{:,N}].$$ Together and define the full MRF acquisition model from the parameter maps $\operatorname*{T1}, \operatorname*{T2}, \delta f$ and $\rho$ to the observed data $Y$. MRF matched filter reconstruction {#sec: MRF reconstruction} --------------------------------- In [@Ma-MRF2013] the image sequence is first reconstructed using the regridding method [@Bydder2007] which approximates the least squares estimate for $X_{i,:}$ given $Y_{i,:}$: $$\label{eq: MRF image BP} \hat{X}_{:,l} = F^H P(t)^T Y_{:,l}$$ or equivalently $\hat{X} = h^H (Y)$. Due to the high level of undersampling, each reconstructed image contains significant aliasing. However, it is argued in [@Ma-MRF2013] that accurate estimates of the parameter maps can still be obtained by matching each voxel sequence to a predicted Bloch response sequence using a set of matched filters. This essentially averages the aliasing across the sequence, treating the aliasing as noise. While the technique provides impressive results, it ignores the main tenet of compressed sensing - that aliasing is interference and under the right circumstances can be completely removed (we explore this idea in detail in §\[sec: CS-MRF\]). Mathematically, it will be convenient to view the matched filter solution as the projection of the voxel sequence onto a discretization of the Bloch response manifold as follows. ### Sampling the Bloch response manifold {#sec: Bloch projection} Suppose that we wished to approximate the projection of the sequence $X_{i,:}$ onto the cone of the Bloch response manifold. One way to do this is to first take a discrete set of samples of the parameter space, $\M$, $\theta_i^{(k)} = \{\operatorname*{T1}_i^{(k)},\operatorname*{T2}_i^{(k)},\delta f_i^{(k)} \}$, ${k = 1, \ldots, P}$ and construct a ‘dictionary’ of magnetization responses, $D = \{D_k\}$, $D_k = B(\theta_i^{(k)}; \alpha, \operatorname*{TR})$, $k = 1, \ldots , P$. The density of such samples controls the accuracy of the final approximation of the projection operator. We can similarly construct a look-up table (LUT) to provide an inverse for $B(\theta_i; \alpha, \operatorname*{TR})$ on the discrete samples such that $\theta_i^{(k)} = \operatorname*{LUT}_B (k)$. The projection onto the cone of the discretized response manifold, $D$, can then be calculated using: $$\label{eq: bloch MF} \hat{k}_i = \operatorname*{argmax}_k \frac{\operatorname*{real}\langle D_k,X_{i,:}\rangle}{\|D_k\|_2}$$ to select the closest sample $D_{\hat{k}_i}$ and $$\label{eq: bloch MF proton density} \hat{\rho}_i = \max\{\operatorname*{real}\langle D_{\hat{k}_i},X_{i,:}\rangle/\|D_{\hat{k}_i}\|_2^2,0\}$$ for the proton density, where the real and max operations are necessary to select only positive correlations since negative $\rho_i$ are not admissible. If we allow $\rho_i$ to be complex valued (see Remark \[rem: proton density\]) then the projection equations become: $$\label{eq: bloch complex MF} \hat{k}_i = \operatorname*{argmax}_k \frac{| \langle D_k,X_{i,:}\rangle|}{\|D_k\|_2}$$ and $$\label{eq: bloch MF complex proton density} \hat{\rho}_i = \langle D_{\hat{k}_i},X_{i,:}\rangle/\|D_{\hat{k}_i}\|_2^2$$ Equations and are precisely the matched filter equations used in [@Ma-MRF2013], applied to the distorted voxel sequences. We therefore see that one interpretation of matched filtering with the MRF dictionary model is to provide an approximate projection onto the cone of the Bloch response manifold for each voxel sequence. A summary of the full MRF parameter map recovery algorithm (with a real valued proton density model) is given in Algorithm \[alg: MRF\]. **Given:** $Y$ Reconstruct $X$: $\hat{X} = h^H (Y)$ MF parameter estimation: $\hat{k}_i = \operatorname*{argmax}_k \operatorname*{real}\langle D_k,\hat{X}_{i,:}\rangle/\|D_{k}\|_2$ $\hat{\theta}_i = \operatorname*{LUT}_{\B}(\hat{k}_i)$ $\hat{\rho}_i = \max\{0,\operatorname*{real}\langle D_{\hat{k}_i},\hat{X}_{i,:}\rangle/\|D_{\hat{k}_i}\|_2^2\}$ **Return:** $\hat{\theta}, \hat{\rho}$ ### Computational cost and accuracy {#computational-cost-and-accuracy .unnumbered} Given that the discretized MRF dictionary can be very large ($\approx 500,000$ samples in [@Ma-MRF2013]), it is useful to consider the computational complexity of the above calculations as a function of parameter accuracy as this is the major computational bottleneck that we will encounter. The accuracy with which we can estimate the parameters for a given voxel will depend on the accuracy of the approximate projection operator and the Lipschitz constants of the inverse mapping, $LUT_{\B}$. We can achieve an approximate projection by generating an $\epsilon$-cover of $\B$ with $D_k$. As the dimension of $\B$ is $3$, this requires choosing $P \sim C\epsilon^{-3}$ atoms in our dictionary. Furthermore, as the projection operation described in takes the form of a nearest neighbour search, we can use fast nearest neighbour search strategies, such as the cover tree method [@Beygelzimer2006], to quickly solve in $\mathcal{O}(L\ln (1/\epsilon))$ computations per voxel, instead of the $\mathcal{O}(L\epsilon^{-3})$ necessary for exhaustive search. This effectively makes the speed of each application of $D$ on a par with that of a traditional fast transform. Similarly, the approximate inverse using $\operatorname*{LUT}_\B$ can also be computed in $\mathcal{O}( \ln (1/\epsilon))$. We could also consider enhancing such an estimate by exploiting the smoothness of the response manifold, either by using local linear approximations of the manifold [@Iwen-Maggioni-2012] or by further locally optimizing the projection numerically around the selected parameter set, once we are assured global convergence. Such an enhancement could allow either for increased accuracy or reduced computation through the use of fewer parameter samples, however, we do not pursue these ideas further here. Compressed Quantitative Imaging {#sec: CS-MRF} =============================== In order to generate a full compressed sensing framework for MRF we will identify sufficient conditions on the excitation pulse sequences and the $k$-space sampling, along with a suitable reconstruction algorithm, to guarantee recovery of the parameter maps from the observed $k$-space samples. As the dimension of our problem is large, $\dim((\R_+\times \M)^N) = 4N$, we do not consider the manifold reconstruction algorithms in [@Iwen-Maggioni-2012] as these scale poorly with the dimension of the manifold. Instead, we propose a CS solution based around the iterative projection algorithm of Blumensath [@TB-2011] which we will see has computational cost that is linear in the voxel dimension. Our approach, which we call BLIP (BLoch response recovery via Iterated Projection), has three key ingredients: a random pulse excitation sequence following the original MRF technique; a random subsampling strategy that can be shown to induce a low distortion embedding of $\R_+^N \times \M^N$ and an efficient iterated projection algorithm [@TB-2011] that imposes consistency with the Bloch equations. Moreover, the projection operation is the same nearest neighbour search described in section \[sec: Bloch projection\]. We first describe the iterative projection method and then consider the implications for the appropriate excitation and sampling strategies. Reconstruction by Iterated Projection ------------------------------------- In [@TB-2011] a general reconstruction algorithm, the Projected Landweber Algorithm (PLA) was proposed as an extension of the popular Iterated Hard Thresholding Algorithm [@TB-IHT-2008; @TB-IHT-2009]. PLA is applicable to arbitrary union of subspace models as long as we have access to a computationally tractable projection operator onto the union of subspace model within the complete signal space. The algorithm is given by: $$X^{(n+1)} = {\cal P}_{\cal A}(X^{(n)}+\mu h^H(Y-hX^{(n)}))$$ where ${\cal P}_{\cal A}$ is the orthogonal projection onto the signal model ${\cal A}$ such that $${\cal P}_{\cal A}(X) = \operatorname*{argmin}_{\tilde{X}\in {\cal A}} \|X-\tilde{X}\|_F$$ and $\mu$ is the step size. The current theory for PLA [@TB-2011] states that a sufficient condition for stable recovery of $X$ given $Y$ is that $h$ is a stable embedding - a so-called Restricted Isometry Property (RIP) or bi-Lipshitz embedding - for the signal model, ${\cal A}$. A mapping, $h$, is said to have the RIP (be a bi-Lipschitz embedding) for the signal model ${\cal A}$ if there exists a sufficiently small constant $\delta >0$ such that: $$\label{eq: bi-lip embedding} (1-\delta) \| X-\tilde{X}\|_2^2 \leq \frac{N}{M}\|h(X-\tilde{X})\|_2^2 \leq (1+\delta) \| X-\tilde{X}\|_2^2$$ for all pairs $X$ and $\tilde{X}$ in ${\cal A}$. How to achieve such an embedding will be considered later in section \[sec: excitation and sampling\]. The theory [@TB-2011] states that it is sufficient that $h$ satisfy the RIP with $\frac{M}{N}(1+\delta) < 1/\mu < \frac{3M}{2N} (1-\delta)$ for the guaranteed recovery. If $h$ is essentially ‘optimal’, e.g. a random ortho-projector, then we should set the step size $\mu \approx N/M$ since in the large system limit $\delta \rightarrow 0$. For our compressed sensing scenario the signal model ${\cal A}$ is the product set $(\R_+ \B)^N$ or, more precisely, its discrete approximation $(\R_+ D)^N$ and the projection operator ${\cal P}_{\cal A}$ can be realized by separately projecting the individual voxel sequences $X_{i,:}^n$ onto the cone of the Bloch response manifold using the equations and . Although $(\R_+ \B)^N$ is not itself a union of subspace model it can easily be extended to $(\R \B)^N$, which forms an uncountably infinite union of lines (1D subspaces). In fact, the theory of [@TB-2011] does not require ${\cal A}$ to be a union of subspace [@TB-2011] and is directly applicable to ${\cal A} = (\R_+ \B)^N$. We therefore appear to have all the ingredients for a full compressed sensing recovery algorithm. This is summarized in Algorithm \[alg:CS-MRF\]. **Given:** $Y$ **Initialization:** $X^{(0)} = \mathbf{0}$, $\mu = N/M$ Image sequence reconstruction Gradient step: $X_{:,l}^{(n+1/2)} = X_{:,l}^{(n)}+\mu F^H P(l)^T(Y_{:,l}-P(l) F X_{:,l}^{(n)})$; Projection step: $\hat{k}_i = \operatorname*{argmax}_k \operatorname*{real}\langle D_k,X_{i,:}^{(n+1/2)}\rangle/\|D_{k}\|_2$ $\hat{\rho}_i = \max\{0,\operatorname*{real}\langle D_{\hat{k}_i},X_{i,:}^{(n+1/2)}\rangle/\|D_{\hat{k}_i}\|_2^2\}$ $X_{i,:}^{(n+1)} = \hat{\rho}_i D_{\hat{k}_i}$ Parameter map estimation: $\hat{\theta}_i = \operatorname*{LUT}_{\B}(\hat{k}_i)$ **Return:** $\hat{\theta}, \hat{\rho}$ \[remark: separating CS and parameter estimation\] Note that the above procedure has separated out the parameter map estimation (by inverting the estimated Bloch responses) and the reconstruction of the magnetization image sequence (via the PLA). Indeed, as long as the partial $k$-space sampling provides a bi-Lipschitz embedding for all possible magnetization responses then the CS component of the imaging is well defined even if the Bloch response is not invertible. ### Step size selection {#sec: adaptive step size} Selection of the correct step size is crucial in order to attain good performance from these iterative projection based algorithms [@TB-NIHT-2010; @TB-2011]. Note that the original parameter estimation in [@Ma-MRF2013] can be interpreted as an application of a single iteration of PLA with a step size $\mu = 1$ and iterating PLA with this step size tends to only deliver a modest improvement over the matched filter (single iteration). The matched filter also has the effect of underestimating the magnitude of $X$, and hence also the proton density map, as $h$ tends to shrink vectors uniformly (when it provides a stable embedding). In contrast, when using the substantially more aggressive step size proposed by the theory we will see that significant improvements are observed in signal recovery and often in a very small number of iterations. In practice, it is also beneficial to select the step size for PLA adaptively to ensure stability. Following the work on adaptive step size selection for IHT [@TB-NIHT-2010] we adopt the following heuristic. We begin each iteration by choosing $\mu = {N/M}$ as is suggested from the CS theory. Then after calculating a new proposed value for $X^{n+1}$ we calculate the quantity: $$\omega = \kappa \frac{\|X^{n+1}-X^{n}\|_2^2}{\|h(X^{n+1}-X^{n})\|_2^2}$$ for some $\kappa < 1$. If $\mu > \omega$ we reject this update, shrink the step size, $\mu \mapsto \mu/2$ and calculate a new proposed value for $X^{n+1}$. As with the Normalized IHT algorithm [@TB-NIHT-2010], this form of line search is sufficient to ensure convergence of the algorithm irrespective of conditions on the measurement operator, and we will use this form of step size selection in all subsequent experiments. Strategies for subsampling $k$-space {#sec: excitation and sampling} ------------------------------------ We now consider what properties of the excitation response sequences and the $k$-space sampling pattern will ensure that the sufficient RIP conditions in the PLA theory are satisfied. First note that, as the signal model treats each voxel as independent, we need to take at least $N \dim(\R_+ \M)$ measurements as this is the dimension of our model. Furthermore, since we only take a small number of measurements at each repetition time, we cannot expect to achieve a stable embedding without imposing further constraints on the excitation response. For example, if the embedding was induced in the first few repetition times and all further responses were non-informative we would not have taken sufficient measurements from the informative portion of the response. Therefore we consider responses that somehow spread the information across the repetition times. We will assume that the excitation sequence induces an embedding for the response map (here random sequences seem to suffice), and identify additional conditions that enable us to develop a random $k$-space subsampling strategy with an appropriate RIP condition. Our approach will follow the technique of random sampling as is common in compressed sensing measurement design, along with a pre-conditioning technique that has been used in the Fast Johnson-Lindenstrauss Transform [@Ailon-FJLT2009] and in spread spectrum compressed sensing [@Puy-2012]. It is also reminiscent of the Rauhut’s bounded orthonormal systems [@Rauhut2010] and has a similar aim of ensuring that information is sufficiently spread within the measurement domain The key vectors of interest are those that discriminate between pairs of possible signals within our model, namely the *chords* of $\R_+\B$, which are the vectors of the form $u = X_{i,:}-\tilde{X}_{i,:}$ with $X_{i,:}, \tilde{X}_{i,:} \in \R_+\B$ and $X_{i,:} \neq \tilde{X}_{i,:}$. We will quantify the pre-conditioning requirement for the excitation response through the *flatness* of such vectors which we define as follows. Let $U$ be a collection of vectors $\{u\}$ in $\C^L$. We denote the *flatness*, $\lambda$, of the these vectors by: $$\lambda := \max_{u \in U} \frac{\| u\|_{\infty}}{\|u\|_2}.$$ Note that from standard norm inequalities $L^{-1/2} \leq \lambda \leq 1$. We will consider the chords of an excitation response to be sufficiently flat up to a log penalty if $\lambda \sim L^{-1/2} \log^\alpha L$ for $U = \{\R_+\B-\R_+\B\}\backslash\{0\}$. In constructing our measurement function we also note that the signal model contains no spatial structure, and therefore we should expect to have to uniformly sample $k$-space in order to achieve a sufficient RIP. Note this is in contrast with the variable density sampling strategy proposed by [@Ma-MRF2013] which concentrated samples at the centre of $k$-space. It turns out that we can achieve this using a remarkably simple random subsampling pattern based on multi-shot Echo-planar Imaging (EPI) [@McKinnon:; @multi-shot; @EPI]. Let $F\in \C^{N \times N}$ denote the 2D discrete Fourier transform (assuming an image size of $\sqrt{N} \times \sqrt{N}$) with $F_{i,:}$, $i = 1, \ldots, N$ denoting the $N$ 2D discrete Fourier basis vectors associated with the spatial frequencies $k_x(i),k_y(i) \in \{0, \ldots, \sqrt{N}-1\}$. Without loss of generality we assume that the vectors are ordered such that $k_x(i) = (i-1) \mod \sqrt{N}$, and $k_y(i) = \lfloor (i-1)/ \sqrt{N} \rfloor$. We can now define a *random Echo-Planar Imaging* measurement operator by $Y_{:,l} = P(\zeta_l)F X_{:,l}$, where $\zeta_l$ is a sequence of independent random variables uniformly drawn from $\{0, \ldots , p-1\}$ and $P(\zeta) \in \R^{M\times N}$ is defined as follows: $$P_{i,j} = \left\{ \begin{array}{ll} 1& \mbox{if~} j = i+\sqrt{N}\bigl(\zeta+(p-1)\lfloor (i-1)/\sqrt{N} \rfloor \bigr) \\ 0& \mbox{otherwise}. \end{array} \right.$$ where for convenience we have assumed that $N$ is exactly divisible by $p$ so that $M=N/p$ is an integer. In words, we uniformly subsample $k_y$ by a factor of $p$ with random shifts across time in $k_y$ of the set of $k$-space samples. This is illustrated in figure \[fig: random EPI pattern\]. {width="0.7\linewidth"} \[fig: random EPI pattern\] Random EPI, along with an excitation response with appropriate chord flatness, $\lambda$, is then sufficient to provide us with a measurement operator, $h$, that is a bi-Lipschitz embedding on our signal model. In appendix \[app: RIP proof\] we prove the following theorem: \[th: randEPI embedding\] Given an excitation response cone, $\R_+ \B$ of dimension $d_\B$, whose chords have a flatness $\lambda$ and a random EPI operator $h: (\R_+ \B)^N \rightarrow \C^{M \times L}$, then, with probability at least $1-\eta$, $h$ is a restricted isometry on $(\R_+ \B)^N - (\R_+ \B)^N$ with constant $\delta$ as long as: $$\lambda^{-2} \geq C \delta^{-2} p^2 d_\B \log ( N /\delta \eta )$$ for some constant $C$ independent of $p, N, d_\B, \delta$ and $\eta$. Specifically, if $\lambda = \mathcal{O}( L^{-1/2} \log^\alpha L)$ then we require: $$L = \mathcal{O} (\delta^{-2} p^2 d_\B \log( N/\delta \eta)\log^{\alpha}(L))$$ excitation pulses. While we might hope to get $L$ of the order of $p d_\B$ it appears that this is not possible, at least for a worst case RIP analysis based on the flatness criterion alone. Indeed, in the experimental section we will provide evidence to suggest that $L\sim p^2$ is indeed the scaling behaviour that we empirically observe. It might seem surprising that the proposed scheme uses uniform random sampling in $k$-space whereas it is usually advisable to use a variable density sampling strategy for compressed sensing solutions for MRI. Indeed, there is good theoretical justification for variable density sampling patterns [@Adcock-2011; @Puy2011]. Our theory above is not inconsistent with such results. Variable density sampling is advantageous because the underlying signal model - sparsity in the wavelet domain - is not incoherent with the Fourier basis [@Puy2011; @Adcock-2011]. However, the Fourier basis is incoherent with a voxel-wise signal model as used above. This is not to say that spatial structure cannot be effectively exploited within a compressed quantitative imaging scheme or that variable density sampling would not then be of benefit. However, as the basic MRF based model does not exploit spatial structure we argue that uniform random sampling is appropriate here. The challenge of incorporating spatial regularity into the signal model is discussed next. Extending the Bloch response model {#sec: model extension} ---------------------------------- Our current compressed sensing model takes no account of additional structure within the parameter maps. This structure could, for example, be the piecewise smoothness of the parameter maps or the magnetization response maps, or an imposed segmentation of the image into different material compositions. In general, it is not clear how such additional regularization can be included in a principled manner, although many heuristic approaches could of course be adopted, as for example in [@Doneva2010]. This is because the parameter values are encoded within the samples of the Bloch response manifold, and therefore the spatial regularity would need to be mapped through the Bloch response leading to a non-separable high dimensional nonlinear signal model. The one exception, which we consider here, is the regularization of the proton density map, or at least a close relative. We note, however, that in this instance the theory relies on the real non-negative proton density model and does not directly extend to the complex case. Let us define the *pseudo-density*, $\tilde{\rho}$ as the proton density map scaled by the norm of the Bloch response vector, so that: $$\tilde{\rho}_i = \rho_i \|B({\bf \theta}_i; \alpha, \operatorname*{TR})\|_2.$$ Similarly we can define the normalized Bloch response as: $$\eta_{i,:} = \tilde{B}({\bf \theta}_i; \alpha, \operatorname*{TR}) \triangleq B({\bf \theta}_i; \alpha, \operatorname*{TR})/\|B({\bf \theta}_i; \alpha, \operatorname*{TR})\|_2$$ and the normalized Bloch response manifold, $\tilde{\B}$ as: $$\tilde{\B} = \left\{\eta_{i,:} = \tilde{B}({\bf \theta}_i; \alpha, \operatorname*{TR}) \mbox{~ for some~} \theta_i \in \M\right\}$$ The pseudo-density will be roughly the same as the density, as long as the Bloch response sequences are all of approximately the same magnitude. The transform to $\{\tilde{\rho},\eta\}$ normalizes the manifold $\tilde{\B}$ so that we can more easily calculate projections onto product signal models of the form $\{\tilde{\rho},\eta\} \in \Sigma \times \tilde{\B}^N$, where $\Sigma$ denotes the set of spatially regularized pseudo-density maps. To do this we will find the following proposition useful: Given an $X \in \C^{N \times L}$, suppose that the projection onto the signal model $\Sigma \times \tilde{\B}^N$ is given by $\hat{\tilde{\rho}} \in \Sigma$ and $\hat{\eta}_{i,:} \in \tilde{\B}$ and results in $\hat{\tilde{\rho}}_i \geq 0$ for all $i$, then: $$\label{eq: eta proj} \hat{\eta}_{i,:} = \operatorname*{argmax}_{\eta_{i,:} \in \tilde{\B}} z_i$$ and $$\label{eq: rho proj} \hat{\tilde{\rho}} = \operatorname*{argmin}_{\tilde{\rho} \in \Sigma} \| \tilde{\rho} - z\|_2^2 $$ where $z_i = \operatorname*{real}\langle \eta_{i,:},X_{i,:} \rangle$. By definition of the orthogonal projection we have: $$\label{eq: orth proj def} \{\hat{\eta},\hat{\tilde{\rho}}\} = \operatorname*{argmin}_{\eta,\tilde{\rho}} \sum_i \sum_j |X_{i,j} - \tilde{\rho}_i \eta_{i,j}|^2$$ Expanding , substituting in $z_i$ and noting that $\|\eta_{i,:}\|_2 = 1$ we have: $$\label{eq:orth proj expansion} \{\hat{\eta},\hat{\tilde{\rho}}\} = \operatorname*{argmin}_{\eta \in {\cal{\B}},\tilde{\rho} \in \Sigma} \sum_i \left( \tilde{\rho}_i^2 - 2\tilde{\rho}_i z_i \right).$$ By assumption $\hat{\tilde{\rho}}_i$ is non-negative so the expression is minimized with respect to $\eta_{i,:}$ by independently of ${\tilde{\rho}}_i$. Finally we note that holds since: $$\sum_i \left( \tilde{\rho}_i^2 - 2\tilde{\rho}_i z_i \right) = \|\tilde{\rho}-z\|_2^2 + \mbox{const.}$$ One way to impose spatial regularity on $\tilde{\rho}$ is to force it to be sparse in the wavelet domain for some appropriate orthogonal wavelet representation, $W$, such that $c = W \tilde{\rho}$. In this case, the projection can be written as $\hat{\tilde{\rho}} = W^T \hat{c}$ with: $$\label{eq: hard threshold} \hat{c} = {\cal H}_k (W z )$$ where ${\cal H}_k$ denotes an element-wise hard thresholding [@TB-IHT-2008; @TB-IHT-2009] that retains only the largest $k$ elements. Under the non-negativity assumption the projection operator can be formed by applying followed by . This results in a simple algorithm for incorporating a degree of spatial regularization within the compressed quantitative imaging framework. In the next section we will see, however, that the inclusion of this addition spatial constraint adds little to the performance of the compressed sensing approach, suggesting that the Bloch equation constraint dominates the performance. The above calculation is only guaranteed to be valid when the resulting pseudo-density map is non-negative. In theory, applying such an operator when we incur negative values of pseudo-density could give a projection that is not optimal. However, in practice we have found that this is not a problem as we always impose non-negativity on both the pseudo-density and the correlations with the Bloch response, $z_i$, in order to ensure that the projection is physically meaningful. Experiments {#sec: experiments} =========== In order to test the efficacy of BLIP for compressed quantitative imaging we performed a set of simulations using an anatomical brain phantom, segmented into various material compositions. This provided a well defined ground truth and enabled us to demonstrate image sequence recovery and parameter map estimation as a function of the $k$-space subsampling factor and the excitation sequence lengths. Experimental Set up ------------------- The key ingredients of the experimental set up are described below. ### Anatomical Brain Phantom {#anatomical-brain-phantom .unnumbered} To develop realistic simulations that also provide a solid ground truth we have adapted the anatomical brain phantom of [@Collins-1998], available at the BrainWeb repository [@Brainweb]. A $217 \times 181$ slice (slice 40) of the crisp segmented anatomical brain was used and restricted to contain only 6 material components, listed in table \[table: material props\]. The phantom was further zero padded to make a $256\times 256$ image to simplify the computations. Since we are using the crisp segmentation the model is somewhat idealized and does not address inaccuracies associated with partial volume effects or many of the other issues with real MRI. However, it serves as a useful test bed to provide a good proof-of-concept for our proposed techniques. The material properties were chosen to be both representative of the correct tissue type [@Hornak-MRI; @webbook] and challenging: the proton densities were fixed to give little discrimination for individual parameters and were set so that there is not an exact match to the sampling of the Bloch response manifold. [|l|c||c|c|c|]{} ------------------------------------------------------------------------ Tissue & index & proton density & $\operatorname*{T1}$ (ms) & $\operatorname*{T2}$ (ms)\ ------------------------------------------------------------------------ Background & 0 & 0 & - & -\ ------------------------------------------------------------------------ CSF & 1 & 100 & 5012 & 512\ ------------------------------------------------------------------------ Grey matter & 2 & 100 & 1545 & 83\ ------------------------------------------------------------------------ White matter & 3 & 80 & 811 & 77\ ------------------------------------------------------------------------ Adipose & 4 & 80 & 530 & 77\ ------------------------------------------------------------------------ Skin/Muscle & $5/6$ & 80 & 1425 & 41\ \[table: material props\] The segmented brain is shown, colored by index, in figure \[fig: segmented brain phantom\]. ![The MNI segmented anatomical brain phantom [@Collins-1998] colored by index: $0=$Background, $1=$CSF, $2=$Grey Matter, $3=$White Matter, $4=$Fat, $5=$Muscle/Skin, $6=$Skin.\[fig: segmented brain phantom\]](brain_phantom_segmentation.eps){width="\linewidth"} ### Pulse excitation {#pulse-excitation .unnumbered} For the excitation sequences we use IR-SSFP sequences (exemplar code can be found in the supplementary material of [@Ma-MRF2013]) with random flip angles drawn from an independent and identically distributed Gaussian distribution: $$\alpha_l \sim \N(0,\sigma_\alpha^2)$$ with a standard deviation, $\sigma_\alpha = 10$ degrees. The repetition times were uniformly spaced at an interval of $10$ ms. While we also experimented with randomizing repetition times, we did not find that these significantly changed the performance of the techniques. Constant repetition time intervals also mean that we can directly assess the imaging speed in terms of the sequence length, $L$. The value of $\sigma_\alpha$ was chosen empirically to provide reasonable persistence of excitation for the expected $\operatorname*{T1}$ and $\operatorname*{T2}$ responses. Figure \[fig: sequence flatness\] (left) shows the magnitude of the response differences for the set of tissue types listed in table \[table: material props\]. It can be seen that the difference in the responses does indeed persist over time. Using these differences we can also estimate their flatness. Figure \[fig: sequence flatness\] (right) shows how the flatness varies as a function of sequence length. We see that $\lambda^{-2}$ roughly scales proportionally to $L$, as desired, with a slight downward sublinear trend. ![Left: examples of the response differences for pairs of tissue types given in table \[table: material props\] when using IR-SSFP pulse sequence excitation with random flip angles. Right: $\lambda^{-2}/L$ as a function of sequence length for the repsonse differences plotted on the left. From this plot it can be deduced that $\lambda^{-2}$ grows roughly proportionally to $L$.[]{data-label="fig: sequence flatness"}](Response_diffs.eps){width="\linewidth"} ![Left: examples of the response differences for pairs of tissue types given in table \[table: material props\] when using IR-SSFP pulse sequence excitation with random flip angles. Right: $\lambda^{-2}/L$ as a function of sequence length for the repsonse differences plotted on the left. From this plot it can be deduced that $\lambda^{-2}$ grows roughly proportionally to $L$.[]{data-label="fig: sequence flatness"}](response_diff_flatness.eps){width="\linewidth"} ### Discretized Bloch response {#discretized-bloch-response .unnumbered} The Bloch response manifold was sampled in a similar manner to [@Ma-MRF2013], however, to simplify things we have only considered variation in $\operatorname*{T1}$ and $\operatorname*{T2}$ here, assuming the off resonance frequency is equal to zero. Similar to [@Ma-MRF2013], discrete samples for $\operatorname*{T1}$ were selected to go between 100 and 2000 in increments of 20 and from 2300 to 6000 in increments of 300. $\operatorname*{T2}$ was sampled between 20 and 100 in increments of 5, from 110 to 200 in increments of 20 and from 400 to 1000 in increments of 200. This results in a dictionary of size $3379\times L$. This range of $\operatorname*{T1}$ and $\operatorname*{T2}$ values clearly spans the anticipated range for the tissue types listed in table \[table: material props\]. ### Subsampling strategy {#subsampling-strategy .unnumbered} For the $k$-space subsampling we use the random EPI sampling scheme detailed in section \[sec: excitation and sampling\]. Specifically, we fully sample the $k$-space in the $k_x$ direction while regularly subsampling the $k_y$ direction by a factor of $p$. This deterministic sampling pattern was then cyclically shifted by a random number of $k_y$ lines at each repetition time. In most experiments $p$ is set to $16$ (sampling at $6.25\%$ of Nyquist). ### Reconstruction algorithms In the experiments below we compare three distinct algorithms for reconstructing the magnetization image sequences. These are: (1) the original MRF algorithm; (2) BLIP algorithm presented in Algorithm \[alg:CS-MRF\]; and (3) BLIP with spatial regularization as detailed in section \[sec: model extension\]. For both iterative algorithms we use the adaptive step size strategy set out in section \[sec: adaptive step size\] with $\kappa = 0.99$. For the spatial regularization we use a Haar wavelet representation with hard thresholding as detailed in section \[sec: model extension\], retaining only the largest 12000 wavelet coefficients at each iteration. As the MRF algorithm (with step size equal to $1$) underestimates the value of the image sequence (and also the proton density) we include in the appropriate plots the performance of a rescaled MRF algorithm where the step size is $\mu = N/M$. Finally, in some of the plots we also include the performance for an oracle estimator. This oracle is given the fully sampled image sequence data as an input and then projects each voxel sequence onto the discretized Bloch response. In this way we can differentiate between errors associated with the Bloch response discretization and the image sequence reconstruction. Results ------- All the experiments were evaluated using a signal-to-error-ratio (SER) in decibels (dBs), calculated as $20 \log_{10} \tfrac{\|x\|_2}{\|x-\hat{x}\|_2}$ for a target signal $x$ with the estimate $\hat{x}$. For $\operatorname*{T1}$ and $\operatorname*{T2}$ this corresponds to the measures $\mbox{T}_1\mbox{NR}$ and $\mbox{T}_2\mbox{NR}$ that has been used to gauge the efficiency of relaxation time acquisition schemes [@Deoni-DESPOT-2003]. To avoid issues of estimates associated with empty voxels the errors are only calculated over regions with a non-zero proton density value. In all experiments, unless stated otherwise, the following parameters were used: the undersampling ratio for the operator $h(\cdot)$ was fixed at $1/16$ and for both the iterative algorithms a maximum of $20$ iterations was allowed, though in many cases fewer iterations would have sufficed. ### Performance as a function of excitation sequence length Our first experiment evaluates the performance of the algorithms in terms of the sequence length, which was varied between $10$ and $1000$ pulses. Here we can separately evaluate the performance of the compressed sensing component and the recovery of the parameter maps. The compressed sensing recovery performance, evaluated by the SER of the image sequence reconstruction, $X$, is shown in figure \[fig: SER vs sequence length\] (a). First, note that the strange behaviour of the oracle estimator for small sequence lengths is probably due to the failure of $f(\cdot)$ to achieve a low distortion embedding. This would result in it being easier to approximate voxel sequences with a given element of the Bloch response approximation. Beyond this the performance reaches a plateau at approximately SER $= 27$ dB which can be considered to be the error associated with the discretization of the Bloch response. The performance of both BLIP algorithms is roughly equivalent. They both sharply increase in performance at a sequence length of $100$ and then tend to a plateau beyond this with an SER of about $0.5$ dB below that of the oracle estimator. This suggests that we can achieve near perfect compressed sensing reconstruction with a sequence containing as few as 100 pulses. In this simulation there was no significant gain from the additional inclusion of the spatial regularization. The performance of MRF is significantly worse. We first highlight that the non-rescaled MRF performance is terrible, however, as noted earlier, this is mainly due to the shrinkage effect of the subsampling operator, $h(\cdot)$. Correcting for this with appropriate rescaling leads to significantly improved estimation. However, we see that the SER increases slowly as a function of sequence length, which is consistent with the argument that the matched filter is averaging over the aliasing rather than cancelling it, as presented in section \[sec: MRF reconstruction\]. Furthermore, even for a sequence length of 1000 the SER still only reaches $12$dB. Subfigures  \[fig: SER vs sequence length\] (b), (c) and (d) show the SER for the estimation of the parameter maps, proton density, $\operatorname*{T1}$ and $\operatorname*{T2}$ respectively, and reflects the combined performance of inverting both $h(\cdot)$ and $f(\cdot)$. In each case the two iterative algorithms approach the oracle performance for sequence lengths of $L \geq 200$, indicating successful parameter map recovery. Furthermore, the performance for the $\rho$ estimates and $\operatorname*{T2}$ estimates do not improve substantially beyond the $L = 200$ value as $L$ is increased reaching a plateau at approximately $16$dB which corresponds to a root mean squared (rms) error of approximately $30$ms. In contrast, the $\operatorname*{T1}$ estimation performance does increase from roughly $20$dB ($213$ms rms error) at $L=200$ to $30$dB ($67$ms rms error) at $L = 1000$. This may be a function of the isometry properties (in the $\operatorname*{T1}$ direction) for the Bloch response embedding, and is possibly related to the longer time constants of $\operatorname*{T1}$. It is an open question as to whether a better excitation sequence can be designed to improve the $\operatorname*{T1}$ estimates for small $L$. ![\[fig: SER vs sequence length\] Reconstruction performance as a function of sequence length. (a) SER for image sequence reconstruction; (b) SER for density map estimation; (c) SER for $\operatorname*{T1}$ map estimation; and (d) SER for $\operatorname*{T2}$ map estimation. Results are shown for the following algorithms: MRF, BLIP, BLIP with spatial regularization. Also shown is the performance of an oracle estimator given the full image sequence data. Finally subfigures (a) and (b) also include the performance of a rescaled MRF estimator.](Image_sequence_SER.eps){width="\linewidth"} [(a)]{} ![\[fig: SER vs sequence length\] Reconstruction performance as a function of sequence length. (a) SER for image sequence reconstruction; (b) SER for density map estimation; (c) SER for $\operatorname*{T1}$ map estimation; and (d) SER for $\operatorname*{T2}$ map estimation. Results are shown for the following algorithms: MRF, BLIP, BLIP with spatial regularization. Also shown is the performance of an oracle estimator given the full image sequence data. Finally subfigures (a) and (b) also include the performance of a rescaled MRF estimator.](Density_map_SER.eps){width="\linewidth"} [(b)]{} ![\[fig: SER vs sequence length\] Reconstruction performance as a function of sequence length. (a) SER for image sequence reconstruction; (b) SER for density map estimation; (c) SER for $\operatorname*{T1}$ map estimation; and (d) SER for $\operatorname*{T2}$ map estimation. Results are shown for the following algorithms: MRF, BLIP, BLIP with spatial regularization. Also shown is the performance of an oracle estimator given the full image sequence data. Finally subfigures (a) and (b) also include the performance of a rescaled MRF estimator.](T1_map_SER.eps){width="\linewidth"} [(c)]{} ![\[fig: SER vs sequence length\] Reconstruction performance as a function of sequence length. (a) SER for image sequence reconstruction; (b) SER for density map estimation; (c) SER for $\operatorname*{T1}$ map estimation; and (d) SER for $\operatorname*{T2}$ map estimation. Results are shown for the following algorithms: MRF, BLIP, BLIP with spatial regularization. Also shown is the performance of an oracle estimator given the full image sequence data. Finally subfigures (a) and (b) also include the performance of a rescaled MRF estimator.](T2_map_SER.eps){width="\linewidth"} [(d)]{} ### Visual Comparison To get a visual indication of the performance of the BLIP approach over the MRF reconstruction at low sequence lengths, images of the 3 different parameter estimates for $L = 300$ are given in figures \[fig: visual comparison density\], \[fig: visual comparison T1\] and \[fig: visual comparison T2\]. The left hand column shows the ground truth parameter maps while the middle row shows the MRF reconstruction (scaled) and the right hand column shows the BLIP estimates (with spatial regularization). While the main aspects of the parameter maps are visible in the MRF reconstructions, there are still substantial aliasing artefacts. These are most prominent in the $\operatorname*{T1}$ and $\operatorname*{T2}$ estimates. In contrast, the BLIP estimates are virtually distortion-free, indicating that good spatial parameter estimates can be obtained with as little as 300 excitation pulses. ![\[fig: visual comparison density\] A visual comparison of the density map estimates from a sequence of length $L=300$. The top plot shows the original density map. The middle image is the MRF estimate and the bottom image is the BLIP estimate.](visual_comparison_density.eps){height="0.99\textheight"} ![\[fig: visual comparison T1\] A visual comparison of the $\operatorname*{T1}$ map estimates from a sequence of length $L=300$. The top plot shows the original $\operatorname*{T1}$ map. The middle image is the MRF estimate and the bottom image is the BLIP estimate.](visual_comparison_T1.eps){height="0.99\textheight"} ![\[fig: visual comparison T2\] A visual comparison of the $\operatorname*{T2}$ map estimates from a sequence of length $L=300$. The top plot shows the original $\operatorname*{T2}$ map. The middle image is the MRF estimate and the bottom image is the BLIP estimate.](visual_comparison_T2.eps){height="0.99\textheight"} ### Convergence rates for BLIP The convergence of the iterative algorithms is shown in figure \[fig: convergence plot\] as a function of the relative data consistency error at each iteration $k$, which we define as $\|Y-h(X^{k})\|_2^2/\|Y\|_2^2$. Results for three different sequence lengths, $100$, $200$ and $500$, are shown in the figure. It is clear that in all cases the algorithms converge rapidly and for sequence lengths of $200$ or more have effectively converged within $20$ iterations (note the log scale along the y-axis). Indeed, this is predicted by the compressed sensing theory for IPA: when the sequence length increases, so that compressed sensing task becomes easier (smaller isometry constant) the rate of convergence also increases. Thus BLIP can be considered to be reasonably computationally efficient. ![\[fig: convergence plot\] Plots of the data consistency error at each iteration for BLIP using a varying sequence length. The convergence rate increases as the sequence length increases. This is consistent with theory as the increased sequence length is likely to reduce the isometry constant.](convergence_plot.eps){width="0.7\linewidth"} Subsampling versus sequence length ---------------------------------- In our next experiment we investigate the dependencies of the undersampling ratio and the sequence length on the reconstruction performance. In this experiment we evaluate the image sequence SER as a function of $L$ and $p$. Recall that the theory presented in section \[sec: excitation and sampling\] suggested that this performance might degrade roughly as a function of $p^2/L$. However, as we noted earlier, the analysis in that section is of a ‘worst case’ type and may be highly conservative. Figure \[fig: L vs p2\] shows a plot of the image sequence SER as a function of $L/p^2$ for three different subsampling rates: $p = 16$ (green), $p = 32$ (red) and $p = 64$ (blue). From the plot we can see that the rapid growth of the SER that we associate with successful recovery occurs in each case at roughly the same value of $L/p^2$. This seems to suggest that the predicted scaling behaviour for $L$ and $p$ in random EPI to achieve RIP is of the right order. This in turn suggests that to maximize efficiency we should attempt to minimize $p$ (all other design criteria being equal). ![\[fig: L vs p2\] A plot of the Image sequence SER (dB) against $L/p^2$ for three different levels of undersampling: $p = 16$ (green), $p = 32$ (red) and $p = 64$ (blue). The rapid increase in SER appears to occur at roughly the same value of $L/p^2$ in each case suggesting that the RIP result in Theorem \[th: randEPI embedding\] is of the right order.](Image_SER_L_vs_p2.eps){width="0.7\linewidth"} Using a complex density model {#sec: complex density experiment} ----------------------------- The simulations, so far, have used the somewhat idealized model that the density map is real and non-negative. In this experiment we demonstrate that the algorithm works just as well when the density map is allowed to be complex and to absorb sensitivity maps and other phase terms. Here we repeat the first experiment but we modify the density map to have a quadratic phase that is zero at the centre of the image and $pi/4$ at the corners. A plot of the phase is shown on the left hand side in figure \[fig: complex\_density\_T2\_SER\]. We then ran the MRF reconstruction algorithm and BLIP with equations and replaced by and in both algorithms. The resulting performance was very similar to that in the real valued case. For brevity we only show a plot of the the $\operatorname*{T2}$ SER in figure \[fig: complex\_density\_T2\_SER\]. We see that the parameter estimation behaves identically to that in the first experiment. Similar behaviour can be observed for the other parameters. Therefore, it seems that there is no significant difference in using the real or complex model for proton density. ![ \[fig: complex\_density\_T2\_SER\] Reconstruction performance for the $\operatorname*{T2}$ map using a complex density model. (a) The quadratic phase applied to the density map; (b) SER for $\operatorname*{T2}$ map estimation as a function of sequence length. Results are shown for the following algorithms: complex MRF, complex BLIP and the complex oracle estimate. ](complex_density_phase.eps){width="\linewidth"} [(a)]{} ![ \[fig: complex\_density\_T2\_SER\] Reconstruction performance for the $\operatorname*{T2}$ map using a complex density model. (a) The quadratic phase applied to the density map; (b) SER for $\operatorname*{T2}$ map estimation as a function of sequence length. Results are shown for the following algorithms: complex MRF, complex BLIP and the complex oracle estimate. ](complex_density_T2_map_SER.eps){width="\linewidth"} [(b)]{} Uniform versus non-uniform sampling ----------------------------------- In §\[sec: excitation and sampling\] we asserted that as the Bloch response model does not include any spatial structure it is preferable to take uniformly random samples of $k$-space in order to achieve the RIP rather than use a variable density scheme. In this final experiment we examine the effect of replacing the (uniform) random EPI sampling with a sampling pattern that weights the lower frequencies more, as is common in compressed sensing schemes for MRI [@Lustig2008]. Specifically, we choose a non-uniform sampling pattern with an equivalent undersampling ratio, $M/N = 1/16$, that always samples $k_y = 0, 1, 2, \sqrt{N}-3, \sqrt{N}-2 \mbox{~and~} \sqrt{N}-1$ (the centre of $k_y$-space), and then samples the remainder of $k$-space uniformly at random (with the remaining $10$ samples). While we have not tried to optimize this non-uniform sampling strategy we have found that other variable density sampling strategies performed similarly. We repeated the first experiment and compared the random EPI sampling to using non-uniform sampling with the sequence length varied between $10$ and $300$. Again we focus on the $\operatorname*{T2}$ reconstruction, although similar behaviour was observed for the density and $\operatorname*{T1}$ estimation (not shown). The $\operatorname*{T2}$ results are plotted in figure \[fig: uniform\_vs\_nonuniform\]. It is clear from the figure that BLIP does not perform well with the non-uniform sampling of $k$-space, and it never achieves the near oracle performance that we observe with the random EPI sampling strategy. Indeed, we observed no non-uniform sampling strategy to do this. Other simulations (not shown) have indicated that uniform i.i.d. undersampling in $k_y$ also performs well, although we have yet to prove this has the RIP. Interestingly, the MRF reconstruction does benefit from the non-uniform sampling, however the reconstruction quality is still very poor. We believe that this can be explained by the fact that in both cases the MRF reconstructions exhibit significant aliasing. However, in the non-uniform case the aliasing is concentrated more in the high frequencies where the signal has less energy and therefore introduces less distortion. ![\[fig: uniform\_vs\_nonuniform\] A plot of the $\operatorname*{T2}$ estimate SER (dB) against $L$ for reconstruction algorithms MRF and BLIP using uniform random (EPI) sampling with $p = 16$ and a non-uniform random sampling with an equivalent undersampling ratio $M/N = 16$. Only in the case of BLIP with uniform random sampling does the $\operatorname*{T2}$ estimate performance approach that of the oracle estimator.](uniform_vs_nonuniform.eps){width="0.7\linewidth"} Conclusions and open questions ============================== We have presented a principled mathematical framework for compressed quantitative MRI based around the recently proposed technique of Magnetic Resonance Fingerprinting [@Ma-MRF2013]. The sensing process can be considered in two separate stages. First, the embedding of the parameter information into the magnetization response sequences through the mapping $f(\cdot)$. Second, the compressive imaging of the induced magnetization image sequence. The key elements of our approach have been: the characterization of the signal model through the Bloch response manifold; the identification of a provably good image sequence reconstruction algorithm based on iterative projection; an excitation response condition based on a newly introduced measure of *flatness* to quantify the persistence of the excitation; and a random EPI $k$-space sampling scheme that can be shown to have the necessary RIP condition when the excitation is suitably flat. The simulations presented in §\[sec: experiments\] show that the proposed technique is capable of achieving good parameter map reconstruction with very short pulse sequences. The next step will be to make a thorough comparison on an MRI scanner with MRF and other existing quantitative MRI techniques such as [@Deoni-DESPOT-2003]. While the current work is specifically targeted at a compressed sensing framework for MRF, we believe that many elements of it should be more broadly applicable. Specifically, the RIP condition for randomized EPI may well have applications in other MR imaging strategies and the characterization of excitation response in terms of flatness could prove a useful tool for the analysis of other compressed sensing schemes involving some form of active sensing. Finally, the use of parametric physical models (through appropriate discretisation) could be applicable to many areas of compressed sensing beyond MRI. The experience we have gained here suggests that such models can be more powerful than traditional spatial image models, such as wavelet sparsity, that are often found in compressive imaging. Open Questions -------------- In setting out this compressed sensing framework a number of questions have arisen that we feel should be addressed. We conclude by briefly describing these below. ### Excitation sequences {#excitation-sequences .unnumbered} What are the key requirements for the excitation sequences? We have introduced the flatness condition, however, we have so far not exploited randomness in the excitation. This raises the question: does the excitation sequence need to be random? Although randomness seems a natural way to obtain flat responses, it is not clear that it is necessary or even preferable. Random excitations may also be able to provide less stringent sampling conditions in order to provide the RIP. Furthermore, whether deterministic or random, how should we optimize the excitation sequences in order to maximise the performance of the parameter map estimation? This seems to be very much a system identification problem. ### Improved signal models {#improved-signal-models .unnumbered} A key question for the Bloch response model is: how densely do we need to sample $\M$? This will depend on the response mapping $f$, the undersampling operator $h$ and the performance of the recovery algorithm. It would be interesting to try to quantify these errors using the existing union of subspace compressed sensing theory [@BD-UoS-09; @TB-2011]. A second question is: how should we best include additional modelling information? It is clearly desirable to include spatial regularization. However, we have seen in §\[sec: experiments\] that the inclusion of our limited spatial regularization within the signal model did not significantly improve performance. On the other hand, this only regularized the density map, whereas, ideally we would like to impose spatial regularity on each of the parameter maps. Unfortunately, a naive construction of such a model would lead to a complex non-separable representation that we cannot easily project onto. Alternatively, we might try to impose block spatial regularity on the image sequence on top of the Bloch response model. This form of spatial regularization was used in [@Doneva2010] and appears to have only provided modest performance improvements. Therefore the question is how to best combine these models to maximize reconstruction performance and can we back this up theoretically? The current signal model is also somewhat idealised. We have treated the proton density values, $\rho_i$, as nonnegative, following the physics. However, in MRI it is more common to treat $\rho_i$ as a complex value, absorbing various phase factors into the quantity. While our framework easily extends to the complex case as highlighted in §\[sec: Bloch projection\] and evaluated in §\[sec: complex density experiment\], it would be interesting to see whether there was a more principled way to deal with such additional phase factors. Another idealization that is made both here and in the original MRF is that the read out time is assumed negligible with respect to the relaxation times. Depending on the level of undersampling this may not be true. This might introduce significant artefacts. If so, can we modify the signal model to account for this? Finally, our model does not account for partial volume effects. These were briefly touched on in the supplementary material of [@Ma-MRF2013], where it was proposed to model individual voxels as a composition of different material components. Such a model is reminiscent of the spatial abundance maps used in hyperspectral imaging. In such a case we are in the realms of compressive source separation [@Golbabaee-2013]. Can we formulate a compressive MRF problem that accounts for partial volume effects in a similar manner? ### Subsampling $k$-space {#subsampling-k-space .unnumbered} We have identified certain conditions that guarantee the RIP for random EPI sampling. This allows us to trade off the $k$-space subsampling factor $p=N/M$ with the length of the excitation sequence, $L$. Unfortunately the trade off scales as $L \sim p^2$. It is not clear whether similar guarantees could be achieved from a deterministic sampling sequence or whether this is indeed optimal. It would be more desirable to have a proportional trade off $L \sim p$. Is such a scaling possible? If so, what is the appropriate combination of excitation sequence and sampling strategy? Finally, if we can successfully incorporate spatial structure into our signal model, as suggested above, it is very likely that a variable density sampling would be preferable. If so, can we leverage existing theory for variable density sampling [@Adcock-2011; @Puy2011] to develop principled designs for variable density sampling for compressive MRF? Dynamics of balanced SSFP sequences {#app: bloch dynamics} =================================== Balanced SSFP sequences are popular in MRI and were the basis of the excitation sequences used in MRF [@Ma-MRF2013], although the term ‘steady state’ is somewhat of a misnomer as this refers to the steady state conditions arrived at following periodic excitation with constant $\alpha$ and $\operatorname*{TR}$ [@Scheffler2003]. In fact, here we are explicitly interested in the transient dynamics of a non-periodic excitation sequence. This is in contrast with traditional SSFP sequences where transient oscillations are seen as undesirable as they can introduce imaging artefacts [@Hargreaves2001]. In this work, as in [@Ma-MRF2013], we will regard the transient behaviour as essential in enabling us to distinguish between different quantitative behaviour. The transient response can be formally described in terms of a $3$-dimensional linear discrete time dynamical system that we summarize below, see [@Hargreaves2001; @Ganter; @Scheffler2003] for further details. To keep things simple we will assume there is no phase increment between pulses and also that the $l$th echo time, $\operatorname*{TE}_l$, is half the $l$th repetition time $\operatorname*{TR}_l$. Following [@Hargreaves2001], let $\m_{l} = (m_{l}^{x},m_{l}^{y},m_{l}^{z})^T \in \R^3$ represent the 3-dimensional magnetization vector for a voxel at the $l$th excitation pulse. In Inversion Recovery SSFP sequences the equilibrium magnetization, $\m_{\mbox{eq}} = [0,0,1]^T$, is initially inverted so that $\m_0 = [0,0,-1]^T$. Then the magnetization after the $l$th RF-excitation is given by the following linear discrete time dynamical system: $$\label{eq: magnetization response} \m_{l+1} = R_x(\alpha_l) R_z(\phi_{l}) E_l \m_{l} + R_x(\alpha_l)(\Id -E_l) \m_{\mbox{eq}}$$ where $R_u(\phi)$ denotes a rotation about the $u \in \{x,y,z\}$ axis by an angle $\phi$, $\phi_l = 2\pi \delta f \operatorname*{TR}_l$ is the off-resonance phase associated with local field variations and chemical shift effects [@Ganter] and $E_l$ is the diagonal matrix characterizing the relaxation process: $$E_l {\mathrel{\mathop:}=}\begin{pmatrix} e^{-\operatorname*{TR}_l/\operatorname*{T2}} &~&~\\ ~& e^{-\operatorname*{TR}_l/\operatorname*{T2}}& ~\\ ~&~&e^{-\operatorname*{TR}_l/\operatorname*{T1}} \end{pmatrix}$$ where the $\operatorname*{T1}$ relaxation time controls the rate of relaxation along the $z$-axis, while the $\operatorname*{T2}$ relaxation time controls the relaxation onto the $z$-axis. Finally let $\hat{\m}_{l}$ denote the magnetization at the echo time, $\operatorname*{TE}_l$. Then this is given by [@Hargreaves2001]: $$\label{eq: TE response} \hat{\m}_{l} = R_z(\phi_{l}/2) E_l^{1/2} \m_{l} + (\Id -E_l^{1/2}) \m_{\mbox{eq}},$$ with the readout coil measuring $\hat{m}_{l}^{x}+ j \hat{m}_{l}^{y}$. Thus the magnetization dynamics in response to a sequence of RF pulses with flip angles, $\alpha_l$, and repetition times, $\operatorname*{TR}_l$, is given by and which apart from the input parameters is solely a function of the tissue parameters $\operatorname*{T1}$, $\operatorname*{T2}$, and the off-resonance frequency, $\delta f$. Proof of Theorem \[th: randEPI embedding\] {#app: RIP proof} ========================================== We first introduce the key lemmas that form the main ingredients of the proof. Our approach will follow the standard route of concentration of measure, $\epsilon$-net and union bound. To this end we will need the following well known Chernoff bound [@Dubhashi2009]: \[th: chernoff\] Let $X = X_1+X_2+\ldots +X_n$, $0 \leq X_i \leq 1$ with $\mu = \Expect (X)$. Then $$\Prob(|X-\mu| > \epsilon \mu) \leq 2 \exp \left( -\frac{\epsilon^2 \mu}{3} \right)$$ The next lemma establishes a near isometry for a single aliased voxel sequence. \[th: single voxel chord isometry\] Let $z \in \C^L$ be a random vector given by: $$z_i = \frac{1}{p}\sum_k U_{k,i} e^{-j 2 \pi \zeta_i k / p}$$ where $\zeta_i$ are independent random variables drawn uniformly from $\{0, \ldots, p-1\}$ and $U \in \C^{p \times L}$ is a matrix whose rows have flatness $\lambda$. Then, with probability at least $1-2 e^{-\epsilon^2 /(3 p \lambda^2)}$, $z$ satisfies: $$(1-\epsilon) \|U\|_F^2 \leq p^2 \|z\|_2^2 \leq (1+\epsilon) \|U\|_F^2$$ We first show that $\Expect \|z\|_2^2 = \frac{1}{p^2}\|U\|_F^2$ and then derive the necessary tail bounds. Let $W_{a,k} = \frac{1}{\sqrt{p}} e^{-j 2 \pi a k/p}$, $a,k = 0, \ldots, p-1$, denote the unitary Discrete Fourier transform in $\C^p$. We can then write $$\begin{aligned} \Expect \| z\|_2^2 &= \sum_{a=0}^{p-1} \frac{1}{p} \left( \sum_{i=1}^L \frac{1}{p} | W_{a,:} U_{:,i}|^2 \right)\\ & = \frac{1}{p^2} \sum_i \sum_a | W_{a,:} U_{:,i}|^2\\ &= \frac{1}{p^2} \sum_i \|U_{:,i}\|_2^2\\ &= \frac{1}{p^2} \|U\|_F^2,\end{aligned}$$ Now note that $\|z\|_2^2$ is the sum of $L$ independent random variables, $\|z\|_2^2 = \sum_i \xi_i$ with $\xi_i = \frac{1}{p} |W_{\zeta_i,:} U_{:,i}|^2$. Furthermore the $\xi_i$ satisfy: $$\begin{split} 0 \leq \xi_i & \leq \frac{1}{p} \|U_{:,i}\|_2^2\\ & \leq \frac{1}{p} \sum_k \max_i |U_{k,i}|^2\\ & \leq \frac{1}{p} \sum_k \lambda^{2} \|U_{k,:}\|_2^2\\ & = \frac{\lambda^2}{p} \|U\|_F^2 \end{split}$$ We can therefore apply the Chernoff bound from Lemma \[th: chernoff\] to $\sum \xi_i$ rescaled by $\frac{\lambda^2}{p} \|U\|_F^2$ to give: $$\Prob(|\|z\|_2^2 - \frac{1}{p^2}\|U\|_F^2| > \epsilon \frac{1}{p^2} \|U\|_F^2) \leq 2 \exp \left( - \frac{\epsilon^2}{3 p \lambda^2}\right)$$ Rearranging this expression completes the proof. Next we extend Lemma \[th: single voxel chord isometry\] to a near isometry for groups of aliased voxels under the action of $h$. Since $h$ is an ortho-projector, $\|h (X)\|_2^2 = \|h^H h(X) \|_2^2$ and so we can equivalently consider the isometry properties of $h^Hh$. Let us denote $Z = h^H(h(X))$ such that $Z_{:,l} = F^H P(\zeta_l)^T Y_{:,l}$. Recall that $h$ is a partially sampled 2D discrete Fourier transform that is fully sampled in the $k_x$ direction and periodically subsampled by a factor of $p = N/M$ in the $k_y$ direction. Therefore each $Z_{i,l}$ is the sum of $p$ aliases taken from $X_{:,l}$: $$Z_{i,l} = \frac{1}{p} \sum_{k=0}^{p-1} X_{\tau_i(k),l}~ e^{-j 2 \pi \zeta_l k / p}$$ where $\tau_i(k)$ gives the index of the $k$th alias for the $i$th voxel (with $\tau_i(0) = i$). We can therefore partition the set $\{1,\ldots,N\}$ into $M$ disjoint index sets $\Lambda_1, \ldots, \Lambda_M$ with each set associated with $p$ aliases, such that $h^Hh$ is separable over $\{\Lambda_i\}$ and $Z_{\Lambda_i,:} = [h^Hh]_{\Lambda_i} X_{\Lambda_i,:}$. Since each $Z_{\Lambda_i,:}$ contains $p$ copies of the same combination of aliases (up to a phase shift) we can conclude that: $$\|Z_{\Lambda_i,:}\|_F^2 = p \|Z_{k,:}\|_2^2, ~\forall k \in \Lambda_i$$ Applying Lemma \[th: single voxel chord isometry\] then gives us: \[th: aliased voxel isometry\] Let $Z_{\Lambda_i,:} = [h^Hh]_{\Lambda_i} X_{\Lambda_i,:}$ for some $X_{\Lambda_i,:} \in \C^{p\times L}$ whose rows have a flatness $\lambda$ where $[h^Hh]_{\Lambda_i}$ is defined above. Then with probability at least $1-2 e^{-\epsilon^2 /(3 p \lambda^2)}$ we have $$\label{eq: single vector isometry} (1-\epsilon) \|X_{\Lambda_i,:}\|_F^2 \leq p \|Z_{\Lambda_i,:}\|_F^2 \leq (1+\epsilon) \|X_{\Lambda_i,:}\|_F^2$$ The final ingredient guarantees a near isometry for low dimensional subsets of the unit sphere (for a more sophisticated but slightly different result in this direction see [@Clarkson-manifolds-2008]) \[th: boxdim isometry\] Let $S \subset \mathbb{S}^{n-1}$ have box counting dimension $d$ such that for any $\epsilon > 0$ there exists an $\epsilon$-cover of $S$ of size $C_S \epsilon^{-d}$. Let $P:\C^n\rightarrow \C^k$ be a random projection such that for any $\delta>0$ and a fixed $x \in S$, $$\label{eq: single point isometry} 1-\delta \leq {\frac{n}{k}}\|Px\|_2^2 \leq 1+\delta$$ holds with probability at least $1-c_0 e^{-c_1 \delta^2}$. Then $P$ satisfies for all $x \in S$ with probability at least $1-\eta$ as long as: $$c_1 \geq 72 \delta^{-2} \left( d \log (36n/\delta k) + \log C_S c_0/\eta \right)$$ Consider an $\epsilon$-cover $S_\epsilon$ of $S$ with $\epsilon = \delta'/(2\sqrt{n/k})$ and suppose that $P$ satisfies $$1-\delta'/2 \leq {\frac{n}{k}}\|Px\|_2^2 \leq 1+\delta'/2$$ for all $x \in S_\epsilon$ with a constant $0<\delta'<1$. Then there exists a $u \in S_\epsilon$ such that: $$\begin{aligned} \sqrt{\frac{n}{k}} \| Px\|_2 &\leq \sqrt{\frac{n}{k}}\| Pu\|_2 +\sqrt{\frac{n}{k}}\| P(x-u)\|_2\\ & \leq 1+ \delta'/2 + \sqrt{\frac{n}{k}} \epsilon \label{eq: squared RIP implies a non-squared RIP}\\ &= 1+ \delta'\end{aligned}$$ where in we have used the fact that $(1+\delta'/2)^2 > (1+\delta'/2)$. We can similarly show that $\sqrt{\frac{n}{k}}\|Px\|_2 \geq 1-\delta'$. Then finally noting that the “non-squared” RIP implies the squared RIP in with $\delta = 3\delta'$ gives us the required isometry. It only remains to bound the probability of failure. Let $p_f$ be the probability that $P$ fails to satisfy on $S$. By the union bound: $$\begin{aligned} pf &\leq |S_\epsilon| c_0 e^{-c_1 (\delta'/2)^2}\\ &\leq C_S c_0 \left(\frac{\delta'}{2\sqrt{n/k}}\right)^{-d} e^{-c_1 (\delta'/2)^2}\end{aligned}$$ Therefore it is sufficient to choose $\eta$ so that: $$\frac{\eta}{C_S c_0} \geq \left(\frac{\delta}{6\sqrt{n/k}}\right)^{-d} e^{-c_1 (\delta/6)^2}$$ Re-arranging the above gives: $$c_1 \geq 72 \delta^{-2} \left( d \log (36n/\delta k) + \log C_S c_0/\eta \right)$$ as required. We are now ready to prove the main theorem. \[Proof of Theorem \[th: randEPI embedding\]\] First, note that $\R_+\B \subset \R \B$ which is an infinite union of subspace model, as is its $p$-product, $(\R\B)^p$ associated with a group of aliased voxels, $\Lambda_i$. To guarantee that $h_{\Lambda_i}$ possesses the necessary RIP on $(\R\B)^p-(\R\B)^p$ it is sufficient to consider the RIP on the normalized difference set $S$ given by: $$S = \{x \in ((\R\B)^p-(\R\B)^p),\|x\|_2 = 1\},$$ due to the linearity of $h$. By construction we have $\dim(S) = 2p d_\B-1$ and we can therefore apply Lemma \[th: boxdim isometry\] to $S$ together with Lemma \[th: aliased voxel isometry\]. This guarantees for all $X_{\Lambda_i,:} \in (\R\B)^p-(\R\B)^p$ that $h$ satisfied with probability at least $1-\eta$ as long as: $$\lambda^{-2} \geq (3p) \times 72 \delta^{-2} \left( (2p d_\B-1) \log (36p/\delta) + \log C_S c_0/\eta \right)$$ To ensure this holds for all aliased voxel groups $\Lambda_i$, $i = 1,\ldots, M$ we can again apply the union bound and replace $\eta$ by $M \eta$. Noting that $p, \delta^{-1}, \eta^{-1} > 1$ we can collect together the constants and simplify to finally give: $$\lambda^{-2} \geq C \delta^{-2} p^2 d_\B \log (N/ \delta \eta )$$ for some constant $C$ independent of $p, N, d_\B, \delta$ and $\eta$ which gives the required conditions of the theorem. [100]{} B. Adcock and A. C. Hansen, Generalized sampling and infinite-dimensional compressed sensing. DAMTP Tech. Rep. 2011/NA12, 2011. N. Ailon and B. Chazelle, The Fast Johnson-Lindenstrauss Transform and Approximate Nearest Neighors. SIAM J. Computing, vol. 39, No. 1, pp. 302-–322, 2009. R. Baraniuk, M. Davenport, R. De Vore, and M. Wakin, A simple proof of the restricted isometry property for random matrices, Constructive Approx., vol. 28, pp. 253–263, 2008. R. G. Baraniuk and M. B. Wakin, Random Projections of Smooth Manifolds. Foundations of Computational Mathematics, vol. 9(1), pp. 51-77, 2009. A Beygelzimer, S. Kakade and J. Langford, Cover Trees for Nearest Neighbor. In Proceedings of the 23rd International Conference on Machine Learning (ICML), Pittsburgh, PA, pp. 97–104, 2006. K. T. Block, M. Uecker, and J. Frahm, Model-based Iterative Reconstruction for Radial Fast Spin-Echo MRI. IEEE Trans. Med. Imag., vol. 28(11), pp. 1759–1769, 2009. T. Blumensath and M. E. Davies, Iterative Hard Thresholding for Sparse Approximation, J. Fourier Analysis and Applications, vol. 14, no. 5, pp. 629–654, 2008. T. Blumensath and M. E. Davies, Sampling Theorems for Signals From the Union of Finite-Dimensional Linear Subspaces. IEEE Trans. Inf. Theory, vol. 55(4), pp. 1872–1882, 2009. T. Blumensath and M. E. Davies, Iterative Hard thresholding for Compressed sensing. Applied Computational Harmonic Analysis, vol. 27, no. 3, pp. 265-274, 2009. T. Blumensath, M. E. Davies, Normalised Iterative Hard Thresholding; guaranteed stability and performance, IEEE Journal of Selected Topics in Signal Processing, vol. 4(2), pp. 298-309, 2010. T. Blumensath, Sampling and Reconstructing Signals From a Union of Linear Subspaces. IEEE Trans. Inf. Theory, vol. 57(7), pp. 4660–4671, 2011. Brainweb data repository, available at: `http://brainweb.bic.mni.mcgill.ca/brainweb/` M. Bydder, A. A. Samsonov, and J. Du, Evaluation of optimal density weighting for regridding. Mag. Res. Im., vol. 25(5), pp. 695–-702, 2007. K. Clarkson, Tighter Bounds for Random Projections of Manifolds. Proceedings of the 24th annual symposium on Computational geometry (SCG’08), pp. 39-48, 2008. D.L. Collins, A.P. Zijdenbos, V. Kollokian, J.G. Sled, N.J. Kabani, C.J. Holmes and A.C. Evans, Design and Construction of a Realistic Digital Brain Phantom. IEEE Trans. on Medical Imaging, vol.17(3), pp.463–468, 1998. S.C.L. Deoni, B.K. Rutt, and T.M. Peters, Rapid Combined T1 and T2 Mapping Using Gradient Recalled Acquisition in the Steady State. Magn. Reson. Med. vol. 49, pp. 515-526, 2003. M. Doneva, P. Bornert, H. Eggers, C. Stehning, J. Senegas and A. Mertins, Compressed sensing reconstruction for magnetic resonance parameter mapping, Magn. Reson. Med., vol 64, pp. 1114–-1120, 2010. D.P. Dubhashi and A. Panconesi, *Concentration of Measure for the Analysis of Randomized Algorithms*. Cambridge University Press, 2009. J. A. Fessler and B. P. Sutton, Nonuniform fast Fourier transform using min-max interpolation. IEEE Trans. Sig. Proc. vol. 51(2) pp. 560–574, 2003. J. A. Fessler, Model-based image reconstruction for MRI. IEEE Sig. Proc. Mag., vol. 27(4), pp. 81–89, 2010. J.P. Hornak, *The Basics of MRI*. Webbook, available on-line at: `http://www.cis.rit.edu/htbooks/mri/`. C. Huang, A. Bilgin, T. Barr and M. I. Altbach, T2 relaxometry with indirect echo compensation from highly undersampled data. Magn Reson Med., vol. 70, pp. 1026-–1037, 2013. M. Iwen and M. Maggioni, Approximation of Points on low-dimensional manifolds via random linear projections. arXiv:1204.3337. E. T. Jaynes, Matrix treatment of nuclear induction. The Physics Review, vol 98(4), pp. 1099–1105, 1955. C. Ganter, Off-Resonance Effects in the Transient Response of SSFP Sequences. Magn. Reson. Med. vol. 52, pp. 368-–375, 2004. M. Golbabaee, S. Arberet and P. Vandergheynst, Compressive Source Separation: Theory and Methods for Hyperspectral Imaging. IEEE Trans. Image Proc., vol. 22(12), pp. 5096–5110, 2013. B.A. Hargreaves, S.S. Vasanawala, J.M. Pauly and D.G. Nishmura, Characterization and Reduction of the Transient Response in Steady-State MR Imaging. Magn, Reson. Med. vol 46, pp. 149–158, 2001. M. Lustig, D. L. Donoho, J.M. Santos, and J.M. Pauly, Compressed sensing MRI. IEEE Sig. Proc. Mag., vol. 25(2), pp. 72-–82, 2008. D. Ma, V. Gulani, N. Seiberlich, K. Liu, J. L. Sunshine, J. L. Duerk and M. A. Griswold, Magnetic Resonance Fingerprinting. Nature, vol. 145, pp. 187–192, 2013. G.C. McKinnon, Ultrafast interleaved gradient-echo-planar imaging on a standard scanner. Magn. Reson. Med. vol. 30, pp. 609–616, 1993. P. Niyogi, S. Smale, and S. Weinberger, Finding the Homology of Submanifolds with High Confidence from Random Samples. Discrete Comput. Geom. vol. 39(1), pp. 419–441, 2008. G. Puy, P. Vandergheynst and Y. Wiaux, On Variable Density Compressive Sampling. IEEE Sig. Proc. Lett., vol. 18(10), pp. 595–598, 2011. G. Puy, P. Vandergheynst, R. Gribonval and Y. Wiaux, Universal and efficient compressed sensing by spread spectrum and application to realistic Fourier imaging techniques. EURASIP Journal on Advances in Signal Processing, 2012, 2012:6. H. Rauhut, Compressive sensing and structured random matrices. Radon Series Comp. Appl. Math., vol. 9, pp. 1–92, 2010. K. Sheffler and S. Lehnhardt, Principles and applications of balanced SSFP techniques. Eur. Radiol. vol. 13, pp. 2409–-2418, 2003. J. Tran-Gia, D. Stab, T. Wech, D. Hahn, and H. Kostler, Model-based Acceleration of Parameter mapping (MAP) for saturation prepared radially acquired data. Magn Reson Med., vol. 70, pp. 1524-–1534, 2013. G. A. Wright, Magnetic Resonance Imaging. IEEE Sig. Proc. Mag., vol.14(1), pp. 56–66, 1997. B. Zhao, F. Lam, W. Luy and Z.-P. Liang, Model-based MR parameter mapping with sparsity constraint. IEEE Int. Symp. Biomed. Imag. (ISBI), pp. 1–4, 2013. [^1]: The off-resonance frequency is a additional parameter that can be incorporated into the Bloch equations and measures local field inhomogeneity and chemical shift effects [@Ganter]. [^2]: As the excitation pulses in MRF are random the term steady state is now somewhat of a misnomer and we should possibly call these Inversion Recovery Randomly Excited Free Precession. [^3]: Strictly speaking we can only consider this to be an embedding for $\rho_i >0$ otherwise $\theta_i$ is not observable.

ArXiv

--- abstract: 'Poincaré’s invariance principle for Hamiltonian flows implies Kelvin’s principle for solution to Incompressible Euler Equation. Iyer-Constantin Circulation Theorem offers a stochastic analog of Kelvin’s principle for Navier-Stokes Equation. Weakly symplectic diffusions are defined to produce stochastically symplectic flows in a systematic way. With the aid of symplectic diffusions, we produce a family of martigales associated with solutions to Navier-Stokes Equation that in turn can be used to prove Iyer-Constantin Circulation Theorem. We also review some basic facts in symplectic and contact geometry and their applications to Euler Equation.' author: - | Fraydoun Rezakhanlou[^1]\ UC Berkeley\ Department of Mathematics\ Berkeley, CA  94720-3840 title: 'Stochastically Symplectic Maps and Their Applications to Navier-Stokes Equation' --- Introduction {#sec1} ============ Hamiltonian systems appear in conservative problems of mechanics governing the motion of particles in fluid. Such a mechanical system is modeled by a Hamiltonian function $H(x,t)$ where $x = (q,p)\in {\mathbb R}^d \times {\mathbb R}^d$, $q = (q_1,\dots,q_d)$, $p = (p_1,\dots,p_d)$ denote the positions and the momenta of particles. The Hamiltonian’s equations of motion are $$\label{eq1.1} {\dot q} = H_p(q,p,t),\ {\dot p} = -H_q(q,p,t)$$ which is of the form $$\label{eq1.2} {\dot x} = J\nabla_xH(x,t),\ \ \ J = \begin{bmatrix} 0 & I_d \\ -I_d & 0 \end{bmatrix}$$ where $I_d$ denotes the $d \times d$ identity matrix. It was known to Poincaré that if $\phi_t$ is the flow of the ODE and $\gamma$ is a closed curve, then $$\label{eq1.3} \frac {d}{dt} \int_{\phi_t(\gamma)} \bar {\lambda}= 0,$$ where $\bar{\lambda}:=p \cdot dq$. We may use Stokes’ theorem to rewrite as $$\label{eq1.4} \frac {d}{dt} \int_{\phi_t(\Gamma)} {d\bar {\lambda}} = 0$$ for every two-dimensional surface $\Gamma$. In words, the $2$-form $${\bar \omega}:=\sum_{i=1}^d dp_i \wedge dq_i,$$ is invariant under the Hamiltonian flow $\phi_t$. Equivalently, $$\label{eq1.5} \phi_t^*\bar \omega=\bar{\omega}.$$ A Hamiltonian system simplifies if we can find a function $u(q,t)$ such that $p(t) = u(q(t),t)$. If such a function $u$ exists, then $q(t)$ solves $$\label{eq1.6} \frac {dq}{dt} = H_p(q,u(q,t),t).$$ The equation for the time evolution of $p$ gives us an equation for the evolution of the velocity function $u$; since $$\begin{aligned} {\dot p} &= (Du){\dot q} + u_t = (Du) H_p(q,u,t) + u_t, \\ {\dot p} &= -H_q(q,u,t),\end{aligned}$$ the function $u(q,t)$ must solve, $$\label{eq1.7} u_t +( Du) H_p(q,u,t) + H_q(q,u,t) = 0.$$ For example, if $H(q,p,t) = \frac {1}{2} |p|^2 + P(q,t)$, then becomes $$\label{eq1.8} u_t + (Du)u + \nabla P(q,t) = 0,$$ and the equation simplifies to $$\label{eq1.9} \frac {dq}{dt} = u(q,t) .$$ Here and below we write $Du$ and $\nabla P$ for the $q$-derivatives of the vector field $u$ and the scalar-valued function $P$ respectively. If the flow of is denoted by $Q_t$, then $\phi_t(q,u(q,0)) = (Q_t(q),u(Q_t(q),t))$. Now means that for any closed $q$-curve $\eta$, $$\label{eq1.10} \frac {d}{dt} \int_{Q_t(\eta)} u(q,t) \cdot dq =\frac {d}{dt} \int_{\eta}(DQ_t)^*\ u\circ Q_t(q,t) \cdot dq= 0,$$ or equivalently $$\label{eq1.11} d (Q_t^*{\alpha}_t)=d{\alpha}_0,$$ where ${\alpha}_t=u(q,t)\cdot dq$. This is the celebrated [*Kelvin’s circulation theorem*]{}. In summary Poincaré’s invariance principle implies Kelvin’s principle for Euler Equation. (Note that the incompressibility condition $\nabla\cdot u=0$ is not needed for .) We may rewrite as $$\label{eq1.12} Q_t^*(d{\alpha}_t)=d{\alpha}_0,$$ and this is equivalent to Euler equation ( the equation with the incompressibility condition $\nabla\cdot u=0$). Moreover, when $d=3$, can be written as $$\label{eq1.13} \xi^t\circ Q_t=(DQ_t)\xi^0,\ \ \ {\text{ or }}\ \ \ \xi^t=\left((DQ_t)\xi^0\right)\circ Q_t^{-1},$$ where $\xi^t(\cdot)=\nabla\times u(\cdot,t)$. The equation is known as [*Weber’s formulation*]{} of Euler Equation and is equivalent to the vorticity equation by differentiating both sides with respect to $t$: $$\label{eq1.14} \xi_t+(D\xi)u=(Du)\xi.$$ Constantin and Iyer \[CI\] discovered a circulation invariance principle for Navier-Stokes equation that is formulated in terms of a diffusion associated with the velocity field. Given a solution $u$ to the Navier-Stokes equation $$\label{eq1.15} u_t +(Du)u + \nabla P(q,t) = \nu {\Delta}u , \ \ \ \nabla\cdot u=0,$$ let us write $Q_t$ for the (stochastic) flow of the SDE $$\label{eq1.16} dq=u(q,t)\ dt+\sqrt{2\nu }\ dW,$$ with $W$ denoting the standard Brownian motion. If we write $A=Q^{-1}$ and $\xi^t=\nabla{\times}u(\cdot,t)$, and assume that $d=3$, then Constantin and Iyer’s circulation formula reads as $$\label{eq1.17} \xi^t={{\mathbb E}}\left( (DQ_t)\xi^0\right)\circ A_t,$$ where ${{\mathbb E}}$ denotes the expected value. We are now ready to state the first result of this article. (To avoid a confusion between stochastic differential and exterior derivative, we use a hat for the latter.) \[th1.1\] Write ${\alpha}_t=u(q,t)\cdot dq$ with $u$ a classical solution of and given $T>0$, set $B_t=Q_{T-t}\circ Q_T^{-1}$. - \(i) Then the process ${\beta}_t=B_t^*\hat d{\alpha}_{T-t}, \ t\in[0,T]$ is a 2-form valued martingale. When $d=3$, this is equivalent to saying that the process $$M_t=\left(\left(DB_t^{-1}\right)\xi^{T-t}\right)\circ B_t,\ \ \ t\in[0,T],$$ is a martingale. - \(ii) Given a surface $\Theta$, the quadratic variation of the martingale ${\beta}_t({\Theta})=\int_\Theta {\beta}_t$ is given by $$\int_0^t\sum_{i=1}^d\left[\int_\Theta B_s^*\zeta_i^{T-s}\right]^2ds,$$ where $$\zeta_i^{\theta}=\sum_{j,k=1}^d u^k_{q_iq_j}(\cdot,{\theta})\ dq_j\wedge dq_k,$$ or equivalently, $\zeta_i^{\theta}(v_1, v_2)={{\mathcal C}}(u_{q_i}(\cdot,{\theta}))v_1\cdot v_2,$ with ${{\mathcal C}}(w)=Dw-(Dw)^*$. - \(iii) For $\Theta$ as in (ii), we have the bound $${{\mathbb E}}\int_0^T\sum_{i=1}^d\left[\int_\Theta B_s^*\zeta_i^{T-s}\right]^2ds\le {{\mathbb E}}\left[\int_\Theta A_T^*\hat d{\alpha}_{0}\right]^2.$$ [**[Remark 1.1]{}**]{} - \(i) In a subsequent paper, we will show how Theorem 1.1 can be extended to certain weak solutions. To make sense of martingales ${\beta}_t$ and $M_t$, we need to make sure that $DQ_t$ exists weakly and belongs to suitable $L^r$ spaces. As it turns out, a natural condition to guarantee $DQ_t\in L^r$ for all $r\in[1,{\infty})$ is $$\int_0^T\left[\int_{{{\mathbb R}}^d}|u(x,t)|^p\ dx\right]^{q/p}\ dt<{\infty},$$ for some $p,q\ge 1$ such that $d/p+2/q\le 1.$ - \(ii) Our result takes a simpler form if $u$ is a solution to backward Navier-Stokes Equation. For such $u$, we simply have that ${\beta}_t=Q_t^*\hat d{\alpha}_t$ is a martingale. When $d=3$, we deduce that $M_t=\big((DA_t)\xi^t\big)\circ Q_t$ is a martingale. 502em $\Box$=0 The organization of the paper is as follows: - In Section 2 we discuss Weber’s formulation of Euler Equation and show how implies . We also discuss two fundamental results in Symplectic Geometry that are related to the so-called [*[Clebsch]{}*]{} variables. - In Section 3 we address some geometric questions for stochastic flows of general diffusions and study [*[symplectic]{}*]{} . - In Section 4 we use symplectic diffusions to establish Theorem 1.1. - In Section 5 we discuss [*[contact]{}*]{} diffusions. Euler Equation {#sec2} ============== In this section we review some basic facts in differential geometry and their applications to Euler Equation. Even though most of the discussion of this section is either well-known or part of folklore, a reader may find our discussion useful as we use similar ideas to prove Theorem 1.1. We also use this section as an excuse to demonstrate/advertise the potential use of symplectic/contact geometric ideas in fluid mechanics. We start with giving the elementary proof of : By Cartan’s formula $$\label{eq2.1} \frac {d}{dt}\phi^*_t\bar{\lambda}=\phi^*_t{{\mathcal L}}_{Z_H}\bar{\lambda}=\phi^*_t dK=d(K\circ \phi_t),$$ where ${{\mathcal L}}_Z$ denotes the Lie derivative with respect to the vector field $Z$, $Z_H=J\nabla_xH$ for $H(q,p,t)=|p|^2/2+P(q,t)$, and and $$\label{eq2.2} K(q,p,t)=p\cdot H_p(q,p,t)-H(q,p,t)=\frac 12|p|^2-P(q,t).$$ If we integrate both sides of over an arbitrary (non-closed) curve of the form $(\eta,u(\eta,t))$, or equivalently restrict the form $\bar{\lambda}$ to the graph of the function $u$, then we obtain $$\label{eq2.3} \frac {d}{dt}\left[(DQ_t)^*\ u\circ Q_t\right]=\nabla (L\circ Q_t),$$ where $L(q,t)=K(q,u(q,t),t)=|u(q,t)|^2/2-P(q,t)$. Here by $A^*$ we mean the transpose of the matrix $A$. Recall $A_t=Q_t^{-1}$, so that $$\left(DQ_t\right)^{-1}=DA_t\circ Q_t.$$ As a consequence of we have $$u(\cdot,t)= (DA_t)^*\ u^0\circ A_t+\nabla (R\circ A_t),$$ for $R=\int_0^t L\circ Q_s \ ds$. As a result, $$\label{eq2.4} u(\cdot,t)={{\mathcal P}}\left[ (DA_t)^*\ u^0\circ A_t\right],$$ where $u^0$ is the initial data and ${{\mathcal P}}$ denotes the Leray-Hodge projection onto the space of divergence-free vector fields. The formula is Weber’s formulation and is equivalent to Euler’s equation. So far we have shown that the Kelvin’s principle is equivalent to the Weber’s formulation of Euler equation. If we use instead, we obtain a new equivalent formulation of Euler equation, namely the vorticity equation or . Recall $$\bar{\omega}(v_1,v_2)=Jv_1\cdot v_2.$$ If we choose $v_1$ and $v_2$ to be tangent to the graph of $u$, i.e. $v_i=(w_i,Du(q,t)w_i)$ for $i=1,2$, then $$\bar{\omega}(v_1,v_2)={{\mathcal C}}(u)w_1\cdot w_2,$$ where ${{\mathcal C}}(u)=Du-(Du)^*$. Hence really means $$\label{eq2.5} {{\mathcal C}}(u(\cdot,t)\circ Q_t)(DQ_t)w_1\cdot(DQ_t)w_2={{\mathcal C}}(u(\cdot,0))w_1\cdot w_2.$$ Let us assume now that $d=3$ so that, ${{\mathcal C}}(u)w=\xi{\times}w$, where $\xi=\nabla{\times}u$ denotes the vorticity. Hence $$\bar{\omega}(v_1,v_2)=(\xi{\times}w_1)\cdot w_2=:[\xi,w_1,w_2].$$ We note that the right-hand side is the volume form evaluated at the triple $(\xi,w_1,w_2)$. Now the invariance becomes $$\label{eq2.6} [\xi^t\circ Q_t,(DQ_t)w_1,(DQ_t)w_2]=[\xi^0,w_1,w_2],$$ where we have written $\xi^t$ for $\xi(\cdot,t)$. Since $u$ is divergence-free, the flow $Q_t$ is volume preserving. As a result, $$[\xi^0,w_1,w_2]=[(DQ_t)\xi^0 ,(DQ_t)w_1, (DQ_t)w_2].$$ From this and we deduce $$[\xi^t\circ Q_t,(DQ_t)w_1,(DQ_t)w_2]=[(DQ_t)\xi^0 ,(DQ_t)w_1,(DQ_t)w_2].$$ Since $w_1$ and $w_2$ are arbitrary, we conclude that is true. [**[Definition 2.1]{}**]{} - \(i) A closed $2$-form ${\omega}$ is symplectic if it is nondegenerate. We say that symplectic forms ${\omega}^1$ and ${\omega}^2$ are [*[isomorphic]{}*]{} if there exists a diffeomorphism $\Psi$ such that $\Psi^*{\omega}^1={\omega}^2$. - \(ii) A $1$-form ${\alpha}$ is contact if $l_x=\{v:d{\alpha}(x;v,w)=0{\text{ for every }}w\}$ is a line and for every $v\in l_x$, we have that ${\alpha}(x;v)\neq 0$. We say that contact forms ${\alpha}^1$ and ${\alpha}^2$ are [*[isomorphic]{}*]{} if there exists a diffeomorphism $\Psi$ such that $\Psi^*{\alpha}^1={\alpha}^2$. We say that contact forms ${\alpha}^1$ and ${\alpha}^2$ are [*[conformally isomorphic]{}*]{} if there exist a diffeomorphism $\Psi$ and a scaler-valued continuous function $f>0$ such that $\Psi^*{\alpha}^1=f{\alpha}^2$. - \(iii) A solution $u$ of Euler equation is called [*[symplectic]{}*]{} if ${\omega}_0=d{\alpha}_0$ is symplectic. - \(iv) A solution $u$ of Euler equation is [*[contact]{}*]{} if there exists a scalar-valued $C^1$ function $f_0$ such that ${\alpha}_0+df_0$ is contact. (Recall ${\alpha}_t=u(\cdot,t)\cdot dx$.) 502em $\Box$=0 [**[Remark 2.1]{}**]{} - \(i) As it is well-known, the degeneracy of a 2-form can only happen when the dimension $d$ is even. Recall ${\alpha}_t=u(\cdot,t)\cdot dx$. If $u$ is a symplectic solution, then ${\omega}_t=d{\alpha}_t$ is symplectic for all $t$ because by , the form ${\omega}_t$ is isomorphic to ${\omega}_0$. - \(ii) When $u$ is a contact solution of Euler equation, then $\tilde {\alpha}_t=Q_t^*{\alpha}_0+df_t$ is contact for all $t$ where $f_t=f_0\circ Q_t$. In general $\tilde {\alpha}_t\neq {\alpha}_t$. However, by equation , we have $d{\alpha}_t=d\tilde{\alpha}_t$. Hence there exists a scalar-valued function $g_t$ such that ${\alpha}_t+dg_t=\tilde {\alpha}_t$ is contact. 502em $\Box$=0 We continue with some general properties of symplectic and contact solutions of Euler Equation. As for symplectic solutions, assume that the dimension $d=2k$ is even and write $$(q_1,\dots,q_d)=(x_1,y_1,\dots,x_k,y_k).$$ A classical theorem of Darboux asserts that all symplectic forms are isomorphic to the standard form $\bar{\omega}=d\bar{\lambda}=\sum_{i=1}^kdy_i\wedge dx_i$. A natural question is whether such an isomorphism exists globally. [**[Definition 2.2]{}**]{} Let $u$ be a symplectic solution of Euler Equation. We say that [*Clebsch*]{} variables exist for $u$ in the interval $[0,T]$, if we can find $C^1$ functions $$X_1,\dots,X_k,Y_1,\dots,Y_k:{{\mathbb R}}^d\times [0,T]\to{{\mathbb R}},\ \ \ F:{{\mathbb R}}^d\times [0,T]\to{{\mathbb R}}$$ such that $\Psi_t=(X_1,Y_1,\dots,X_k,Y_k)(\cdot,t)$ is a diffeomorphism, and $$u(x,t)=\left(\sum_{i=1}^k Y_i\cdot \nabla X_i\right)(x,t)+\nabla F(x,t),$$ for every $t\in[0.T]$. Alternatively, we may write ${\alpha}_t=\Psi_t^*\bar {\lambda}+dF$ or $d{\alpha}_t=\Psi_t^*\bar {\omega}$. 502em $\Box$=0 \[prop2.1\] Let $u$ be a symplectic solution to Euler Equation. - \(i) If Clebsch variables exist for $t=0$, then they exist in the interval $[0,{\infty})$. - \(ii) If $d=4$ and Clebsch variables exist for $t=0$ outside some ball $B_r=\{x:|x|\le r\}$, then they exist globally in the interval $[0,{\infty})$. [**Proof. (i)**]{} This is an immediate consequence of : If $\Psi_0^*\bar{\omega}={\omega}_0=d{\alpha}_0$, then $$\left(Q_t\circ \Psi_0^{-1}\right)^*d{\alpha}_t=\Psi_0^{-1*}Q_t^*d{\alpha}_t=\Psi_0^{-1*}d{\alpha}_0=\bar{\omega},$$ which means that we can choose $\Psi_t=\Psi_0\circ A_t$ for the Clebsch change of variables. [**(ii)**]{} This is a consequence of a deep theorem of Gromov \[Gr\]: When $d=4$, a symplectic form is isomorphic to standard form $\bar{\omega}$, if this is the case outside a ball $B_r$. 502em $\Box$=0 Observe that Euler Equation can be rewritten as $$\label{eq2.7} \frac {d }{dt}{\alpha}_t+i_u(d{\alpha}_t)=-dH,$$ where $H(q,t)=P(q,t)+|u(q,t)|^2/2$ is the Hamiltonian function. For a steady solution, ${\alpha}_t$ is independent of $t$ and we simply get $$i_u(d{\alpha})=-dH.$$ If $u$ is a symplectic steady solution of Euler Equation, then $i_u(d{\alpha})=-dH$ means that $u$ is a [*[Hamiltonian vector field]{}*]{} with respect to the symplectic form $d{\alpha}$. Of course the associated the Hamiltonian function is $H$. Alternatively, we may write $$\label{eq2.7} u=-{{\mathcal C}}(u)^{-1}\nabla H.$$ \[prop2.2\] Let $u$ be a steady symplectic solution to Euler Equation, and let $c$ be a regular level set of $H(q,t)=P(q,t)+|u(q,t)|^2$ i.e. $\nabla H(q)\neq 0$ whenever $H(q)=c$. Then the restriction of the form ${\alpha}$ to the submanifold $H=c$ is contact. In words, regular level sets of $H$ are contact submanifolds. [**Proof.**]{} By a standard fact in Symplectic Geometry (see for example \[R\]), the level set $H=c$ is contact if and only if we can find a [*Liouville* ]{} vector field $X$ that is transversal to $M_c=\{H=c\}$. More precisely, $${{\mathcal L}}_X\ d{\alpha}=d{\alpha},\ \ \ X(q)\notin T_qM_c,$$ for every $q\in M_c$. Here $T_qM_c$ denotes the tangent fiber to $M_c$ at $q$. The first condition means that $di_Xd{\alpha}=d{\alpha}$. This is satisfied if $i_Xd{\alpha}={\alpha}$. This really means that ${{\mathcal C}}(u)X=u$ and as a result, we need to choose $X={{\mathcal C}}(u)^{-1}u$. It remains to show that $X$ is never tangent to $M_c$. For this, it suffices to check that $X\cdot \nabla H\neq 0$. Indeed, when $H=c$, $$X\cdot \nabla H={{\mathcal C}}(u)^{-1}u\cdot \nabla H=-u\cdot {{\mathcal C}}(u)^{-1}\nabla H=|u|^2\neq 0,$$ by because by assumption $\nabla H\neq 0$. We are done. 502em $\Box$=0 [**[Example 2.1]{}**]{} In this example we describe some simple solutions when the dimension is even. We use polar coordinates to write $x_i=r_i\cos{\theta}_i,$ $y_i=r_i\cos{\theta}_i$, and let $e_i$ (respectively $f_i$) denote the vector for which the $x_i$-th coordinate (respectively $y_i$-th coordinate) is $1$ and any other coordinate is $0$. Set $$e_i({\theta}_i)=(\cos{\theta}_i) e_i+(\sin{\theta}_i) f_i,\ \ \ e'_i({\theta}_i)=(\sin{\theta}_i) e_i-(\cos{\theta}_i) f_i.$$ We may write $$u=\sum_{i=1}^k\left(a^i e_i({\theta}_i)+b^ie'_i({\theta}_i)\right).$$ The form ${\alpha}=u\cdot dx$ can be written as $${\alpha}=\sum_{i=1}^k\left(a^i dr_i-r_ib^id{\theta}_i\right)=:\sum_{i=1}^k\left(a^i dr_i-B^id{\theta}_i\right).$$ For a simple solution, let us assume that all $a^i$s and $b^i$s depend on $r=(r_1,\dots,r_d)$ only. We then have $$d{\alpha}=\sum_{i<j}(a^i_{r_j}-a^j_{r_i})\ dr_i\wedge dr_j-\sum_{i,j}r_i^{-1}B^i_{r_j}\ dr_j\wedge (r_id{\theta}_i).$$ Now $u$ solves Euler Equation if the vector fields $a=(a^1,\dots,a^d)$ and $b=(b^1,\dots,b^d)$ satisfy $$\begin{aligned} &a_t+{{\mathcal C}}(a)a-E(b)^*b+\nabla_r H=0,\nonumber\\ &b_t+E(b)a=0,\label{eq2.9}\\ &\sum_{i=1}^d (r_ia^i)_{r_i}/r_i=0,\nonumber\end{aligned}$$ for some scalar function $H(r)$. Here $E(b)$ denotes a $d{\times}d$ matrix with entries $E_{ij}=r_i^{-1}B^i_{r_j}$. Note that if $\bar b=\sum_j (b^j)^2/2$, then $$E(b)^*b=[r_i^{-1}(b^i)^2]_i+\nabla \bar b=:\hat b+\nabla \bar b.$$ Hence, by changing $H$ to $H'=H-\bar b$, we may rewrite the first equation in with $$\label{eq2.10} a_t+{{\mathcal C}}(a)a-\hat b+\nabla_r H'=0.$$ When $u$ is a steady solution, the first two equations in simplifies to $$\label{eq2.11} {{\mathcal C}}(a)a-E(b)^*b+\nabla_r H=0,\ \ \ E(b)a=0.$$ We can readily show that $u$ is a symplectic solution if and only if the matrix $E(b)$ is invertible. Moreover, by taking the dot product of both sides of the first equation in , and using the second equation we learn $$\label{eq2.12} a\cdot \nabla H=0.$$ Also, the equation $E(b)a=0$ really means $$\label{eq2.13} a\cdot \nabla B^i=0 \ \ {\text{for }}i=1,\dots,d.$$ When $d=4$ and $u$ is independent of time, it is straight forward to solve : From the last equation in we learn that there exists a function $\psi(r_1,r_2)$ such that $$a^1=\psi_{r_2}/(r_1r_2),\ \ \ a^2=-\psi_{r_1}/(r_1r_2).$$ From this, and we learn that $\nabla H$, $\nabla B^1$, $\nabla B^2$ and $\nabla \psi$ are all parallel. So we may write $$H=\mu(\psi), \ \ B^1=\mu_1(\psi),\ \ B^2=\mu_2(\psi),$$ for some $C^1$ functions $\mu, mu_1,\mu_2:{{\mathbb R}}\to{{\mathbb R}}$. Finally we go back to the first equation in to write $$a^2(a^2_{r_1}-a^1_{r_2})+\frac {B^1B^1_{r_1}}{r_1^2}+\frac {B^2B^2_{r_1}}{r_2^2}+H_{r_1}=0.$$ Expressing this equation in terms of $\psi$ yields the elliptic PDE $$r_1r_2\left[\left(\frac {\psi_{r_1}}{r_1r_2}\right)_{r_1}+\left(\frac {\psi_{r_2}}{r_1r_2}\right)_{r_2}\right] =\left(\frac {\mu'_1}{r_1^2}+\frac {\mu'_2}{r_2^2}-\mu'\right)(\psi).$$ This equation may be compared to the Bragg-Hawthorne Equation that is solved to obtain axi-symmetric steady solutions in dimension three. 502em $\Box$=0 We now turn to the odd dimensions. assume that $d=2k+1$ for $k\in {{\mathbb N}}$. We write $(q_1,\dots,q_n)=(x_1,y_1,\dots,x_k,y_k,z)$ and when $k=1$ we simply write $(q_1,q_2,q_3)=(x,y,z)$. In this case, the standard contact form is $\bar{\lambda}=\sum_{i=1}^ky_idx_i+dz$. Again, locally all contact forms are isomorphic to $\bar{\lambda}$. [**[Definition 2.3]{}**]{} Let $u$ be a solution of Euler Equation. We say that [*Clebsch*]{} variables exist for $u$ in the interval $[0,T]$, if we can find $C^1$ functions $$X_1,\dots,X_k,Y_1,\dots,Y_k:{{\mathbb R}}^d\times [0,T]\to{{\mathbb R}},\ \ \ f,Z:{{\mathbb R}}^d\times [0,T]\to{{\mathbb R}}$$ such that $\Psi_t=(X_1,Y_1,\dots,X_k,Y_k,Z)(\cdot,t)$ is a diffeomorphism, $f>0$, and $$(fu)(x,t)=\left(\sum_{i=1}^k Y_i\cdot \nabla X_i\right)(x,t)+\nabla Z(x,t),$$ for every $t\in[0.T]$. Alternatively, we may write $f{\alpha}_t=\Psi_t^*\bar {\lambda}$. 502em $\Box$=0 As we recalled in the proof of Proposition 2.1(ii), if $d=4$ and a symplectic form is isomorphic to the standard form at infinity, then the isomorphism can be extended to the whole $R^d$. This is no longer true when $d=3$; in fact there is countable collection of non-isomorphic forms ${\lambda}^n$ in ${{\mathbb R}}^3$ such that each ${\lambda}^n$ is isomorphic to $\bar {\lambda}$ at infinity but not globally. A fundamental result of Eliashberg gives a complete classification of contact forms. According to Eliashberg’s Theorem \[El\], any contact form in ${{\mathbb R}}^3$ is conformally isomorphic to one of the following forms - \(i) The standard form $\bar{\lambda}$. - \(ii) The form $\hat {\lambda}=\frac {\sin r}{2r}(x_1 dx_2-x_2 dx_1)+{\cos r} \ dx_3$, where $r^2=x_1^2+x_2^2$. - \(iii) A countable collection of pairwise non-isomorphic forms $\{{\lambda}^n:n\in{{\mathbb Z}}\}$, where each ${\lambda}^n$ is isomorphic to $\bar {\lambda}$ outside the ball $B_1$ but not globally in ${{\mathbb R}}^d$. The above classification is related to the important notion of [*[overtwisted]{}*]{} contact forms. In fact $\hat {\lambda}$ is globally overtwisted whereas ${\lambda}^n$ are overtwisted only in a neighborhood of the origin. (We refer to \[El\] or \[Ge\] for the definition of overtwisted forms). [**[Example 2.2]{}**]{} When $d=3$, we may use cylindrical coordinates $x_1=r\cos{\theta},x_2=r\sin{\theta}$ to write $u=ae({\theta})+be'({\theta})+ce_3,$ where $$e({\theta})=r(\cos{\theta},\sin{\theta},0), \ \ e'({\theta})=r(\sin{\theta},-\cos{\theta},0),\ \ e_3=(0,0,1).$$ A solution is called axisymmetric if $a,b,$ and $c$ do not depend on ${\theta}$. It turns out that any $$u=b(r)e({\theta})+c(r)e_3,\ \ {\alpha}=u\cdot dx=c(r)\ dz-rb(r)\ d{\theta},$$ is a steady solution to Euler equation. Such a solution is contact if $$u\cdot\xi=r^{-1}(c(r)B'(r)-c'(r)B(r))\neq 0,$$ where $B(r)=rb(r)$. For example, if $b(r)=r,c(r)=1$, then we get $${\alpha}=r^2\ d{\theta}+dz=xdy-ydx+dz,$$ is isomorphic to $\bar{\lambda}$. On the other hand, choosing $b(r)=r\sin r, c(r)=\cos r$ would yield exactly $\hat{\lambda}$. 502em $\Box$=0 Symplectic Diffusions {#sec3} ===================== We study stochastic flows associated with diffusions. More precisely consider SDE $$\label{eq3.1} dx(t)=V_0(x(t),t)dt+\sum_{i=1}^kV_i(x(t),t){\circ}dW^i(t),$$ where $(W^i:i=1,\dots,k)$ are standard one dimensional Brownian motions on some filtered probability space $({\Omega},\{{{\mathcal F}}_t\},{{\mathbb P}})$, and $V_0,\dots,V_k$ are $C^r$– vector fields in ${{\mathbb R}}^n$. Here we are using Stratonovich stochastic differentials for the second term on the right-hand of and a solution to the SDE is a diffusion with the infinitesimal generator $$L=V_0\cdot\nabla+\frac 12 \sum_{i=1}^k(V_i\cdot\nabla)^2,$$ or in short $L=V_0+\frac 12\sum_{i=1}^kV_i^2$, where we have simply written $V$ for the $V$-directional derivative operator $V\cdot \nabla$. We assume that the random flow $\phi_{s,t}$ of is well defined almost surely. More precisely for ${{\mathbb P}}-$almost all realization of ${\omega}$, we have a flow $\{ \phi_{s,t}(\cdot,{\omega}):0\le s\le t\}$ where $\phi_{s,t}(\cdot,{\omega}):{{\mathbb R}}^n\to{{\mathbb R}}^n$ is a $C^{r-1}$ diffeomorphism and $\phi_{s,t}(a,{\omega})=:x(t)$ is a solution of subject to the initial condition $x(s)=a$. (We also write $\phi_t$ for $\phi_{0,t}$.) For example a uniform bound on the $C^r$-norm of the coefficients $V_0,\dots,V_k$ would guarantee the existence of such a stochastic flow provided that $r\ge 2$. We also remark that we can formally differentiate with respect to the initial condition and derive a SDE for ${{\Lambda}}_{s,t}(x)={{\Lambda}}_t(x):=D_x\phi_{s,t}(x)$: $$\label{eq3.2} d{\Lambda}_t(x)=D_xV_0(\phi_{s,t}(x),t){{\Lambda}}_t(x)dt+\sum_{i=1}^kD_xV_i(\phi_{s,t}(x),t)\ {{\Lambda}}_t(x){\circ}dW^i(t).$$ Given a differential $\ell$-form ${\alpha}(x;v_1,\dots,v_\ell)$, we define $$\left(\phi_{s,t}^*{\alpha}\right)(x;v_1,\dots,v_\ell)={\alpha}(\phi_{s,t}(x);{{\Lambda}}_{s,t}(x)v_1,\dots,{{\Lambda}}_{s,t}(x)v_\ell).$$ Given a vector field $V$, we write ${{\mathcal L}}_V$ for the Lie derivative in the direction $V$. More precisely, for every differential form ${\alpha}$, $$\label{eq3.3} {{\mathcal L}}_V{\alpha}=(\hat d{\circ}i_V+i_V{\circ}\hat d){\alpha},$$ where $\hat d$ and $i_V$ denote the exterior derivative and $V-$contraction operator respectively. (To avoid a confusion between the stochastic differential and exterior derivative, we are using a hat for the latter.) We are now ready to state a formula that is the stochastic analog of Cartan’s formula and it is a rather straight forward consequence of . We refer to Kunita \[K2\] for a proof. \[prop3.1\] Set ${\bf {V}}=(V_0,V_1,\dots,V_m)$ and $${{\mathcal A}}_{\bf V}={{\mathcal L}}_{V_0}+\frac 12\sum_{i=1}^k{{\mathcal L}}_{V_i}^2.$$ We also $\eta_t$ for $\phi_{s,t}^*\eta$ for any form $\eta$. We have $$\begin{aligned} d{\alpha}_t&=\left({{\mathcal L}}_{V_0}{\alpha}\right)_t\ dt+\sum_{i=1}^k\left({{\mathcal L}}_{V_i}{\alpha}\right)_t {\circ}dW^i(t)\label{eq3.4}\\ &=\left({{\mathcal A}}_{\bf V}{\alpha}\right)_t \ dt+\sum_{i=1}^k\left({{\mathcal L}}_{V_i}{\alpha}\right)_t \ dW^i(t)\nonumber.\end{aligned}$$ [**Example 3.1**]{} - \(i) If ${\alpha}=f$ is a $0$-form, then ${{\mathcal A}}_{\bf V}f=Lf$ is simply the infinitesimal generator of the underlying diffusion. - \(ii) If ${\alpha}=\rho\ dx_1\wedge\dots\wedge dx_n$, is a volume form, then ${{\mathcal A}}_{\bf V}{\alpha}=(L^*\rho)\ dx_1\wedge\dots\wedge dx_n$, where $L^*$ is the adjoint of the operator $L$. - \(iii) If $W=(W^1,\dots,W^n)$ is a $n$-dimensional standard Brownian motion and $$dx=V_0(q,t)dt+dW,$$ then for a volume form ${\alpha}=\rho\ dx_1\wedge\dots\wedge dx_n$, we write ${\alpha}_t=\rho^t\ dx_1\wedge\dots\wedge dx_n$, and becomes $$d\rho^t=L^*\rho^t\ dt+\nabla\rho^t\cdot dW.$$ In particular, when $\nabla\cdot V_0=0$ and $\rho=\rho^0=1$, then $\rho^t=1$ is a solution. In other words, the standard volume $dx_1\wedge\dots\wedge dx_n$ is preserved for such a diffusion if the drift $V_0$ is divergence free. 502em $\Box$=0 We now make two definitions: \[def3.1\] Let ${\alpha}$ be a symplectic form. - \(i) We say that the diffusion is (strongly) ${\alpha}$-symplectic if its flow is symplectic with respect ${\alpha}$, almost surely. That is $\phi_{t}^*{\alpha}={\alpha}$, a.s. - \(ii) We say that the diffusion is weakly symplectic if ${\alpha}_t:=\phi_{t}^*{\alpha}$, is a martingale. Using Proposition 3.1 it is not hard to deduce \[prop3.2\] - \(i) The diffusion is (strongly) ${\alpha}$-symplectic if and only if the vector fields $V_0,V_1,\dots,V_k$ are ${\alpha}$-Hamiltonian, i.e. ${{\mathcal L}}_{V_0}{\alpha}={{\mathcal L}}_{V_1}{\alpha}=\dots={{\mathcal L}}_{V_k}{\alpha}=0$. - \(ii) The diffusion is weakly ${\alpha}$-symplectic if and only if ${{\mathcal A}}_{\bf V}{\alpha}=0$. We discuss two systematic ways of producing weakly symplectic diffusions. [**[Recipe (i)]{}**]{} Given a symplectic form ${\alpha}$, we write $X_H=X_H^{\alpha}$ for the Hamiltonian vector field associated with the Hamiltonian function $H$. Note that by non-degeneracy of ${\alpha}$, there exists a unique vector field $X={{\mathcal X}}^{\alpha}(\nu)$ such that $i_X {\alpha}=\nu$ for every $1$-form $\nu$ and $X_H=-{{\mathcal X}}^{\alpha}(dH)$. In the following proposition, we show that given $V_1,V_2,\dots,V_k$, we can always find a unique $\hat V_0$ such that the diffusion associated with ${\bf V} =(X_H+\hat V_0,V_1,\dots,V_k)$ is weakly ${\alpha}$-symplectic. \[prop3.3\] The diffusion is weakly ${\alpha}$-symplectic if and only if there exists a Hamiltonian function $H$, such that $$\label{eq3.5} V_0=X_H-\frac 12\sum_{j=1}^k{{\mathcal X}}^{\alpha}\left(i_{V_j}\ \hat d\ i_{V_j}{\alpha}\right).$$ [*[Proof.]{}*]{} By definition, $${{\mathcal A}}_{{\bf V}}{\alpha}=\hat d\left[i_{V_0}{\alpha}+\frac 12\sum_{j=1}^k\left(i_{V_j}\ \hat d\ i_{V_j}{\alpha}\right)\right].$$ Hence ${{\mathcal A}}_{{\bf V}}{\alpha}=0$ means that for some function $H$, $$i_{V_0}{\alpha}+\frac 12\sum_{j=1}^k\left(i_{V_j}\ \hat d\ i_{V_j}{\alpha}\right)=-dH.$$ From this we can readily deduce . 502em $\Box$=0 [**[Recipe (ii)]{}**]{} We now give a useful recipe for constructing $\bar{\omega}$-diffusions where $\bar{\omega}$ is the standard symplectic form and $n=2d$. \[prop3.3\] Given a Hamiltonian function $H$, consider a diffusion $x(t)=(q(t),p(t))$ that solves $$\begin{aligned} dq&=H_p(q,p)\ dt+\sum_{j=1}^kA_j(x,t)\ dW^j,\label{eq3.6}\\ dp&=-H_q(q,p)\ dt+\sum_{j=1}^kB_j(x,t)\ dW^j,\nonumber\end{aligned}$$ with $A_j=(A_j^1,\dots,A_j^n)$, and $B_j=(B_j^1,\dots,B_j^n)$. Then $x(t)$ is weakly $\bar{\omega}$-symplectic if and only if $Z_1=(Z_1^i:i=1,\dots,d)=0$ and $Z_2=(Z_2^i:i=1,\dots,d)=0$, where $$\begin{aligned} Z_1^i&=\sum_{r,j}\left(\frac{\partial A^r_j}{\partial q_i}B_j^r -\frac{\partial B^r_j}{\partial q_i} A_j^r\right),\label{eq3.7}\\ Z_2^i&=\sum_{r,j}\left(\frac{\partial A^r_j}{\partial p_i}B^r_j-\frac{\partial B^r_j}{\partial p_i}A^r_j\right). \nonumber\end{aligned}$$ [*[Proof.]{}*]{} The Stratonovich differential is related to Itô differential by $$a\circ dW=a\ dW+\frac 12 [da,dW].$$ As a result, the diffusion $x(t)$ satisfies for $V_j=\bmatrix A_j\\B_j\endbmatrix$, $j=1,\dots,k$, and $V_0=J\nabla H-\frac 12\hat V_0$ with $\hat V_0=\bmatrix A_0\\B_0\endbmatrix$, where $$\begin{aligned} A_0^i&=\sum_{r,j}\left(\frac {\partial A_j^i}{\partial q_r }A^r_j+ \frac {\partial A^i_j}{\partial p_r }B_j^r\right) ,\label{eq2.8}\\ B_0^i&=\sum_{r,j}\left(\frac {\partial B_j^i}{\partial q_r }A^r_j+ \frac {\partial B^i_j}{\partial p_r }B_j^r\right) .\nonumber\end{aligned}$$ We need to show that is satisfied if and only if $Z_1=Z_2=0$. For this, let us write ${\beta}( F)$ for the $1$-form $F\cdot dx$ and observe $$i_V\bar{\omega}={\beta}(JV),\ \ \ \hat d{\beta}(F)(v,w)={{\mathcal C}}(F)v\cdot w,\ \ \ {{\mathcal X}}^{\bar{\omega}}({\beta}(F))=-J F,$$ where ${{\mathcal C}}(F)=DF-(DF)^*$ with $DF$ denoting the matrix of the partial derivatives of $F$ with respect to $x$. From this we deduce $$\sum_{j=1}^k{{\mathcal X}}^{\bar{\omega}}\left(i_{V_j}\ \hat d\ i_{V_j}\bar{\omega}\right)=-\sum_{j=1}^kJ{{\mathcal C}}(JV_j)V_j.$$ On account of this formula and Proposition 3.3, it remains to verify $$\label{eq3.9} \hat V_0=-\sum_{j=1}^kJ{{\mathcal C}}(JV_j)V_j.$$ A straight forward calculation yields $${{\mathcal C}}(JV_j)=\bmatrix X_{11}^j & X_{12}^j \\ X_{21}^j & X_{22}^j\endbmatrix$$ where $$\begin{aligned} X_{11}^j&=\left[\frac{\partial B^i_j}{\partial q_r}-\frac{\partial B^r_j}{\partial q_i}\right]_{i,r=1}^n, \ \ \ X_{12}^j=\left[\frac{\partial B^i_j}{\partial p_r}+\frac{\partial A^r_j}{\partial q_i}\right]_{i,r=1}^n,\\ X_{21}^j&=\left[-\frac{\partial A^i_j}{\partial q_r}-\frac{\partial B^r_j}{\partial p_i}\right]_{i,r=1}^n, \ \ \ X_{22}^j=\left[-\frac{\partial A^i_j}{\partial p_r}+\frac{\partial A^r_j}{\partial p_i}\right]_{i,r=1}^n.\end{aligned}$$ From this we deduce $$J{{\mathcal C}}(JV_j)V_j=\bmatrix Y^j_1\\Y^j_2\endbmatrix,$$ where $$\begin{aligned} Y_1^j&=\left[-\sum_r\left(\frac{\partial A^i_j}{\partial q_r}+\frac{\partial B^r_j}{\partial p_i}\right)A^r_j +\sum_r\left(\frac{\partial A^r_j}{\partial p_i}-\frac{\partial A^i_j}{\partial p_r}\right)B^r_j\right]_{i=1}^n,\\ Y_2^j&=\left[\sum_r\left(\frac{\partial B^r_j}{\partial q_i}-\frac{\partial B^i_j}{\partial q_r} \right)A_j^r -\sum_r\left(\frac{\partial B^i_j}{\partial p_r}+\frac{\partial A^r_j}{\partial q_i}\right)B_j^r\right]_{i=1}^n.\end{aligned}$$ Summing these expressions over $j$ yields $$-\sum_jJ{{\mathcal C}}(JV_j)V_j=\hat V_0-\bmatrix Z_2\\-Z_1\endbmatrix=V_0-J\bmatrix Z_1\\Z_2\endbmatrix,$$ where $Z_1$ and $Z_2$ are defined by . From this we learn that is valid if and if $Z_1=Z_2=0$. This completes the proof. 502em $\Box$=0 An immediate consequence of Proposition 3.4 is Corollary 3.1. \[cor3.1\] Let $x(t)=(q(t),p(t))$ be a diffusion satisfying $$\begin{aligned} dq&=H_p(q,p)\ dt+\sqrt{2\nu}\ dW ,\label{eq3.10}\\ dp&=-H_q(q,p)\ dt +\sqrt{2\nu}{\Gamma}(q,t) dW.\nonumber\end{aligned}$$ where ${\Gamma}$ is a continuously differentiable $d{\times}d$-matrix valued function and $W=(W^1,\dots,W^d)$ is a standard Brownian motion in ${{\mathbb R}}^d$. The process $x(t)$ is weakly $\bar{\omega}$-symplectic if and only if the trace of ${\Gamma}$ is independent of $q$. [*[Proof.]{}*]{} Observe that $x(t)$ satisfies for $A=I_d$ and $B$ that is independent of $p$. From this we deduce that $Z_2=0$ and $Z_1=-\sqrt{2\nu}\nabla_q ( tr {\Gamma})$. We are done. 502em $\Box$=0 Martingale Circulation {#sec4} ====================== [*[Proof of Theorem 1.1.]{}*]{} [**[Step 1.]{}**]{} As in Section 1, we write $D$ and $\nabla$ for $q$-differentiation. For $x$-differentiation however, we write $D_x$ and $\nabla_x$ instead. Let us write $x'(t)=(q'(t),p'(t))$ for a diffusion that satisfies $$\begin{aligned} dq'(t)&=p'(t)\ dt+\sqrt{2\nu}\ d\bar W\nonumber\\ dp'(t)&=-\nabla P(q'(t),t) \ dt+\sqrt{2\nu}\ Dw(q'(t),t)\ d\bar W.\label{eq4.1}\end{aligned}$$ for a time dependent $C^1$ vector field $w$ in ${{\mathbb R}}^d$ and a standard Brownian motion $\bar W$. The flow of this diffusion is denoted by $\phi_t$. We then apply Corollary 3.1 for $H(q,p,t)=\frac 12 |p|^2+P(q,t)$ and ${\Gamma}=Dw$, to assert that the diffusion $x'$ is weakly $\bar{\omega}$-symplectic if $\nabla \cdot w=0$. Let us now assume that $w$ satisfies the backward Navier-Stokes equation $$\label{eq4.2} w_t+(Dw)w+\nabla P+\nu{\Delta}w=0,\ \ \ \nabla\cdot w=0.$$ We observe that if the process $q'(t)$ is a diffusion satisfying $$\label{eq4.3} dq'(t)=w(q'(t),t)\ dt+\sqrt{2\nu}\ d\bar W.$$ and $p'(t)=w(q'(t),t)$, then by Ito’s formula, $$\begin{aligned} dp'(t)&=\left[w_t+(Dw)w+\nu {\Delta}w\right](q'(t),t) \ dt+\sqrt{2\nu}Dw(q'(t),t)\ d\bar W\nonumber\\ &=-\nabla P(q'(t),t) \ dt+\sqrt{2\nu}Dw(q'(t),t)\ d\bar W.\label{eq4.4}\end{aligned}$$ This means that if $\bar Q_t$ denotes the flow of the SDE , then $$\begin{aligned} \label{eq4.5} \phi_t(q,w(q,0))&=(\bar Q_t(q),w(\bar Q_t(q),t)),\\ D_x\phi_t{(q,w(q,0))}\bmatrix a\\Dw(q,0)a\endbmatrix&=\bmatrix (D\bar Q_t(q))a\\(Dw(\bar Q_t(q),t)) (D\bar Q_t(q))a\endbmatrix.\nonumber\end{aligned}$$ By the conclusion of Corollary 3.1, the process $$\hat M_t(x;v_1,v_2)=\left[J(D\phi_t(x))v_1\right]\cdot \left[(D\phi_t(x))v_2\right]$$ is a $2$-form valued martingale. This means that for any surface ${\gamma}:D\to{{\mathbb R}}^d{\times}{{\mathbb R}}^d$, the process $$\hat M_t({\gamma})=\iint_D \hat M_t({\gamma};{\gamma}_{{\theta}_1},{\gamma}_{{\theta}_2})\ d{\theta}_1 d{\theta}_2,$$ is a martingale. We consider a surface that lies on the graph of $w(\cdot,0)$. That is, $${\gamma}({\theta}_1,{\theta}_2)=(\tau({\theta}_1,{\theta}_2), w(\tau({\theta}_1,{\theta}_2),0)),$$ for a surface $\tau:D\to{{\mathbb R}}^d$. We now use to assert that $$\bar M_t(\tau):=\hat M_t({\gamma})=\iint_D \bar M_t(\tau;\tau_{{\theta}_1},\tau_{{\theta}_2})\ d{\theta}_1 d{\theta}_2,$$ where $$\begin{aligned} \bar M_t(q;a_1,a_2)&=J\bmatrix (D\bar Q_t(q))a_1\\(Dw(\bar Q_t(q),t))(D\bar Q_t(q))a_1\endbmatrix\cdot \bmatrix (D\bar Q_t(q))a_2\\(Dw(\bar Q_t(q),t))(D\bar Q_t(q))a_2\endbmatrix\\ &=\left[(Dw-(Dw)^*)( \bar Q_t(q),t)\right](D\bar Q_t(q))a_1\cdot (D\bar Q_t(q))a_2\\ &=\bar Q_t^*\hat d\bar {\alpha}_t(q;a_1,a_2),\end{aligned}$$ where $\bar{\alpha}_t=w(q,t)\cdot dq$. In summary, $\bar M_t=\bar Q_t^*\hat d\bar {\alpha}_t$ is a martingale. When $d=3$, $$\begin{aligned} \bar Q_t^*\hat d\bar {\alpha}_t(q;a_1,a_2) &=\left[(\eta^t\circ \bar Q_t(q)){\times}(D\bar Q_t(q))a_1\right]\cdot (D\bar Q_t(q))a_2\\ &=\left[\eta^t\circ \bar Q_t(q),(D\bar Q_t(q))a_1,(D\bar Q_t(q))a_2\right],\end{aligned}$$ where $\eta^t(\cdot)=\nabla{\times}w(\cdot,t)$ and $[a,b,c]$ is the determinant of a matrix with column vectors $a,b$ and $c$. Since $w$ is divergence-free, the flow $\bar Q_t$ is volume preserving (see Example 3.1(iii)). Hence $$\bar M_t(q;a_1,a_2)=[(D\bar A_t\circ \bar Q_t(q))\eta^t\circ \bar Q_t(q),a_1,a_2],$$ where $\bar A_t=\bar Q_t^{-1}$. Since $\hat M_t$ is a martingale, we deduce that the process $$\tilde M_t(q)=(D\bar A^t\circ \bar Q_t(q))(\eta^t\circ \bar Q_t(q)),$$ is a martingale. [**[Step 2.]{}**]{} Suppose that now $u$ is a solution to the forward Navier-Stokes equation and recall that when $d=3$, we write $\xi=\nabla{\times}u$. We set $w(q,t)=-u(q,T-t)$ for $t\in[0,T]$. Then $w$ satisfies in the interval $t\in[0,T]$. Recall that $q(t)$ is the solution of SDE with the flow $Q_t$. We choose $\bar W(t)=W(T-t)-W(T)$ in the equation . According to a theorem of Kunita (see Theorem 13.15 in page 139 of \[RW\] and \[K1\]), the flows $Q$ and $\bar Q$ are related by the formula $$\bar Q_t=Q_{T-t}\circ Q_T^{-1}=B_t.$$ Observe that $\bar {\alpha}_t=-{\alpha}_{T-t}$ and $$\bar M_t=\bar Q_t^*\hat d\bar{\alpha}_t=-B_t^*\hat d{\alpha}_{T-t}=-{\beta}_t.$$ Hence $({\beta}_t: t\in[0,T])$ is a martingale because $\bar M_t$ is a martingale by Step 1. Also, when $d=3$, $$\tilde M_t=\big((D\bar A_t)\eta^t\big)\circ \bar Q_t(q))=- \left(\left(DB_t^{-1}\right)\xi^{T-t}\right) \circ B_t.$$ This completes the proof of Part $(i)$. [**[Step 3.]{}**]{} The process $x'(t)$ is a diffusion of the form with $k=d$ and $$V_i(x',t)=V_i(q,t)=\bmatrix e_i\\w_{q_i}\endbmatrix,$$ where $e_i=[{\delta}_i^j]_{j=1}^d$ is the unit vector in the $i$-th direction. A straight forward calculation yields that for the standard symplectic form $\bar{\omega}=\sum_j dp_j\wedge dq_j$, $$\begin{aligned} i_{V_i}\bar{\omega}&=w_{q_i}\cdot dq-dp_i=:{\gamma}^i-dp_i,\\ {{\mathcal L}}_{V_i}\ \bar{\omega}&=\hat d{\gamma}^i=\sum_{j,k}w^k_{q_iq_j}\ dq_j\wedge dq_k,\\ \zeta^i(v_1,v_2)&=\hat d{\gamma}^i(v_1,v_2)={{\mathcal C}}(w_{q_i})v_1\cdot v_2,\end{aligned}$$ where $w=(w^1,\dots,w^d)$. From this and we deduce if $$z_t=\int_\Theta\ (\hat d\bar {\alpha})_t,\ \ \ y_i(t)=\int_\Theta\ \zeta^i_t,$$ then $$dz_t=\sum_{i=1}^d y_i(s)\ dW^i(t),$$ because by Step 1, we know that ${{\mathcal A}}_{\bf V}{\omega}=0$. From this, we readily deduce that the quadratic variation of the process $z_t$ is given by $$\int_0^t \sum_i y_i(s)^2\ ds.$$ We now reverse time as in Step 2 to complete the proof of Part $(ii)$. Part $(iii)$ is an immediate consequence of the identity $${{\mathbb E}}z_T^2={{\mathbb E}}z_0^2+{{\mathbb E}}\int_0^T\sum_i y_i(s)^2\ ds.$$ 502em $\Box$=0 Contact Diffusions {#sec5} ================== Recall that contact forms are certain $1$-forms that are non-degenerate in some rather strong sense. To explain this, recall that when ${\alpha}$ is a contact form in dimension $n=2d+1$, then the set $l_x=\{v:d{\alpha}_x(v,w)=0 \text{ for all }w\in T_xM\}$ is a line. Also, if we define the kernel of ${\alpha}$ by $$\eta_x^{\alpha}=\eta_x=\{v: {\alpha}_x(v)=0\},$$ then the contact condition really means that $l_x$ and $\eta_x$ give a decomposition of ${{\mathbb R}}^{n}$ that depends solely on ${\alpha}$: $$\label{eq5.1} {{\mathbb R}}^{n}=\eta_x\oplus l_x.$$ We also define the Reeb vector field $R(x)=R^{\alpha}(x)$ to be the unique vector such that $$R(x)\in l_x,\ \ \ {\alpha}_x(R(x))=1.$$ The role of Hamiltonian vector fields in the contact geometry are played by contact vector field. [**[Definition 5.1]{}**]{} A vector field $X$ is called an ${\alpha}$-contact vector field if ${{\mathcal L}}_X{\alpha}=f{\alpha}$ for some scalar-valued continuous function $f$. 502em $\Box$=0 It is known that for a given a “Hamiltonian” $H:M\to{{\mathbb R}}$, there exists a unique contact ${\alpha}$-vector field $X_H=X_{H,{\alpha}}$ such that $i_{X_H}{\alpha}={\alpha}({X_H})=H$. The function $f$ can be expressed in terms of $H$ with the aid of the Reeb’s vector field $R=R^{\alpha}$; indeed, $f=dH(R^{\alpha})$, and as a result, $${{\mathcal L}}_{X_H}{\alpha}=dH(R^{\alpha}){\alpha}.$$ In our Euclidean setting, we consider a form ${\alpha}=u\cdot dx$ for a vector field $u$ and $${\beta}(v_1,v_2):=d{\alpha}(v_1,v_2)={{\mathcal C}}(u)v_1\cdot v_2,$$ where ${{\mathcal C}}(u)=Du-(Du)^*$. (Recall that we are writing $A^*$ for the transpose of $A$.) Since $C^*=-C$, we have that $\det {{\mathcal C}}=(-1)^n \det {{\mathcal C}}$. This implies that $C$ cannot be invertible if the dimension is odd. Hence the null space $l_x$ of ${{\mathcal C}}(u)(x)$ is never trivial and our assumption $dim\ell_x=1$ really means that this null space has the smallest possible dimension. Now simply means that $u(x)\cdot R(x)\neq 0$. Of course $R$ is chosen so that $u(x)\cdot R(x)\equiv 1$. Writing $u^{\perp}$ and $R^{\perp}$ for the space of vectors perpendicular to $u$ and $R$ respectively, then $\eta=u^\perp$, and we may define a matrix ${{\mathcal C}}'(u)$ which is not exactly the inverse of ${{\mathcal C}}(u)$ (because ${{\mathcal C}}(u)$ is not invertible), but it is specified uniquely by two requirements: - \(i) ${{\mathcal C}}'(u)$ restricted to $R^\perp$ is the inverse of ${{\mathcal C}}(u):u^\perp\to R^\perp$. - \(ii) ${{\mathcal C}}'(u)R=0$. The contact vector field associated with $H$ is given by $$X_H=-{{\mathcal C}}'(u) \nabla H+H R.$$ In particular, when $n=3$, the form ${\alpha}=u\cdot dx$ is contact if and only if $u\cdot \xi$ is never $0$, where $\xi=\nabla {\times}u$ is the curl (vorticity) of $u$. In this case the Reeb vector field is given by $R=\xi/(u\cdot \xi)$, and $${{\mathcal L}}_Z u=\nabla(u\cdot X)+\xi{\times}Z,$$ We also write $\bar u=u/\rho$. The contact vector field associated with $H$ is given by $$X_H=\bar u{\times}\nabla H+H R.$$ Let $x(t)$ be a diffusion satisfying and assume that this diffusion has a random flow $\phi_t$. Given a contact form ${\alpha}={\alpha}_0$, set ${\alpha}_t=\phi_t^*{\alpha}_0$ as before. [**Definition 5.2.**]{} - \(i) We say that the diffusion is strongly ${\alpha}$-contact, if for some scaler-valued semimartingale $Z_t$ of the form, $$\label{eq5.2} dZ_t=g_0(x(t),t)\ dt+\sum_{i=1}^kg_i(x(t),t)\circ dW^i(t),$$ we have $$d{\alpha}_t={\alpha}_t\ dZ_t.$$ - \(ii) We say that the diffusion is weakly ${\alpha}$-contact, if there exists a continuous scalar-valued function $f(x,t)$ such that $$M_t={\alpha}_t-\int_0^tf(x(s),s){\alpha}_s\ ds,$$ is a martingale. 502em $\Box$=0 We end this section with two proposition. \[prop5.1\] The following statements are equivalents: - \(i) The diffusion is strongly ${\alpha}$-contact. - \(ii) There exists a scaler-valued process $A_t$ of the form $$dA_t=h_0(x(t),t)\ dt+\sum_{i=1}^kh_i(x(t),t)\circ dW^i(t).$$ such that ${\alpha}_t=e^{A_t}{\alpha}.$ (Recall ${\alpha}_t=\phi_t^*{\alpha}$ with $\phi_t=\phi_{0,t}$ representing the flow of the diffusion .) - \(iii) The vector fields $V_0,\dots,V_k$ are ${\alpha}$-contact. \[prop5.2\] The following statements are equivalents: - \(i) The diffusion is weakly ${\alpha}$-contact. - \(ii) For some scalar-valued function $f(x,t)$, we have ${{\mathcal A}}_{\bf V}{\alpha}=f{\alpha}$. The proof of Proposition 5.2 is omitted because it is an immediate consequence of and the definition. [**[Proof of Proposition 5.1.]{}**]{} Suppose that the vector fields $V_0,\dots,V_k$ are ${\alpha}$-contact. Then there exist scalar-valued functions $g_0(x,t),\dots,g_k(x,t)$ such that ${{\mathcal L}}_{V_i}{\alpha}=g_i{\alpha}$. From this and Proposition 3.1 we learn that $d{\alpha}_t={\alpha}_t \ dZ_t$ for $Z_t$ as in . Hence $(iii)$ implies $(i)$. Now assume $(i)$ and set $$Y_t=\exp\left(-Z_t+\frac 12[Z]_t\right).$$ We have $$\begin{aligned} dY_t&=Y_t(- dZ_t+d[Z]_t),\\ d\left(Y_t {\alpha}_t\right)&={\alpha}_tY_t(- dZ_t+d[Z]_t)+Y\ d{\alpha}_t+ d[Y,{\alpha}]_t\\ &= {\alpha}_tY_t\ d[Z]_t+ d[Y,{\alpha}]_t= {\alpha}_tY_t\ d[Z]_t- {\alpha}_tY_t\ d[Z]_t=0.\end{aligned}$$ Hence $Y_t{\alpha}_t={\alpha}$ and we have $(ii)$ for $A_t=Z_t-\frac 12[Z]_t$. We now assume $(ii)$. We certainly have $$\begin{aligned} d{\alpha}_t&={\alpha}e^{A_t}\left(dA_t+\frac 12d[A]_t\right)={\alpha}_t\left(dA_t+\frac 12d[A]_t\right)\\ &={\alpha}_t\left(g_0(x(t),t)\ dt+\sum_{i=1}^kg_i(x(t),t)\circ dW^i(t)\right),\end{aligned}$$ for $g_0=h_0+\frac 12(\sum_i h_i^2)$ and $g_i=h_i$ for $i=1,\dots,k$. Comparing this to yields ${{\mathcal L}}_{V_i}{\alpha}=g_i{\alpha}$ for $i=0,\dots,k$. Hence $(iii)$ is true and this completes the proof. 502em $\Box$=0 [HR]{} P. Constantin and G. Iyer, A stochastic Lagrangian representation of the three-dimensional incompressible Navier-Stokes equations. Commun. Pure Appl. Math. [**LXI**]{}, 330-345, (2008). Y. Eliashberg, Classification of contact structures on ${{\mathbb R}}^3$. Internat. Math. Res. Notices , no. 3, 87–91, (1993). G. L. Eyink, Turbulent diffusion of lines and circulations. Physics Letters A [**368**]{}, 486-490, (2007). G. L. Eyink, A stochastic least-action principle for the incompressible Navier-Stokes equation, submitted to Physica D. arxiv:0810.0817 \[math-ph\], (2009). H. Geiges, An Introduction to Contact Topology. Cambridge university Press, Cambridge, 2008. M. Gromov, Pseudoholomorphic curves in symplectic manifolds. Invent. Math. [**82**]{}, 307–347, (1985). H. Hofer and E. Zehnder, Symplectic Invariants and Hamiltonian Dynamics. Birkhäuser Verlag, Basel, 2011. H. Kunita, On backward stochastic differential equations. Stochastica, [**6**]{}, 293-313 (1982). H. Kunita, Stochastic Flows and Stochastic Differential Equations. Cambridge university Press, Cambridge, 1990. F. Rezakhanlou, Lectures on Symplectic Geometry, www.math.berkeley.edu/rezakhan. L. C. G. Rogers and D. Williams, Diffusions, Markov processes, and martingales. Vol. 2. Itô calculus. Cambridge University Press, Cambridge, 2000. [^1]: This work is supported in part by NSF Grant DMS-1106526.

ArXiv

--- abstract: 'The Force Concept Inventory (FCI) was developed by Hestenes, Wells and Swackhamer, in order to assess student understanding of the concept of force. FCI has been used for over 20 years and in different countries. When applying the inventory in a new context it is important to evaluate the reliability and discrimination power of this assessment tool. In this study the reliability and discrimination power are evaluated in the context of Engineering education at a Norwegian university, using statistical tests, focusing on both item analysis and on the entire test. The results indicate that FCI is a reliable and discriminating tool in most cases. As there are exceptions, statistical tests should always be done when FCI is administered in a new context.' author: - | J.R. Persson\ Programme for Teacher Education\ Norwegian University of Science and Technology\ NO-7491 Trondheim\ Norway\ title: Evaluating the Force Concept Inventory for different student groups at the Norwegian University of Science and Technology --- Introduction ============ Standardised multiple-choice tests can be used as a tool in physics education to assess student learning. A number of such tests have been developed covering a range of different domains in physics. One of the most commonly used test, the Force Concept Inventory (FCI), was introduced by Hestenes, Wells and Swackhamer [@Hestenes1992]. FCI has since then been used as a tool in assessing the efficiency of a number of developed teaching methods (see for example [@Coletta2007]). FCI is limited to the understanding of the concept of force, but an increased understanding of force should work as a more general indication of learning in mechanics as a whole. Considering the extended use of FCI, it should also work as a tool in assessing the learning of Norwegian students in their introduction physics courses. In order to investigate the reliability and discrimination power of FCI at a Norwegian university, a number of statistical tests focusing both on individual items and on the test as a whole, has been performed. There exist two aspects of test reliability; consistency and discriminatory power. A test is said to be reliable if it is consistent within itself and over time. If a test is shown to be reliable, one can assume that the same students would get the same score if they would take the test again after a period of time. A large variance in the test score of a reliable test will then depend on a systematic variation in the student population, where different levels of understanding or mastery will give different scores on the test. Both these aspects of test reliability can be assessed statistically. In order to evaluate the reliability of the FCI in a Norwegian context, the test was administrated to different student groups at the Norwegian University of Science and Technology (NTNU) in Trondheim. Even if the test is intended to be of general use, the level of the students understanding or mastery will affect the usefulness of the test, especially when the group has a higher degree of understanding or mastery. Background ========== A concept inventory is a criterion-referenced test designed to evaluate if students have an accurate knowledge of a specific set of concepts within a defined area. Concept inventories are typically organized as multiple-choice tests in order to ensure that they are objectively scored in a reproducible manner and possible to administrate in large classes. Unlike a teacher-made multiple-choice test, questions and response choices in concept inventories are a subject of extensive research and development. The aims of the research may include ascertaining (a) the range of what individuals think a particular question is asking and (b) the most common responses to the questions. In the concept inventory, each question includes one correct answer and several distractors. The distractors are incorrect answers that are usually (but not always) based on students’ commonly held misconceptions. Ideally, the scores should reflect the amount of content knowledge students has mastered. The purpose of a criterion-referenced test is to ascertain whether students master a predetermined amount of content knowledge. The distractors are often based on ideas commonly held by students, as determined by years of research on misconceptions. The Force Concept Inventory (FCI) [@Hestenes1992] is a multiple-choice test, designed to assess student understanding of the most basic concepts in Newtonian physics, particular forces. The test has 30 questions covering six areas of understanding: kinematics, Newton’s First, Second, and Third Laws, the superposition principle, and types of forces (such as gravitation, friction). Each question has only one correct Newtonian answer, with distractors based on student’s common misconceptions. A low score indicates that the student has an Aristotelian view while a high score (typically around 60% correct or higher) indicates a Newtonian understanding. The Norwegian version of FCI used, was translated and developed by Angell and collaborators at University of Oslo [@Angell2012]. Student groups ============== The test was given in three different courses with different student groups, both as a pre-(instruction) and post-(instruction) test in the Fall semester 2012. The Courses were traditional calculus-based introductory physics courses. As all engineering students at NTNU have to take at least one course in physics, it was possible to administer FCI to both physics masters and non-physics masters. However, different physics courses are given to different masters programs, but all courses contain about the same amount of content relevant for the FCI survey during lectures and are using the same textbook as the main source. The three groups consisted of students in different physics courses; Mechanical Physics (TFY4145/FY1001) for Physics Masters; Physics (TFY4104) for Master students in Marine Technology, Industrial Economics and Technology Management and Mechanical Engineering; and Physics (TFY4115) for Master students in Electronics, Engineering Cybernetics and Nanotechnology. It should also be noted that TFY4104 and TFY4115 include electromagnetics and thermodynamics, respectively, in addition to mechanics. The test was voluntary with no extra credit given. The numbers of students taking the tests are given in Table \[TableKey1\]. The result of the tests with respect to understanding will be presented elsewhere as we focus on the reliability of the test in this paper. The students in the different groups have a similar background, but one can assume that the Physics Masters has a more explicit interest and knowledge in physics and will subsequently score higher on the FCI. The Physics masters and Nanotechnology students are generally believed to be high-achieving students as admission grades are higher compared with the other Master programs. The Physics masters and Nanotechnology students are first year students while the others are second year students. By examining the results in the different groups it is possible to establish the reliability within each group. Using the data from the individual groups we performed five statistical tests: three focusing on individual items (item difficulty index, item discrimination index , item point biserial coefficient) and two on the test a whole (Kuder-Richardson test reliability and test Ferguson’s $\delta$). Pre-test Post-test --------- ---------- ----------- TFY4104 182 105 TFY4115 91 58 TFY4145 140 91 : Number of students taking the FCI.[]{data-label="TableKey1"} Item difficulty index ===================== The item difficulty index (P) is a measure of the difficulty of each test item and of the test as a whole. It is defined as the ratio of the total number $N_{1}$ of correct answers to the total number $N$ of students who answered the specific item: $$\mathbf{P=}\frac{N_{1}}{N} \label{1}$$ The difficulty index is, however, somewhat misnamed, since it is simply the proportion of correct answers to a particular item, where the name easiness index would be more appropiate. The greater P value, the higher percentage of correct answers and consequently the easier the item is for the population. The difficulty index will thus depend on the population, something which is the case in this study. There are a number of different criteria for acceptable values of the difficulty index for a test [@Doran1980]. The optimum value for an item should be $P=0.5$, while it is useful to have a sensible range. A widely adopted criterion requires the difficulty index to be between 0.3 and 0.9 for each question. For a test with a large number ($M$) of items it is more sensible to consider the test difficulty as the average difficulty index ($\bar{P}$) of all the items ($P_{i}$): $$\mathbf{\bar{P}=}\frac{1}{M}\sum P_{i}$$ Figures \[fig:Figure1\] and \[fig:Figure2\] plots the difficulty index $P $ values for each question in FCI, for the three different student groups and the pre-test and post-test respectively. The difficulty index in the pre-tests, fall, in most cases, within the desired range of 0.3-0.9. There are maximum 4 items with difficulty index above 0.9 in the pre-test, something that is acceptable. In the post-test the number of questions with a difficulty above 0.9 rises to 14 and 7, for the TFY4145 and TFY4115 groups, respectively. The average difficulty indexes for the tests are given in table \[TableKey2\]. The average difficulty indexes range from 0.70 to 0.78 in the pre-test to 0.72 to 0.86 in the post-test. Even if the results fall within the acceptable range, the values are very close to the limit of the acceptable range. Taking in to account that the number of items with a difficulty index over 0.9 is large, the use of FCI in its original form as a post-test for a student group such as the TFY4145 group, is very questionable. However, if one want to study the weaker part of the student population, the test can still be used. ![Difficulty index pre-test[]{data-label="fig:Figure1"}](Figure1){width="14.4cm" height="8.1cm"} ![Difficulty index post-test[]{data-label="fig:Figure2"}](Figure2){width="11cm" height="6.55cm"} Pre-test (%) Post-test (%) --------- -------------- --------------- TFY4104 70 72 TFY4115 72 81 TFY4145 78 86 : Average difficulty index[]{data-label="TableKey2"} Item discrimination index ========================= The item discrimination index (D) is a measure of the discriminatory power for the individual items in a test. That is, the extent to which an individual test item distinguishes a student who know the material well from those who do not. A high discrimination index will indicate a higher probability for students with a higher level of knowledge to answer the item correctly, while those with less knowledge will get the wrong answer. The item discrimination index (D) is calculated by dividing the sample into two groups of equal size, a high (H) score group and a low (L) score group based on their individual total scores on the test. For each specific item, one counts the number of correct answers in both the high and low groups: $N_{H}$ and $N_{L}$. Using the total number of students taking the test (N), the discrimination index for a specific item can be calculated as $$D=\frac{N_{H}-N_{L}}{N/K}$$ where K is a numerical factor based on how the division into the high and low group is made. If we split the sample in two, using the median, the high and low groups consist each of 50% of the total sample, giving K=2. However, it is possible to use other groupings, for example taking the top 25% as the high group and the bottom 25% as the low group. The 50%-50% grouping may underestimate the discrimination power, since it takes all students into account even those where the difference is small. To reduce the probability of underestimating the discrimination power we use a 25%-25% grouping. The discrimination index is then expressed as: $$D=\frac{N_{H}-N_{L}}{N/4}$$ The range of the item discrimination index D is \[-1,+1\], where +1 is the best value and -1 the worst. In the case where all students in the high score group and none in the low score group get the correct answer the discrimination index would be +1. If none in the high score group and all in the low score group get the correct answer the discrimination index would be -1. These extremes are very unlikely, but shows that items with a negative discrimination index should be removed. A question is typical considered to provide a good discrimination if D$>$0.3 [@Doran1980], lower values indicate that students resort to guessing on that item. In a test with a large number of items it is possible to allow a few items with a lower discrimination index, but the majority should have higher discrimination indices in order to ensure that the test can distinguish students with strong and weak mastery. It is useful to calculate the averaged discrimination index ($\bar{D}$) for all items in the test. $$\bar{D}=\frac{1}{M}\sum D_{i}$$ Figures \[fig:Figure3\] and \[fig:Figure4\] plots the discrimination index D values for the items in FCI, for the three different student groups and the pre-test and post-test respectively. A majority of the items in the pre-test has a discrimination index D$>$0.3, only a few has a lower value, with variations between different groups. The average discrimination indices are 0.49, 0.49 and 0.45, for the different student groups (TFY4145, TFY4104, and TFY4115, respectively). This indicates that the FCI has a good discriminating power in the pre-test situation. In the post-test the number of questions with the discrimination index D$<$0.3, rises to 15 and 11, out of 30, for the TFY4145 and TFY4115 groups, respectively, while the averaged discrimination index decreases to 0.36 and 0.42, respectively. This raises serious doubts as how applicable the post-test is for the TFY4145 group. The discrimination power for the TFY4115 group is lower than in the pre-test but still within the accepted range. In the TFY4104 group, the discrimination index remains almost the same (0.48). Questions 6, 16 and 29, and to some extent question 19 are especially doubtful as they combine a high difficulty index, that is being quite simple, with a low discrimination index, not distinguishing the high score and low score groups in these student groups. ![Item Discriminating index pre-test[]{data-label="fig:Figure3"}](Figure3){width="14.4cm" height="8.1cm"} ![Item Discrimination index post-test[]{data-label="fig:Figure4"}](Figure4){width="14.4cm" height="8.53cm"} Point biserial coefficient ========================== The point biserial coefficient is another measure of the individual item reliability. It reflects the correlation between the total score and the score on individual items in the test. A positive coefficient indicates that a student with a high total score is more likely to answer the item correctly than a student with a low total score. Thus giving a complementary measure to the item discrimination index. In order to calculate the point biserial coefficient for an item, one obtain the correlation between the score for a question and the total scores. If the number of items in the test is sufficiently large, $>$20, the test can be viewed as continuous. The point biserial coefficient can then be defined as: $$r_{pbc}=\frac{\bar{X}_{1}-\bar{X}_{0}}{\sigma_{x}}\sqrt{\frac{P}{1-P}}$$ Where $\bar{X}_{1}$ is the average total score for those who answered a item correctly, $\bar{X}_{0}$ is the average total score for all participants, $\sigma _{x}$ is the standard deviation of the total scores and P is the difficulty index for this specific item. For an item to be considered as reliable it should be consistent with the whole test, a high correlation between individual item scores and the total score is desirable. A satisfactory point biserial coefficient is $r_{pbc}$$>$ 0.2[@Doran1980]. Items with lower values may be used, as long as the number of these items is small, but the test as a whole should have an average higher than 0.2. The average point biserial coefficients for the different student groups are given in table \[TableKey3\]. All values are greater than 0.2 so the overall items has a fairly high correlation with the whole test. Figures [fig:Figure5]{} and \[fig:Figure6\] shows the point biserial coefficients for the individual items in the pre- and post-tests for different student groups, respectively. It should be noted that questions 16, 19 and 29 overall show a lower degree of correlation than the others. As these questions also show a lower degree of discrimination and these might be subject to revision, at least in the context of the student groups in this study. There is a course-dependent variation of the point biserial coefficient for the post-test. These variations might be due to statistical variations or different course context. ![Point biserial coefficient pre-test[]{data-label="fig:Figure5"}](Figure5){width="14.4cm" height="8.1cm"} ![Point biserial index post-test[]{data-label="fig:Figure6"}](Figure6){width="14.4cm" height="8.56cm"} Pre-test Post-test --------- ---------- ----------- TFY4104 0.45 0.48 TFY4115 0.42 0.49 TFY4145 0.50 0.34 : Average point biserial coefficients[]{data-label="TableKey3"} Test analysis ============= The reliability of single items in the test is measured by the point biserial coefficient. In order to examine the reliability of the test as a whole, other methods have to be used. In this work we use two measures of the reliability for the test as a whole: Kuder-Richardson reliability index and Ferguson’s delta ($\delta$). Kuder-Richardson reliability index ---------------------------------- A not very practical way to evaluate the reliability of a test, is to administer it twice to the same sample. In such a case we would expect a significant correlation between the two test scores, provided the students’ performance is stable and the test conditions are the same. The correlation coefficient between the two sets of scores will be defining the reliability index of the test. It is obvious that this method is not practical to use. In the case of a test that has been designed specifically for a certain knowledge domain and with parallel questions, the Spearman-Brown formula [@Ghiselli1981] can be used to calculate the reliability index. This equation connects the reliability index with the correlation between any parallel equally sized subsets in the test. Kuder and Richardson [Kuder1937]{} developed this idea further by dividing the test into the smallest possible subsets, that is individual items. This means that each item is considered as a single parallel test, and assuming that the means, variance and standard deviation is the same for all items in the whole test. The result derived gives the reliability index as: $$r_{test}=\frac{M}{M-1}\left( 1-\frac{\sum\sigma_{xi}^{2}}{\sigma_{x}^{2}}\right)$$ where M is the number of items in the whole test, $\sigma _{xi}$ is the standard deviation for the ith item score and $\sigma _{x}$ is the standard deviation of the total test score. This expression takes the different variances of the individual items into account, relaxing the assumption that all items must have the same means, variance and standard deviation. For multiple-choise tests the formula can be rewritten as: $$r_{test}=\frac{M}{M-1}\left( 1-\frac{\sum P_{i}(1-P_{i})}{\sigma_{x}^{2}}\right)$$ where M is the number of items in the test, $P_{i}$ is the difficulty index for each item and $\sigma _{x}$ is the standard deviation of the total test score. These are the Kuder-Richardson reliability formulas, often referred to as KR-20 and KR-21 as being formula 20 and 21 in Kuder and Richardson’s original paper [@Kuder1937] The possible range of the Kuder-Richardson reliability index is between 0 and 1, where a value greater than 0.7 would make the test reliable for group measurements and a value over 0.8 for assessing individuals [@Doran1980]. In this study the obtained Kuder-Richardson reliability indices are all over 0.8 (Table \[TableKey4\]).Something that also open up for individual assessment. Pre-test Post-test --------- ---------- ----------- TFY4104 0.87 0.88 TFY4115 0.84 0.86 TFY4145 0.90 0.87 : Kuder-Richardson reliability index[]{data-label="TableKey4"} Ferguson’s delta ---------------- Ferguson’s delta is another widely used whole test statistic. It measures the discriminatory power of the whole test by investigating how the students’ individual scores are distributed. In a test one aims at a broad distribution of total scores, as this is supposwed to show a better discrimination. The expression of Ferguson’s delta can be written as [@Kline1986 p 150]: $$\delta=\frac{N^{2}-\sum f_{i}^{2}}{N^{2}-\left( N^{2}/\left( M+1\right) \right) }$$ where N is the number of students taking the test, M is the number of items in the test and $f_{i}$ is the frequency of cases with the same score. One should be aware that Ferguson’s delta is more a measure of the population than the test itself, since a change in population will change the result of the Ferguson’s delta formula, while not testing the test itself. If a test and population combined has a Ferguson’s delta greater than 0.90, it is considered to provide a good discrimination for this population [Kline1986]{}. In our study the Ferguson’s delta is greater than 0.90, in all cases as is shown in table \[TableKey5\]. Pre-test Post-test --------- ---------- ----------- TFY4104 0.98 0.97 TFY4115 0.97 0.94 TFY4145 0.96 0.91 : Ferguson’s delta[]{data-label="TableKey5"} Discussion ========== The reliability and discriminatory power of the Force Concept Inventory has been evaluated using five statistical tests in three different student groups, both in pre-instructional and post-instructional tests, at the Norwegian University of Science and Technology (NTNU) . The aim of this study was to test the applicability of the FCI in different contexts, as made possible with Physics majors and engineering students required to take at least one physics course at NTNU. We have found that the FCI is reliable and discriminating enough for pre-tests in all student groups. The post-test for Physics majors (TFY4145) can not be considered as applicable in the present form for the full group, the average difficulty index (86%) has reached a level where ceiling effects will cause problems. The average discrimination index, though still over the 0.3 level, is not a good indicator as half of the questions have a discrimination index below that level. A similar but not as serious problem can also be seen in the TFY4115 group. However, it is still possible to use the test for specific subgroups, that is low achieving students, in order to investigate their understanding. The Force Concept Inventory is a widely used instrument, but as has been shown here, it can not be used without taking the context and student groups into account. Used on a high-achieving group, there is a substantial risk of encountering ceiling effects, with a decrease in discriminatory power. It will still be useful for the students within this group that has not obtained an understanding of the fundamental concepts. Questions 6, 16, 19 and 29 in the FCI are somewhat problematic and might be replaced with other questions in a high-achieving group, such as TFY4145. However, one might also consider constructing a special high-achieving FCI suitable for Physics majors. [9]{} Hestenes, D., Wells, M. and Swackhamer, G. (1992) Force concept inventory, Phys. Teach. 30, 141. Coletta, V. P., Phillips, J. A. and Steinert, J. J. (2007) Interpreting force concept inventory scores: Normalized gain and SAT scores, Phys. Rev. ST Phys. Educ. Res. 3, 010106. Angell, C (2012) Private communication Doran, R. (1980) Basic Measurement and Evaluation of Science Instruction. (National Science Teachers Association, Washington DC) Ghiselli, E., Campbell, J. and Zedeck, S. (1981) Measurement theory for the behavioural sciences (Freeman, San Francisco) Kuder, G. and Richardson, M. (1937) The Theory of the estimation of psychometrika test reliability, Psychometrika 2, 151 Kline, P. (1986) A Handbook of Test Construction: Introduction to Psychometic Design (Methueen, London)

ArXiv

--- abstract: 'Used to estimate the risk of an estimator or to perform model selection, cross-validation is a widespread strategy because of its simplicity and its apparent universality. Many results exist on the model selection performances of cross-validation procedures. This survey intends to relate these results to the most recent advances of model selection theory, with a particular emphasis on distinguishing empirical statements from rigorous theoretical results. As a conclusion, guidelines are provided for choosing the best cross-validation procedure according to the particular features of the problem in hand.' author: - | Sylvain Arlot,\ CNRS ; Willow Project-Team,\ Laboratoire d’Informatique de l’Ecole Normale Superieure\ (CNRS/ENS/INRIA UMR 8548)\ 45, rue d’Ulm, 75230 Paris, France\ `[email protected]`\  \ Alain Celisse,\ Laboratoire Paul Painlevé, UMR CNRS 8524,\ Université des Sciences et Technologies de Lille 1\ F-59655 Villeneuve d’Ascq Cedex, France\ `[email protected]`\ bibliography: - 'surveyCV.bib' - 'bibliosyl.bib' title: 'A survey of cross-validation procedures for model selection' --- Introduction {#sec.intro} ============ Many statistical algorithms, such as likelihood maximization, least squares and empirical contrast minimization, rely on the preliminary choice of a model, that is of a set of parameters from which an estimate will be returned. When several candidate models (thus algorithms) are available, choosing one of them is called the model selection problem. Cross-validation (CV) is a popular strategy for model selection, and more generally algorithm selection. The main idea behind CV is to split the data (once or several times) for estimating the risk of each algorithm: Part of the data (the training sample) is used for training each algorithm, and the remaining part (the validation sample) is used for estimating the risk of the algorithm. Then, CV selects the algorithm with the smallest estimated risk. Compared to the resubstitution error, CV avoids overfitting because the training sample is independent from the validation sample (at least when data are [*i.i.d.*]{}). The popularity of CV mostly comes from the generality of the data splitting heuristics, which only assumes that data are [*i.i.d.*]{}. Nevertheless, theoretical and empirical studies of CV procedures do not entirely confirm this “universality”. Some CV procedures have been proved to fail for some model selection problems, depending on the goal of model selection: estimation or identification (see Section \[sec.modsel\]). Furthermore, many theoretical questions about CV remain widely open. The aim of the present survey is to provide a clear picture of what is known about CV, from both theoretical and empirical points of view. More precisely, the aim is to answer the following questions: What is CV doing? When does CV work for model selection, keeping in mind that model selection can target different goals? Which CV procedure should be used for each model selection problem? The paper is organized as follows. First, the rest of Section \[sec.intro\] presents the statistical framework. Although non exhaustive, the present setting has been chosen general enough for sketching the complexity of CV for model selection. The model selection problem is introduced in Section \[sec.modsel\]. A brief overview of some model selection procedures that are important to keep in mind for understanding CV is given in Section \[sec.modselproc\]. The most classical CV procedures are defined in Section \[sec.def\]. Since they are the keystone of the behaviour of CV for model selection, the main properties of CV estimators of the risk for a fixed model are detailed in Section \[sec.riskestim\]. Then, the general performances of CV for model selection are described, when the goal is either estimation (Section \[sec.cvefficient\]) or identification (Section \[sec.cvconsistent\]). Specific properties of CV in some particular frameworks are discussed in Section \[sec.specific\]. Finally, Section \[sec.complex\] focuses on the algorithmic complexity of CV procedures, and Section \[sec.conclu\] concludes the survey by tackling several practical questions about CV. Statistical framework {#sec.def.cadre.pb} --------------------- Assume that some data $\xi_1, \ldots, \xi_n \in \Xi$ with common distribution $P$ are observed. Throughout the paper—except in Section \[sec.dependent\]—the $\xi_i$ are assumed to be independent. The purpose of statistical inference is to estimate from the data $\paren{\xi_i}_{1 \leq i \leq n}$ some target feature ${s}$ of the unknown distribution $P$, such as the mean or the variance of $P$. Let ${\mathbb{S}}$ denote the set of possible values for $s$. The quality of $t \in {\mathbb{S}}$, as an approximation of ${s}$, is measured by its loss ${\ensuremath{\mathcal{L}\paren{t}}}$, where $\mathcal{L}: {\mathbb{S}}{\mapsto}\R$ is called the [*loss function*]{}, and is assumed to be minimal for $t={s}$. Many loss functions can be chosen for a given statistical problem. Several classical loss functions are defined by $$\label{def.loss} {\ensuremath{\mathcal{L}\paren{t}}} = {\ensuremath{\mathcal{L}_{P}\paren{t}}} {:=}\E_{\xi \sim P} \croch{ \gamma\paren{ t ; \xi }} \enspace ,$$ where $\gamma: {\mathbb{S}}\times \Xi {\mapsto}[0,\infty)$ is called a [*contrast function*]{}. Basically, for $t \in {\mathbb{S}}$ and $\xi \in \Xi$, $\gamma(t;\xi)$ measures how well $t$ is in accordance with observation of $\xi$, so that the loss of $t$, defined by , measures the average accordance between $t$ and new observations $\xi$ with distribution $P$. Therefore, several frameworks such as transductive learning do not fit definition . Nevertheless, as detailed in Section \[sec.def.cadre.ex\], definition includes most classical statistical frameworks. Another useful quantity is the [*excess loss*]{} $$\perte{t} {:=}{\ensuremath{\mathcal{L}_{P}\paren{t}}} - {\ensuremath{\mathcal{L}_{P}\paren{{s}}}} \geq 0\enspace,$$ which is related to the risk of an estimator ${\,\widehat{s}\,}$ of the target $s$ by $$\begin{aligned} R({\,\widehat{s}\,})= \E_{\xi_1,\ldots,\xi_n\sim P}\croch{\perte{{\,\widehat{s}\,}}}\enspace .\end{aligned}$$ Examples {#sec.def.cadre.ex} -------- The purpose of this subsection is to show that the framework of Section \[sec.def.cadre.pb\] includes several important statistical frameworks. This list of examples does not pretend to be exhaustive. #### Density estimation aims at estimating the density ${s}$ of $P$ with respect to some given measure $\mu$ on $\Xi$. Then, ${\mathbb{S}}$ is the set of densities on $\Xi$ with respect to $\mu$. For instance, taking $\gamma(t;x) = - \ln(t(x))$ in , the loss is minimal when $t = {s}$ and the excess loss $$\perte{t} = {\ensuremath{\mathcal{L}_{P}\paren{t}}} - {\ensuremath{\mathcal{L}_{P}\paren{{s}}}} = \E_{\xi \sim P} \croch{ \ln \paren{ \frac{ {s}(\xi)} {t(\xi)} } } = \int {s}\ln \paren{ \frac{s}{t} } \textup{d}\mu$$ is the Kullback-Leibler divergence between distributions $t \mu$ and ${s}\mu$. #### Prediction aims at predicting a quantity of interest $Y \in \Y$ given an explanatory variable $X \in \X$ and a sample of observations $(X_1,Y_1), \ldots, (X_n, Y_n)$. In other words, $\Xi = \X \times \Y$, ${\mathbb{S}}$ is the set of measurable mappings $\X {\mapsto}\Y$ and the contrast $\gamma(t;(x,y))$ measures the discrepancy between the observed $y$ and its predicted value $t(x)$. Two classical prediction frameworks are regression and classification, which are detailed below. #### Regression corresponds to continuous $\Y$, that is $\Y \subset \R$ (or $\R^k$ for multivariate regression), the feature space $\X$ being typically a subset of $\R^{\ell}$. Let ${s}$ denote the regression function, that is ${s}(x) = \E_{(X,Y) \sim P} \croch{ Y {\, \right| \left. \,}X = x}$, so that $$\forall i, \quad Y_i = {s}(X_i) + \epsilon_i \qquad \mbox{with} \quad \E\croch{\epsilon_i {\, \right| \left. \,}X_i} = 0 \enspace .$$ A popular contrast in regression is the [*least-squares contrast*]{} $ \gamma\paren{t;(x,y)} = (t(x) - y)^2 $, which is minimal over ${\mathbb{S}}$ for $t = {s}$, and the excess loss is $$\perte{t} = \E_{(X,Y) \sim P} \croch{ \paren{ s(X) - t(X) }^2 } \enspace .$$ Note that the excess loss of $t$ is the square of the $L^2$ distance between $t$ and $s$, so that prediction and estimation are equivalent goals. #### Classification corresponds to finite $\Y$ (at least discrete). In particular, when $\Y = {\left\{ \left. 0,1 \right. \right\}}$, the prediction problem is called [*binary (supervised) classification*]{}. With the 0-1 contrast function $\gamma(t;(x,y)) = \1_{t(x) \neq y}$, the minimizer of the loss is the so-called Bayes classifier ${s}$ defined by $$s(x) = \1_{\eta(x) \geq 1/2} \enspace ,$$ where $\eta$ denotes the regression function $\eta(x) = {\mathbb{P}}_{(X,Y) \sim P}\paren{ Y=1 {\, \right| \left. \,}X = x}$. Remark that a slightly different framework is often considered in binary classification. Instead of looking only for a classifier, the goal is to estimate also the confidence in the classification made at each point: ${\mathbb{S}}$ is the set of measurable mappings $\X {\mapsto}\R$, the classifier $x {\mapsto}\1_{t(x) \geq 0}$ being associated to any $t \in {\mathbb{S}}$. Basically, the larger ${\left\lvert t(x) \right\rvert}$, the more confident we are in the classification made from $t(x)$. A classical family of losses associated with this problem is defined by with the contrast $\gamma_{\phi} \paren{t; (x,y)} = \phi \paren{ - (2y-1) t(x) }$ where $\phi : \R {\mapsto}[0, \infty)$ is some function. The 0-1 contrast corresponds to $\phi(u) = \1_{u \geq 0}$. The convex loss functions correspond to the case where $\phi$ is convex, nondecreasing with $\lim_{- \infty} \phi = 0$ and $\phi(0)=1$. Classical examples are $\phi(u) = \max{\left\{ \left. 1 + u, 0 \right. \right\}}$ (hinge), $\phi(u) = \exp(u)$, and $\phi(u) = \log_2 \paren{ 1 + \exp(u)}$ (logit). The corresponding losses are used as objective functions by several classical learning algorithms such as support vector machines (hinge) and boosting (exponential and logit). Many references on classification theory, including model selection, can be found in the survey by [@Bou_Bou_Lug:2005]. Statistical algorithms {#sec.def.cadre.sol} ---------------------- In this survey, a [*statistical algorithm*]{} $\A$ is any (measurable) mapping $\A: \bigcup_{n \in \N} \Xi^n {\mapsto}S$. The idea is that data $D_n = \paren{\xi_i}_{1 \leq i \leq n} \in \Xi^n$ will be used as an input of $\A$, and that the output of $\A$, $\A(D_n)={\widehat{s}^{\mathcal{A}}}(D_n) \in {\mathbb{S}}$, is an estimator of $s$. The quality of $\A$ is then measured by ${\ensuremath{\mathcal{L}_{P}\paren{{\widehat{s}^{\mathcal{A}}}(D_n)}}}$, which should be as small as possible. In the sequel, the algorithm $\A$ and the estimator ${\widehat{s}^{\mathcal{A}}}(D_n)$ are often identified when no confusion is possible. [*Minimum contrast estimators*]{} form a classical family of statistical algorithms, defined as follows. Given some subset $S$ of ${\mathbb{S}}$ that we call a [*model*]{}, a minimum contrast estimator of $s$ is any minimizer of the empirical contrast $$t {\mapsto}{\ensuremath{\mathcal{L}_{P_n}\paren{t}}} = \frac{1}{n} \sum_{i=1}^n \gamma\paren{t ; \xi_i}, \qquad \mbox{where} \quad P_n = \frac{1}{n} \sum_{i=1}^n \delta_{\xi_i}\enspace,$$ over $S$. The idea is that the empirical contrast ${\ensuremath{\mathcal{L}_{P_n}\paren{t}}}$ has an expectation ${\ensuremath{\mathcal{L}_{P}\paren{t}}}$ which is minimal over ${\mathbb{S}}$ at $s$. Hence, minimizing ${\ensuremath{\mathcal{L}_{P_n}\paren{t}}}$ over a set $S$ of candidate values for $s$ hopefully leads to a good estimator of $s$. Let us now give three popular examples of empirical contrast minimizers: - [*Maximum likelihood estimators*]{}: take $\gamma(t;x) = - \ln(t(x))$ in the density estimation setting. A classical choice for $S$ is the set of piecewise constant functions on a regular partition of $\Xi$ with $K$ pieces. - [*Least-squares estimators*]{}: take $\gamma(t;(x,y)) = (t(x) - y)^2$ the least-squares contrast in the regression setting. For instance, $S$ can be the set of piecewise constant functions on some fixed partition of $\X$ (leading to regressograms), or a vector space spanned by the first vectors of wavelets or Fourier basis, among many others. Note that regularized least-squares algorithms such as the Lasso, ridge regression and spline smoothing also are least-squares estimators, the model $S$ being some ball of a (data-dependent) radius for the $L^1$ (resp. $L^2$) norm in some high-dimensional space. Hence, tuning the regularization parameter for the LASSO or SVM, for instance, amounts to perform model selection from a collection of models. - [*Empirical risk minimizers*]{}, following the terminology of [@Vap:1982]: take any contrast function $\gamma$ in the prediction setting. When $\gamma$ is the 0-1 contrast, popular choices for $S$ lead to linear classifiers, partitioning rules, and neural networks. Boosting and Support Vector Machines classifiers also are empirical contrast minimizers over some data-dependent model $S$, with contrast $\gamma = \gamma_{\phi}$ for some convex functions $\phi$. Let us finally mention that many other classical statistical algorithms can be considered with CV, for instance local average estimators in the prediction framework such as $k$-Nearest Neighbours and Nadaraya-Watson kernel estimators. The focus will be mainly kept on minimum contrast estimators to keep the length of the survey reasonable. Model selection {#sec.modsel} =============== Usually, several statistical algorithms can be used for solving a given statistical problem. Let $\paren{\ERM_{\lambda}}_{\lambda \in \Lambda}$ denote such a family of candidate statistical algorithms. The [*algorithm selection problem*]{} aims at choosing from data one of these algorithms, that is, choosing some ${\ensuremath{\widehat{\lambda}}}(D_n) \in \Lambda$. Then, the final estimator of $s$ is given by $\ERM_{{\ensuremath{\widehat{\lambda}}}(D_n)} (D_n)$. The main difficulty is that the same data are used for training the algorithms, that is, for computing $\paren{\ERM_{\lambda} (D_n) }_{\lambda \in \Lambda}$, and for choosing ${\ensuremath{\widehat{\lambda}}}(D_n)$ . The model selection paradigm {#sec.modsel.paradigm} ---------------------------- Following Section \[sec.def.cadre.sol\], let us focus on the [*model selection problem*]{}, where candidate algorithms are minimum contrast estimators and the goal is to choose a model $S$. Let $\paren{S_m}_{{m \in \M}_n}$ be a family of models, that is, $S_m \subset {\mathbb{S}}$. Let $\gamma$ be a fixed contrast function, and for every ${m \in \M}_n$, let $\ERM_m$ be a minimum contrast estimator over model $S_m$ with contrast $\gamma$. The goal is to choose $\mh(D_n) \in \M_n$ from data only. The choice of a model $S_m$ has to be done carefully. Indeed, when $S_m$ is a [“small”]{} model, $\ERM_m$ is a poor statistical algorithm except when $s$ is very close to $S_m$, since $$\perte{\ERM_m} \geq \inf_{t \in S_m} {\left\{ \left. \perte{t} \right. \right\}} {:=}\perte{S_m} \enspace .$$ The lower bound $\perte{S_m}$ is called the [*bias*]{} of model $S_m$, or [*approximation error*]{}. The bias is a nonincreasing function of $S_m$. On the contrary, when $S_m$ is [“huge”]{}, its bias $\perte{S_m}$ is small for most targets $s$, but $\ERM_m$ clearly overfits. Think for instance of $S_m$ as the set of all continuous functions on $[0,1]$ in the regression framework. More generally, if $S_m$ is a vector space of dimension $D_m$, in several classical frameworks, $$\label{eq.biais-var} \E\croch{ \perte{\ERM_m(D_n)} } \approx \perte{S_m} + \lambda D_m$$ where $\lambda>0$ does not depend on $m$. For instance, $\lambda = 1 / (2n)$ in density estimation using the likelihood contrast, and $\lambda = \sigma^2 /n$ in regression using the least-squares contrast and assuming $\operatorname{var}\paren{ Y{\, \right| \left. \,}X } = \sigma^2$ does not depend on $X$. The meaning of is that a good model choice should balance the bias term $\perte{S_m}$ and the [*variance*]{} term $\lambda D_m$, that is solve the so-called [*bias-variance trade-off*]{}. By extension, the variance term, also called [*estimation error*]{}, can be defined by $$\E\croch{ \perte{\ERM_m(D_n)} } - \perte{S_m} = \E\croch{{\ensuremath{\mathcal{L}_{P}\paren{{\,\widehat{s}_m\,}}}}} - \inf_{t\in S_m}{\ensuremath{\mathcal{L}_{P}\paren{t}}}\enspace,$$ even when does not hold. The interested reader can find a much deeper insight into model selection in the Saint-Flour lecture notes by [@Mas:2003:St-Flour]. Before giving examples of classical model selection procedures, let us mention the two main different goals that model selection can target: estimation and identification. Model selection for estimation {#sec.modsel.estim} ------------------------------ On the one hand, the goal of model selection is [*estimation*]{} when $\ERM_{\mh(D_n)} (D_n)$ is used as an approximation of the target $s$, and the goal is to minimize its loss. For instance, AIC and Mallows’ $C_p$ model selection procedures are built for estimation (see Section \[sec.modselproc.unbiased\]). The quality of a model selection procedure $D_n {\mapsto}\mh(D_n)$, designed for estimation, is measured by the excess loss of $\ERM_{\mh(D_n)} (D_n)$. Hence, the best possible model choice for estimation is the so-called [*oracle*]{} model $S_{{\ensuremath{m^{\star}}}}$, defined by $$\label{eq.mo} {\ensuremath{m^{\star}}}= {\ensuremath{m^{\star}}}(D_n) \in \arg\min_{{m \in \M}_n} {\left\{ \left. \perte{ \ERM_m (D_n) } \right. \right\}} \enspace .$$ Since ${\ensuremath{m^{\star}}}(D_n)$ depends on the unknown distribution $P$ of data, one cannot expect to select $\mh(D_n)={\ensuremath{m^{\star}}}(D_n)$ almost surely. Nevertheless, we can hope to select $\mh(D_n)$ such that ${\,\widehat{s}\,}_{\mh(D_n)}$ is almost as close to $s$ as ${\,\widehat{s}\,}_{{\ensuremath{m^{\star}}}(D_n)}$. Note that there is no requirement for $s$ to belong to $\bigcup_{m\in\mathcal{M}_n}S_m$. Depending on the framework, the optimality of a model selection procedure for estimation is assessed in at least two different ways. First, in the asymptotic framework, a model selection procedure $\mh$ is called [*efficient*]{} (or asymptotically optimal) when it leads to $\mh$ such that $$\frac{ \perte{ \ERM_{\mh(D_n)} (D_n) } } { \inf_{{m \in \M}_n} {\left\{ \left. \perte{ \ERM_m (D_n) } \right. \right\}} } \xrightarrow[n \rightarrow \infty]{a.s.} 1 \enspace .$$ Sometimes, a weaker result is proved, the convergence holding only in probability. Second, in the non-asymptotic framework, a model selection procedure satisfies an [*oracle inequality*]{} with constant $C_n \geq 1$ and remainder term $R_n \geq 0$ when $$\label{eq.oracle-ineq} \perte{ \ERM_{\mh(D_n)} (D_n) } \leq C_n \inf_{{m \in \M}_n} {\left\{ \left. \perte{ \ERM_m (D_n) } \right. \right\}} + R_n$$ holds either in expectation or with large probability (that is, a probability larger than $1 - C^{\prime}/ n^2$, for some positive constant $C^{\prime}$). Note that if holds on a large probability event with $C_n$ tending to 1 when $n$ tends to infinity and $R_n \ll \perte{\ERM_{{\ensuremath{m^{\star}}}}(D_n)}$, then the model selection procedure $\mh$ is efficient. In the estimation setting, model selection is often used for building [*adaptive estimators*]{}, assuming that $s$ belongs to some function space $\mathcal{T}_{\alpha}$ [@Bar_Bir_Mas:1999].Then, a model selection procedure $\mh$ is optimal when it leads to an estimator $\ERM_{\mh(D_n)} (D_n)$ (approximately) minimax with respect to $\mathcal{T}_{\alpha}$ without knowing $\alpha$, provided the family $\paren{S_m}_{{m \in \M}_n}$ has been well-chosen. Model selection for identification {#sec.modsel.identif} ---------------------------------- On the other hand, model selection can aim at identifying the “true model” $S_{m_0}$, defined as the “smallest” model among $\paren{S_m}_{{m \in \M}_n}$ to which $s$ belongs. In particular, $s \in \bigcup_{{m \in \M}_n} S_m$ is assumed in this setting. A typical example of model selection procedure built for identification is BIC (see Section \[sec.modselproc.identif\]). The quality of a model selection procedure designed for identification is measured by its probability of recovering the true model $m_0$. Then, a model selection procedure is called [*(model) consistent*]{} when $${\mathbb{P}}\paren{ \mh(D_n) = m_0 } \xrightarrow[n \rightarrow \infty]{} 1 \enspace .$$ Note that identification can naturally be extended to the general algorithm selection problem, the “true model” being replaced by the statistical algorithm whose risk converges at the fastest rate [see for instance @Yan:2007b]. Estimation [*vs.*]{} identification {#sec.modsel.estim-vs-identif} ----------------------------------- When a true model exists, model consistency is clearly a stronger property than efficiency defined in Section \[sec.modsel.estim\]. However, in many frameworks, no true model does exist so that efficiency is the only well-defined property. Could a model selection procedure be model consistent in the former case (like BIC) and efficient in the latter case (like AIC)? The general answer to this question, often called the AIC-BIC dilemma, is negative: [@Yan:2005a] proved in the regression framework that no model selection procedure can be simultaneously model consistent and minimax rate optimal. Nevertheless, the strengths of AIC and BIC can sometimes be shared; see for instance the introduction of a paper by [@Yan:2005a] and a recent paper by [@Erv_Gru_Roo:2008]. Overview of some model selection procedures {#sec.modselproc} =========================================== Several approaches can be used for model selection. Let us briefly sketch here some of them, which are particularly helpful for understanding how CV works. Like CV, all the procedures considered in this section select $$\label{eq.uerp} \mh(D_n) \in \arg\min_{{m \in \M}_n} {\left\{ \left. \crit(m; D_n) \right. \right\}} \enspace ,$$ where $\forall {m \in \M}_n$, $\crit(m; D_n)=\crit(m) \in \R$ is some data-dependent criterion. A particular case of is [*penalization*]{}, which consists in choosing the model minimizing the sum of empirical contrast and some measure of complexity of the model (called penalty) which can depend on the data, that is, $$\label{eq.pen} \mh (D_n) \in \arg\min_{{m \in \M}_n} {\left\{ \left. {\ensuremath{\mathcal{L}_{P_n}\paren{\ERM_{m}}}} + \pen(m; D_n) \right. \right\}} \enspace .$$ This section does not pretend to be exhaustive. Completely different approaches exist for model selection, such as the Minimum Description Length (MDL) [@Riss83], and the Bayesian approaches. The interested reader will find more details and references on model selection procedures in the books by [@Bur_And:2002] or [@Mas:2003:St-Flour] for instance. Let us focus here on five main categories of model selection procedures, the first three ones coming from a classification made by [@Sha:1997] in the linear regression framework. The unbiased risk estimation principle {#sec.modselproc.unbiased} -------------------------------------- When the goal of model selection is estimation, many model selection procedures are of the form where $\crit(m;D_n)$ unbiasedly estimates (at least, asymptotically) the loss ${\ensuremath{\mathcal{L}_{P}\paren{\ERM_m}}}$. This general idea is often called unbiased risk estimation principle, or Mallows’ or Akaike’s heuristics. In order to explain why this strategy can perform well, let us write the starting point of most theoretical analysis of procedures defined by : By definition , for every ${m \in \M}_n$, $$\label{eq.start-oracle} \perte{\ERM_{\mh}} + \crit(\mh) - {\ensuremath{\mathcal{L}_{P}\paren{\ERM_{\mh}}}} \leq \perte{\ERM_{m}} + \crit(m) - {\ensuremath{\mathcal{L}_{P}\paren{\ERM_{m}}}} \enspace .$$ If $\E\croch{ \crit(m) - {\ensuremath{\mathcal{L}_{P}\paren{\ERM_{m}}}} } = 0$ for every ${m \in \M}_n$, then concentration inequalities are likely to prove that $\varepsilon_n^-, \varepsilon_n^+ > 0$ exist such that $$\forall {m \in \M}_n, \quad \varepsilon_n^+ \geq \frac{ \crit(m) - {\ensuremath{\mathcal{L}_{P}\paren{\ERM_{m}}}} }{ \perte{\ERM_m} } \geq - \varepsilon_n^- > -1$$ with high probability, at least when $\card(\M_n) \leq C n^{\alpha}$ for some $C,\alpha \geq 0$. Then, directly implies an oracle inequality like with $C_n = (1 + \varepsilon_n^+) / (1 - \varepsilon_n^-)$. If $\varepsilon_n^+, \varepsilon_n^- \rightarrow 0$ when $n \rightarrow \infty$, this proves the procedure defined by is efficient. Examples of model selection procedures following the unbiased risk estimation principle are FPE [Final Prediction Error, @Aka:1969], several cross-validation procedures including the Leave-one-out (see Section \[sec.def\]), and GCV [Generalized Cross-Validation, @Cra_Wah:1979 see Section \[sec.def.classex.other\]]. With the penalization approach , the unbiased risk estimation principle is that $\E\croch{\pen(m)}$ should be close to the “ideal penalty” $${\pen_{\mathrm{id}}}(m) {:=}{\ensuremath{\mathcal{L}_{P}\paren{\ERM_m}}} - {\ensuremath{\mathcal{L}_{P_n}\paren{\ERM_m}}} \enspace .$$ Several classical penalization procedures follow this principle, for instance: - With the log-likelihood contrast, AIC [Akaike Information Criterion, @Aka:1973] and its corrected versions [@Sug:1978; @Hur_Tsa:1989]. - With the least-squares contrast, Mallows’ $C_p$ [@Mal:1973] and several refined versions of $C_p$ [see for instance @Bar:2002]. - With a general contrast, covariance penalties [@Efr:2004]. AIC, Mallows’ $C_p$ and related procedures have been proved to be optimal for estimation in several frameworks, provided $\card(\M_n) \leq C n^{\alpha}$ for some constants $C,\alpha \geq 0$ [see the paper by @Bir_Mas:2006 and references therein]. The main drawback of penalties such as AIC or Mallows’ $C_p$ is their dependence on some assumptions on the distribution of data. For instance, Mallows’ $C_p$ assumes the variance of $Y$ does not depend on $X$. Otherwise, it has a suboptimal performance [@Arlo08]. Several resampling-based penalties have been proposed to overcome this problem, at the price of a larger computational complexity, and possibly slightly worse performance in simpler frameworks; see a paper by [@Efr:1983] for bootstrap, and a paper by [@Arlo08c] and references therein for generalization to exchangeable weights. Finally, note that all these penalties depend on multiplying factors which are not always known (for instance, the noise-level, for Mallows’ $C_p$). [@Bir_Mas:2006] proposed a general data-driven procedure for estimating such multiplying factors, which satisfies an oracle inequality with $C_n \rightarrow 1$ in regression [see also @Arl_Mas:2009:pente]. Biased estimation of the risk {#sec.modselproc.biased} ----------------------------- Several model selection procedures are of the form where $\crit(m)$ does not unbiasedly estimate the loss ${\ensuremath{\mathcal{L}_{P}\paren{\ERM_m}}}$: The weight of the variance term compared to the bias in $\E\croch{\crit(m)}$ is slightly larger than in the decomposition of ${\ensuremath{\mathcal{L}_{P}\paren{\ERM_m}}}$. From the penalization point of view, such procedures are [*overpenalizing*]{}. Examples of such procedures are FPE$_{\alpha}$ [@Bha_Dow:1977] and GIC$_{\lambda}$ [Generalized Information Criterion, @Nis:1984; @Sha:1997] with $\alpha,\lambda>2$, which are closely related. Some cross-validation procedures, such as Leave-$p$-out with $p/n \in (0,1)$ fixed, also belong to this category (see Section \[subsec.classical.examples.exhaust\]). Note that FPE$_{\alpha}$ with $\alpha=2$ is FPE, and GIC$_{\lambda}$ with $\lambda=2$ is close to FPE and Mallows’ $C_p$. When the goal is estimation, there are two main reasons for using “biased” model selection procedures. First, experimental evidence show that overpenalizing often yields better performance when the signal-to-noise ratio is small [see for instance @Arlo07 Chapter 11]. Second, when the number of models $\card(\M_n)$ grows faster than any power of $n$, as in the complete variable selection problem with $n$ variables, then the unbiased risk estimation principle fails. From the penalization point of view, [@Bir_Mas:2006] proved that when $\card(\M_n) = e^{\kappa n}$ for some $\kappa >0$, the minimal amount of penalty required so that an oracle inequality holds with $C_n = \grandO(1)$ is much larger than ${\pen_{\mathrm{id}}}(m)$. In addition to the FPE$_{\alpha}$ and GIC$_{\lambda}$ with suitably chosen $\alpha, \lambda$, several penalization procedures have been proposed for taking into account the size of $\M_n$ [@Bar_Bir_Mas:1999; @Bar:2002; @Bir_Mas:2002; @Sau:2006]. In the same papers, these procedures are proved to satisfy oracle inequalities with $C_n$ as small as possible, typically of order $\ln(n)$ when $\card(\M_n) = e^{\kappa n}$. Procedures built for identification {#sec.modselproc.identif} ----------------------------------- Some specific model selection procedures are used for identification. A typical example is BIC [Bayesian Information Criterion, @Sch:1978]. More generally, [@Sha:1997] showed that several procedures identify consistently the correct model in the linear regression framework as soon as they overpenalize within a factor tending to infinity with $n$, for instance, GIC$_{\lambda_n}$ with $\lambda_n \rightarrow + \infty$, FPE$_{\alpha_n}$ with $\alpha_n \rightarrow +\infty$ [@Shi:1984], and several CV procedures such as Leave-$p$-out with $p=p_n\sim n$. BIC is also part of this picture, since it coincides with GIC$_{\ln(n)}$. In another paper, [@Sha:1996] showed that $m_n$-out-of-$n$ bootstrap penalization is also model consistent as soon as $m_n \sim n$. Compared to Efron’s bootstrap penalties, the idea is to estimate ${\pen_{\mathrm{id}}}$ with the $m_n$-out-of-$n$ bootstrap instead of the usual bootstrap, which results in overpenalization within a factor tending to infinity with $n$ [@Arlo08c]. Most MDL-based procedures can also be put into this category of model selection procedures [see @Gru:2007]. Let us finally mention the Lasso [@Tibs96] and other $\ell^1$ penalization procedures, which have recently attracted much attention [see for instance @Hes_etal:2008]. They are a computationally efficient way of identifying the true model in the context of variable selection with many variables. Structural risk minimization {#sec.modselproc.Vapnik} ---------------------------- In the context of statistical learning, [@Vap_Cer:1974] proposed the structural risk minimization approach [see also @Vap:1982; @Vap:1998]. Roughly, the idea is to penalize the empirical contrast with a penalty (over)-estimating $$\penidglo(m) {:=}\sup_{t \in S_m} {\left\{ \left. {\ensuremath{\mathcal{L}_{P}\paren{t}}} - {\ensuremath{\mathcal{L}_{P_n}\paren{t}}} \right. \right\}} \geq {\pen_{\mathrm{id}}}(m) \enspace .$$ Such penalties have been built using the Vapnik-Chervonenkis dimension, the combinatorial entropy, (global) Rademacher complexities [@Kol:2001; @Bar_Bou_Lug:2002], (global) bootstrap penalties [@Fro:2007], Gaussian complexities or the maximal discrepancy [@Bar_Men:2002]. These penalties are often called [*global*]{} because $\penidglo(m)$ is a supremum over $S_m$. The localization approach [see @Bou_Bou_Lug:2005] has been introduced in order to obtain penalties closer to ${\pen_{\mathrm{id}}}$ (such as local Rademacher complexities), hence smaller prediction errors when possible [@Bar_Bou_Men:2005; @Kol:2006]. Nevertheless, these penalties are still larger than ${\pen_{\mathrm{id}}}(m)$ and can be difficult to compute in practice because of several unknown constants. A non-asymptotic analysis of several global and local penalties can be found in the book by [@Mas:2003:St-Flour] for instance; see also [@Kol:2006] for recent results on local penalties. [*Ad hoc*]{} penalization {#sec.modselproc.adhoc} ------------------------- Let us finally mention that penalties can also be built according to particular features of the problem. For instance, penalties can be proportional to the $\ell^p$ norm of $\ERM_m$ (similarly to $\ell^p$-regularized learning algorithms) when having an estimator with a controlled $\ell^p$ norm seems better. The penalty can also be proportional to the squared norm of $\ERM_m$ in some reproducing kernel Hilbert space (similarly to kernel ridge regression or spline smoothing), with a kernel adapted to the specific framework. More generally, any penalty can be used, as soon as $\pen(m)$ is larger than the estimation error (to avoid overfitting) and the best model for the final user is not the oracle ${\ensuremath{m^{\star}}}$, but more like $$\arg\min_{{m \in \M}_n} {\left\{ \left. \perte{ S_m } + \kappa \pen(m) \right. \right\}}$$ for some $\kappa >0$. Where are cross-validation procedures in this picture? {#sec.modselproc.CV} ------------------------------------------------------ The family of CV procedures, which will be described and deeply investigated in the next sections, contains procedures in the first three categories. CV procedures are all of the form , where $\crit(m)$ either estimates (almost) unbiasedly the loss ${\ensuremath{\mathcal{L}_{P}\paren{\ERM_m}}}$, or overestimates the variance term (see Section \[sec.modsel.paradigm\]). In the latter case, CV procedures either belong to the second or the third category, depending on the overestimation level. This fact has two major implications. First, CV itself does not take into account prior information for selecting a model. To do so, one can either add to the CV estimate of the risk a penalty term (such as ${\left \lVert \ERM_m \right\rVert}_p$), or use prior information to pre-select a subset of models $\Mt(D_n) \subset \M_n$ before letting CV select a model among $\paren{S_m}_{m \in \Mt(D_n)}$. Second, in statistical learning, CV and resampling-based procedures are the most widely used model selection procedures. Structural risk minimization is often too pessimistic, and other alternatives rely on unrealistic assumptions. But if CV and resampling-based procedures are the most likely to yield good prediction performances, their theoretical grounds are not that firm, and too few CV users are careful enough when choosing a CV procedure to perform model selection. Among the aims of this survey is to point out both positive and negative results about the model selection performance of CV. Cross-validation procedures {#sec.def} =========================== The purpose of this section is to describe the rationale behind CV and to define the different CV procedures. Since all CV procedures are of the form , defining a CV procedure amounts to define the corresponding CV estimator of the risk of an algorithm $\A$, which will be $\crit(\cdot)$ in . Cross-validation philosophy {#sec.def.philo} --------------------------- As noticed in the early 30s by [@Lars31], training an algorithm and evaluating its statistical performance on the same data yields an overoptimistic result. CV was raised to fix this issue [@MoTu68; @Sto:1974; @Gei:1975], starting from the remark that testing the output of the algorithm on new data would yield a good estimate of its performance [@Brei96c]. In most real applications, only a limited amount of data is available, which led to the idea of [*splitting the data*]{}: Part of the data (the training sample) is used for training the algorithm, and the remaining data (the validation sample) is used for evaluating its performance. The validation sample can play the role of new data as soon as data are [*i.i.d.*]{}. Data splitting yields the [*validation*]{} estimate of the risk, and averaging over several splits yields a [*cross-validation*]{} estimate of the risk. As will be shown in Sections \[sec.def.simple2cross\] and \[subsec.classical.examples\], various splitting strategies lead to various CV estimates of the risk. The major interest of CV lies in the universality of the data splitting heuristics, which only assumes that data are identically distributed and the training and validation samples are independent, two assumptions which can even be relaxed (see Section \[sec.dependent\]). Therefore, CV can be applied to (almost) any algorithm in (almost) any framework, for instance regression [@Sto:1974; @Gei:1975], density estimation [@Rude82; @Ston84] and classification [@DeWa79; @Bar_Bou_Lug:2002], among many others. On the contrary, most other model selection procedures (see Section \[sec.modselproc\]) are specific to a framework: For instance, $C_p$ [@Mal:1973] is specific to least-squares regression. From validation to cross-validation {#sec.def.simple2cross} ----------------------------------- In this section, the hold-out (or validation) estimator of the risk is defined, leading to a general definition of CV. ### Hold-out {#sec.def.simple2cross.ho} The [*hold-out*]{} [@DeWa79] or (simple) [*validation*]{} relies on a single split of data. Formally, let ${\ensuremath{I^{(t)}}}$ be a non-empty proper subset of ${\left\{ \left. 1 , \ldots, n \right. \right\}}$, that is, such that both ${\ensuremath{I^{(t)}}}$ and its complement ${\ensuremath{I^{(v)}}}= \paren{{\ensuremath{I^{(t)}}}}^c = {\left\{ \left. 1 , \ldots, n \right. \right\}} \backslash {\ensuremath{I^{(t)}}}$ are non-empty. The [*hold-out*]{} estimator of the risk of $\A(D_n)$ with training set ${\ensuremath{I^{(t)}}}$ is defined by $$\label{def.validation} {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}\paren{ \A ; D_n ; {\ensuremath{I^{(t)}}}} {:=}\frac{1}{n_v} \sum_{i \in {\ensuremath{D_n^{(v)}}}} \gamma\paren{ \A({\ensuremath{D_n^{(t)}}}) ; (X_i,Y_i)} \enspace ,$$ where ${\ensuremath{D_n^{(t)}}}{:=}(\xi_i )_{i \in {\ensuremath{I^{(t)}}}}$ is the [*training sample*]{}, of size $n_t = \card({\ensuremath{I^{(t)}}})$, and ${\ensuremath{D_n^{(v)}}}{:=}(\xi_i )_{i \in {\ensuremath{I^{(v)}}}}$ is the [*validation sample*]{}, of size $n_v = n-n_t$; ${\ensuremath{I^{(v)}}}$ is called the validation set. The question of choosing $n_t$, and ${\ensuremath{I^{(t)}}}$ given its cardinality $n_t$, is discussed in the rest of this survey. ### General definition of cross-validation {#sec.def.simple2cross.CVgal} A general description of the CV strategy has been given by [@Gei:1975]: In brief, CV consists in averaging several hold-out estimators of the risk corresponding to different splits of the data. Formally, let $B \geq 1$ be an integer and ${\ensuremath{I^{(t)}}}_1, \ldots, {\ensuremath{I^{(t)}}}_B$ be a sequence of non-empty proper subsets of ${\left\{ \left. 1 , \ldots, n \right. \right\}}$. The CV estimator of the risk of $\A(D_n)$ with training sets $\paren{{\ensuremath{I^{(t)}}}_j}_{1 \leq j \leq B}$ is defined by $$\label{def.vc.gal} {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}\paren{ \A ; D_n ; \paren{{\ensuremath{I^{(t)}}}_j}_{1 \leq j \leq B} } {:=}\frac{1}{B} \sum_{j=1}^B {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}\paren{ \A ; D_n ; {\ensuremath{I^{(t)}}}_j } \enspace .$$ All existing CV estimators of the risk are of the form , each one being uniquely determined by the way the sequence $\paren{{\ensuremath{I^{(t)}}}_j}_{1 \leq j \leq B}$ is chosen, that is, the choice of the splitting scheme. Note that when CV is used in model selection for identification, an alternative definition of CV was proposed by [@Yan:2006; @Yan:2007b] and called [*CV with voting*]{} (CV-v). When two algorithms $\A_1$ and $\A_2$ are compared, $\A_1$ is selected by CV-v if and only if $ {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}( \A_1 ; D_n ; {\ensuremath{I^{(t)}}}_j ) < {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}( \A_2 ; D_n ; {\ensuremath{I^{(t)}}}_j )$ for a majority of the splits $j = 1, \ldots, B$. By contrast, CV procedures of the form can be called “CV with averaging” (CV-a), since the estimates of the risk of the algorithms are averaged before their comparison. Classical examples {#subsec.classical.examples} ------------------ Most classical CV estimators split the data with a fixed size $n_t$ of the training set, that is, $\card({\ensuremath{I^{(t)}}}_j) \approx n_t$ for every $j$. The question of choosing $n_t$ is discussed extensively in the rest of this survey. In this subsection, several CV estimators are defined. Two main categories of splitting schemes can be distinguished, given $n_t$: exhaustive data splitting, that is considering all training sets ${\ensuremath{I^{(t)}}}$ of size $n_t$, and partial data splitting. ### Exhaustive data splitting {#subsec.classical.examples.exhaust} #### Leave-one-out [LOO, @Sto:1974; @All:1974; @Gei:1975] is the most classical exhaustive CV procedure, corresponding to the choice $n_t = n-1 \,$: Each data point is successively “left out” from the sample and used for validation. Formally, LOO is defined by with $B=n$ and ${\ensuremath{I^{(t)}}}_j = {\left\{ \left. j \right. \right\}}^c$ for $j=1, \ldots, n \,$: $$\label{def.loo} {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LOO}}}}}\paren{ \A ; D_n } = \frac{1}{n} \sum_{j=1}^n \gamma\paren{ \A \paren{D_n^{(-j)}} ; \xi_j }$$ where $D_n^{(-j)} = \paren{\xi_i}_{i \neq j}$. The name LOO can be traced back to papers by [@Pic_Coo:1984] and by [@Bre_Spe:1992], but LOO has several other names in the literature, such as [*delete-one CV*]{} [see @KCLi:1987], [*ordinary CV*]{} [@Sto:1974; @Bur:1989], or even only [*CV*]{} [@Efr:1983; @KCLi:1987]. #### Leave-$p$-out [LPO, @Sha:1993] with $p \in {\left\{ \left. 1, \ldots, n \right. \right\}}$ is the exhaustive CV with $n_t = n-p\,$: every possible set of $p$ data points are successively “left out” from the sample and used for validation. Therefore, LPO is defined by with $B = {n\choose p}$ and $\sparen{ {\ensuremath{I^{(t)}}}_j}_{1 \leq j \leq B}$ are all the subsets of ${\left\{ \left. 1, \ldots, n \right. \right\}}$ of size $p$. LPO is also called [*delete-$p$ CV*]{} or [*delete-$p$ multifold CV*]{} [@Zha:1993]. Note that LPO with $p=1$ is LOO. ### Partial data splitting {#subsec.classical.examples.partial} Considering ${n \choose p}$ training sets can be computationally intractable, even for small $p$, so that partial data splitting methods have been proposed. #### [*V*]{}-fold CV (VFCV) with $V \in {\left\{ \left. 1, \ldots, n \right. \right\}}$ was introduced by [@Gei:1975] as an alternative to the computationally expensive LOO [see also @Bre_etal:1984 for instance]. VFCV relies on a preliminary partitioning of the data into $V$ subsamples of approximately equal cardinality $n/V$; each of these subsamples successively plays the role of validation sample. Formally, let $A_1, \ldots, A_V$ be some partition of ${\left\{ \left. 1, \ldots, n \right. \right\}}$ with $\card\paren{A_j}\approx n/V$. Then, the VFCV estimator of the risk of $\A$ is defined by with $B=V$ and ${\ensuremath{I^{(t)}}}_j = A_j^c$ for $j = 1, \ldots, B$, that is, $$\begin{aligned} \label{def.vfcv} \hspace*{-.15cm} {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{VF}}}}}\paren{ \ERM ; D_n ; \paren{ A_j }_{1 \leq j \leq V}} = \frac{1}{V} \sum_{j=1}^V \croch{ \frac{1}{\card(A_j)} \sum_{i \in A_j} \gamma\paren{ \ERM \paren{D_n^{(-A_j)}} ; \xi_i } }\end{aligned}$$ where $D_n^{(-A_j)} = \paren{\xi_i}_{i \in A_j^c}$. By construction, the algorithmic complexity of VFCV is only $V$ times that of training $\A$ with $n-n/V$ data points, which is much less than LOO or LPO if $V \ll n$. Note that VFCV with $V=n$ is LOO. #### Balanced Incomplete CV [BICV, @Sha:1993] can be seen as an alternative to VFCV well-suited for small training sample sizes $n_t$. Indeed, BICV is defined by with training sets $\paren{ A^c }_{A \in \mathcal{T}}$, where $\mathcal{T}$ is a balanced incomplete block designs [BIBD, @Joh:1971], that is, a collection of $B>0$ subsets of ${\left\{ \left. 1, \ldots, n \right. \right\}}$ of size $n_v=n-n_t$ such that: 1. $\card{\left\{ \left. A \in \mathcal{T} {\, \mbox{ s.t. } \,}k \in \A \right. \right\}}$ does not depend on $k \in {\left\{ \left. 1, \ldots, n \right. \right\}}$. 2. $\card{\left\{ \left. A \in \mathcal{T} {\, \mbox{ s.t. } \,}k,\ell \in \A \right. \right\}}$ does not depend on $k \neq \ell \in {\left\{ \left. 1, \ldots, n \right. \right\}}$. The idea of BICV is to give to each data point (and each pair of data points) the same role in the training and validation tasks. Note that VFCV relies on a similar idea, since the set of training sample indices used by VFCV satisfy the first property and almost the second one: Pairs $(k,\ell)$ belonging to the same $A_j$ appear in one validation set more than other pairs. #### Repeated learning-testing (RLT) was introduced by [@Bre_etal:1984] and further studied by [@Bur:1989] and by [@Zha:1993] for instance. The RLT estimator of the risk of $\A$ is defined by with any $B>0$ and $\sparen{{\ensuremath{I^{(t)}}}_j}_{1 \leq j \leq B}$ are $B$ different subsets of ${\left\{ \left. 1, \ldots, n \right. \right\}}$, chosen randomly and independently from the data. RLT can be seen as an approximation to LPO with $p = n - n_t$, with which it coincides when $B = {n \choose p}$. #### Monte-Carlo CV [MCCV, @Pic_Coo:1984] is very close to RLT: $B$ independent subsets of ${\left\{ \left. 1, \ldots, n \right. \right\}}$ are randomly drawn, with uniform distribution among subsets of size $n_t$. The only difference with RLT is that MCCV allows the same split to be chosen several times. ### Other cross-validation-like risk estimators {#sec.def.classex.other} Several procedures have been introduced which are close to, or based on CV. Most of them aim at fixing an observed drawback of CV. #### Bias-corrected versions of VFCV and RLT risk estimators have been proposed by [@Bur:1989; @Bur:1990], and a closely related penalization procedure called $V$-fold penalization has been defined by [@Arl:2008a], see Section \[sec.riskestim.bias.correction\] for details. #### Generalized CV [GCV, @Cra_Wah:1979] was introduced as a rotation-invariant version of LOO in least-squares regression, for estimating the risk of a linear estimator $\ERM = M {\bf Y}$ where ${\bf Y} = (Y_i)_{1 \leq i \leq n} \in \R^n$ and $M$ is an $n \times n$ matrix independent from ${\bf Y}$: $$\begin{aligned} \crit_{\mathrm{GCV}} (M,{\bf Y}) {:=}\frac{ n^{-1} {\left \lVert {\bf Y} - M {\bf Y} \right\rVert}^2 } { \paren{ 1 - n^{-1} \tr(M)}^2} \quad \mbox{where} \quad \forall t \in \R^n, \, {\left \lVert t \right\rVert}^2 = \sum_{i=1}^n t_i^2 \enspace .\end{aligned}$$ GCV is actually closer to $C_L$ [@Mal:1973] than to CV, since GCV can be seen as an approximation to $C_L$ with a particular estimator of the variance [@Efr:1986]. The efficiency of GCV has been proved in various frameworks, in particular by [@Li85; @KCLi:1987] and by [@Cao_Gol:2006]. #### Analytic Approximation When CV is used for selecting among linear models, [@Sha:1993] proposed an analytic approximation to LPO with $p \sim n$, which is called APCV. #### LOO bootstrap and .632 bootstrap The bootstrap is often used for stabilizing an estimator or an algorithm, replacing $\A(D_n)$ by the average of $\A(D_n^{\star})$ over several bootstrap resamples $D_n^{\star}$. This idea was applied by [@Efr:1983] to the LOO estimator of the risk, leading to the [*LOO bootstrap*]{}. Noting that the LOO bootstrap was biased, [@Efr:1983] gave a heuristic argument leading to the [*$.632$ bootstrap*]{} estimator of the risk, later modified into the [*$.632+$ bootstrap*]{} by [@Efr_Tib:1997]. The main drawback of these procedures is the weakness of their theoretical justifications. Only empirical studies have supported the good behaviour of $.632+ $ bootstrap [@Efr_Tib:1997; @Mol_Sim_Pfe:2005]. Historical remarks ------------------ Simple validation or hold-out was the first CV-like procedure. It was introduced in the psychology area [@Lars31] from the need for a reliable alternative to the [*resubstitution error*]{}, as illustrated by [@AnAB72]. The hold-out was used by [@Herz69] for assessing the quality of predictors. The problem of choosing the training set was first considered by [@Sto:1974], where “controllable” and “uncontrollable” data splits were distinguished; an instance of uncontrollable division can be found in the book by [@Simo71]. A primitive LOO procedure was used by [@Hill66] and by [@LaMi68] for evaluating the error rate of a prediction rule, and a primitive formulation of LOO can be found in a paper by [@MoTu68]. Nevertheless, LOO was actually introduced independently by [@Sto:1974], by [@All:1974] and by [@Gei:1975]. The relationship between LOO and the jackknife [@Que:1949], which both rely on the idea of removing one observation from the sample, has been discussed by [@Sto:1974] for instance. The hold-out and CV were originally used only for estimating the risk of an algorithm. The idea of using CV for model selection arose in the discussion of a paper by [@EfMo73] and in a paper by [@Geis74]. The first author to study LOO as a model selection procedure was [@Sto:1974], who proposed to use LOO again for estimating the risk of the selected model. Statistical properties of cross-validation estimators of the risk {#sec.riskestim} ================================================================= Understanding the behaviour of CV for model selection, which is the purpose of this survey, requires first to analyze the performances of CV as an estimator of the risk of a single algorithm. Two main properties of CV estimators of the risk are of particular interest: their bias, and their variance. Bias {#sec.riskestim.bias} ---- Dealing with the bias incurred by CV estimates can be made by two strategies: evaluating the amount of bias in order to choose the least biased CV procedure, or correcting for this bias. ### Theoretical assessment of the bias The independence of the training and the validation samples imply that for every algorithm $\A$ and any ${\ensuremath{I^{(t)}}}\subset {\left\{ \left. 1, \ldots, n \right. \right\}}$ with cardinality $n_t$, $$\E\croch{ {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}\paren{ \A ; D_n ; {\ensuremath{I^{(t)}}}} } = \E\croch{\gamma\paren{ \A \paren{{\ensuremath{D_n^{(t)}}}} ; \xi }} = \E \croch{ {\ensuremath{\mathcal{L}_{P}\paren{ \A\paren{D_{n_t}} }}} } \enspace .$$ Therefore, assuming that $\card({\ensuremath{I^{(t)}}}_j) = n_t$ for $j=1, \ldots, B$, the expectation of the CV estimator of the risk only depends on $n_t \,$: $$\label{eq.val.esperance} \E\croch{ {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}\paren{ \A ; D_n ; \paren{{\ensuremath{I^{(t)}}}_j}_{1 \leq j \leq B} } } = \E \croch{ {\ensuremath{\mathcal{L}_{P}\paren{ \A\paren{D_{n_t}} }}} } \enspace .$$ In particular shows that the bias of the CV estimator of the risk of $\A$ is the difference between the risks of $\A$, computed respectively with $n_t$ and $n$ data points. Since $n_t < n$, the bias of CV is usually nonnegative, which can be proved rigorously when the risk of $\A$ is a decreasing function of $n$, that is, when $\A$ is a smart rule; note however that a classical algorithm such as 1-nearest-neighbour in classification is not smart [@Dev_Gyo_Lug:1996 Section 6.8]. Similarly, the bias of CV tends to decrease with $n_t$, which is rigorously true if $\A$ is smart. More precisely, has led to several results on the bias of CV, which can be split into three main categories: asymptotic results ($\A$ is fixed and the sample size $n$ tends to infinity), non-asymptotic results (where $\A$ is allowed to make use of a number of parameters growing with $n$, say $n^{1/2}$, as often in model selection), and empirical results. They are listed below by statistical framework. #### Regression The general behaviour of the bias of CV (positive, decreasing with $n_t$) is confirmed by several papers and for several CV estimators. For LPO, non-asymptotic expressions of its bias were proved by [@Cel:2008:phd] for projection estimators, and by [@ArCe09] for regressograms and kernels estimators when the design is fixed. For VFCV and RLT, an asymptotic expansion of their bias was yielded by [@Bur:1989] for least-squares estimators in linear regression, and extended to spline smoothing [@Bur:1990]. Note finally that [@Efr:1986] proved non-asymptotic analytic expressions of the expectations of the LOO and GCV estimators of the risk in regression with binary data [see also @Efr:1983 for some explicit calculations]. #### Density estimation shows a similar picture. Non-asymptotic expressions for the bias of LPO estimators for kernel and projection estimators with the quadratic risk were proved by [@CeRo08] and by [@Celi08]. Asymptotic expansions of the bias of the LOO estimator for histograms and kernel estimators were previously proved by [@Rude82]; see [@Bowm84] for simulations. [@Hall87] derived similar results with the log-likelihood contrast for kernel estimators, and related the performance of LOO to the interaction between the kernel and the tails of the target density $s$. #### Classification For the simple problem of discriminating between two populations with shifted distributions, [@Dav_Hal:1992] compared the asymptotical bias of LOO and bootstrap, showing the superiority of the LOO when the shift size is $n^{-1/2} \, $: As $n$ tends to infinity, the bias of LOO stays of order $n^{-1}$, whereas that of bootstrap worsens to the order $n^{-1/2}$. On realistic synthetic and real biological data, [@Mol_Sim_Pfe:2005] compared the bias of LOO, VFCV and .632+ bootstrap: The bias decreases with $n_t$, and is generally minimal for LOO. Nevertheless, the $10$-fold CV bias is nearly minimal uniformly over their experiments. In the same experiments, .632+ bootstrap exhibits the smallest bias for moderate sample sizes and small signal-to-noise ratios, but a much larger bias otherwise. #### CV-calibrated algorithms When a family of algorithm $\paren{\A_{\lambda}}_{\lambda \in \Lambda}$ is given, and ${\ensuremath{\widehat{\lambda}}}$ is chosen by minimizing ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}(\A_{\lambda};D_n)$ over $\lambda$, ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}(\A_{{\ensuremath{\widehat{\lambda}}}};D_n)$ is biased for estimating the risk of $\A_{{\ensuremath{\widehat{\lambda}}}}(D_n)$, as reported from simulation experiments by [@Sto:1974] for the LOO, and by [@JoKM00] for VFCV in the variable selection setting. This bias is of different nature compared to the previous frameworks. Indeed, ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}(\A_{{\ensuremath{\widehat{\lambda}}}},D_n)$ is biased simply because ${\ensuremath{\widehat{\lambda}}}$ was chosen using the same data as ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}(\A_{\lambda},D_n)$. This phenomenon is similar to the optimism of ${\ensuremath{\mathcal{L}_{P_n}\paren{{\,\widehat{s}\,}(D_n)}}}$ as an estimator of the loss of ${\,\widehat{s}\,}(D_n)$. The correct way of estimating the risk of $\A_{{\ensuremath{\widehat{\lambda}}}}(D_n)$ with CV is to consider the full algorithm $\A': D_n {\mapsto}\A_{{\ensuremath{\widehat{\lambda}}}(D_n)}(D_n)$, and then to compute ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}\paren{\A';D_n}$. The resulting procedure is called “double cross” by [@Sto:1974]. ### Correction of the bias {#sec.riskestim.bias.correction} An alternative to choosing the CV estimator with the smallest bias is to correct for the bias of the CV estimator of the risk. [@Bur:1989; @Bur:1990] proposed a corrected VFCV estimator, defined by $$\begin{aligned} {\ensuremath{\widehat{\mathcal{L}}^{\mathrm{corrVF}}}} (\A; D_n) &= {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{VF}}}}}\paren{ \ERM ; D_n} + {\ensuremath{\mathcal{L}_{P_n}\paren{\A(D_n)}}} - \frac{1}{V} \sum_{j=1}^V {\ensuremath{\mathcal{L}_{P_n}\paren{\A(D_n^{(-A_j)})}}} \enspace , $$ and a corrected RLT estimator was defined similarly. Both estimators have been proved to be asymptotically unbiased for least-squares estimators in linear regression. When the $A_j$s have exactly the same size $n/V$, the corrected VFCV criterion is equal to the sum of the empirical risk and the $V$-fold penalty [@Arl:2008a], defined by $$\penVF (\A;D_n) = \frac{V-1}{V} \sum_{j=1}^V \croch{ {\ensuremath{\mathcal{L}_{P_n}\paren{\A(D_n^{(-A_j)})}}} - {\ensuremath{\mathcal{L}_{P_n^{(-A_j)}}\paren{\A(D_n^{(-A_j)})}}} } \enspace .$$ The $V$-fold penalized criterion was proved to be (almost) unbiased in the non-asymptotic framework for regressogram estimators. Variance {#sec.riskestim.var} -------- CV estimators of the risk using training sets of the same size $n_t$ have the same bias, but they still behave quite differently; their variance $\operatorname{var}({{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}(\A;D_n;({\ensuremath{I^{(t)}}}_j)_{1 \leq j \leq B}))$ captures most of the information to explain these differences. ### Variability factors Assume that $\card({\ensuremath{I^{(t)}}}_j) = n_t$ for every $j$. The variance of CV results from the combination of several factors, in particular $(n_t,n_v)$ and $B$. #### Influence of $(n_t,n_v)$ Let us consider the hold-out estimator of the risk. Following in particular [@NaBe03], $$\begin{aligned} & \qquad \notag \operatorname{var}\croch{ {{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}\paren{\A;D_n;{\ensuremath{I^{(t)}}}}} \\ \notag &= \E\croch{\operatorname{var}\paren{ {\ensuremath{\mathcal{L}_{P_n^{(v)}}\paren{\A({\ensuremath{D_n^{(t)}}})}}} {\, \right| \left. \,}{\ensuremath{D_n^{(t)}}}}} + \operatorname{var}\croch{ {\ensuremath{\mathcal{L}_{P}\paren{\A(D_{n_t})}}} } \\ &= \frac{1}{n_v}\E\croch{\operatorname{var}\paren{\gamma\paren{{\,\widehat{s}\,},\xi}{\, \right| \left. \,}{\,\widehat{s}\,}= \A({\ensuremath{D_n^{(t)}}})}} + \operatorname{var}\croch{ {\ensuremath{\mathcal{L}_{P}\paren{\A(D_{n_t})}}} } \label{exp.variance} \enspace .\end{aligned}$$ The first term, proportional to $1/n_v$, shows that more data for validation decreases the variance of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}$, because it yields a better estimator of ${\ensuremath{\mathcal{L}_{P}\paren{\A({\ensuremath{D_n^{(t)}}})}}}$. The second term shows that the variance of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}$ also depends on the distribution of ${\ensuremath{\mathcal{L}_{P}\paren{\A({\ensuremath{D_n^{(t)}}})}}}$ around its expectation; in particular, it strongly depends on the [*stability*]{} of $\A$. #### Stability and variance When $\A$ is unstable, ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LOO}}}}}\paren{\A}$ has often been pointed out as a variable estimator [Section 7.10, @Has_Tib_Fri:2001; @Bre:1996]. Conversely, this trend disappears when $\A$ is stable, as noticed by [@Mol_Sim_Pfe:2005] from a simulation experiment. The relation between the stability of $\A$ and the variance of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}\paren{\A}$ was pointed out by [@DeWa79] in classification, through upper bounds on the variance of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LOO}}}}}\paren{\A}$. [@BoEl02] extended these results to the regression setting, and proved upper bounds on the maximal upward deviation of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LOO}}}}}\paren{\A}$. Note finally that several approaches based on the bootstrap have been proposed for reducing the variance of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LOO}}}}}\paren{\A}$, such as LOO bootstrap, .632 bootstrap and .632+ bootstrap [@Efr:1983]; see also Section \[sec.def.classex.other\]. #### Partial splitting and variance When $(n_t,n_v)$ is fixed, the variability of CV tends to be larger for partial data splitting methods than for LPO. Indeed, having to choose $B < {n \choose n_t}$ subsets $({\ensuremath{I^{(t)}}}_j)_{1 \leq j \leq B}$ of ${\left\{ \left. 1, \ldots, n \right. \right\}}$, usually randomly, induces an additional variability compared to ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LPO}}}}}$ with $p = n - n_t$. In the case of MCCV, this variability decreases like $B^{-1}$ since the ${\ensuremath{I^{(t)}}}_j$ are chosen independently. The dependence on $B$ is slightly different for other CV estimators such as RLT or VFCV, because the ${\ensuremath{I^{(t)}}}_j$ are not independent. In particular, it is maximal for the hold-out, and minimal (null) for LOO (if $n_t = n-1$) and LPO (with $p = n-n_t$). Note that the dependence on $V$ for VFCV is more complex to evaluate, since $B$, $n_t$, and $n_v$ simultaneously vary with $V$. Nevertheless, a non-asymptotic theoretical quantification of this additional variability of VFCV has been obtained by [@CeRo08] in the density estimation framework [see also empirical considerations by @JoKM00]. ### Theoretical assessment of the variance Understanding precisely how $\operatorname{var}({{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}(\A))$ depends on the splitting scheme is complex in general, since $n_t$ and $n_v$ have a fixed sum $n$, and the number of splits $B$ is generally linked with $n_t$ (for instance, for LPO and VFCV). Furthermore, the variance of CV behaves quite differently in different frameworks, depending in particular on the stability of $\A$. The consequence is that contradictory results have been obtained in different frameworks, in particular on the value of $V$ for which the VFCV estimator of the risk has a minimal variance [@Bur:1989; @Has_Tib_Fri:2001 Section 7.10]. Despite the difficulty of the problem, the variance of several CV estimators of the risk has been assessed in several frameworks, as detailed below. #### Regression In the linear regression setting, [@Bur:1989] yielded asymptotic expansions of the variance of the VFCV and RLT estimators of the risk with homoscedastic data. The variance of RLT decreases with $B$, and in the case of VFCV, in a particular setting, $$\begin{aligned} \operatorname{var}\paren{{{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{VF}}}}}(\A)} = \frac{2 \sigma^2}{n} + \frac{4 \sigma^4}{n^{2}} \croch{ 4 + \frac{4}{V-1}+\frac{2}{(V-1)^2}+\frac{1}{(V-1)^3}}+o\paren{n^{-2}} \enspace .\end{aligned}$$ The asymptotical variance of the VFCV estimator of the risk decreases with $V$, implying that LOO asymptotically has the minimal variance. Non-asymptotic closed-form formulas of the variance of the LPO estimator of the risk have been proved by [@Cel:2008:phd] in regression, for projection and kernel estimators for instance. On the variance of RLT in the regression setting, see the asymptotic results of [@Gir:1998] for Nadaraya-Watson kernel estimators, as well as the non-asymptotic computations and simulation experiments by [@NaBe03] with several learning algorithms. #### Density estimation Non-asymptotic closed-form formulas of the variance of the LPO estimator of the risk have been proved by [@CeRo08] and by [@Celi08] for projection and kernel estimators. In particular, the dependence of the variance of ${\ensuremath{\widehat{\mathcal{L}}^{\mathrm{LPO}}}}$ on $p$ has been quantified explicitly for histogram and kernel estimators by [@CeRo08]. #### Classification For the simple problem of discriminating between two populations with shifted distributions, [@Dav_Hal:1992] showed that the gap between asymptotic variances of LOO and bootstrap becomes larger when data are noisier. [@NaBe03] made non-asymptotic computations and simulation experiments with several learning algorithms. [@Has_Tib_Fri:2001] empirically showed that VFCV has a minimal variance for some $2<V<n$, whereas LOO usually has a large variance; this fact certainly depends on the stability of the algorithm considered, as showed by simulation experiments by [@Mol_Sim_Pfe:2005]. ### Estimation of the variance {#sec.riskestim.var.estim} There is no universal—valid under all distributions—unbiased estimator of the variance of RLT [@NaBe03] and VFCV estimators [@Ben_Gra:2004]. In particular, [@Ben_Gra:2004] recommend the use of variance estimators taking into account the correlation structure between test errors; otherwise, the variance of CV can be strongly underestimated. Despite these negative results, (biased) estimators of the variance of ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{CV}}}}}$ have been proposed by [@NaBe03], by [@Ben_Gra:2004] and by [@Mar_etal:2005], and tested in simulation experiments in regression and classification. Furthermore, in the framework of density estimation with histograms, [@CeRo08] proposed an estimator of the variance of the LPO risk estimator. Its accuracy is assessed by a concentration inequality. These results have recently been extended to projection estimators by [@Celi08]. Cross-validation for efficient model selection {#sec.cvefficient} ============================================== This section tackles the properties of CV procedures for model selection when the goal is estimation (see Section \[sec.modsel.estim\]). Relationship between risk estimation and model selection {#sec.cvefficient.link-risk-mod-sel} -------------------------------------------------------- As shown in Section \[sec.modselproc.unbiased\], minimizing an unbiased estimator of the risk leads to an efficient model selection procedure. One could conclude here that the best CV procedure for estimation is the one with the smallest bias and variance (at least asymptotically), for instance, LOO in the least-squares regression framework [@Bur:1989]. Nevertheless, the best CV estimator of the risk is not necessarily the best model selection procedure. For instance, [@Bre_Spe:1992] observed that uniformly over the models, the best risk estimator is LOO, whereas 10-fold CV is more accurate for model selection. Three main reasons for such a difference can be invoked. First, the asymptotic framework ($\A$ fixed, $n \rightarrow \infty$) may not apply to models close to the oracle, which typically has a dimension growing with $n$ when $s$ does not belong to any model. Second, as explained in Section \[sec.modselproc.biased\], estimating the risk of each model with some bias can be beneficial and compensate the effect of a large variance, in particular when the signal-to-noise ratio is small. Third, for model selection, what matters is not that every estimate of the risk has small bias and variance, but more that $$\sign\paren{\crit(m_1) - \crit(m_2)} = \sign\paren{{\ensuremath{\mathcal{L}_{P}\paren{\ERM_{m_1}}}} - {\ensuremath{\mathcal{L}_{P}\paren{\ERM_{m_2}}}}}$$ with the largest probability for models $m_1,m_2$ near the oracle. Therefore, specific studies are required to evaluate the performances of the various CV procedures in terms of model selection efficiency. In most frameworks, the model selection performance directly follows from the properties of CV as an estimator of the risk, but not always. The global picture {#subsubsec.asymptotic.equiv.} ------------------ Let us start with the classification of model selection procedures made by [@Sha:1997] in the linear regression framework, since it gives a good idea of the performance of CV procedures for model selection in general. Typically, the efficiency of CV only depends on the asymptotics of $n_t/n \,$: - When $n_t \sim n$, CV is asymptotically equivalent to Mallows’ $C_p$, hence asymptotically optimal. - When $n_t \sim \lambda n$ with $\lambda \in (0,1)$, CV is asymptotically equivalent to GIC$_{\kappa}$ with $\kappa = 1+\lambda^{-1}$, which is defined as AIC with a penalty multiplied by $\kappa/2$. Hence, such CV procedures are overpenalizing by a factor $(1+\lambda)/(2 \lambda) > 1$. The above results have been proved by [@Sha:1997] for LPO [see also @KCLi:1987 for the LOO]; they also hold for RLT when $B \gg n^2$ since RLT is then equivalent to LPO [@Zha:1993]. In a general statistical framework, the model selection performance of MCCV, VFCV, LOO, LOO Bootstrap, and .632 bootstrap for selection among minimum contrast estimators was studied in a series of papers [@vdL_Dud:2003; @vdL_Dud_Kel:2004; @vdL_Dud_vdV:2006; @vdV_Dud_vdL:2006]; these results apply in particular to least-squares regression and density estimation. It turns out that under mild conditions, an oracle-type inequality is proved, showing that up to a multiplying factor $C_n \rightarrow 1$, the risk of CV is smaller than the minimum of the risks of the models with a sample size $n_t$. In particular, in most frameworks, this implies the asymptotic optimality of CV as soon as $n_t \sim n$. When $n_t \sim \lambda n$ with $\lambda \in (0,1)$, this naturally generalizes Shao’s results. Results in various frameworks ----------------------------- This section gathers results about model selection performances of CV when the goal is estimation, in various frameworks. Note that model selection is considered here with a general meaning, including in particular bandwidth choice for kernel estimators. #### Regression First, the results of Section \[subsubsec.asymptotic.equiv.\] suggest that CV is suboptimal when $n_t$ is not asymptotically equivalent to $n$. This fact has been proved rigorously for VFCV when $V = \grandO(1)$ with regressograms [@Arl:2008a]: with large probability, the risk of the model selected by VFCV is larger than $1 + \kappa(V)$ times the risk of the oracle, with $\kappa(V)>0$ for every fixed $V$. Note however that the best $V$ for VFCV is not the largest one in every regression framework, as shown empirically in linear regression [@Bre_Spe:1992; @HeTs86]; [@Bre:1996] proposed to explain this phenomenon by relating the stability of the candidate algorithms and the model selection performance of LOO in various regression frameworks. Second, the “universality” of CV has been confirmed by showing that it naturally adapts to heteroscedasticity of data when selecting among regressograms. Despite its suboptimality, VFCV with $V = \grandO(1)$ satisfies a non-asymptotic oracle inequality with constant $C>1$ [@Arl:2008a]. Furthermore, $V$-fold penalization (which often coincides with corrected VFCV, see Section \[sec.riskestim.bias.correction\]) satisfies a non-asymptotic oracle inequality with $C_n \rightarrow 1$ as $n \rightarrow +\infty$, both when $V = \grandO(1)$ [@Arl:2008a] and when $V=n$ [@Arlo08c]. Note that $n$-fold penalization is very close to LOO, suggesting that it is also asymptotically optimal with heteroscedastic data. Simulation experiments in the context of change-point detection confirmed that CV adapts well to heteroscedasticity, contrary to usual model selection procedures in the same framework [@ArCe09]. The performances of CV have also been assessed for other kinds of estimators in regression. For choosing the number of knots in spline smoothing, [@Bur:1990] proved that corrected versions of VFCV and RLT are asymptotically optimal provided $n/(B n_v) = \grandO(1)$. Furthermore, in kernel regression, several CV methods have been compared to GCV in kernel regression by [@Har_Hal_Mar:1988] and by [@Gir:1998]; the conclusion is that GCV and related criteria are computationally more efficient than MCCV or RLT, for a similar statistical performance. Finally, note that asymptotic results about CV in regression have been proved by [@Gyo_etal:2002], and an oracle inequality with constant $C>1$ has been proved by [@Wegk03] for the hold-out, with least-squares estimators. #### Density estimation CV performs similarly than in regression for selecting among least-squares estimators [@vdL_Dud_Kel:2004]: It yields a risk smaller than the minimum of the risk with a sample size $n_t$. In particular, non-asymptotic oracle inequalities with constant $C>1$ have been proved by [@Cel:2008:phd] for the LPO when $p/n\in [a,b]$, for some $0<a<b<1$. The performance of CV for selecting the bandwidth of kernel density estimators has been studied in several papers. With the least-squares contrast, the efficiency of LOO was proved by [@Hal:1983] and generalized to the multivariate framework by [@Ston84]; an oracle inequality asymptotically leading to efficiency was recently proved by [@Dale05]. With the Kullback-Leibler divergence, CV can suffer from troubles in performing model selection [see also @ScGr81; @ChGW87]. The influence of the tails of the target $s$ was studied by [@Hall87], who gave conditions under which CV is efficient and the chosen bandwidth is optimal at first-order. #### Classification In the framework of binary classification by intervals (that is, with $\X=[0,1]$ and piecewise constant classifiers), [@KMNR97] proved an oracle inequality for the hold-out. Furthermore, empirical experiments show that CV yields (almost) always the best performance, compared to deterministic penalties [@KMNR97]. On the contrary, simulation experiments by [@Bar_Bou_Lug:2002] in the same setting showed that random penalties such as Rademacher complexity and maximal discrepancy usually perform much better than hold-out, which is shown to be more variable. Nevertheless, the hold-out still enjoys quite good theoretical properties: It was proved to adapt to the margin condition by [@Bla_Mas:2006], a property nearly unachievable with usual model selection procedures [see also @Mas:2003:St-Flour Section 8.5]. This suggests that CV procedures are naturally adaptive to several unknown properties of data in the statistical learning framework. The performance of the LOO in binary classification was related to the stability of the candidate algorithms by [@KeRo99]; they proved oracle-type inequalities called “sanity-check bounds”, describing the worst-case performance of LOO [see also @BoEl02]. An experimental comparison of several CV methods and bootstrap-based CV (in particular .632+ bootstrap) in classification can also be found in papers by [@Efr:1986] and [@Efr_Tib:1997]. Cross-validation for identification {#sec.cvconsistent} =================================== Let us now focus on model selection when the goal is to identify the “true model” $S_{m_0}$, as described in Section \[sec.modsel.identif\]. In this framework, asymptotic optimality is replaced by (model) consistency, that is, $${\mathbb{P}}\paren{ \mh(D_n) = m_0 } \xrightarrow[n \rightarrow \infty]{} 1 \enspace .$$ Classical model selection procedures built for identification, such as BIC, are described in Section \[sec.modselproc.identif\]. General conditions towards model consistency -------------------------------------------- At first sight, it may seem strange to use CV for identification: LOO, which is the pioneering CV procedure, is actually closely related to the unbiased risk estimation principle, which is only efficient when the goal is estimation. Furthermore, estimation and identification are somehow contradictory goals, as explained in Section \[sec.modsel.estim-vs-identif\]. This intuition about inconsistency of some CV procedures is confirmed by several theoretical results. [@Sha:1993] proved that several CV methods are inconsistent for variable selection in linear regression: LOO, LPO, and BICV when $\liminf_{n \rightarrow \infty} (n_t/n) > 0$. Even if these CV methods asymptotically select all the true variables with probability 1, the probability that they select too much variables does not tend to zero. More generally, [@Sha:1997] proved that CV procedures behave asymptotically like GIC$_{\lambda_n}$ with $\lambda_n = 1 + n/n_t$, which leads to inconsistency as soon as $n / n_t = \grandO(1)$. In the context of ordered variable selection in linear regression, [@Zha:1993] computed the asymptotic value of the probability of selecting the true model for several CV procedures. He also numerically compared the values of this probability for the same CV procedures in a specific example. For LPO with $p/n \rightarrow \lambda \in ( 0 , 1)$ as $n$ tends to $+\infty$, ${\mathbb{P}}\paren{\mh = m_0}$ increases with $\lambda$. The result is slightly different for VFCV: ${\mathbb{P}}\paren{\mh = m_0}$ increases with $V$ (hence, it is maximal for the LOO, which is the worst case of LPO). The variability induced by the number $V$ of splits seems to be more important here than the bias of VFCV. Nevertheless, ${\mathbb{P}}\paren{\mh = m_0}$ is almost constant between $V=10$ and $V=n$, so that taking $V>10$ is not advised for computational reasons. These results suggest that if the training sample size $n_t$ is negligible in front of $n$, then model consistency could be obtained. This has been confirmed theoretically by [@Sha:1993; @Sha:1997] for the variable selection problem in linear regression: CV is consistent when $n \gg n_t \rightarrow \infty$, in particular RLT, BICV (defined in Section \[subsec.classical.examples.partial\]) and LPO with $p=p_n \sim n$ and $n-p_n \rightarrow \infty$. Therefore, when the goal is to identify the true model, a larger proportion of the data should be put in the validation set in order to improve the performance. This phenomenon is somewhat related to the [*cross-validation paradox*]{} [@Yan:2006]. Refined analysis for the algorithm selection problem ---------------------------------------------------- The behaviour of CV for identification is better understood by considering a more general framework, where the goal is to select among statistical algorithms the one with the fastest convergence rate. [@Yan:2006; @Yan:2007b] considered this problem for two candidate algorithms (or more generally any finite number of algorithms). Let us mention here that [@Sto:1977b] considered a few specific examples of this problem, and showed that LOO can be inconsistent for choosing the best among two “good” estimators. The conclusion of Yang’s papers is that the sufficient condition on $n_t$ for the consistency in selection of CV strongly depends on the convergence rates $\paren{r_{n,i}}_{i=1,2}$ of the candidate algorithms. Let us assume that $r_{n,1}$ and $r_{n,2}$ differ at least by a multiplicative constant $C>1$. Then, in the regression framework, if the risk of $\ERM_i$ is measured by $\E {\left \lVert \ERM_i - s \right\rVert}_2$, [@Yan:2007b] proved that the hold-out, VFCV, RLT and LPO with voting (CV-v, see Section \[sec.def.simple2cross.CVgal\]) are consistent in selection if $$\label{eq.cond-consist.reg} n_v, n_t \rightarrow \infty \quad \mbox{and} \quad \sqrt{n_v} \max_i r_{n_t,i} \rightarrow \infty \enspace ,$$ under some conditions on ${\left \lVert \ERM_i - s \right\rVert}_p$ for $p=2,4,\infty$. In the classification framework, if the risk of $\ERM_i$ is measured by ${\mathbb{P}}\paren{\ERM_i \neq s}$, [@Yan:2006] proved the same consistency result for CV-v under the condition $$\label{eq.cond-consist.classif} n_v, n_t \rightarrow \infty \quad \mbox{and} \quad \frac{ n_v \max_i r_{n_t,i}^2 } { s_{n_t}} \rightarrow \infty \enspace ,$$ where $s_n$ is the convergence rate of ${\mathbb{P}}\paren{\ERM_1(D_n) \neq \ERM_2(D_n)}$. Intuitively, consistency holds as soon as the uncertainty of each estimate of the risk (roughly proportional to $n_v^{-1/2}$) is negligible in front of the risk gap ${\left\lvert r_{n_t,1} - r_{n_t,2} \right\rvert}$ (which is of the same order as $\max_i r_{n_t,i}$). This condition holds either when at least one of the algorithms converges at a non-parametric rate, or when $n_t \ll n$, which artificially widens the risk gap. Empirical results in the same direction were proved by [@Die:1998] and by [@Alp:1999], leading to the advice that $V=2$ is the best choice when VFCV is used for comparing two learning procedures. See also the results by [@NaBe03] about CV considered as a testing procedure comparing two candidate algorithms. The sufficient conditions and can be simplified depending on $\max_i r_{n,i}$, so that the ability of CV to distinguish between two algorithms depends on their convergence rates. On the one hand, if $\max_i r_{n,i} \propto n^{-1/2}$, then or only hold when $n_v \gg n_t \rightarrow \infty$ (under some conditions on $s_n$ in classification). Therefore, the cross-validation paradox holds for comparing algorithms converging at the parametric rate (model selection when a true model exists being only a particular case). Note that possibly stronger conditions can be required in classification where algorithms can converge at fast rates, between $n^{-1}$ and $n^{-1/2}$. On the other hand, and are milder conditions when $\max_i r_{n,i} \gg n^{-1/2}$: They are implied by $n_t / n_v = \grandO(1)$, and they even allow $n_t \sim n$ (under some conditions on $s_n$ in classification). Therefore, non-parametric algorithms can be compared by more usual CV procedures ($n_t>n/2$), even if LOO is still excluded by conditions and . Note that according to a simulation experiments, CV with averaging (that is, CV as usual) and CV with voting are equivalent at first but not at second order, so that they can differ when $n$ is small [@Yan:2007b]. Specificities of some frameworks {#sec.specific} ================================ Originally, the CV principle has been proposed for [*i.i.d.*]{} observations and usual contrasts such as least-squares and log-likelihood. Therefore, CV procedures may have to be modified in other specific frameworks, such as estimation in presence of outliers or with dependent data. Density estimation ------------------ In the density estimation framework, some specific modifications of CV have been proposed. First, [@Hal_Mar_Par:1992] defined the “smoothed CV”, which consists in pre-smoothing the data before using CV, an idea related to the smoothed bootstrap. They proved that smoothed CV yields an excellent asymptotical model selection performance under various smoothness conditions on the density. Second, when the goal is to estimate the density at one point (and not globally), [@Hal_Sch:1989] proposed a local version of CV and proved its asymptotic optimality. Robustness to outliers ---------------------- In presence of outliers in regression, [@Leu:2005] studied how CV must be modified to get both asymptotic efficiency and a consistent bandwidth estimator [see also @LeMW93]. Two changes are possible to achieve robustness: Choosing a “robust” regressor, or choosing a robust loss-function. In presence of outliers, classical CV with a non-robust loss function has been shown to fail by [@Haerd84]. [@Leu:2005] described a CV procedure based on robust losses like $L^1$ and Huber’s [@Hube64] ones. The same strategy remains applicable to other setups like linear models in [@RoFB97]. Time series and dependent observations {#sec.dependent} -------------------------------------- As explained in Section \[sec.def.philo\], CV is built upon the heuristics that part of the sample (the validation set) can play the role of [*new data*]{} with respect to the rest of the sample (the training set). “New” means that the validation set is independent from the training set with the same distribution. Therefore, when data $\xi_1, \ldots, \xi_n$ are not independent, CV must be modified, like other model selection procedures [in non-parametric regression with dependent data, see the review by @Ops_Wan_Yan:2001]. Let us first consider the statistical framework of Section \[sec.intro\] with $\xi_1, \ldots, \xi_n$ identically distributed but not independent. Then, when for instance data are positively correlated, [@Har_Weh:1986] proved that CV overfits for choosing the bandwidth of a kernel estimator in regression [see also @Chu_Mar:1991; @Ops_Wan_Yan:2001]. The main approach used in the literature for solving this issue is to choose ${\ensuremath{I^{(t)}}}$ and ${\ensuremath{I^{(v)}}}$ such that $\min_{i \in {\ensuremath{I^{(t)}}}, \, j \in {\ensuremath{I^{(v)}}}} {\left\lvert i-j \right\rvert} > h > 0$, where $h$ controls the distance from which observations $i$ and $j$ are independent. For instance, the LOO can be changed into: ${\ensuremath{I^{(v)}}}= {\left\{ \left. J \right. \right\}}$ where $J$ is uniformly chosen in ${\left\{ \left. 1, \ldots, n \right. \right\}}$, and ${\ensuremath{I^{(t)}}}= {\left\{ \left. 1, \ldots, J-h-1, J+h+1, \ldots, n \right. \right\}}$, a method called “modified CV” by [@Chu_Mar:1991] in the context of bandwidth selection. Then, for short range dependences, $\xi_i$ is almost independent from $\xi_j$ when ${\left\lvert i-j \right\rvert} > h$ is large enough, so that $\paren{\xi_j}_{j \in {\ensuremath{I^{(t)}}}}$ is almost independent from $\paren{\xi_j}_{j \in {\ensuremath{I^{(v)}}}}$. Several asymptotic optimality results have been proved on modified CV, for instance by [@Har_Vie:1990] for bandwidth choice in kernel density estimation, when data are $\alpha$-mixing (hence, with a short range dependence structure) and $h=h_n \rightarrow \infty$ “not too fast”. Note that modified CV also enjoys some asymptotic optimality results with long-range dependences, as proved by [@Hal_Lah_Pol:1995], even if an alternative block bootstrap method seems more appropriate in such a framework. Several alternatives to modified CV have also been proposed. The “$h$-block CV” [@Bur_Cho_Nol:1994] is modified CV plus a corrective term, similarly to the bias-corrected CV by [@Bur:1989] (see Section \[sec.riskestim.bias\]). Simulation experiments in several (short range) dependent frameworks show that this corrective term matters when $h/n$ is not small, in particular when $n$ is small. The “partitioned CV” has been proposed by [@Chu_Mar:1991] for bandwidth selection: An integer $g>0$ is chosen, a bandwidth ${\ensuremath{\widehat{\lambda}}}_k$ is chosen by CV based upon the subsample $\paren{\xi_{k+g j}}_{j \geq 0}$ for each $k = 1, \ldots, g$, and the selected bandwidth is a combination of $\sparen{{\ensuremath{\widehat{\lambda}}}_k}$. When a parametric model is available for the dependency structure, [@Har:1994] proposed the “time series CV”. An important framework where data often are dependent is time-series analysis, in particular when the goal is to predict the next observation $\xi_{n+1}$ from the past $\xi_1, \ldots, \xi_n$. When data are stationary, $h$-block CV and similar approaches can be used to deal with (short range) dependences. Nevertheless, [@Bur_Nol:1992] proved in some specific framework that unaltered CV is asymptotic optimal when $\xi_1, \ldots, \xi_n$ is a stationary Markov process. On the contrary, using CV for non-stationary time-series is a quite difficult problem. The only reasonable approach in general is the hold-out, that is, ${\ensuremath{I^{(t)}}}= {\left\{ \left. 1, \ldots, m \right. \right\}}$ and ${\ensuremath{I^{(v)}}}= {\left\{ \left. m+1, \ldots, n \right. \right\}}$ for some deterministic $m$. Each model is first trained with $\paren{\xi_j}_{j \in {\ensuremath{I^{(t)}}}}$. Then, it is used for predicting successively $\xi_{m+1}$ from $\paren{\xi_j}_{j \leq m}$, $\xi_{m+2}$ from $\paren{\xi_j}_{j \leq m+1}$, and so on. The model with the smallest average error for predicting $\paren{\xi_j}_{j \in {\ensuremath{I^{(v)}}}}$ from the past is chosen. Large number of models {#sec.specific.manymodels} ---------------------- As mentioned in Section \[sec.modselproc\], model selection procedures estimating unbiasedly the risk of each model fail when, in particular, the number of models grows exponentially with $n$ [@Bir_Mas:2006]. Therefore, CV cannot be used directly, except maybe with $n_t \ll n$, provided $n_t$ is well chosen [see Section \[sec.cvefficient\] and @Cel:2008:phd Chapter 6]. For least-squares regression with homoscedastic data, [@Wegk03] proposed to add to the hold-out estimator of the risk a penalty term depending on the number of models. This method is proved to satisfy a non-asymptotic oracle inequality with leading constant $C >1$. Another general approach was proposed by [@ArCe09] in the context of multiple change-point detection. The idea is to perform model selection in two steps: First, gather the models $\paren{S_m}_{{m \in \M}_n}$ into meta-models $\sparen{\St_D}_{D \in \D_n}$, where $\D_n$ denotes a set of indices such that $\card(\D_n)$ grows at most polynomially with $n$. Inside each meta-model $\St_D = \bigcup_{{m \in \M}_n(D)} S_m$, $\ERM_D$ is chosen from data by optimizing a given criterion, for instance the empirical contrast ${\ensuremath{\mathcal{L}_{P_n}\paren{t}}}$, but other criteria can be used. Second, CV is used for choosing among $\paren{\ERM_D}_{D \in \D_n}$. Simulation experiments show this simple trick automatically takes into account the cardinality of $\M_n$, even when data are heteroscedastic, contrary to other model selection procedures built for exponential collection of models which all assume homoscedasticity of data. Closed-form formulas and fast computation {#sec.complex} ========================================= Resampling strategies, like CV, are known to be time consuming. The naive implementation of CV has a computational complexity of $B$ times the complexity of training each algorithm $\A$, which is usually intractable for LPO, even with $p=1$. The computational cost of VFCV or RLT can still be quite costly when $B>10$ in many practical problems. Nevertheless, closed-form formulas for CV estimators of the risk can be obtained in several frameworks, which greatly decreases the computational cost of CV. In density estimation, closed-form formulas have been originally derived by [@Rude82] and by [@Bowm84] for the LOO risk estimator of histograms and kernel estimators. These results have been recently extended by [@CeRo08] to the LPO risk estimator with the quadratic loss. Similar results are more generally available for projection estimators as settled by [@Celi08]. Intuitively, such formulas can be obtained provided the number $N$ of values taken by the $B={n\choose n_v}$ hold-out estimators of the risk, corresponding to different data splittings, is at most polynomial in the sample size. For least-squares estimators in linear regression, [@Zha:1993] proved a closed-form formula for the LOO estimator of the risk. Similar results have been obtained by [@Wahb75; @Wahb77a], and by [@Cra_Wah:1979] in the spline smoothing context as well. These papers led in particular to the definition of GCV (see Section \[sec.def.classex.other\]) and related procedures, which are often used instead of CV (with a naive implementation) because of their small computational cost, as emphasized by [@Gir:1998]. Closed-form formulas for the LPO estimator of the risk were also obtained by [@Cel:2008:phd] in regression for kernel and projection estimators, in particular for regressograms. An important property of these closed-form formulas is their additivity: For a regressogram associated to a partition $({I_{\lambda}})_{{\lambda \in \Lambda_m}}$ of $\X$, the LPO estimator of the risk can be written as a sum over ${\lambda \in \Lambda_m}$ of terms which only depend on observations $(X_i,Y_i)$ such that $X_i \in {I_{\lambda}}$. Therefore, dynamic programming [@BeDr62] can be used for minimizing the LPO estimator of the risk over the set of partitions of $\X$ in $D$ pieces. As an illustration, [@ArCe09] successfully applied this strategy in the change-point detection framework. Note that the same idea can be used with VFCV or RLT, but for a larger computational cost since no closed-form formulas are available for these CV methods. Finally, in frameworks where no closed-form formula can be proved, some efficient algorithms exist for avoiding to recompute ${{\ensuremath{\widehat{\mathcal{L}}^{\mathrm{H-O}}}}}\sparen{\A;D_n;{\ensuremath{I^{(t)}}}_j}$ from scratch for each data splitting ${\ensuremath{I^{(t)}}}_j$. These algorithms rely on updating formulas such as the ones by [@Ripl96] for LOO in linear and quadratic discriminant analysis; this approach makes LOO as expensive to compute as the empirical risk. Very similar formulas are also available for LOO and the $k$-nearest neighbours estimator in classification [@DaMa08]. Conclusion: which cross-validation method for which problem? {#sec.conclu} ============================================================ This conclusion collects a few guidelines aiming at helping CV users, first interpreting the results of CV, second appropriately using CV in each specific problem. The general picture {#sec.conclu.gal} ------------------- Drawing a general conclusion on CV methods is an impossible task because of the variety of frameworks where CV can be used, which induces a variety of behaviors of CV. Nevertheless, we can still point out the three main criteria to take into account for choosing a CV method for a particular model selection problem: - [*Bias*]{}: CV roughly estimates the risk of a model with a sample size $n_t < n$ (see Section \[sec.riskestim.bias\]). Usually, this implies that CV overestimates the variance term compared to the bias term in the bias-variance decomposition with sample size $n$.\ When the goal is estimation and the signal-to-noise ratio (SNR) is large, the smaller bias usually is the better, which is obtained by taking $n_t \sim n$. Otherwise, CV can be asymptotically suboptimal. Nevertheless, when the goal is estimation and the SNR is small, keeping a small upward bias for the variance term often improves the performance, which is obtained by taking $n_t \sim \kappa n$ with $\kappa \in (0,1)$. See Section \[sec.cvefficient\].\ When the goal is identification, a large bias is often needed, which is obtained by taking $n_t \ll n$; depending on the framework, larger values of $n_t$ can also lead to model consistency, see Section \[sec.cvconsistent\]. - [*Variability*]{}: The variance of the CV estimator of the risk is usually a decreasing function of the number $B$ of splits, for a fixed training size. When the number of splits is fixed, the variability of CV also depends on the training sample size $n_t$. Usually, CV is more variable when $n_t$ is closer to $n$. However, when $B$ is linked with $n_t$ (as for VFCV or LPO), the variability of CV must be quantified precisely, which has been done in few frameworks. The only general conclusion on this point is that the CV method with minimal variability seems strongly framework-dependent, see Section \[sec.riskestim.var\] for details. - [*Computational complexity*]{}: Unless closed-form formulas or analytic approximations are available (see Section \[sec.complex\]), the complexity of CV is roughly proportional to the number of data splits: 1 for the hold-out, $V$ for VFCV, $B$ for RLT or MCCV, $n$ for LOO, and $\binom{n}{p}$ for LPO. The optimal trade-off between these three factors can be different for each problem, depending for instance on the computational complexity of each estimator, on specificities of the framework considered, and on the final user’s trade-off between statistical performance and computational cost. Therefore, no “optimal CV method” can be pointed out before having taken into account the final user’s preferences. Nevertheless, in density estimation, closed-form expressions of the LPO estimator have been derived by [@CeRo08] with histograms and kernel estimators, and by [@Celi08] for projection estimators. These expressions allow to perform LPO without additional computational cost, which reduces the aforementioned trade-off to the easier bias-variability trade-off. In particular, [@CeRo08] proposed to choose $p$ for LPO by minimizing a criterion defined as the sum of a squared bias and a variance terms [see also @Pol_Rom_Wol:1999 Chapter 9]. How the splits should be chosen? {#sec.choix-It} -------------------------------- For hold-out, VFCV, and RLT, an important question is to choose a particular sequence of data splits. First, should this step be random and independent from $D_n$, or take into account some features of the problem or of the data? It is often recommended to take into account the structure of data when choosing the splits. If data are stratified, the proportions of the different strata should (approximately) be the same in the sample and in each training and validation sample. Besides, the training samples should be chosen so that $\ERM_m({\ensuremath{D_n^{(t)}}})$ is well defined for every training set; in the regressogram case, this led [@Arl:2008a] and [@ArCe09] to choose carefully the splitting scheme. In supervised classification, practitioners usually choose the splits so that the proportion of each class is the same in every validation sample as in the sample. Nevertheless, [@Bre_Spe:1992] made simulation experiments in regression for comparing several splitting strategies. No significant improvement was reported from taking into account the stratification of data for choosing the splits. Another question related to the choice of $({\ensuremath{I^{(t)}}}_j)_{1 \leq j \leq B}$ is whether the ${\ensuremath{I^{(t)}}}_j$ should be independent (like MCCV), slighly dependent (like RLT), or strongly dependent (like VFCV). It seems intuitive that giving similar roles to all data points in the $B$ “training and validation tasks” should yield more reliable results as other methods. This intuition may explain why VFCV is much more used than RLT or MCCV. Similarly, [@Sha:1993] proposed a CV method called BICV, where every point and pair of points appear in the same number of splits, see Section \[subsec.classical.examples.partial\]. Nevertheless, most recent theoretical results on the various CV procedures are not accurate enough to distinguish which one may be the best splitting strategy: This remains a widely open theoretical question. Note finally that the additional variability due to the choice of a sequence of data splits was quantified empirically by [@JoKM00] and theoretically by [@CeRo08] for VFCV. V-fold cross-validation {#sec.practice.VFCV} ----------------------- VFCV is certainly the most popular CV procedure, in particular because of its mild computational cost. Nevertheless, the question of choosing $V$ remains widely open, even if indications can be given towards an appropriate choice. A specific feature of VFCV—as well as exhaustive strategies—is that choosing $V$ uniquely determines the size of the training set $n_t=n(V-1)/V$ and the number of splits $B=V$, hence the computational cost. Contradictory phenomena then occur. On the one hand, the bias of VFCV decreases with $V$ since $n_t=n(1-1/V)$ observations are used in the training set. On the other hand, the variance of VFCV decreases with $V$ for small values of $V$, whereas the LOO ($V=n$) is known to suffer from a high variance in several frameworks such as classification or density estimation. Note however that the variance of VFCV is minimal for $V=n$ in some frameworks like linear regression (see Section \[sec.riskestim.var\]). Furthermore, estimating the variance of VFCV from data is a difficult problem in general, see Section \[sec.riskestim.var.estim\]. When the goal of model selection is estimation, it is often reported in the literature that the optimal $V$ is between $5$ and $10$, because the statistical performance does not increase much for larger values of $V$, and averaging over 5 or 10 splits remains computationally feasible [@Has_Tib_Fri:2001 Section 7.10]. Even if this claim is clearly true for many problems, the conclusion of this survey is that better statistical performance can sometimes be obtained with other values of $V$, for instance depending on the SNR value. When the SNR is large, the asymptotic comparison of CV procedures recalled in Section \[subsubsec.asymptotic.equiv.\] can be trusted: LOO performs (nearly) unbiased risk estimation hence is asymptotically optimal, whereas VFCV with $V=\grandO(1)$ is suboptimal. On the contrary, when the SNR is small, overpenalization can improve the performance. Therefore, VFCV with $V<n$ can yield a smaller risk than LOO thanks to its bias and despite its variance when $V$ is small [see simulation experiments by @Arl:2008a]. Furthermore, other CV procedures like RLT can be interesting alternatives to VFCV, since they allow to choose the bias (through $n_t$) independently from $B$, which mainly governs the variance. Another possible alternative is $V$-fold penalization, which is related to corrected VFCV (see Section \[sec.def.classex.other\]). When the goal of model selection is identification, the main drawback of VFCV is that $n_t \ll n$ is often required for choosing consistently the true model (see Section \[sec.cvconsistent\]), whereas VFCV does not allow $n_t < n/2$. Depending on the frameworks, different (empirical) recommandations for choosing $V$ can be found in the literature. In ordered variable selection, the largest $V$ seems to be the better, $V=10$ providing results close to the optimal ones [@Zha:1993]. On the contrary, [@Die:1998] and [@Alp:1999] recommend $V=2$ for choosing the best learning procedures among two candidates. Future research {#sec.conclu.futur} --------------- Perhaps the most important direction for future research would be to provide, in each specific framework, precise quantitative measures of the variance of CV estimators of the risk, depending on $n_t$, the number of splits, and how the splits are chosen. Up to now, only a few precise results have been obtained in this direction, for some specific CV methods in linear regression or density estimation (see Section \[sec.riskestim.var\]). Proving similar results in other frameworks and for more general CV methods would greatly help to choose a CV method for any given model selection problem. More generally, most theoretical results are not precise enough to make any distinction between the hold-out and CV methods having the same training sample size $n_t$, because they are equivalent at first order. Second order terms do matter for realistic values of $n$, which shows the dramatic need for theory that takes into account the variance of CV when comparing CV methods such as VFCV and RLT with $n_t = n(V-1)/V$ but $B \neq V$.

ArXiv

--- author: - | Lingda Li Ari B. Hayes Stephen A. Hackler Eddy Z. Zhang\ Mario Szegedy Shuaiwen Leon Song$^\star$\ \ \ bibliography: - 'references.bib' title: 'A Graph-based Model for GPU Caching Problems' --- Conclusion {#sec:conclusion} ========== In this paper, we propose a task partition technique to improve data sharing among different GPU threads. We use data-affinity graphs to model data sharing and map task scheduling problem to an edge partition problem. This is the first time the *edge-centric* model is used for GPU cache performance modeling. We propose a novel partition algorithm based on the *edge-centric* model, and our algorithm provides high quality task schedule and yet is low-overhead. We also provide rigorous proof for the analytical bound of our algorithm. Our experiments show that our method can improve data sharing and thus performance significantly for various GPU applications.

ArXiv

--- abstract: 'The renormalization that relates a coupling “a" associated with a distinct renormalization group beta function in a given theory is considered. Dimensional regularization and mass independent renormalization schemes are used in this discussion. It is shown how the renormalization $a^*=a+x_2a^2$ is related to a change in the mass scale $\mu$ that is induced by renormalization. It is argued that the infrared fixed point is to be a determined in a renormalization scheme in which the series expansion for a physical quantity $R$ terminates.' author: - 'F.A. Chishtie[^1]' - 'D.G.C. McKeon [^2]' title: | Renormalization Scheme Dependence\ and the Renormalization Group Beta Function --- 10.0in 9.0in -0.60in PACS No.: 11.10Hi Introduction ============ Elimination of divergences arising in the computation of radiative effects in quantum field theory results in the introduction of an unphysical mass scale parameter $\mu$. Explicit dependence on $\mu$ must be offset by implicit dependence on $\mu$ through the parameters (couplings, masses and field strengths) characterizing the theory, resulting the renormalization group (RG) equations \[1-3\]. If dimensional regularization \[4-6\] is used in conjunction with mass independent renormalization \[7, 8\], it is possible to make additional finite renormalizations that lead to further RG equations. We will examine the effect of finite renormalizations in the QCD calculation of the process $e^+e^- \rightarrow$ hadrons with cross section $R_{e^+e^-}$. The strong fine structure constant is $\alpha = a\pi$ and all quarks are taken to be massless. A variety of renormalization schemes (RS) will be considered; minimal subtraction (MS) \[8\], a scheme due to ’t Hooft in which the RG function $\beta$ associated with $a$ has no contribution beyond second order \[9, 10, 11\], and a scheme in which no radiative corrections beyond second order contribute to the expansion of $R_{e^+e^-}$ in powers of $a$ \[12, 13\]. The latter two schemes only involve RS invariant quantities. In each of these schemes $a = 0$ is an ultraviolet fixed (UV) points; that is, as the centre of mass energy scale $Q$ increases, the couplant $a$ goes to zero (“asymptotic freedom”) \[8, 14-17\]. We will consider the possibility of there also being an infrared (IR) fixed point, in which $a$ goes to some finite value as $Q$ decreases to zero. The scheme in which the expansion of $R_{e^+e^-}$ in powers of $a$ is finite is argued to be the only RS to be of relevance, as in this scheme the behaviour of the infinite series in powers of $a$ that occurs in other schemes is not a problem. Finite Renormalization ====================== The cross section $R_{e^+e^-}$ can be expressed as a power series in the couplant $a$ $$R_{e^+e^-} = 3\left( \sum_i q^2_i\right) (1 + R)$$ where the $n$ loop contribution to $R$ in perturbation theory is given by the term of order $a^{n+1}$ in the expansion $$R = \sum_{n=0}^\infty \sum_{m=0}^n T_{n,m} a^{n+1} L^m \quad \left( T_{0,0} = 1, \; L = b \ln \frac{\mu}{Q} \right).$$ The explicit dependence of $R$ on $\mu$ through $L$ is cancelled by its implicit dependence through $a\left( \ln \frac{\mu}{\Lambda}\right)$ where $$\begin{aligned} \mu \frac{da}{d\mu} &= -ba^2 (1 + ca + c_2 a^2 + \ldots )\\ &\equiv \beta (a). \nonumber\end{aligned}$$ The solution to eq. (3) is taken to be \[18\] $$\ln \left( \frac{\mu}{\Lambda}\right) = \int_0^{a(\ln\frac{\mu}{\Lambda})} \frac{dx}{\beta(x)} + \int_0^\infty \frac{dx}{bx^2(1+cx)}.$$ The general relation between the bare couplant $a_B$ appearing in the initial QCD Lagrangian and the renormalized couplant $a$ when using a mass independent renormalization scheme with dimensional regularization is \[8\] $$a_B = \mu^{-\epsilon} \left[ A_0(a) + \frac{A_1(a)}{\epsilon} + \frac{A_2(a)}{\epsilon^2} + \ldots \right]$$ where $\epsilon = -4 + n$, $n$ being the number of dimensions. Since $a_B$ is independent of $\mu$ $$\mu \frac{da_B}{d\mu} = 0 = \left( \mu \frac{\partial}{\partial \mu} +\beta (a) \frac{\partial}{\partial a}\right) \mu^{-\epsilon} \left[ A_0 + \frac{A_1}{\epsilon} + \frac{A_2}{\epsilon^2} + \ldots \right].$$ Eq. (6) is satisfied at orders $\epsilon$, $\epsilon^0$ provided $$\begin{aligned} \beta (a) &= \frac{1}{A_0^\prime} \left( A_1 - \frac{A_0}{A_0^\prime} A_1^\prime \right) + \frac{A_0}{A_0^\prime} \epsilon \nonumber \\ & = - \left( \frac{A_0}{A_0^\prime}\right)^2 \frac{d}{da} \left( \frac{A_1}{A_0}\right) + \left( \frac{A_0}{A_0^\prime}\right)\epsilon \;.\end{aligned}$$ Terms of order $\epsilon^{-n}$ $(n = 1,2 \ldots)$ in eq. (6) fix $A_2, A_3 \ldots$ in terms of $A_0$ and $A_1$. The function $A_0$ in eq. (5) is not fixed; the MS RS corresponds to selecting $A_0 (a) = a$. Using $\overline{a}$ to denote the MS fine structure constant, then by eq. (7) $$\overline{\beta}(\overline{a}) = \left( \overline{A}_1 (\overline{a}) - \overline{a}\; \overline{A}^\prime_1 (\overline{a}) \right) + \overline{a}\epsilon$$ is the MS $\beta$-function. We now can expand a general function $A_0$ as $$A_0 (a) = a + x_2 a^2 + x_3 a^3 + \ldots \; ;$$ the identification $$\overline{a} = A_0 (a)$$ constitutes a finite renormalization of $\overline{a}$. In general, if we have two different couplings $a$ and $a^*$ such that they are related by the renormalization $$\begin{aligned} a^* &= a + y_2 a^2 + y_3 a^3 + \ldots \\ &\equiv \rho (a) \nonumber\end{aligned}$$ then from the relation $$\mu \frac{da^*}{d\mu} = \mu \frac{da}{d\mu} \frac{da^*}{da} \Longrightarrow \beta^* (a^*) = \frac{d\rho (a)}{da} \beta (a)$$ or by eqs. (3,11) $$\tag{13} -b^*a^{*^{2}} \left( 1 + c^* a^* + c_2^* a^{*^{2}} + \ldots\right) = -ba^2 (1 + ca + c_2a^2 + \ldots) (1 + 2y_2a + 3y_3 a^2 + \ldots )$$ we see that \[19\] $$\tag{14a} b^* = b$$ $$\tag{14b} c^* = c$$ $$\tag{14c} c^*_2 = c_2 - cy_2 + y_3 - y_2^2$$ $$\tag{14d} c_3^* = c_3 - 3cy_2^2 + 2(c_2 - 2c_2^*) y_2 + 2y_4 - 2y_2y_3$$ $$\begin{aligned} \tag{14c} c_4^* & = c_4 - 2y_4y_2 - y_3^2 + c\left(y_4 - y_2^3 - 6y_2y_3\right) + 3y_3 c_2 - 4y_3c_2^* \\ &\hspace{1cm}-6y_2^2 c_2^* + 2y_2c_3 -5y_2 c_3^* + 3y_5 \nonumber\end{aligned}$$ etc . Consequently, $b$ and $c$ are RS invariants while $c_2, c_3 \ldots$ are RS dependent. The RG function $\beta$ is thus not unique; in addition to the one associated with MS in eq. (8), there is the ’t Hooft scheme in which $c_n = 0\; (n \geq 2)$ so that $\beta$ consists of two terms $$\tag{15} \beta (a) = - ba^2 (1 + ca)$$ or the particular $\beta$-function associated with $N = 1$ supersymmetric gauge theory \[20, 21\] $$\tag{16} \beta (a) = \frac{-ba^2}{1-ca}.$$ In principle, the result of eq. (16) could be altered upon making a finite renormalization of a. It was noted in ref. \[18\] that a RS can be characterized by the RS dependent coefficients $c_2, c_3 \ldots$ in eq. (3). From eq. (13) it would appear that these coefficients could be identified with $y_3, y_4 \ldots$ appearing in $\rho (a)$ in eq. (11); in ref. \[18\] it was suggested that $y_2$ should be associated with the scale parameter $\mu$. (In ref. \[19\] this coefficient is taken to be arbitrary and not related to $\mu$; dependence of $a$ on $y_2$ is analyzed as if $y_2$ were a free parameter independent on $\mu$.) Identifying $y_2$ with $\mu$ is reasonable, as by eqs. (3, 12) $$\tag{17} \frac{da}{\beta(a)} = \frac{da^*}{\beta^*(a^*)}$$ and so by eq. (4) $$\tag{18} \int_0^{a\left(\ln \frac{\mu}{\Lambda}\right)} \frac{dx}{\beta(x)} - \int_0^{a^*(\ln \frac{\mu^*}{\Lambda})} \frac{dx}{\beta^*(x)} = \ln \left(\frac{\mu}{\mu^*}\right).$$ However, if $\mu = \mu^*$ this extra arbitrariness disappears and $y_2$ should vanish. To see this connection between $\mu$ and $y_2$ more explicitly, let us consider $$\tag{19} \frac{da}{dc_i} = \beta_i(a) \qquad (i \geq 2);$$ from $$\tag{20} \mu \frac{\partial^2a}{\partial\mu \partial c_i} - \mu \frac{\partial^2a}{ \partial c_i\partial\mu} = 0$$ it follows that \[18\] $$\begin{aligned} \tag{21} \beta_i (a) & = -b\beta (a) \int_0^a dx \frac{x^{i+2}}{\beta^2(x)}\\ & \hspace{1cm}\approx \frac{a^{i+1}}{i-1} \Bigg[ 1 + \left( \frac{(-i+2)c}{i}\right) a + \left( \frac{(i^2 -3i+2)c^2 + (-i^2+3i)c_2}{(i+1)i}\right)a^2 \nonumber \\ & \hspace{.5cm}+\left( \frac{(-i^3 + 3i^2 + 4i)c_3 + (2i^3-6i^2 + 4)cc_2 + (-i^3 + 3i^2 - 2i)c^3}{(i+2)(i+1)i}\right)a^3 + \ldots \Bigg].\nonumber \end{aligned}$$ (RS dependency in theories with renormalized masses is considered in refs. \[22, 23\].) If now $a$ and $a^*$ at the same value of $\mu$ are expanded as $$\tag{22} a^* = a + \lambda_2 (c_i^*, c_i) a^2 + \lambda_3 (c_i^*, c_i) a^3 + \ldots$$ then the equation $$\tag{23} \frac{da^*}{dc_i} = 0 = \left( \frac{\partial}{\partial c_i} + \beta_i \frac{\partial}{\partial a}\right) \left( a + \lambda_2 a^2 + \ldots \right)$$ with the boundary condition $\lambda_n(c_i, c_i) = 0$ can be used to show that \[12, 13 \] $$\begin{aligned} \tag{24} a^* &= a+ \left(c_2^* - c_2\right) a^3 + \frac{1}{2} \left(c_3^* - c_3\right) a^4 + \big[ \frac{1}{6} \left(c_2^{*2} - c_2^2\right)\\ &+ \frac{3}{2} \left(c_2^* - c_2\right)^2 - \frac{c}{6} \left(c_3^* -c_3\right) + \frac{1}{3} \left(c_4^* - c_4\right)\big] a^5 +\ldots \nonumber\end{aligned}$$ Eqs. (11,14) are consistent with eq. (24) only if $y_2 = 0$. If now from eq. (14b) we see that $$\tag{25a} y_3 = c_2^* - c_2 + cy_2 + y_2^2$$ so that from eqs. (14d, 25a) we obtain $$\tag{25b} y_4 = \frac{1}{2} \left[ c_3^* - c_3 + \left(6c_2^* - 4c_2\right) y_2 + 5cy_2^2 + 2y_2^3\right];$$ eqs. (14e, 25a, 25b) now lead to $$\begin{aligned} \tag{25c} y_5 &= \frac{1}{3}\Big\{\left( c_4^* - c_4\right)+ y_2\left(5c_3^* - 2c_3\right) + \left(4c_2^* - 3c_2+6y_2c\right)\left(c_2^* - c_2+cy_2+y_2^2\right)\nonumber \\ &\hspace{1cm}+\left(c_2^* - c_2+cy_2+y_2^2\right)^2+ 6y_2^2c_2^* +y_2^3c+ \left( 2y_2 - c\right) \left[\frac{1}{2}\left( c_3^* - c_3\right) + y_2\left(3 c_2^* -2 c_2\right)+ \frac{5}{2}cy_2^2 +y_2^3 \right] \nonumber \Big\} \end{aligned}$$ etc.\ We now take $a^*$ to have the same $\beta$ function as $a$, but evaluated with mass scale $\mu^*$ rather than $\mu$. If now we expand $a^*$ in terms of $a$ \[12, 13, 24\] so that $$\tag{26} a^* = a + \left(\sigma_{21} \ell \right) a^2 + \left(\sigma_{31}\ell + \sigma_{32} \ell^2 \right) a^3 + \ldots\qquad \left( \ell = b \ln \frac{\mu}{\mu^*}\right)$$ then as $$\tag{27} \mu \frac{da^*}{d\mu} = 0 = \left( \mu \frac{\partial}{\partial \mu} + \beta (a) \frac{\partial}{\partial a} \right) \left[ a + \left(\sigma_{21} \ell \right) a^2 + \left(\sigma_{31} \ell+ \sigma_{32} \ell^2 \right)a^3 + \ldots \right]$$ we find that $$\begin{aligned} \tag{28} a^* &= a + (\ell) a^2 + \left( c\ell + \ell^2\right) a^3 + \left(c_2\ell + \frac{5}{2} c\ell^2 + \ell^3\right) a^4 \\ &\hspace{2cm}+ \left[ c_3 \ell + \left( 3c_2 + \frac{3}{2} c^2\right) \ell^2 + \frac{13}{3} c\ell^3 + \ell^4\right] a^5 \nonumber \\ & \hspace{2cm}+\left[ c_4 \ell + \frac{7}{2}\left( c_3 + c_2 c\right)\ell^2 + \left( 6c_2 + \frac{35}{6} c^2\right) \ell^3 + \frac{77}{12} c\ell^4 + \ell^5\right] a^6 + \ldots \nonumber\end{aligned}$$ This is identical to what is obtained from eqs. (11, 25) in the limit $c_i^* = c_i$ provided $$\tag{29} y_2 = \ell.$$ This is consistent with eq. (18) and with the observation in ref. \[18\] that $y_2$ is to be identified with the mass scale parameter $\mu$. It is also posible to use eqs. (24, 28) together to expand $a^*(\mu^*, c{_i}^*)$ in powers of $a(\mu, c{_i})$; the result is the same as eq. (11) with eqs. (25, 29). We now make the expansion $$\tag{30} \overline{A}_1(\overline{a}) = \overline{\lambda}_2 \overline{a}^2 + \overline{\lambda}_3 \overline{a}^3 + \overline{\lambda}_4 \overline{a}^4 + \ldots$$ where $\overline{A}_1(\overline{a})$ is associated with the MS RS. We find from eq. (8) that in this scheme $$\tag{31a,b,c} \overline{b} = \overline{\lambda}_2, \qquad \overline{c} = 2 \overline{\lambda}_3/ \overline{\lambda}_2, \qquad \overline{c}_n = (n+1) \overline{\lambda}_{n+2}/ \overline{\lambda}_2.$$ If under a finite renormalization given by eq. (9) we end up with a coupling $a$ and a RG function $\beta(a)$ given by eq. (7), then the expansions of eqs. (3,9) show that if $$\tag{32} A_1(a) = \lambda_2 a^2 + \lambda_3 a^3 + \ldots$$ then $$\tag{33a} \lambda_2 = b$$ $$\tag{33b} \lambda_3 = \frac{1}{2} b(c + 4x_2)$$ $$\tag{33c} \lambda_4 = \frac{b}{3} \left( 7 x_3 + 2 x_2^2 + \frac{7}{2} cx_2 +c_2\right)$$ $$\tag{33d} \lambda_5 = \frac{b}{4} \left( 10 x_4 + \frac{22}{3} x_2x_3 - \frac{4}{3} x_2^3 + \frac{5}{3} cx_2^2 + 6cx_3+ \frac{10}{3} x_3c_2 +c_3\right)$$ etc. Upon identifying $y_i$ with $x_i$, eq. (25) can be used to convert eqs. (33c, 33d) to $$\tag{34a} \lambda_4 = b \left( \frac{7}{3}\overline{c}_2 + \frac{7}{2} x_2 c + 3x_2^2 -2c_2\right)$$ $$\tag{34b} \lambda_5 = \frac{b}{4} \left[\frac{112}{3} \overline{c}_2x_2-24x_2c_2+ 40cx_2^2 + 16 x_2^3+ 6 c^2x_2+ 6c(\overline{c}_2-c_2) + 5 \overline{c}_3 - 4c_3\right]$$ etc.\ Together, eqs. (9,14, 31, 33) can be used to show that $$\tag{35a} \lambda_2 a^2 + \lambda_3 a^3 + \lambda_4 a^4 + \ldots = \overline{\lambda}_2\overline{a}^2 + \overline{\lambda}_3 \overline{a}^3 + \overline{\lambda}_4 \overline{a}^4 + \ldots$$ so that $$\tag{35b} A_1 (a) = \overline{A}_1 (\overline{a}).$$ We thus see that with $\beta(a)$ and $\overline{\beta}(\overline{a})$ being given by eqs. (7) and (8) respectively, eqs. (10) and (14) are consistent. We now can examine the presence of an IR fixed point in light of finite renormalizations. Infrared Fixed Points ===================== Eq. (12) implies that if $\beta(a) = 0$, then $\beta^* (a^*) = 0$ as well, if $a^* = \rho (a)$. In addition, from eq. (12) it also follows that $$\tag{36} \frac{d\beta^*(a^*)}{da^*} = \left( \frac{d^2\rho(a)}{da^2} \beta(a) + \frac{d\rho(a)}{da}\frac{d\beta(a)}{da} \right) \Big/ \frac{d\rho(a)}{da}$$ and so if $\beta(a) = 0$, then \[25, 26\] $$\tag{37} \frac{d\beta^*(a^*)}{da^*} = \frac{d\beta(a)}{da} .$$ These arguments rely on $\rho(a)$ being a well defined function when $\beta(a) = 0$; for ill behaved functions $\rho(a)$, it may turn out that $\beta^*(a^*)$ is non zero or that eq. (37) is not satisfied \[27, 28\]. In refs. \[12, 13\] the problem of RS dependence was considered in the context of the cross section $R_{e^+e^-}$ of eq. (1). It was shown that by applying the RG equation $$\tag{38} \left( \mu \frac{\partial}{\partial\mu} + \beta (a) \frac{\partial}{\partial a}\right) R = 0$$ it is possible sum the logarithm in eq. (2) so that the explicit dependence of $R$ on $\mu$ (through $L$) and its implicit dependence (through $a(\mu)$) cancel, leaving us with $$\tag{39} R = \sum_{n=0}^\infty T_n a^n \left( \ln \frac{Q}{\Lambda}\right) \qquad (T_n \equiv T_{n,0})$$ where $Q$ is centre of mass energy and $\Lambda$ is a scale parameter introduced in eq. (4). The behaviour of the sum in eq. (39) can be affected by three things. There is first the behaviour of $a \left( \ln \frac{Q}{\Lambda}\right)$ as $Q$ itself evolves, secondly the behaviour of $T_n$ as $n$ becomes large, and thirdly the convergence behaviour of the infinite sum itself. But $R$ is invariant under the finite renormalization of eq. (11); we will now examine how $R$ is affected by two particular renormalization schemes. We will then consider their implications for the IR fixed point. Of course this IR fixed point cannot incorporate non-perturbative effects in QCD; our discussion is entirely in the context of the perturbative expansion of eq. (2). Non-perturbative effects in the low energy regime would necessarily involve the emergence of low energy Goldstone Bosons (pions). Changes in RS lead to compensating changes in $T_n$ and $a$ so that $$\tag{40} \frac{d}{dc_i} R = 0 = \left(\frac{\partial}{\partial c_i} + \beta_i (a) \frac{\partial}{\partial a}\right) \sum_{n=0}^\infty T_n a^{n+1} .$$ This leads \[12, 13\] to a set of nested equations for $T_n$ whose solutions are $$\begin{aligned} \tag{41a-f} T_0 &= \tau_0 =1, \quad T_1 = \tau_1, \quad T_2 = -c_2 + \tau_2, \quad T_3 = -2c_2\tau_1 - \frac{1}{2} c_3 + \tau_3\\ T_4 &= -\frac{1}{3} c_4 - \frac{c_3}{2} \left( - \frac{1}{3} c + 2\tau_1\right) + \frac{4}{3} c_2^2 - 3c_2 \tau_2 + \tau_4\nonumber\\ T_5 &= \left[ \frac{1}{3} cc_2^2 + \frac{3}{2} c_2c_3 + \frac{11}{3} c_2^2 \tau_1 - 4c_2\tau_3\right] - \frac{1}{2} \left[ \frac{1}{6} c^2c_3 - \frac{2}{3} c_3c \tau_1 + 3 c_3 \tau_2\right] \nonumber \\ & \hspace{1cm} - \frac{1}{3} \left[ - \frac{1}{2} c_4 c + \frac{1}{2} c_4\tau_1\right] - \frac{1}{4} c_5 + \tau_5\nonumber\end{aligned}$$ etc.\ In eq. (41), the $\tau_n$ are constants of integration and hence are RS invariants. They can be determined by evaluating the $T_n$ and $b$, $c$, $c_n$ in one particular RS (such as MS) upon explicitly computing the relevant Feynman diagrams. Two particular RS are of special interest. In the first scheme, the $c_i$ are selected so that $T_n = 0 (n \geq 2)$. From eq. (41) this means that $$\begin{aligned} \tag{42a-c} c_2 &= \tau_2 \\ c_3 &= 2(-2c_2\tau_1 + \tau_3)\nonumber \\ &= 2 (-2 \tau_1\tau_2 + \tau_3)\nonumber \\ c_4 &= - \frac{3}{2}c_3 \left( - \frac{1}{3} c + 2\tau_1\right) + 4c_2^2 - 9c_2\tau_2 + 3\tau_4\nonumber \\ &= c\left( \tau_3 - 2\tau_1\tau_2\right) + 12 \tau_1^2 \tau_2 - 6\tau_1 \tau_3 - 5 \tau_2^2 + 3\tau_4 \nonumber \end{aligned}$$ etc.\ In the second scheme due to ’t Hooft we set $c_n = 0 (n \geq 2)$ \[9-11\], so that $$\tag{43} T_n = \tau_n.$$ In the first instance, the series in eq. (39) collapses down to two terms $$\tag{44a} R_{(1)} = a_{(1)} + \tau_1 a_{(1)}^2 \left(\ln \frac{Q}{\Lambda}\right)$$ while in the second case we have the infinite series $$\tag{44b} R_{(2)} = \sum_{n=0}^\infty \tau_n a_{(2)}^{n+1} \left(\ln \frac{Q}{\Lambda}\right).$$ In the first case $a_{(1)}$ “runs” according to eq. (4) with the $c_i$ of eq. (3) being given by eq. (42). In the second case, $a_{(2)}$ runs according to the ’t Hooft $\beta$-function of eq. (14). In eq. (44a), there is no possible problem associated with there being a divergent series for $R_{(1)}$ or of having diverging behaviour for the coefficients of $a_{(1)}^n$; one need only discuss how $a_{(1)}\left(\ln \frac{Q}{\Lambda}\right)$ behaves as $Q$ varies. On the other hand, in the infinite series for $a_{(2)}$, the behaviour of $a_{(2)}\left(\ln \frac{Q}{\Lambda}\right)$ is completely known as upon integrating eq. (4) we obtain the Lambert W function \[29-32\]. With $\beta_{(2)}(a_{(2)})$ given by eq. (14) it is evident that $a_{(2)}$ has an UV fixed point at $a_{(2)} = 0$ (if $b > 0$) and an IR fixed point at $a_{(2)} = \frac{-1}{c}$ (if $ c < 0$). As one \[14-17\] and two \[33, 34\] loop calculations show that for an SU(N) gauge theory with $N_f$ flavours of quarks, $$\tag{45a,b} b = \frac{33 - 2N_f}{6}, \quad c = \frac{153 - 19N_f}{2(33-2N_f)}$$ which means that in order to have asymtotic freedom $(b < 0)$ and a positive IR fixed point $(c < 0)$ we must have $$\tag{46} 8 \leq N_f \leq 16.$$ However, even if $N_f$ satisfies the restrictions of eq. (46), the series of eq. (44b) could be badly behaved, or alternatively if eq. (46) is not satisfied, this series could be well behaved. We are thus led to consider the finite sum of eq. (44a). The equality $$\tag{47} R_{(1)} = R_{(2)}$$ is consistent with eq. (24). However, since $R_{(2)}$ involves an infinite series whose convergence is not known, the IR behaviour of $a_{(2)}$ is not necessarily a reflection on the IR behaviour of $a_{(1)}$ since $R_{(1)}$ only involves a finite series in which convergence is no longer an issue. We contend that in analyzing the IR behaviour of $R_{e^+e^-}$ it is the behaviour of $a_{(1)}$ that is of relevance. We will now consider the IR behaviour of $a_{(1)}$ up to four loop order and thereby find out how $R_{(1)}$ behaves as ${Q\to 0}$. Explicit calculation shows that with $N_f = 3$ active flavours of quarks \[13\], $$\tag{48a} \hspace{-1.2cm}\tau_1 = 1.6401$$ $$\tag{48b} \tau_2 = -5.812885185$$ $$\tag{48c} \tau_3 = -81.73499303$$ etc.\ These values follow from the four-loop calculations of $T_n$ and $\beta(a)$ done in the $\overline{MS}$ scheme \[35, 36\]. With these values of $\tau_1$, $\tau_2$, $\tau_3$ it follows from eq. (42) that in the RS associated with $a_{(1)}$, $$\tag{49a} \hspace{-.5cm}c_2 = -5.812885185$$ $$\tag{49b} c_3 = -125.3351844092$$ etc.\ Since with $N_f = 3$ \[8, 14-17\], $$\tag{50a} b = 9/4$$ and \[33, 34\] $$\tag{50b} c = 16/9$$ we see that the function $\beta_{(1)}(a_{(1)})$ is given by $$\tag{51} \beta_{(1)}\left(a_{(1)}\right) = -2.25 a_{(1)}^2 \left( 1 + 1.77778 a_{(1)} - 5.812885185 a_{(1)}^2 - 125.3351844092 a_{(1)}^3 + \ldots \right).$$ We now consider only the four-loop $\left(\mathcal{O}\left(a_{(1)}^5\right)\right)$ contribution to $\beta_{(1)}$ in eq. (51). The function $$\tag{52} f(x) = 1 + 1.77778x - 5.812885185x^2 - 125.3351844092x^3$$ has only one positive zero, and that is found by Newton’s Method to be at $$\tag{53} x \approx .20743211594.$$ Consequently $\beta_{(1)}$ has an IR fixed point at $$\tag{54} a_{(1)} \approx .20743$$ if we use only up to the four-loop contributions to $\beta_{(1)}$. (In MS, $\overline{\beta}(\overline{a})$ has been computed to five-loop order \[37\], but to get $\beta_{(1)}(a_{(1)})$ to this order we would also require $R_{e^+e^-}$ to five-loop order; this is as yet unknown.) From eqs. (1), (44a) and (54) it follows that $$\tag{55} R_{e^+e^-}\Big/\left(3 \sum_i q_i^2\right) = 1 + (.20743) + (1.6401)(.20743)^2 \approx 1.278.$$ This is the limit of $R_{e^+e^-}$ as $Q\rightarrow 0$ when we use $R_{(1)}$ in eq. (44a) and keep contributions only up to four-loop order. This is consistent with $R_{e^+e^-}$ as presented in ref. \[13\]. Both $c_2$ and $c_3$ when using $\beta_{(1)}$ are negative (see eq. (49)) and so it is not unreasonable to anticipate that $c_4$ is also negative. In fig. (1) we have plotted possible values of $c_4$ versus zero of $\beta_{(1)}$ to five loop order (using eqs. (49, 50) for $b$, $c$, $c_2$, $c_3$) when there are three active flavours of quarks. We see that for $-2000 \leq c_4 \leq 300$, $\beta_{(1)}$ to five loop order has zeros lying between $.1$ and $.3$ which is quite reasonable. We note that by eq. (41f), $\tau_5$ and $T_5$ both vary linearly with $c_4$. If we use eq. (24) to relate $a_{(1)}$ to $a_{(2)}$ we see from eq. (42) that $$\tag{56} a_{(2)} = a_{(1)} - \tau_2 a_{(1)}^3 + \left(2\tau_1\tau_2 - \tau_3\right) a_{(1)}^4 + \ldots \; .$$ Using eq. (48), we find that the value of $a_{(2)}$ corresponding to $a_{(1)} = .20743$ is $$\tag{57} a_{(2)} = .37533\; .$$ This clearly is not an IR fixed point for the function $\beta_{(2)}(a_{(2)})$; with $\beta_{(2)}$ given by eq. (15), the only value of $a_{(2)}$ for which $\beta_{(2)}$ vanishes is given by $a_{(2)} = - \frac{1}{c}$ which for $N_f = 3$ is, by eq. (50b), negative – an unacceptable value. However, by eq. (47), we see that as $Q \rightarrow 0$, $1 + R_{(1)} \rightarrow 1.278$, then $1 + R_{(2)}$ must also approach this value even though $a_{(2)}$ approaches a value given by eq. (57) which is not an IR fixed point of $\beta_{(2)}$. Actually $a_{(2)}$ approaches this value only if we keep the three terms of eq. (24) given in eq. (56); if we were to simply consider integrating eq. (2) to find $a_{(2)}\left(\ln \frac{Q}{\Lambda}\right)$ exactly, it is evident that as $Q \rightarrow 0$, $a_{(2)} \rightarrow \infty$ if $b > 0, c > 0$. This indicates that the full series of eq. (56) diverges as $Q \rightarrow 0$ (when $a_{(1)}$ approaches an exact IR fixed point and $a_{(2)}$ diverges). We also note that with $N_f = 3$, the $\beta$ function in the $\overline{MS}$ scheme at four and five loop order has no positive roots, indicating that in the $\overline{MS}$ scheme, there is no IR fixed point, at least to this order in perturbation theory. This is discussed in ref. \[39\]. An interesting third RS is one in which $c_2$ is allowed to vary while $c_i (i > 2)$ vanishes. By eq. (41), in this scheme $R$ is given by $$\begin{aligned} R_{(3)} = a_{(3)} &+ \tau_1 a_{(3)}^2 + (-c_2 + \tau_2) a_{(3)}^3\nonumber \\ &\qquad + (-2 c_2 \tau_1 + \tau_3) a_{(3)}^4 + \left(\frac{4}{3} c_2^2 -3 c_2 \tau_2 + \tau_4\right)a_{(3)}^5 + \ldots \nonumber\end{aligned}$$ while $$\beta_3(a_{(3)}) = - b a_{(3)}^2 \left( 1 + c a_{(3)} + c_2 a_{(3)}^2\right).\nonumber$$ It appears to be possible to have a value of $c_2$ and an acceptable value of $a_{(3)}$ such that $\beta_{(3)} = 0$ and at order $a_{(3)}^5$, $R_{(3)} = 1.278$; we find that $c_2 = -13.106$ and $a_{(3)} = .3523$. We conclude that the perturbative expression for $1 + R_{(2)}$ appearing in eq. (44b) is an infinite series whose behaviour as $Q \rightarrow 0$ is such that it approaches a fixed value even though $a_{(2)}\left(\ln \frac{Q}{\Lambda}\right)$ is not an IR fixed point in this limit. This fixed value for $1 + R_{(2)}$ is trivial to compute using eq. (44a) provided we can find an IR fixed point for $\beta_{(1)}$. Upon only employing the first four terms of the expansion of eq. (51) for $\beta_{(1)}$, this IR fixed point is given by eq. (54). Discussion ========== We have examined various aspects of RS dependency in perturbative QCD. First of all, we have considered within the context of mass independent renormalization schemes how the parameters $y_i$ appearing in the finite renormalization of eq. (11) are related to the parameters $\mu$ and $c_i$ of eq. (3); we led to eqs. (25) (for $y_n$, $n > 2$) and eq. (29) (for $y_2$). Treating $y_2$ as an independent parameter as in ref. \[19\] should involve the mass scale parameter $\mu$. Secondly, the behaviour of QCD predictions for physical processes in the IR limit are considered. After using RG summation to eliminate $\mu$ dependence, the RS scheme in which perturbative calculations give rise to just a finite number of contributions in powers of the coupling is taken to be the scheme that can be used to consider this IR limit. This is because in this scheme one is not confronted with an infinite series in powers of the coupling whose behaviour is unknown. (Indeed, it has been argued that this series contains “renormalons” \[9\] and is at best asymptotic.) The IR behaviour when using this finite series is controlled by the IR behaviour of $a_{(1)}$ in eq. (51) (for $N_f = 3$) if we only use results up to four-loop order. We note that since the values of $\tau_i$ are determined by eq. (41) when $T_i$ and $c_i$ are computed in a particular renormalization scheme, it follows that the $\tau_i$ must be computed separately for distinct processes. Thus, by eq. (42), the IR fixed point for $\beta_{(1)}$ is not the same in different processes. We also would like to outline how the “Principle of maximum conformality” \[38, 40\] (PMC) or its improvement \[41\] is related to the renormalization group summation employed in this paper. Let us consider the sum of eq. (39) when using the RS of eq. (43) so that $R$ is given by eq. (44b). It is now possible to expand $a_{(2)} \left(\ln \frac{Q}{\Lambda}\right)$ in terms of $a \left(\ln \frac{Q}{\Lambda}\right)$ where $a$ is the couplant in some other RS by using eq. (24) $$\tag{58} a_{(2)}\left(\ln \frac{Q}{\Lambda}\right) = a\left(\ln \frac{Q}{\Lambda}\right) + (-c_2) a^3\left(\ln \frac{Q}{\Lambda}\right) + \frac{1}{2}(-c_3) a^4\left(\ln \frac{Q}{\Lambda}\right)$$ $$\hspace{2cm}+ \left[ \frac{1}{6}(-c_2^2) + \frac{3}{2}(-c_2)^2 - \frac{c}{6}(-c_3) + \frac{1}{3}(-c_4)\right] a^5\left(\ln \frac{Q}{\Lambda}\right) + \ldots \; .\nonumber$$ Now in turn, $a\left(\ln \frac{Q}{\Lambda}\right)$ can be expanded in terms of $a\left(\ln \frac{\mu_n}{\Lambda}\right)$ using eq. (28) when $a_{(2)}^{n}\left(\ln \frac{Q}{\Lambda}\right)$ appears in eq. (44b). This results in (with $a_n = a\left(\ln \frac{\mu_n}{\Lambda}\right)$ and $\ell_n = b \ln \frac{\mu_n}{Q}$) $$\begin{aligned} \tag{59} R_{(2)} & = \tau_0 \left[ a_1 + \ell_1 a_1^2 + \left( c\ell_1 + \ell_1^2\right)a_1^3 + \left( c_2 \ell_1 + \frac{5}{2} c \ell_1^2 + \ell_1^3 \right) a_1^4 + \ldots \right] \nonumber \\ & +\tau_1 \left[ a_2 + \ell_2 a_2^2 + \left( c\ell_2 + \ell_2^2\right)a_2^3 + \left( c_2 \ell_2 + \frac{5}{2} c \ell_2^2 + \ell_2^3 \right) a_2^4 + \ldots \right]^2\nonumber \\ & + \tau_2 \left[ a_3 + \ell_3 a_3^2 + \left( c\ell_3 + \ell_3^2\right)a_3^3 + \left( c_2 \ell_3 + \frac{5}{2} c \ell_3^2 + \ell_3^3 \right) a_3^4 + \ldots \right]^3\nonumber \\ & + \ldots . \nonumber\end{aligned}$$ Upon grouping terms in eq. (59) in ascending powers of $a$ we obtain $$\begin{aligned} \tag{60} R_{(2)} & = \tau_0 a_1 + \left[ (\tau_0\ell_1) a_1^2 + \tau_1 a_2^2\right]\nonumber \\ & + \left\lbrace \left[ \tau_0 \left(c_1 + \ell_1^2\right) a_1^3 + \tau_1 (2\ell_2) a_2^3 \right] + \tau_2 a_3^3 \right\rbrace \nonumber \\ & + \Big\{ \Big[ \tau_0 \left(c_2 \ell_1 + \frac{5}{2} c \ell_1^2 + \ell_1^3 \right) a_1^4 + \tau_1 \left( \ell_2^2 + 2 (c_2\ell_2 + \ell_2^2)\right) a_2^4 \nonumber \\ &+ \tau_2 (3\ell_3) a_3^4 \Big] + \tau_3 a_{4}^4 \Big\} + \ldots .\nonumber\end{aligned}$$ It is now possible to select $\mu_1, \mu_2, \mu_3 \ldots$ so that eq. (60) reduces to $$\tag{61} R_{(2)} = \sum_{n=0}^\infty \tau_n a_{n+1}^{n+1};$$ this entails solving $$\tag{62a} \tau_0 \ell_1 = 0$$ $$\tag{62b} \tau_0 (c_1 + \ell_1^2) a_1^3 + \tau_1 (2\ell_2) a_2^3 = 0$$ $$\tag{62c} \tau_0 \left(c_2 + \ell_1 + \frac{5}{2}c \ell_1^2 + \ell_1^3\right) a_1^4 + \tau_1 \left(\ell^2_2 + 2 ( c_2 \ell_2 + \ell_2^2)\right) a_2^4 + \tau_2 (3\ell_3) a_3^4 = 0$$ etc.\ From eq. (62a), we see that $\mu_1 = Q$; solving for $\mu_2,\mu_3 \ldots$ etc. becomes progressively more difficult. The resulting expression for $R_{(2)}$ in eq. (61) is now expressed in terms of scheme independent quantities $\tau_n$ and all dependence on $\mu$ has disappeared; $\mu_1, \mu_2 \ldots$ contain all of the explicit dependence on $c_2, c_3 \ldots$. As a result, eq. (61) is equivalent to what is obtained using the approach to PMC used in ref. \[38\]. We wish to note that the low energy behaviour of $a(Q)$ has also been examined using light cone holography \[40\]. Our considerations have been limited to examining RS ambiguities that arise in using conventional perturbative evaluation of physical quantities in QCD. In the future we hope to extend these considerations of perturbative expansions in quantum field theory to processes involving non-trivial masses and/or multiple couplings \[23\]. Acknowledgements {#acknowledgements .unnumbered} ================ A correspondence with A. Kataev and A. Deur is gratefully acknowledged and R. Macleod made a helpful comment. [99]{} E.C.G. Stueckelberg and A. Peterman, *Helv. Phys. Acta* **26**, 499 (1953). M. Gell-Mann and F.E. Low, *Phys. Rev.* **95**, 1300 (1954). N.N. Bogoliubov and D.V. Shirkov, *Nuovo Cimento* **3**, 845 (1956). C.G. Bollini and J.J. Giambiagi, *Phys. Lett.* **40B**, 566 (1972). J.F. Ashmore, *Nuovo Cimento Lett.* **4**, 289 (1972). G. ’t Hooft and M. Veltman, *Nucl. Phys.* **B44**, 189 (1972). S. Weinberg, *Phys. Rev.* **D8**, 3497 (1973). G. ’t Hooft, *Nucl. Phys.* **B61**, 455 (1973). G. ’t Hooft, “The Whys of Subnuclear Physics” Erice 1977, edited by A. Zichichi (Plenum, New York, 1979). T. Banks and A. Zaks, *Nucl. Phys.* **B196**,189 (1982). N.N. Khuri and O.A. McBryan, *Phys. Rev.* **D20**,881 (1979). D.G.C. McKeon, *Phys. Rev.* **D92**, 045031 (2015). F.A. Chishtie, D.G.C. McKeon and T.N. Sherry, *Phys. Rev.* **D94**, 054031 (2016). H.D. Politzer, *Phys. Rev. Lett.* **30**, 1346 (1973). D.J. Gross and F. Wilczek, *Phys. Rev. Lett.* **30**, 1343 (1973). V.S. Vanyashin and M.V. Terentev, *JETP* **21**, 375 (1965). I.B. Khriplovich, *Sov. J. Nucl. Phys.* **10**, 235 (1970). P.M. Stevenson, *Phys. Rev.* **D23**, 2916 (1981). R. Shrock, *Phys. Rev.* **D90**, 045011 (2014); G. Choi and R. Shrock, *Phys. Rev.* **D90**, 125029 (2014). D.R.T. Jones, *Phys. Lett.* **123B**, 45 (1983). V. Novikov, V. Shifman, A. Vainshtein and V. Zakharov, *Nucl. Phys.* **B229**, 381 (1983). D.G.C. McKeon, *Can. J. Phys.* **61**, 564 (1983); **59**, 1327 (1981). F.A. Chishtie, D.G.C. McKeon, T.N. Sherry and C. Zhao (in preparation). X.G. Wu, S.J. Brodsky and M. Mojaza, *Prog. Part. Nucl. Phys.* **72**, 44 (2013); M.R. Ahmady, V. Elias, D.G.C. McKeon, A. Squires and T.G. Steele, *Nucl. Phys.* **B655**, 221 (2003); J.M. Chung and B.K. Chung, *Phys. Rev.* **D60**, 105001 (1999).. D.J. Gross in “Methods in Field Theory” ed. R. Balian and J. Zinn-Justin (North-Holland, Amsterdam, 1976). A. Peterman, *Phys. Rep.* **53C**, 157 (1979). J. Chyla, *Phys. Rev.* **D38**, 3845 (1988). P.M. Stevenson, hep-ph 1606.06951. B.A. Magradze, hep-th 9808247. E. Gardi, G. Grunberg and A. Karliner, *JHEP* **07**, 007 (1998). D.V. Shirkov, *Theor. Math. Phys.* **119**, 438 (1999). R.M. Corless, G.H. Gonnet, D.E.G. Hare, D.J. Jeffrey and D.E. Knuth, *Adv. Comp. Math.* **5**, 329 (1996). D.R. T. Jones, *Nucl. Phys.* **B75**, 531 (1974). W.E. Caswell, *Phys. Rev. Lett.* **33**, 244 (1974). P.A. Baikov, K.G. Chetyrkin and J.H. Kuhn, *Phys. Rev. Lett.* **101**, 012002 (2008). T. van Ritbergen, J. Vermaseren and S.A. Larin, *Phys. Lett.* **B400**, 379 (1997). P.A. Baikov, K.G. Chetyrkin and J.H. Kuhn,hep-ph 1606-08659. M. Mojaza, S.J. Brodsky and X.G. Wu, *Phys. Rev. Lett.* **110**, 192001 (2013). C. Pica and F. Sannino, *Phys. Rev.* **D83**, 035013 (2011). A. Deur, S.J. Brodsky and G.F. de Teramond, *Prog. Part. Nucl. Phys.* **90**, 1 (2016). A.L. Kataev and S.V. Mikhailov,*Phys. Rev.* **D91**, 014007 (2015). ![Infrared Fixed Points (IRFPs) of the scheme-1 $\beta_{(1)}$-function at five loop order versus $c_4$ []{data-label="Fig. 1"}](Fig1.eps) [^1]: [email protected] [^2]: [email protected]

ArXiv

--- abstract: 'En utilisant l’approche de $\mathbb R$-filtration en géométrie d’Arakelov, on établit des majorations explicites des fonctions de Hilbert-Samuel géométrique et arithmétique pour les fibrés inversibles sur une variété projective et les fibrés inversibles hermitiens sur une variété projective arithmétique.' address: 'Université Grenoble Alpes, Institut Fourier (UMR 5582), F-38402 Grnoble, France' author: - Huayi Chen bibliography: - 'chen.bib' title: 'Majorations explicites des fonctions de Hilbert-Samuel géométrique et arithmétique' --- By using the $\mathbb R$-filtration approach of Arakelov geometry, one establishes explicit upper bounds for geometric and arithmetic Hilbert-Samuel function for line bundles on projective varieties and hermitian line bundles on arithmetic projective varieties. Introduction ============ Soient $X$ un schéma projectif et intègre défini sur un corps $k$, et $L$ un $\mathcal O_X$-module inversible. Rappelons que la fonction de Hilbert-Samuel de $L$ est définie comme l’application de $\mathbb N$ vers $\mathbb N$ qui envoie $n\in\mathbb N$ en $h^0(L^{\otimes n})$, le rang de l’espace des sections globales $H^0(X,L^{\otimes n})$ sur $k$. Le théorème de Riemann-Roch et le théorème d’annihilation de Serre montrent que, si le faisceau $L$ est ample, alors la relation suivante est vérifiée : $$\label{Equ:RRH}h^0(L^{\otimes n})=c_1(L)^d\frac{n^d}{d!}+o(n^d),$$ où $d$ est la dimension de Krull du schéma $X$, et $c_1(L)^d$ est le nombre d’auto-intersection de $L$. Le théorème d’approximation de Fujita [@Fujita94; @Takagi07] montre que la relation est vérifiée en générale, quitte à remplacer le nombre d’intersection $c_1(L)^d$ par le volume de $L$, défini comme $$\mathrm{vol}(L):=\limsup_{n\rightarrow+\infty}\frac{h^0(L^{\otimes n})}{n^d/d!}.$$ En d’autres termes, la limite supérieure définissant la fonction volume est en fait une limite. Il est cependant plus délicat d’étudier l’estimation explicite de la fonction de Hilbert-Samuel qui sont valables pour tout entier $n\geqslant 1$. Dans la littérature, différentes approches ont été proposées, souvent sous des conditions de positivité forte pour le faisceau inversible $L$. On peut consulter par exemple les travaux de Nesterenko [@Nesterenko85], Chardin [@Chardin89] et Sombra [@Sombra97], où on suppose que le faisceau $L$ est très ample et fixe un plongement de la variété polarisée $(X,L)$ dans un espace projectif. Une majoration explicite de $h^0(L^{\otimes n})$ est ensuite obtenue par récurrence sur la dimension $d$ du schéma $X$, en utilisant l’intersection avec des hyperplanes de l’espace projectif. Cette approche a une nature algébrique car le choix d’un plongement de la variété polarisée correspond à un système de générateurs homogènes de l’algèbre graduée des sections globales des puissances tensorielles de $L$. On renvoie les lecteurs vers l’article de Bertrand dans [@LNM1752 chapitre 9] pour une présentation détaillée de cette méthode. L’approche de Kollár et Matsusaka [@Kollar_Matsusaka] repose sur la comparaison entre la fonction $h^0$ et la caractéristique d’Euler-Poincaré (somme alternée des rangs des espaces de cohomologie). Cette méthode est relativement plus proche de l’esprit du théorème de Riemann-Roch. On suppose que le schéma $X$ est régulier et que le faisceau $L$ est semi-ample, i.e., une puissance tensorielle de $L$ est sans lieu de base. Un encadrement effectif mais assez compliqué dans le cas de dimension supérieure a été obtenu pour $h^0(L^{\otimes n})$. L’encadrement ne dépend que du nombre d’auto-intersection $c_1(L)^d$ et le nombre d’intersection $c_1(L)^{d-1}c_1(\omega_X)$, où $\omega_X$ est le fibré canonique de $X$. La fonction de Hilbert-Samuel peut être généralisée dans le cadre de la géométrie d’Arakelov, où on considère une variété arithmétique projective $\mathscr X$ (i.e., un schéma intègre, projectif et plat sur $\operatorname{Spec}\mathbb Z$) et un faisceau inversible $\mathscr L$ sur $\mathscr X$ muni d’une métrique continue sur $\mathscr L(\mathbb C)$, invariante par la conjugaison complexe (ces données sont appelées un *faisceau inversible hermitien* sur $\mathscr X$ et notées comme $\overline{\mathscr L}$). Similairement à la situation géométrique, on définit la fonction de Hilbert-Samuel arithmétique de $\overline{\mathscr L}$ comme la fonction de $\mathbb N$ vers $[0,+\infty[$ qui envoie $n$ en $\widehat{h}^0(\overline{\mathscr L}{}^{\otimes n})$, le logarithme du nombre des sections globales de $\mathscr L^{\otimes n}$ dont la norme sup est majorée par $1$. La suite $$\frac{\widehat{h}^0(\overline{\mathscr L}{}^{\otimes n})}{n^{d+1}/(d+1)!},\quad n\geqslant 1$$ possède une limite[^1] que l’on note comme $\widehat{\operatorname{vol}}(\overline{\mathscr L})$, et on l’appelle le *volume arithmétique* de $\overline{\mathscr L}$. Ici $d$ désigne la dimension relative de $\mathscr X\rightarrow\operatorname{Spec}\mathbb Z$ (et donc la dimension de $\mathscr X$ est $d+1$). On peut aussi exprimer ce résultat comme une formule asymptotique $$\widehat{h}^0(\overline{\mathscr L}{}^{\otimes n})=\widehat{\operatorname{vol}}(\overline{\mathscr L})\frac{n^{d+1}}{(d+1)!}+o(n^{d+1}).$$ Du point de vue birationnel, il est naturel de se demander si on peut obtenir une estimation de $h^0(L^{\otimes n})$ en fonction des invariants birationnels de $L$ (comme par exemple le volume de $L$). La même question se pose aussi pour la fonction $\widehat{h}^0$ dans le cadre arithmétique. Cependant, les outils que l’on dispose, comme par exemple le théorème d’approximation de Fujita (géométrique ou arithmétique) sous forme actuelle, ne permet pas de traiter ce problème de façon effective. Il est encore peu probable que les méthodes que l’on a résumé plus haut se généralisent dans la situation birationnelle ou s’adaptent facilement dans le cadre de la géométrie arithmétique. Les résultats arithmétiques sont rares dans la littérature et portent notamment sur les cas où la dimension de la variété arithmétique est petite. On peut consulter par exemple les résultats de Blichfeld [@Blichfeld14], Henk [@Henk02] et Gaudron [@Gaudron09] pour le cas d’une courbe arithmétique (ces résultats sont basés sur la géométrie des nombres) et le travail de Yuan et Zhang [@Yuan_Zhang13] pour le cas d’une surface arithmétique. Le but principal de cet article est d’établir une majoration effective pour la fonction de Hilbert-Samuel arithmétique en dimension quelconque. Dans le cas de surface arithmétique, ce résultat est asymptotiquement plus précis que la majoration obtenue dans [@Yuan_Zhang13] (cf. la remarque \[Rem:comparaisonYuanZhang\] [*infra.*]{}). \[Thm:majorationgeom\] Soit $\mathscr X$ une variété arithmétique projective définie sur l’anneau des entiers algébriques d’un corps de nombres $K$. Il existe une application $\widehat{\varepsilon}$ que l’on explicitera, de l’ensemble des faisceaux inversibles hermitiens gros sur $\mathscr X$ vers $[0,+\infty[$, qui vérifie les conditions suivantes : (a) si $\overline{\mathscr L}$ et $\overline{\mathscr L}{}'$ sont deux faisceaux inversibles hermitiens gros tels que $\overline{\mathscr L}{}^\vee\otimes\overline{\mathscr L}{}'$ possède au moins une section effective non-nulle, alors $\widehat{\varepsilon}(\overline{\mathscr L})\leqslant\widehat{\varepsilon}(\overline{\mathscr L}{}')$; (b) pour tout faisceau inversible hermitien gros $\overline{\mathscr L}$ sur $\mathscr X$, on a $$\widehat{h}^0(\overline{\mathscr L})\leqslant\frac{\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})}{(d+1)!}+\widehat{\varepsilon}(\overline{\mathscr L}),$$ où $d+1$ est la dimension de Krull du schéma $\mathscr X$; (c) pour tout faisceau inversible hermitien gros $\overline{\mathscr L}$, on a $$\widehat{\varepsilon}(\overline{\mathscr L}{}^{\otimes n})= [K:\mathbb Q]\frac{\operatorname{vol}(\mathscr L_K)}{(d-1)!}n^d\ln(n)+O(n^d).$$ [Le terme d’erreur $\widehat{\varepsilon}(\overline{\mathscr L})$ sera rendu explicite plus loin dans §\[Sec:majroationarith\] (cf. le théorème \[Thm:majorationh0\]), il dépend de la pente maximale asymptotique de $\overline{\mathscr L}$ (qui est un invariant arithmétique birationnel), ainsi que des invariants birationnels de $\mathscr L_K$. Si on compare ce théorème aux résultats dans la littérature, il y a deux nouveautés essentielles. Premièrement l’inégalité dans (b) est valable pour le fibré inversible hermitien $\overline{\mathscr L}$. On n’a pas besoin de passer à une puissance tensorielle d’exposant suffisamment grand, comme par exemple dans le théorème de Hilbert-Samuel arithmétique de [@Abbes-Bouche]. Deuxièmement, dans le théorème \[Thm:majorationgeom\] on ne demande aucune condition de positivité sur la métrique du fibré inversible hermitien $\overline{\mathscr L}$.]{} La démonstration du théorème repose sur la méthode de $\mathbb R$-filtration introduite dans [@Chen10b; @Chen10]. On considère chaque $\mathcal E_n=H^0(\mathscr X,\mathscr L^{\otimes n})$ ($n\in\mathbb N$) comme un réseau dans un $\mathbb R$-espace vectoriel muni de la norme sup. Les minima successifs du réseau correspondent à une $\mathbb R$-filtration $\mathcal F$ sur le $\mathbb Q$-espace vectoriel $E_n=H^0(\mathscr X_K,\mathscr L_K)$ telle que $$\forall\,t\in\mathbb R,\quad \mathcal F^t(E_n)=\mathrm{Vect}\big\{s\in\mathcal E_n\,:\,\|s\|_{\sup}\leqslant\mathrm{e}^{-t}\big\}.$$ Rappelons que, pour tout $i\in\{1,\ldots,R_n\}$, où $R_n=\operatorname{rg}_{\mathbb Q}(E_n)$, le $i^{\text{\`eme}}$ minimum logarithmique du réseau $\mathcal E_n$ est défini comme $$\lambda_i(\mathcal E_n):=\sup\{t\in\mathbb R\,:\,\operatorname{rg}_{\mathbb Q}(\mathcal F^t(E_n))\geqslant i\}.$$ En particulier, on peut interpréter $\sum_{i}\max(\lambda_i(\mathcal E_1),0)$ sous forme d’une intégrale : $$\sum_{i=1}^{R_1}\max(\lambda_i(\mathcal E_1),0)=\int_{0}^{+\infty}\operatorname{rg}(\mathcal F^tE_1)\,\mathrm{d}t.$$ Il s’avère que cette somme est étroitement liée à $\widehat{h}^0(\overline{\mathscr L}{}^{\otimes n})$, compte tenu du deuxième théorème de Minkowski et des résultats en géométrie des nombres comme par exemple [@Blichfeld14]. En outre, pour tout $t\in\mathbb R$, la somme directe $E^t_\sbullet=\bigoplus_{n\geqslant 0}\mathcal F^{nt}(E_n)$ est un système linéaire gradué de $\mathscr L_K$ (où on considère $\mathscr X_K$ comme un $\mathbb Q$-schéma projectif). Si on désigne par $\operatorname{vol}(E_\sbullet^t)$ son volume, défini comme $$\operatorname{vol}(E_\sbullet^t)=\limsup_{n\rightarrow+\infty}\frac{\operatorname{rg}_{\mathbb Q}(E_n^t)}{n^d/d!},$$ on peut aussi exprimer le volume arithmétique $\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})$ comme une intégrale $$\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})=(d+1)\int_0^{+\infty}\operatorname{vol}(E_\sbullet^t)\,\mathrm{d}t.$$ Ainsi on peut ramener le problème à la majoration de $\operatorname{rg}_{\mathbb Q}(\mathcal F^tE_1)=\operatorname{rg}_{\mathbb Q}(E_1^t)$ en fonction de $\operatorname{vol}(E_\sbullet^t)$. Cela peut être considéré comme une généralisation du problème de majoration explicite de la fonction de Hilbert-Samuel géométrique dans le cadre des systèmes linéaires gradués. Ce problème est résolu par le théorème suivant, qui peut être vu comme un avatar géométrique du théorème \[Thm:majorationgeom\]. \[Thm:HMgeometryique\] Soit $X$ un schéma projectif et intègre défini sur un corps $k$. On désigne par $S$ l’ensemble des systèmes linéaires gradués de faisceaux inversibles sur $X$, qui contiennent des diviseurs amples. Il existe une application $\varepsilon:S\rightarrow [0,+\infty[$ que l’on explicitera, qui vérifie les conditions suivantes : (a) si $V_\sbullet$ et $W_\sbullet$ sont des systèmes linéaires gradués dans $S$, des faisceaux inversibles $L$ et $M$ respectivement et si le faisceau inversible $L^\vee\otimes M$ admet une section effective non-nulle $s$ telle que la multiplication par des puissances de $s$ envoie $V_\sbullet$ dans $W_\sbullet$, alors on a $\varepsilon(V_\sbullet)\leqslant\varepsilon(W_\sbullet)$; (b) pour tout système linéaire gradué $V_\sbullet$ dans $S$, on a $$\operatorname{rg}_k(V_1)\leqslant\frac{\operatorname{vol}(V_\sbullet)}{d!}+\varepsilon(V_\sbullet),$$ où $d$ est la dimension de Krull de $X$; (c) pour tout système linéaire gradué $V_\sbullet$ dans $S$ et tout entier $n\geqslant 1$, on a $$\varepsilon(V_\sbullet^{(n)})\leqslant n^{d-1}\varepsilon(V_\sbullet),$$ où $V_\sbullet^{(n)}=\bigoplus_{m\geqslant 0}V_{nm}$. Comparé aux résultats dans la littérature, le théorème \[Thm:HMgeometryique\] s’applique à des systèmes linéaires gradués très généraux, et on ne demande pas la condition d’amplitude (ou de semi-amplitude) pour les faisceaux inversibles en question. [Le terme d’erreur $\varepsilon(.)$ sera précisé dans §\[Sec:HS\] et dépend du choix d’une chaîne de sous-extensions du corps des fonctions rationnelles $k(X)$ sur le corps de base $k$ dont les extensions successives sont de degré de transcendance $1$.]{} Pour un système linéaire gradué $V_\sbullet$ fixé, si on se contente d’obtenir l’existence d’une fonction $F_{V_\sbullet}:\mathbb N\rightarrow+\infty$ telle que $F_{V_\sbullet}(n)=O(n^{d-1})$ pour $n\rightarrow+\infty$ et que $$\operatorname{rg}_k(V_n)\leqslant \frac{\operatorname{vol}(V_\sbullet)}{d!}n^d+F_{V_\sbullet}(n),$$ on peut utiliser la théorie des corps d’Okounkov développée dans [@Lazarsfeld_Mustata08; @Kaveh_Khovanskii] pour relier $V_\sbullet$ à un corps convexe $\Delta$ dans $\mathbb R^d$. On peut majorer le rang de $V_n$ par le nombre de points à coordonnées entiers dans $n\Delta$ et ensuite faire appel à un résultat de Betke et Böröczky [@Betke_Boroczky99] pour obtenir la majoration asymptotique. Cependant, cette méthode est inadéquate pour l’application dans la situation arithmétique. En effet, pour obtenir une majoration de la fonction de Hilbert-Samuel arithmétique, il faut appliquer la majoration de la fonction de Hilbert-Samuel géométrique à une famille continue de systèmes linéaires gradués. Cependant, le terme sous-dominant dans la majoration de la fonction de Hilbert-Samuel géométrique obtenue par cette méthode dépend du bord du corps convexe associé au système linéaire gradué. Il est difficile d’obtenir un contrôle explicite et uniforme pour la famille de systèmes linéaires gradués qui apparaissent dans l’étude de la fonction de Hilbert-Samuel arithmétique. En outre, cette méthode ne peut pas être directement appliquée dans la situation arithmétique car dans l’analogue arithmétique du corps d’Okounkov, il n’y a pas de lien entre l’ensemble des sections effectives du faisceau inversible hermitien avec l’ensemble des points de coordonnées entières dans le corps d’Okounkov arithmétique associé. Pour démontrer le théorème \[Thm:HMgeometryique\], le point clé est d’adopter un point de vue arithmétique. En effet, l’approche de $\mathbb R$-filtration s’applique également dans le cadre de la géométrie arithmétique sur le corps de fonctions, où considère $X$ comme une fibration au-dessus d’une courbe projective régulière sur $k$. Une telle fibration est toujours réalisable, quitte à remplacer $X$ par une modification birationnelle, où la fonction volume reste invariante. On utilise ainsi un argument de nature arithmétique comme dans la stratégie de démonstration du théorème \[Thm:majorationgeom\] et ramène le problème à un problème similaire pour la fibre générique de $X$, qui est un schéma projectif et intègre de dimension $\dim(X)-1$ définie sur le corps de fonction de la courbe de base. La majoration est obtenue par récurrence sur la dimension de $X$, et le majorant dépend du choix d’un tour de fibrations sur courbes d’une modification birationnelle de $X$. Dans le cas où la caractéristique de $k$ est zéro, on peut utiliser la $\mathbb R$-filtration de Harder-Narasimhan. Cependant, dans le cas où la caractéristique de $k$ est positif, il faut utiliser la filtration par minima. Le majorant est légèrement plus grand, mais toujours de même ordre de grandeur. Cette approche de $\mathbb R$-filtration, qui s’applique à la fois aux cas géométrique et arithmétique, combine les avantages de plusieurs méthodes mentionnées plus haut. D’abord le majorant de la fonction de Hilbert-Samuel géométrique ou arithmétique est obtenu par une formule de récurrence sur la dimension de $X$, qui rend le calcul explicit. Deuxièmement, la contribution arithmétique du système linéaire gradué par rapport aux courbes projectives régulières figurant dans le tour de fibrations ressemble beaucoup à la contribution du faisceau inversible dualisant dans l’approche de Kollár et Matsusaka. Enfin, cette méthode peut être naturellement généralisée dans le cadre de système linéaire gradué filtré comme dans [@Boucksom_Chen], qui permet de découvrir de nouveaux phénomènes en géométrie arithmétique. On établit par exemple le résultat suivant. \[Thm:majorationdeminima\] Soit $\mathscr X$ une variété arithmétique projective définie sur l’anneau des entiers algébriques d’un corps de nombres $K$. Il existe une application $\widehat{\epsilon}$ que l’on explicitera, de l’ensemble des faisceaux inversibles hermitiens gros sur $\mathscr X$ vers $[0,+\infty[$, qui vérifie les conditions suivantes : (a) si $\overline{\mathscr L}$ et $\overline{\mathscr L}{}'$ sont deux faisceaux inversibles hermitiens gros tels que $\overline{\mathscr L}{}^\vee\otimes\overline{\mathscr L}{}'$ possède au moins une section effective non-nulle, alors $\widehat{\epsilon}(\overline{\mathscr L})\leqslant\widehat{\epsilon}(\overline{\mathscr L}{}')$; (b) pour tout faisceau inversible hermitien gros $\overline{\mathscr L}$ sur $\mathscr X$, on a $$\sum_{i=1}^{r}\max\big(\lambda_i(H^0(\mathscr X,\mathscr L),\|.\|_{\sup}),0\big)\leqslant\frac{\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})}{(d+1)!}+\widehat{\epsilon}(\overline{\mathscr L}),$$ où $d+1$ est la dimension de Krull du schéma $\mathscr X$, et $r=\operatorname{rg}_{\mathbb Z}H^0(\mathscr X,\mathscr L)$; (c) pour tout faisceau inversible hermitien gros $\overline{\mathscr L}$, on a $\widehat{\epsilon}(\overline{\mathscr L}{}^{\otimes n})\leqslant n^d\widehat{\epsilon}(\overline{\mathscr L})$. La différence principale entre ce théorème et le théorème \[Thm:majorationgeom\] est dans la condition (c). Au lieu d’avoir un terme d’erreur d’ordre $n^d\ln(n)$, le terme d’erreur $\widehat{\epsilon}(\overline{\mathscr L}{}^{\otimes n})$ ici ([qui sera précisé dans la démonstration du théorème \[Thm:HSarithmetic\]]{}) est d’ordre $n^d$ lorsque $n\rightarrow+\infty$. [Un résultat similaire pour les minima successifs absolus est établi dans le théorème \[Thm:HMarithmetique\].]{} Ce résultat est frappant car dans une formule de développement d’une fonction arithmétique de type Hilbert-Samuel, on attend souvent que le terme sous-dominant soit d’ordre $O(n^d\ln(n))$ quand $n$ tend vers l’infini. L’estimation (c) dans le théorème \[Thm:majorationdeminima\] suggère que le terme d’ordre $O(n^d\ln(n))$ dans le développement de la fonction de Hilbert-Samuel arithmétique provient notamment de la comparaison entre différents invariants arithmétiques de fibrés vectoriels normés sur la courbe arithmétique, ou de la distorsion entre les choix de différents métriques. La contribution géométrique pourrait agir plutôt sur le terme suivant d’ordre $O(n^d)$. On espère que ce nouveau point de vue nous aidera à mieux comprendre le rôle de la géométrie du schéma $\mathscr X_K$ dans l’étude de la fonction de Hilbert-Samuel arithmétique. Pendant la rédaction de l’article, Xinyi Yuan et Tong Zhang m’ont communiqué leur travaux [@Yuan_Zhang14], où ils ont obtenu indépendemment des résultats similaires aux théorèmes \[Thm:majorationgeom\] et \[Thm:majorationdeminima\]. Leur approche a certaines similitudes comparée à celle adoptée dans cet article, notamment l’arguement de récurrence sur la dimension de la variété géométrique ou arithmétique. La différence majeure entre les deux approches repose sur la réalisation du procédé de récurrence. Dans [@Yuan_Zhang14], l’argument de Yuan et Zhang est basé sur la positivité du fibré inversible et les termes d’erreur dans leurs théorèmes dépendent des nombres d’intersections de certains fibrés inversibles auxiliaires qui contrôlent la positivité du fibré inversible dont on veut borner la fonction de Hilbert-Samuel. Cependant, dans l’article présent, on choisit de généralizer le problème dans le cadre des systèmes linéaires gradués, munis des structures de métriques et puis ramener le problème à la fibre générique de la variété (arithmetique ou fibrée sur une courbe) afin de réduire la dimension. Il est une question délicate de comparer les termes d’erreur obtenus par ces méthodes différentes car la liaison entre les deux approches est encore obscure, mais il n’est pas exclu qu’une combinaison astucieuse de ces méthodes conduira à une majoration de la fonction de Hilbert-Samuel géométrique ou arithmétique, où le terme d’erreur ne dépend que du volume du fibré inversible et le produit d’intersection positif du fibré inversible avec le faisceau dualisant. L’article est organisé comme la suite. Dans le deuxième paragraphe, on établit un lien entre la valeur maximale du polygone de Harder-Narasimhan d’un fibré vectoriel sur une courbe avec la dimension de l’espace vectoriel des sections globales du fibré vectoriel. Cette comparaison sera utile plus loin dans la majoration de la fonction de Hilbert-Samuel géométrique. Le troisième paragraphe est consacré à un rappel de la notion de pente maximale asymptotique pour les systèmes linéaires gradués. C’est un invariant birationnel qui interviendra dans le terme d’erreur de la majoration. Dans le quatrième paragraphe, on propose une nouvelle notion : tour de fibrations sur courbes, où on considère une variété projective comme des fibrations successives sur les courbes projectives régulières définies sur des corps de plus en plus gros. C’est un outil essentiel pour la majoration de la fonction de Hilbert-Samuel. En utilisant cet outil et la $\mathbb R$-filtration de Harder-Narasimhan, on établit la majoration explicite de la fonction de Hilbert-Samuel géométrique dans le cinquième paragraphe, sous condition que le corps de base est de caractéristique zéro. Le siximème paragraphe est consacré à un rappel sur la notion de fibré vectoriel adélique sur un corps de nombres, et la filtration par minima absolus. On obtient dans le septième paragraphe la majoration explicite de la fonction de Hilbert-Samuel arithmétique. Enfin, dans le dernier paragraphe, on démontre la majoration de la fonction de Hilbert-Samuel géométrique dans la cas de caractéristique positif, en utilisant la méthode arithmétique en considérant la $\mathbb R$-filtration par minima. [**Remerciements :**]{} Je voudrais remercier Éric Gaudron pour des remarques qui m’ont aidé à améliorer la rédaction de l’article. Pendant la préparation et la rédaction de l’article, j’ai bénéficié des discussions avec Sebastien Boucksom, je tiens à lui exprimer mes gratitudes. Enfin, je suis reconnaissant à Xinyi Yuan et Tong Zhang pour m’avoir communiqué leur article et pour des discussions très intéressantes. Degré positif d’un fibré vectoriel ================================== Soient $k$ un corps et $C$ une courbe projective régulière définie sur $k$. Rappelons que la formule de Riemann-Roch montre que, pour tout fibré vectoriel $E$ sur $C$, on a $$\label{Equ:RR} h^0(E)-h^1(E)=\deg(E)+\operatorname{rg}(E)(1-g),$$ où $h^0(E)$ et $h^1(E)$ sont respectivement la dimension sur le corps $k$ des espaces de cohomologie $H^0(X,E)$ et $H^1(X,E)$, et $g$ désigne le genre de $C$. En utilisant cette formule, on peut relier $h^0(E)$ à la valeur maximale du polygone de Harder-Narasimhan de $E$. Étant donné un fibré vectoriel non-nul $E$ sur $C$, la *pente* de $E$ est définie comme le quotient du degré de $E$ par son rang, notée comme $\mu(E)$. Le fibré vectoriel $E$ est dit *semi-stable* si chaque sous-fibré vectoriel non-nul de $E$ admet une pente $\leqslant\mu(E)$. Si $E$ est un fibré vectoriel non-nul qui n’est pas nécessairement semi-stable, il existe un unique sous-fibré vectoriel $E_{\operatorname{des}}$ de $E$ qui vérifie les conditions suivantes (a) pour tout sous-fibré vectoriel non-nul $F$ de $E$, on a $\mu(F)\leqslant\mu(E_{\operatorname{des}})$; (b) si $F$ est un sous-fibré vectoriel non-nul de $E$ tel que $\mu(F)=\mu(E_{\operatorname{des}})$, alors $F\subset E_{\operatorname{des}}$. Le sous-fibré vectoriel $E_{\operatorname{des}}$ est appelé le *sous-fibré déstabilisant* de $E$. Sa pente est appelée la *pente maximale* de $E$, notée comme $\mu_{\max}(E)$. La condition (b) plus haut implique que le quotient $E/E_{\operatorname{des}}$ est sans torsion, donc est un fibré vectoriel sur $C$. Ainsi on peut construire par récurrence un drapeau de sous-fibrés vectoriels de $E$ : $$0=E_0\subsetneq E_1\subsetneq \ldots\subsetneq E_n=E$$ tel que $E_i/E_{i-1}=(E/E_{i-1})_{\operatorname{des}}$ pour tout $i\in\{1,\ldots,n\}$. Ce drapeau est appelé le *drapeau de Harder-Narasimhan* de $E$. Chaque sous-quotient $E_i/E_{i-1}$ est un fibré vectoriel semi-stable sur $C$. En outre, si on désigne par $\alpha_i$ la pente du sous-quotient $E_i/E_{i-1}$, alors les inégalités $\alpha_1>\ldots>\alpha_n$ sont vérifiées. Il s’avère que le drapeau de Harder-Narasimhan est le seul\[page:HNunicite\] drapeau de sous-fibrés vectoriels de $E$ tel que les sous-quotients soient semi-stables et de pentes strictement décroissantes (cf. [@Huyb théorème 1.3.4] pour une démonstration). La dernière pente $\alpha_n$ est appelée la *pente minimale* de $E$, notée comme $\mu_{\min}(E)$. C’est aussi la valeur minimale des pentes des fibrés vectoriels quotients de $E$. En particulier, les pentes maximale et minimale sont reliées par la formule de dualité suivante : pour tout fibré vectoriel non-nul $E$ sur $C$, on a $$\mu_{\max}(E)+\mu_{\min}(E^\vee)=0.$$ On désigne par $P_E$ la fonction concave et affine par morceau définie sur l’intervalle $[0,\operatorname{rg}(E)]$, qui est affine sur chaque intervalle $[\operatorname{rg}(E_{i-1}),\operatorname{rg}(E_i)]$ et de pente $\alpha_i$. Rappelons que le graphe de $P_E$ s’identifie au bord supérieur de l’enveloppe convexe de l’ensemble des points de la forme $(\operatorname{rg}(F),\deg(F))\in\mathbb R^2$, où $F$ parcourt l’ensemble des sous-fibrés vectoriels de $E$. La fonction $P_E$ est appelée le *polygone de Harder-Narasimhan* de $E$. Soit $E$ un fibré vectoriel non-nul sur la courbe projective $C$. On désigne par $\deg_+(E)$ la valeur maximale de la fonction $P_E$ sur l’intervalle $[0,\operatorname{rg}(E)]$, appelée le *degré positif* of $E$. On voit aussitôt de la définition que, si la pente minimale de $E$ est positive, alors $\deg_+(E)$ s’identifie au degré de $E$. On convient que le degré positif du fibré vectoriel nul est zéro. \[Lem:h0\] Soient $C$ une courbe projective régulière de genre $g$ définie sur un corps $k$, et $E$ un fibré vectoriel non-nul sur $C$. (a) Si $\mu_{\max}(E)<0$, alors $h^0(E)=0$. (b) Si $\mu_{\min}(E)>2g-2$, alors $h^0(E)=\deg(E)+\operatorname{rg}(E)(1-g)$. (c) Si $\mu_{\min}(E)>0$, alors $|h^0(E)-\deg(E)|\leqslant \operatorname{rg}(E)|g-1|$. \(a) Supposons que $E$ possède une section globale non-nul. Elle correspond à un homomorphisme non-nul de $\mathcal O_C$ vers $E$. Donc on a $$0=\mu(\mathcal O_C)\leqslant\mu_{\max}(E).$$ \(b) D’après la formule de Riemann-Roch et la dualité de Serre $h^1(E)=h^0(E^\vee\otimes\omega_C)$, où $\omega_C$ est le faisceau dualisant sur $C$, on obtient $$h^0(E)-h^0(E^\vee\otimes\omega_C)=\deg(E)+\operatorname{rg}(E)(1-g).$$ Comme $$\mu_{\max}(E^\vee\otimes\omega_C)=\mu_{\max}(E^\vee)+\deg(\omega_C)=2g-2-\mu_{\min}(E),$$ si $\mu_{\min}(E)>2g-2$, alors on a $\mu_{\max}(E^\vee\otimes\omega_C)<0$. Donc $h^0(E^\vee\otimes\omega_C)=0$ compte tenu de (a). Par conséquent, l’égalité $h^0(E)=\deg(E)-\operatorname{rg}(E)(1-g)$ est vérifiée. \(c) D’après (b), l’inégalité est vérifiée lorsque $g\leqslant 1$. Dans la suite, on suppose $g\geqslant 2$. Comme $\mu_{\min}(E)>0$, on a $\mu_{\min}(E\otimes\omega_C)=\mu_{\min}(E)+2g-2>2g-2$. D’après (b), on obtient $$h^0(E\otimes\omega_C)=\deg(E\otimes\omega_C)+\operatorname{rg}(E)(1-g)=\deg(E)+\operatorname{rg}(E)(g-1).$$ Comme $h^0( \omega_C)>0$, on a $$h^0(E)\leqslant h^0(E\otimes\omega_C)\leqslant\deg(E)+\operatorname{rg}(E)(g-1).$$ En outre, d’après la formule de Riemann-Roch , on a $h^0(E)\geqslant \deg(E)+\operatorname{rg}(E)(1-g)$. Donc l’inégalité est démontrée. \[Lem:majh0ss\] Soient $C$ une courbe projective régulière définie sur un corps $k$, et $E$ un fibré vectoriel non-nul sur $C$. Si $E$ est semi-stable et de pente $0$, alors $h^0(E)\leqslant\operatorname{rg}(E)$. On peut supposer que $E$ possède une section globale non-nulle, sinon le résultat est trivial. Cette section définit un homomorphisme non-nul de $\mathcal O_C$ vers $E$. Comme $E$ est semi-stable de pente $0$, le faisceau quotient $E/\mathcal O_C$ est ou bien nul, ou bien un fibré vectoriel semi-stable de pente $0$. De plus, la suite exacte longue de groupes de cohomologie associée à la suite exacte courte $0\rightarrow\mathcal O_C\rightarrow E\rightarrow E/\mathcal O_C\rightarrow 0$ montre que $$h^0(E)\leqslant h^0(\mathcal O_C)+h^0(E/\mathcal O_C)=h^0(E/\mathcal O_C)+1.$$ Par récurrence sur le rang de $E$, on obtient le résultat. \[Thm:h0etdegplus\] Soient $C$ une courbe projective régulière de genre $g$ définie sur un corps $k$, et $E$ un fibré vectoriel non-nul sur $C$. On a $$\label{Equ:comparison} |{h}^0(E)-\deg_+(E)|\leqslant \operatorname{rg}(E)\max(g-1,1).$$ Soit $$0=E_0\subsetneq E_1\subsetneq\ldots\subsetneq E_n=E$$ le drapeau de Harder-Narasimhan de $E$. Pour tout $i\in\{1,\ldots,n\}$, soit $\alpha_i$ la pente de $E_{i}/E_{i-1}$. Soit $j$ le plus grand indice tel que $\alpha_j\geqslant 0$. Si un tel indice n’existe pas, on note $j=0$ par convention. On a $\deg_+(E)=\deg(E_j)$ par définition. En outre, comme $E/E_j$ est ou bien nul ou bien de pente maximale strictement négative, on a $h^0(E/E_j)=0$ et donc $h^0(E)=h^0(E_j)$. Si $j=0$, alors on a $h^0(E)=0=\deg_+(E)$ et l’inégalité devient triviale. Dans la suite, on suppose $j\in\{1,\ldots,n\}$. On traite d’abord le cas où $g\geqslant 1$. Si $\alpha_j=\mu_{\min}(E_j)>0$, d’après le lemme \[Lem:h0\].(c), on obtient $$|h^0(E)-\deg_+(E)|=|h^0(E_j)-\deg(E_j)|\leqslant\operatorname{rg}(E_j)(g-1),$$ qui implique . Il reste le cas où $\alpha_j=0$. Dans ce cas-là $E_j/E_{j-1}$ est un fibré vectoriel semi-stable de pente $0$. D’après le lemme \[Lem:majh0ss\] on a $h^0(E_{j}/E_{j-1})\leqslant\operatorname{rg}(E_{j}/E_{j-1})$, qui implique que $$\begin{split}\label{Equ:encardh0} h^0(E_{j-1})\leqslant h^0(E)=h^0(E_j)&\leqslant h^0(E_{j-1})+h^0(E_j/E_{j-1})\\ &\leqslant h^0(E_{j-1})+\operatorname{rg}(E_j/E_{j-1}).\end{split}$$ En outre, on a $\deg_+(E)=\deg(E_{j-1})$. Le fibré vectoriel $E_{j-1}$ est ou bien nul, ou bien de pente minimale $>0$. D’après le lemme \[Lem:h0\].(c) on obtient (l’inégalité est triviale lorsque $E_{j-1}=0$) $$|h^0(E_{j-1})-\deg_+(E)|=|h^0(E_{j-1})-\deg(E_{j-1})|\leqslant \operatorname{rg}(E_{j-1})(g-1).$$ Si on combine cette inégalité avec , on obtient . Dans la suite, on suppose que $g=0$. Comme $\alpha_j=\mu_{\min}(E_j)\geqslant 0>2g-2$, d’après le lemme \[Lem:h0\].(b) on obtient $h^0(E_j)-\deg(E_j)=\operatorname{rg}(E_j)$. Comme on a $ h^0(E_j)=h^0(E)$ et $\deg(E_j)=\deg_+(E)$, le résultat est aussi vrai dans ce cas-là. À l’aide de $\mathbb R$-filtration de Harder-Narasimhan introduite dans [@Chen10b], on peut interpréter la fonction $\deg_+(.)$ comme une intégrale. Soit $E$ un fibré vectoriel non-nul sur $C$. On suppose que son drapeau de Harder-Narasimhan est $$0=E_0\subsetneq E_1\subsetneq \ldots\subsetneq E_n=E.$$ Pour tout $i\in\{1,\ldots,n\}$, soit $\alpha_i$ la pente du sous-quotient $E_i/E_{i-1}$. On définit une famille $(\mathcal F^tE)_{t\in\mathbb R}$ de sous-fibrés vectoriels de $E$ comme $$\mathcal F^tE=E_i\;\text{si $\alpha_i\geqslant t>\alpha_{i-1}$},$$ où par convention $\alpha_{0}=+\infty$ et $\alpha_{n+1}=-\infty$. Pour tout nombre réel $t$, on désigne par $\mathcal F^{t+}E$ le sous-fibré vectoriel $\sum_{a>0}\mathcal F^{t+a}E$ de $E$ et on définit $\operatorname{sq}^t(E)$ le quotient $\mathcal F^tE/\mathcal F^{t+}E$, appelé le *sous-quotient* d’indice $t$ de la filtration $\mathcal F$. D’après la définition du drapeau de Harder-Narasimhan, on obtient que chaque sous-quotient $\mathrm{sq}^t(E)$ est ou bien le fibré vectoriel nul, ou bien un fibré vectoriel semi-stable de pente $t$. En outre, l’unicité du drapeau de Harder-Narasimhan que l’on a mentionnée dans la page conduit au critère suivant de la $\mathbb R$-filtration de Harder-Narasimhan. \[Pro:criteredeHN\] Soit $E$ un fibré vectoriel non-nul sur $C$ et $(\mathcal G^tE)_{t\in\mathbb R}$ une $\mathbb R$-filtration décroissante en sous-fibrés vectoriels de $E$ telle que $\mathcal G^tE=0$ pour $t$ suffisamment positif, $\mathcal G^tE=E$ pour $t$ suffisamment négatif et $\bigcap_{a>0}\mathcal G^{t-a}E=\mathcal G^tE$ pour tout $t\in\mathbb R$. Alors $\mathcal G$ est la $\mathbb R$-filtration de Harder-Narasimhan si et seulement si, pour tout $t\in\mathbb R$, le sous-quotient d’indice $t$ de la filtration $\mathcal G$ est ou bien nul, ou bien un fibré vectoriel semi-stable de pente $t$. Soit $E$ un fibré vectoriel non-nul sur $C$. On suppose que son drapeau de Harder-Narasimhan est $$0=E_0\subsetneq E_1\subsetneq \ldots\subsetneq E_n=E.$$ Pour tout $i\in\{1,\ldots,n\}$, soit $\alpha_i$ la pente du sous-quotient $E_i/E_{i-1}$. On désigne par $\nu_E$ la mesure de probabilité borélienne sur $\mathbb R$ définie comme $$\nu_E(\mathrm{d}t)=-\mathrm{d}\frac{\operatorname{rg}(\mathcal F^tE)}{\operatorname{rg}(E)}=\sum_{i=1}^n\frac{\operatorname{rg}(E_i/E_{i-1})}{\operatorname{rg}(E)}\delta_{\alpha_i}.$$ Avec cette notation, on peut réécrire $\deg_+(E)$ comme $$\label{Equ:degplus} \deg_+(E)=\operatorname{rg}(E)\int_0^{+\infty}t\,\nu_E(\mathrm{d}t)=\int_0^{\mu_{\max}(E)}\operatorname{rg}(\mathcal F^tE)\,\mathrm{d}t,$$ où la dernière égalité provient de l’intégration par partie et du fait que $\mathcal F^tE=0$ quand $t>\mu_{\max}(E)$. Le théorème \[Thm:h0etdegplus\] montre alors que $$\bigg|h^0(E)-\int_0^{\mu_{\max}(E)}\operatorname{rg}(\mathcal F^tE)\,\mathrm{d}t\bigg|\leqslant\operatorname{rg}(E)\max(g-1,1).$$ Dans le cas où le corps de base $k$ est de caractéristique zéro, d’après un résultat de Narasimhan et Seshadri [@Nara_Se65], le produit tensoriel de deux fibrés vectoriels semi-stables sur $C$ est encore semi-stable. On renvoie les lecteurs dans [@Bost_Chen §1.1] pour un survol succinct de différentes approches autour de la semi-stabilité tensorielle dans la littérature. Ce résultat implique que la $\mathbb R$-filtration de Harder-Narasimhan du produit tensoriel de deux fibrés vectoriels s’identifie à la filtration produit. \[Pro:produitdefiltrationHN\] On suppose que le corps $k$ est de caractéristique zéro. Soient $E$ et $F$ deux fibrés vectoriels non-nuls sur $C$. Si $t$ est un nombre réel, alors on a $$\label{Equ:HNfiltratio}\mathcal F^{t}(E\otimes F)=\sum_{\begin{subarray}{c} (a,b)\in\mathbb R^2\\ a+b=t \end{subarray}}\mathcal F^a(E)\otimes\mathcal F^b(F),$$ où $\mathcal F$ désigne la $\mathbb R$-filtration de Harder-Narasimhan. On désigne par $\mathcal G$ la $\mathbb R$-filtration de $E\otimes F$ telle que $\mathcal G^t(E\otimes F)$ soit défini comme le membre à droite de la formule . Notre but est de démontrer que la filtration $\mathcal G$ s’identifie à la $\mathbb R$-filtration de Harder-Narasimhan de $E\otimes F$. Pour tout $t\in\mathbb R$, le sous-quotient d’indice $t$ de la filtration $G$ s’écrit sous la forme $$\operatorname{sq}_{\mathcal G}^t(E\otimes F)=\bigoplus_{a+b=t}\operatorname{sq}^a(E)\otimes\operatorname{sq}^b(F).$$ Le fibré vectoriel $\operatorname{sq}^a(E)$ (resp. $\operatorname{sq}^b(F)$) est ou bien nul, ou bien semi-stable de pente $a$ (resp. $b$). D’après le résultat de Narasimhan et Seshadri, le produit tensoriel $\operatorname{sq}^a(E)\otimes\operatorname{sq}^b(F)$ est un fibré vectoriel nul ou semi-stable de pente $a+b$. Cela montre que le sous-quotient $\operatorname{sq}_{\mathcal G}^t(E\otimes F)$ est nul ou semi-stable de pente $t$. D’après la proposition \[Pro:criteredeHN\], on obtient que la filtration $\mathcal G$ est la $\mathbb R$-filtration de Harder-Narasimhan de $E$. \[Cor:filtrationdeHNenalg\] On suppose que le corps $k$ est de caractéristique zéro. Soit $E_\sbullet=\bigoplus_{n\geqslant 0}E_n$ une $\mathcal O_C$-algèbre graduée. On suppose que chaque composante homogène $E_n$ est un fibré vectoriel sur $C$. Alors la $\mathbb R$-filtration de Harder-Narasimhan est compatible à la structure de $\mathcal O_C$-algèbre de $E_\sbullet$. Autrement dit, pour tout couple d’entiers $(n,m)\in\mathbb N^2$ et tout $(a,b)\in\mathbb R^2$, on a $$(\mathcal F^aE_n)\cdot(\mathcal F^bE_m)\subset\mathcal F^{a+b}E_{n+m}.$$ Par définition $(\mathcal F^aE_n)\cdot(\mathcal F^bE_m)$ est l’image canonique de $(\mathcal F^aE_n)\otimes(\mathcal F^bE_m)$ par l’homomorphisme $\varphi_{n,m}:E_n\otimes E_m\rightarrow E_{n+m}$ de la structure de $\mathcal O_C$-algèbre de $E_\sbullet$. La proposition précédente montre que $(\mathcal F^aE_n)\otimes(\mathcal F^bE_m)$ est contenu dans $\mathcal F^{a+b}(E_n\otimes E_m)$. En outre, d’après [@Chen10b proposition 2.2.4], tout homomorphisme de fibrés vectoriels sur $C$ préserve les $\mathbb R$-filtrations de Harder-Narasimhan. En particulier, on a $$\varphi_{n,m}(\mathcal F^{a+b}(E_n\otimes E_m))\subset\mathcal F^{a+b}(E_{n+m}),$$ d’où le résultat. Pente maximal asymptotique ========================== Soient $C$ une courbe projective régulière définie sur un corps $k$, et $\pi:X\rightarrow C$ un morphisme projectif et plat d’un schéma intègre $X$ vers $C$. On s’intéresse à des invariants birationnels de faisceaux inversibles sur $X$. Rappelons que le *volume* d’un faisceau inversible $L$ sur $X$ est défini comme $$\operatorname{vol}(L):=\limsup_{n\rightarrow+\infty}\frac{\operatorname{rg}_kH^0(X,L^{\otimes n})}{n^{\dim(X)}/\dim(X)!}.$$ On dit que le faisceau inversible $L$ est *gros* si son volume est strictement positif. La fonction volume est un invariant birationnel (cf. [@LazarsfeldI proposition 2.2.43]) : si $p:X'\rightarrow X$ est un morphisme projectif birationnel d’un schéma intègre $X'$ vers $X$, alors on a $\operatorname{vol}(p^*(L))=\operatorname{vol}(L)$. En outre, le faisceau inversible $L$ est gros\[Page:criteregors\] si et seulement si une puissance tensorielle de $L$ peut être décomposée en le produit tensoriel d’un faisceau inversible ample et un faisceau inversible effectif (i.e. qui possède au moins une section globale non-nulle). On renvoie les lecteurs dans [@LazarsfeldI corollaire 2.2.7] pour une démonstration. Ce critère montre en particulier que les faisceaux inversibles gros forment un cône ouvert dans le groupe de Picard de $X$ : si $L$ est un faisceau inversible gros et si $M$ est un faisceau inversible quelconque sur $X$, alors pour tout entier $n$ assez positif, le produit tensoriel $L^{\otimes n}\otimes M$ est un faisceau inversible gros. Soit $K$ le corps des fonctions rationnelles sur la courbe $C$. On désigne par $\eta:\operatorname{Spec}K\rightarrow C$ le point générique de $C$. Soit $L$ un faisceau inversible sur $X$ tel que $L_\eta$ soit gros. Dans [@Chen10b théorème 4.3.6], il est démontré que la suite $\big(\mu_{\max}(\pi_*(L^{\otimes n}))/n\big)_{n\geqslant 1}$ converge dans $\mathbb R$. On désigne par $\mu_{\max}^{\pi}(L)$ la limite de cette suite, appelée la *pente maximale asymptotique* de $L$ relativement à $\pi$. Par définition, on a $\mu_{\max}^{\pi}(L^{\otimes n})=n\mu_{\max}^{\pi}(L)$ pour tout entier $n\geqslant 1$. En outre, si $M$ est un faisceau inversible sur $C$, alors on a (cf. [@Chen10b proposition 4.3.8]) $$\label{Equ:mupimaxtordu}\mu_{\max}^{\pi}(L\otimes\pi^*(M))=\mu_{\max}^{\pi}(L)+\deg(M).$$ Soit $E$ un fibré vectoriel non-nul sur la courbe $C$. Si $\mu_{\max}(E)>|g-1|$, où $g$ est le genre de $C$, alors on a $h^0(E)>0$. Soit $E_{\operatorname{des}}$ le sous-fibré déstabilisant du fibré vectoriel $E$. On a $\mu_{\min}(E_{\operatorname{des}})=\mu(E_{\operatorname{des}})=\mu_{\max}(E)>0$. D’après le lemme \[Lem:h0\] (c), on obtient $$h^0(E)\geqslant h^0(E_{\operatorname{des}})\geqslant \operatorname{rg}(E_{\operatorname{des}})\mu_{\max}(E)-\operatorname{rg}(E_{\operatorname{des}})|g-1|,$$ d’où le résultat. \[Rem:grosseur\] Soit $L$ un faisceau inversible sur $X$ qui est génériquement gros. Le lemme précédent montre que, si $\mathrm{\mu}_{\max}^{\pi}(L)>0$, alors pour tout entier $n$ suffisamment positif, le faisceau inversible $L^{\otimes n}$ est effectif. En effet, d’après la définition de $\mu_{\max}^{\pi}(L)$, la pente maximale de $\mu_{\max}(\pi_*(L^{\otimes n}))$ croît linéairement par rapport à $n$ lorsque $n$ tend vers l’infini. Elle dépasse $|g-1|$ lorsque $n$ est assez positif. \[Pro:criteredegros\] Soit $L$ un faisceau inversible sur $X$. Alors $L$ est gros si et seulement si $L_\eta$ est gros et $\mu_{\max}^{\pi}(L)>0$. “$\Longrightarrow$”: Soit $L$ un faisceau inversible gros sur $X$. D’après le critère de grosseur que l’on a mentionné plus haut dans la page , il existe un entier $n\geqslant 1$, un faisceau inversible ample $A$ et un faisceau inversible effectif $L'$ sur $X$ tels que $L^{\otimes n}\cong A\otimes L'$. La restriction de cette formule à la fibre générique de $X$ donne une décomposition de $L_\eta^{\otimes n}$ en produit tensoriel d’un faisceau inversible ample et un faisceau inversible effectif. Cela montre que $L_\eta$ est gros. Comme $L$ est un faisceau inversible gros, pour tout entier $n$ assez positif, $L^{\otimes n}$ possède au moins une section globale non-nulle (cf. la remarque \[Rem:grosseur\]). Par conséquent, on a $\mu_{\max}(\pi_*(L^{\otimes n}))\geqslant 0$ pour tout entier $n$ assez positif. On en déduit $\mu_{\max}^{\pi}(L)\geqslant 0$. Soit $M$ un faisceau inversible sur $C$ tel que $\deg(M)>0$. Comme $L$ est gros, il existe un entier $n\geqslant 1$ tel que $L^{\otimes n}\otimes\pi^*( M^\vee)$ soit gros. On a alors $\mu_{\max}^{\pi}(L^{\otimes n}\otimes \pi^*(M^\vee))\geqslant 0$. D’après , cela implique que $$n\mu_{\max}^{\pi}(L)\geqslant\deg(M)>0,$$ d’où $\mu_{\max}^{\pi}(L)>0$. “$\Longleftarrow$”: On suppose que $L$ est un faisceau inversible sur $X$ qui est génériquement gros et tel que $\mu_{\max}^{\pi}(L)>0$. On fixe un faisceau inversible ample $A$ sur $X$. Comme $L_\eta$ est supposé être gros, il existe un entier $d\geqslant 1$ tel que $L_\eta^{\otimes d}\otimes A_\eta^\vee$ possède une section globale non-nulle $s$. La section $s$ se relève en une section rationnelle de $L^{\otimes d}\otimes A^\vee$ dont le diviseur est effectif à un diviseur vertical près. Il existe alors un faisceau inversible ample $M$ sur $C$ tel que $s$ se prolonge en une section globale non-nulle de $L^{\otimes d}\otimes A^\vee\otimes\pi^*(M)$. En outre, comme $\mu_{\max}^{\pi}(L)>0$, il existe un entier $n\geqslant 1$ tel que $$\mu_{\max}^{\pi}(L^{\otimes n}\otimes\pi^*(M^\vee))=n\mu_{\max}^{\pi}(L)-\deg(M)>0.$$ D’après le lemme précédent (voir aussi la remarque qui le suit), il existe alors un entier $m\geqslant 1$ tel que $L^{\otimes nm}\otimes\pi^*(M^\vee)^{\otimes m}$ soit effectif. On en déduit que le faisceau inversible $$L^{md+nm}\otimes (A^{\otimes m})^\vee\cong ( L^{\otimes d}\otimes A^\vee\otimes\pi^*(M))^{\otimes m}\otimes(L^{\otimes nm}\otimes\pi^*(M^\vee)^{\otimes m})$$ admet une section globale non-nulle. Cela montre que $L$ est un faisceau inversible gros. \[Cor:mumaxbir\] La pente maximale asymptotique est un invariant birationnel pour les faisceaux inversibles génériquement gros sur $X$. Soit $M$ un faisceau inversible ample sur $C$. On affirme que, pour tout faisceau inversible $L$ sur $X$ qui est génériquement gros, la valeur $\mu_{\max}^{\pi}(L)$ est égale à $$\sup\Big\{\frac{n\deg(M)}{m}\,\Big|\,(n,m)\in\mathbb N_{\geqslant 1}^2,\,L^{\otimes m}\otimes\pi^*(M^\vee)^{\otimes n}\text{ est gros}\Big\}.$$ En effet, d’après la proposition précédente, $L^{\otimes m}\otimes \pi^*(M^\vee)^{\otimes n}$ est gros si et seulement si $$\mu_{\max}^{\pi}(L^{\otimes m}\otimes \pi^*(M^\vee)^{\otimes n})=m\mu_{\max}^{\pi}(L)-n\deg(M)>0,$$ ou de façon équivalente, $\mu_{\max}^{\pi}(L)>(n/m)\deg(M)$. Comme la fonction volume est un invariant birationnel, on obtient que, pour tout morphisme projectif et birationnel $p:X'\rightarrow X$, le faisceau inversible $L^{\otimes m}\otimes\pi^*(M^\vee)^{\otimes n}$ est gros si et seulement si $$p^*(L^{\otimes m}\otimes\pi^*(M^{\vee})^{\otimes n})\cong p^*(L^{\otimes m})\otimes (\pi p)^*(M^\vee)^{\otimes n}$$ l’est. D’où $\mu_{\max}^{\pi}(L)=\mu_{\max}^{\pi}(p^*L)$. Dans la suite, on généralise la construction de pente maximale asymptotique aux systèmes gradués en fibrés vectoriels. Soit $L$ un faisceau inversible sur $X$. On entend par *système gradué en fibrés vectoriels* de $L$ toute sous-$\mathcal O_C$-algèbre graduée de $\bigoplus_{n\geqslant 0}\pi_*(L^{\otimes n})$. Si $E_\sbullet$ est un système gradué en fibrés vectoriels de $L$, alors sa fibre générique $E_{\sbullet,\eta}:=\bigoplus_{n\geqslant 0}E_{n,\eta}$ est un système linéaire gradué de $L_\eta$. On dit que $E_{\sbullet,\eta}$ *contient un diviseur ample* si les conditions suivantes sont satisfaites (cf. la condition (C) dans [@Lazarsfeld_Mustata08 définition 2.9]) : (a) l’espace vectoriel $E_{n,\eta}$ sur $K$ est non-nul pour tout entier $n$ suffisamment positif, (b) il existe un faisceau inversible ample $A_\eta$ sur $X_\eta$, un entier $p\geqslant 1$ et une section globale non-nulle $s$ de $L^{\otimes p}_\eta\otimes A_\eta^\vee$ tels que, pour tout $n\in\mathbb N$, l’image de l’homomorphisme $$\xymatrix{\relax H^0(X_\eta,A_\eta^{\otimes n})\ar[rr]^-{\cdot s^{n}}&&H^0(X_\eta,L_\eta^{\otimes np})}$$ soit contenue dans $E_{np,\eta}$. Cette condition revient à la grosseur de $L_\eta$ lorsque $E_\sbullet=\bigoplus_{n\geqslant 0}\pi_*(L^{\otimes n})$ est le système gradué total. Si la fibre générique de $E_\sbullet$ contient un diviseur ample, alors la suite de pentes maximales normalisée $(\mu_{\max}(E_n)/n)_{n\geqslant 1}$ converge dans $\mathbb R$ (cf. [@Chen10b théorème 4.3.1]). On désigne par $\mu_{\max}^{\mathrm{asy}}(E_\sbullet)$ sa limite. Pour tout entier $n\geqslant 0$, le fibré vectoriel $E_n$ est un sous-fibré vectoriel de $\pi_*(L^{\otimes n})$. On obtient donc $\mu_{\max}(E_n)\leqslant\mu_{\max}(\pi_*(L^{\otimes n}))$. Cela implique que $\mu_{\max}^{\mathrm{asy}}(E_\sbullet)\leqslant\mu_{\max}^{\pi}(L)$. \[Rem:systemegradue\] Comme la fibre générique de $E_\sbullet$ est un anneau intègre, on obtient du corollaire \[Cor:filtrationdeHNenalg\] que, dans le cas où le corps $k$ est de caractéristique $0$, si $E_n$ et $E_m$ sont non-nuls, alors on a[^2] $$\mu_{\max}(E_{n+m})\geqslant\mu_{\max}(E_n)+\mu_{\max}(E_m).$$ Cela montre que $\mu_{\max}^{\mathrm{asy}}(E_\sbullet)\geqslant \mu_{\max}(E_n)/n$ dès que $E_n$ est non-nul ($n\geqslant 1$). Soient $L$ un faisceau inversible sur $X$ et $E_\sbullet$ un système gradué en fibrés vectoriels de $L$. On définit le *volume* de $E_\sbullet$ comme $$\operatorname{vol}(E_\sbullet):=\limsup_{n\rightarrow+\infty}\frac{h^0(E_n)}{n^{\dim(X)}/\dim(X)!}.$$ Lorsque $E_\sbullet$ est le système gradué total $\bigoplus_{n\geqslant 0}\pi^*(L^{\otimes n})$, son volume s’identifie au volume de $L$. Soient $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$ et $L$ un faisceau inversible gros sur $X$. Soit $V_\sbullet=\bigoplus_{n\geqslant 0}V_n$ un système linéaire gradué de $L$ (i.e. une sous-algèbre graduée de $\bigoplus_{n\geqslant 0}H^0(X,L^{\otimes n})$). On suppose que $V_\sbullet$ contient un diviseur ample, c’est-à-dire que $V_n\neq 0$ pour $n$ assez positif, et qu’il existe un faisceau inversible ample $A$ sur $X$, un entier $p\geqslant 1$ et une section globale non-nulle $s$ de $L^{\otimes p}\otimes A^\vee$ tels que, pour tout $n\in\mathbb N$, on a $$\operatorname{Im}(\xymatrix{\relax H^0(X,A^{\otimes n})\ar[r]^-{\cdot s^{n}}&H^0(X,L^{\otimes np}))}\subset V_{np}.$$ Rappelons que le *volume* de $V_\sbullet$ est défini comme $$\operatorname{vol}(V_\sbullet)=\limsup_{n\rightarrow+\infty}\frac{\operatorname{rg}_k(V_n)}{n^{\dim(X)}/(\dim X)!}.$$ Pour tout entier $n\geqslant 0$, on désigne par $V_{n,K}$ le sous-$K$-espace vectoriel de $H^0(X_\eta,L_\eta^{\otimes n})$ engendré par l’image canonique $V_n$. Il s’avère que $V_{\sbullet,K}:=\bigoplus_{n\geqslant 0}V_{n,K}$ est un système linéaire gradué de $L_\eta$ qui contient un diviseur ample. Dans la suite, on construit un système gradué en fibrés vectoriels $E_\sbullet$ tel que $E_{\sbullet,\eta}$ coincide à $V_{\sbullet,K}$ et que $\operatorname{vol}(E_\sbullet)=\operatorname{vol}(V_\sbullet)$. \[Thm:passageauxfibres\] Soit $V_\sbullet$ un système linéaire gradué de $L$ qui contient un diviseur ample. Pour tout entier $n\geqslant 0$, soit $E_n$ le sous-$\mathcal O_C$-module de $\pi_*(L^{\otimes n})$ engendré[^3] par $V_n$. Alors $E_\sbullet=\bigoplus_{n\geqslant 0}E_n$ est un système gradué en fibrés vectoriels de $L$, dont la fibre générique contient un diviseur ample. De plus, on a $\mathrm{vol}(E_\sbullet)=\mathrm{vol}(V_\sbullet)$. Sans perte de généralité, on peut supposer $X$ normal. En effet, par passage à la normalisation $\nu:\widetilde X\rightarrow X$, on peut considérer $V_\sbullet$ comme un système linéaire gradué de $\nu^*(L)$, qui contient un diviseur ample. On désigne par $\varphi:C\rightarrow\operatorname{Spec}k$ le morphisme structurel. Par définition $E_\sbullet$ est l’image de l’homomorphisme de $\mathcal O_C$-algèbre graduée $\bigoplus_{n\geqslant 0}\varphi^*(V_n)\rightarrow\bigoplus_{n\geqslant 0}\pi_*(L^{\otimes n})$. Donc il est une sous-$\mathcal O_C$-algèbre graduée de $\bigoplus_{n\geqslant 0}\pi_*(L^{\otimes n})$, i.e., un système gradué en fibrés vectoriels de $L$. En outre, il existe un entier $p\geqslant 1$ et un faisceau inversible ample $A$ sur $X$ tels que le faisceau inversible $L^{\otimes p}\otimes A^\vee$ possède une section globale non-nulle $s$ vérifiant $$\label{Equ:contientundivamp}\operatorname{Im}(\xymatrix{\relax H^0(X,A^{\otimes n})\ar[r]^-{\cdot s^{n}}&H^0(X,L^{\otimes np})})\subset V_{np}$$ pour tout entier $n\geqslant 1$. Comme $A$ est ample, pour tout entier $m$ suffisamment positif, le $K$-espace vectoriel $H^0(X_\eta,A_\eta^{\otimes m})$ (où $\eta$ est le point générique de $C$) est engendré[^4] par $H^0(X,A^{\otimes m})$. Quitte à remplacer $A$ par l’une de ses puissance tensorielle, on peut supposer que cette propriété est vérifiée pour tout entier $m\geqslant 1$. On déduit alors de la relation que $$\operatorname{Im}(\xymatrix{\relax H^0(X_\eta,A_\eta^{\otimes n})\ar[r]^-{\cdot s_\eta^{n}}&H^0(X_\eta,L_\eta^{\otimes np})})\subset E_{np,K}.$$ Cela montre que le système linéaire $E_{\sbullet,\eta}$ contient un diviseur ample. Comme le système linéaire gradué $V_\sbullet$ contient un diviseur ample, on obtient que, pour tout entier $n$ assez positif, le morphisme rationnel de $X$ vers $\mathbb P(V_n)$ défini par le système linéaire est birationnel. Si $p\geqslant 1$ est un entier, on désigne par $u_p:X_p\rightarrow X$ l’éclatement de $X$ le long du lieu de base de $V_p$, défini comme $$X_p=\mathrm{Proj}\Big(\bigoplus_{n\geqslant 0}(\varphi\pi)^*(\mathrm{Sym}^n(V_p))\longrightarrow L^{\otimes np}\Big).$$ Soient en outre $j_p:X_p\rightarrow\mathbb P(V_p)$ le morphisme canonique et $L_p$ le tire en arrière du faisceau universel $\mathcal O_{V_p}(1)$ à $X_p$. Il existe alors $N\in\mathbb N$ tel que $j_p$ définisse un morphisme birationnel entre $X_p$ et son image dans $\mathbb P(V_p)$ dès que $p>N$. Soit $p$ un tel entier. Considérons les homomorphisme de $k$-espaces vectoriels comme ci-dessous $$\xymatrix{\relax\mathrm{Sym}^n(V_p)\ar[r]&H^0(j_p(X_p),\mathcal O_{V_p}(n))\ar[r]&H^0(X_p,L_p^{\otimes n})\ar[r]&H^0(X_p,u_p^*(L^{\otimes np}))}.$$ Le premier morphisme est surjectif pour $n$ assez positif car on peut identifier $\mathrm{Sym}^n(V_p)$ à $H^0(\mathbb P(V_p),\mathcal O_{V_p}(n))$. Le deuxième homomorphisme est injectif car $j_p:X_p\rightarrow j_p(X_p)$ est un morphisme birationnel et $L_p^{\otimes n}\cong j_p^*\mathcal O_{V_p}(n)$. Le dernier homomorphisme est défini comme la multiplication par la $k^{\text{i\`eme}}$ puissance de la section qui détermine le diviseur exceptionnel de l’éclatement $u_p:X_p\rightarrow X$, donc est aussi injectif. Si on identifie $H^0(X_p,u_p^*(L^{\otimes np}))$ à $H^0(X,L^{\otimes np})$ (on peut faire ça car le schéma $X$ est supposé être normal, cf. [@EGAIII_1 corollaire 4.3.12]), l’image de l’homomorphisme composé s’identifie à $\operatorname{Im}(\mathrm{Sym}^pV_n\rightarrow V_{np})$. En outre, comme le morphisme $j_p:X_p\rightarrow j_p(X_p)$ est birationnel, on a $\mathrm{vol}(L_p)=\mathrm{vol}(\mathcal O_{V_p}(n)|_{j_p(X_p)})$. Cela montre que le volume du faisceau inversible $L_p$ est égale à celui du système linéaire gradué $$V^{[p]}_\sbullet:=\bigoplus_{n\geqslant 0}\operatorname{Im}(\mathrm{Sym}^n(V_p)\longrightarrow V_{np}).$$ Soit $E^{[p]}_{\sbullet}:=\bigoplus_{n\geqslant 0}\operatorname{Im}(\mathrm{Sym}^n(E_p)\rightarrow E_{np})$. C’est un système gradué en fibrés vectoriels de $L^{\otimes p}$. Pour tout entier $n$ assez positif, on a $E_n^{[p]}\subset(\pi u_p)_*(L_p^{\otimes n})$. On obtient alors $$\mathrm{vol}(V_\sbullet)\geqslant\frac{\mathrm{vol}(V^{[p]}_{\sbullet})}{p^{\dim(X)}}=\frac{\mathrm{vol}(L_p)}{p^{\dim(X)}}\geqslant\frac{\mathrm{vol}(E^{[p]}_{\sbullet})}{p^{\dim(X)}}.$$ D’après le théorème d’approximation de Fujita pour les systèmes linéaires gradués en fibrés adéliques (qui sont plus généraux que les fibrés vectoriels, cf. [@Boucksom_Chen théorème 2.9]), on a $$\sup_{p\geqslant 1}\frac{\mathrm{vol}(E_\sbullet^{[p]})}{p^{\dim(X)}}=\mathrm{vol}(E_\sbullet).$$ On obtient donc $\operatorname{vol}(V_\sbullet)\geqslant\operatorname{vol}(E_\sbullet)$. Enfin, comme $V_n\subset H^0(C,E_n)$ pour tout $n\in\mathbb N$, on obtient $\operatorname{vol}(V_\sbullet)\leqslant\operatorname{vol}(E_\sbullet)$. La démonstration est donc achevée. \[Def:imagedirecte\] Soient $X$ un schéma intègre projectif défini sur $k$ et $L$ un faisceau inversible gros sur $X$. Soit $V_\sbullet$ un système linéaire gradué de $L$ qui contient un diviseur ample. Pour tout entier $n\geqslant 0$, soit $\pi_*(V_n)$ le sous-$\mathcal O_C$-module de $\pi_*(L^{\otimes n})$ engendré par $V_n$. On désigne par $\pi_*(V_\sbullet)$ le système gradué en fibrés vectoriels $\bigoplus_{n\geqslant 0}\pi_*(V_n)$ et par $\mu_{\max}^\pi(V_\sbullet)$ la quantité $\mu_{\max}^{\mathrm{asy}}(\pi_*(V_\sbullet))$, appelée la *pente maximale asymptotique* de $V_\sbullet$ relativement à $\pi$. \[Rem:comparaisonmu\] Soient $L$ et $M$ deux faisceau inversibles gros sur $X$, et $V_\sbullet$ et $W_\sbullet$ des systèmes linéaires gradués de $L$ et $M$ respectivement. On suppose que $V_\sbullet$ et $W_\sbullet$ contiennent des diviseurs amples. S’il existe une section non-nulle $s$ de $L^\vee\otimes M$ telle que $$\forall\,n\geqslant 1,\quad \operatorname{Im}(\xymatrix{\relax V_n\ar[r]^-{s^n\cdot}&H^0(X,M^{\otimes n})})\subset W_n,$$ on dit que $V_\sbullet$ est *contenu* dans $W_\sbullet$ (via la section $s$). Il s’avère que la multiplication par $s^n$ définit aussi un homomorphisme injectif de $\pi_*(V_n)$ vers $\pi_*(W_n)$. On obtient donc $\mu_{\max}(V_n)\leqslant\mu_{\max}(W_n)$, qui implique la relation $\mu_{\max}^{\pi}(V_\sbullet)\leqslant\mu_{\max}^{\pi}(W_\sbullet)$. Similairement, on a $\operatorname{vol}(V_\sbullet)\leqslant \operatorname{vol}(W_\sbullet)$. Tour de fibrations sur courbes {#Sec:tourdefibration} ============================== Soient $k$ un corps et $X$ un schéma projectif et intègre de dimension $d+1$ sur $\operatorname{Spec}k$, où $d$ est un entier, $d\geqslant 0$. Par *tour de fibrations sur courbes* de $X$, on entend toute donnée $(p_i:X_i\rightarrow C_i)_{i=0}^{d}$ où (a) lorsque $d=0$, $C_0$ et $X_0$ sont tous les deux la normalisation du schéma $X$ et $p_0:X_0\rightarrow C_0$ est le morphisme d’identité, (b) lorsque $d\geqslant 1$, $C_0$ est une courbe projective régulière sur $\operatorname{Spec}k$, $X_0=X$ et $p_0$ est un $k$-morphisme projectif et plat[^5] de $X$ vers $C_0$, (c) de façon récursive, pour tout $i\in\{1,\ldots,d-1\}$, $C_{i}$ est une courbe projective régulière définie sur le corps $R(C_{i-1})$ des fonctions rationnelles sur $C_{i-1}$, $X_{i}$ est la fibre générique de $p_{i-1}$ et $p_{i}:X_{i}\rightarrow C_{i}$ est un morphisme projectif et plat de $R(C_{i-1})$-schémas, (d) $C_{d}$ est la normalisation de la fibre générique de $p_{d-1}$ et $p_d:C_{d}\rightarrow C_d$ est le morphisme d’identité. Si $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^{d-1}$ est un tour de fibrations sur courbes du schéma $X$, on désigne par $g(\Theta)$ le vecteur $(g(C_0),\ldots,g(C_d))\in\mathbb N^{d+1}$, où $g(C_i)$ est le genre de la courbe $C_i$. Le vecteur $g(\Theta)$ est appelé le *genre* de $\Theta$. Soient $X$ un schéma projectif et intègre de dimension $d+1$ sur $\operatorname{Spec}k$, où $d\geqslant 2$. Si $(p_i:X_i\rightarrow C_i)_{i=0}^{d}$ est un tour de fibrations sur courbes du $k$-schéma $X$, alors $(p_j:X_j\rightarrow C_j)_{j=i}^d$ est un tour de fibrations sur courbes du $R(C_{i-1})$-schéma $X_i$. Si le schéma $X$ est normal, alors $C_d$ s’identifie à la fibre générique de $p_{d-1}$ (cela provient de la préservation de la clôture intégrale par la localisation). On entend par *modification birationnelle* d’un $k$-schéma projectif et intègre $X$ tout morphisme $f:X'\rightarrow X$ d’un schéma projectif et intègre $X'$ vers $X$ qui est birationnel (autrement dit, $f$ induit un isomorphisme entre les corps des fonctions rationnelles de $X$ et de $X'$). La proposition suivante montre que l’existence d’un tour de fibrations sur courbes de genre fixé est une propriété invariante par toute modification birationnelle. Soit $X$ un schéma projectif et intègre de dimension $d+1$ sur $\operatorname{Spec}k$. On suppose que le schéma $X$ admet un tour de fibrations sur courbes de genre $(g_0,\ldots,g_d)$. Alors, pour toute modification birationnelle $f:X'\rightarrow X$, le schéma $X'$ admet aussi un tour de fibrations sur courbes de même genre. On suppose que $C_0$ est une courbe projective régulière de genre $g_0$ sur $\operatorname{Spec}k$ et que $p_0:X\rightarrow C_0$ est un $k$-morphisme projectif et plat. Alors le morphisme composé $fp_0$ est un morphisme projectif et plat de $X'$ vers $C_0$. De plus, le morphisme canonique de la fibre générique de $fp_0$ vers celle de $f$ est un $R(C_0)$-morphisme projectif et birationnel, où $R(C_0)$ désigne le corps des fonctions rationnelles sur $C_0$. Par récurrence sur la dimension de $X$, on obtient le résultat. Soient $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$ et $\Theta$ un tour de fibrations sur courbes du schéma $X$. La proposition précédente non seulement montre que toute modification birationnelle admet un tour de fibrations sur courbes de même genre que celui de $\Theta$, sa démonstration construit effectivement un tel tour de fibrations sur courbes pour toute modification birationnelle $f:X'\rightarrow X$, que l’on notera comme $f^*(\Theta)$. Si $\Theta$ est de la forme $(p_i:X_i\rightarrow C_i)_{i=0}^d$, alors le $i^{\text{\`eme}}$ morphisme dans $f^*(\Theta)$ est obtenu comme le composé de $p_i$ avec une modification birationnelle de $X_i$. En outre, on peut vérifier que, si $f_1:X'\rightarrow X$ et $f_2:X''\rightarrow X'$ sont des modifications birationnelles successives, alors on a $(f_1f_2)^*\Theta=f_2^*(f_1^*\Theta)$. Étant donné un schéma projectif et intègre $X$ défini sur $k$, le choix d’un tour $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^d$ de fibrations sur courbes de $X$ définit une chaîne $$k\subset R(C_0)\subset\ldots\subset R(C_d)=R(X)$$ de sous-extension de $R(X)/k$, où $R(X)$ est le corps des fonctions rationnelles sur $X$. Chaque extension consécutive dans la chaîne est transcendante de degré de transcendance $1$. Réciproquement, si on fixe une chaîne $$k=k_{-1}\subset k_0\subset\ldots \subset k_d=R(X)$$ de sous-extension de $R(X)/k$ de sorte que chaque extension $k_i/k_{i-1}$ est transcendante de degré de transcendance $1$, alors il existe une modification birationnelle $X'$ de $X$ qui possède un tour de fibrations sur courbes $\Theta=(p_i:X_i'\rightarrow C_i)_{i=0}^d$ tel que $k_i=R(C_i)$ quel que soit $i\in\{0,\ldots,d\}$. En effet, on peut choisir $C_0$ comme la courbe projective régulière définie sur $\operatorname{Spec}k$ telle que $R(C_0)=k_0$. L’inclusion de $k_0$ dans $R(X)$ définit un $k$-morphisme rationnel de $X$ vers $C_0$. Quitte à éclater le lieu où ce morphisme rationnel n’est pas défini, on obtient une modification birationnelle de $X$ muni d’un $k$-morphisme projectif et plat vers $C_0$. Par un procédé de récurrence, on peut construire une modification birationnelle de $X$ qui possède un tour de fibrations sur courbes vérifiant les propriétés comme ce que l’on a décrit plus haut. Soit $X$ un schéma projectif et intègre de dimension $d+1$ sur $\operatorname{Spec}k$. On suppose que $X$ admet un tour de fibrations sur courbes $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^d$. Soient $L$ un faisceau inversible gros sur $X$ et $V_\sbullet$ un système linéaire gradué de $L$, qui contient un diviseur ample. Si $d=0$, on désigne par $\operatorname{vol}^{\Theta}(V_\sbullet)$ le volume de $V_\sbullet$ et par $\mu^{\Theta}_{\max}(V_\sbullet)$ la pente maximale asymptotique de $V_\sbullet$ relativement au morphisme d’identité de la normalisation de $X$, où on considère $V_\sbullet$ comme un système linéaire gradué du tire en arrière du faisceau inversible $L$ sur la normalisation de $X$. Si $d\geqslant 1$, de façon récursive, on désigne par $\operatorname{vol}^{\Theta}(V_\sbullet)$ le vecteur $$\big(\operatorname{vol}(V_\sbullet),\operatorname{vol}^{\Theta'}(p_{0*}(V_\sbullet)_{\eta_0})\big),$$ où $\Theta'=(p_i:X_i\rightarrow C_i)_{i=1}^d$ (qui est un tour de fibrations sur courbes de $X_1$), $\eta_0$ est le point générique de la courbe $C_0$ et $p_{0*}(V_\sbullet)_{\eta_0}$ est la fibre générique de $p_{0*}(V_\sbullet)$, qui est un système linéaire gradué du tire en arrière de $L$ sur $X_1$ contenant un diviseur ample. De façon similaire, on désigne par $\mu_{\max}^{\Theta}(V_\sbullet)$ le vecteur[^6] $$\big(\mu_{\max}^{p_0}(V_\sbullet),\mu_{\max}^{\Theta'}(p_{0*}(V_\sbullet)_{\eta_0})\big).$$ Si $f:X'\rightarrow X$ est une modification birationnelle du schéma $X$, alors on a $$\operatorname{vol}^{f^*\Theta}(V_\sbullet)=\operatorname{vol}^{\Theta}(V_\sbullet)\quad\text{et}\quad \mu_{\max}^{f^*\Theta}(V_\sbullet)=\mu_{\max}^{\Theta}(V_\sbullet),$$ où on considère $V_\sbullet$ comme un système linéaire gradué de $f^*(L)$. Soient $X$ un schéma projectif et intègre de dimension $d+1$ sur $\operatorname{Spec}k$ et $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^d$ un tour de fibration sur courbes de $X$. Soient $L$ est un faisceau inversible gros sur $X$ et $V_\sbullet$ un système linéaire gradué de $L$. Soient $(v_0,\ldots,v_d)$ le vecteur $\operatorname{vol}^{\Theta}(V_\sbullet)$ et $(\mu_0,\ldots,\mu_d)$ le vecteur $\mu_{\max}^{\Theta}(V_\sbullet)$. Alors on a $$\forall\,i\in\{0,\ldots,d\},\quad v_i\leqslant\operatorname{vol}(L|_{X_i})\quad\text{et}\quad\mu_i\leqslant\mu^{p_i}_{\max}(L|_{X_i})$$ En outre, si $V_\sbullet$ est le système linéaire gradué total $\bigoplus_{n\geqslant 0}H^0(X,L^{\otimes n})$, alors on a $\mu_i=\mu_{\max}^{p_i}(L|_{X_i})$ pour tout $i\in\{0,\ldots,d\}$. Cependant, en général l’égalité $v_i=\operatorname{vol}(L|_{X_i})$ n’est pas vraie lorsque $i\geqslant 1$. En effet, le système gradué en fibrés vectoriels $p_{0*}(V_\sbullet)$ ne tient compte que la partie positive du système gradué total $\bigoplus_{n\geqslant 0}p_{0*}(L^{\otimes n})$ (qui permet cependant de retrouver le volume de $V_\sbullet$). Les égalités $v_i=\operatorname{vol}(L|_{X_i})$ ($i\in\{0,\ldots,d\}$) sont vérifiées notamment lorsque $L$ est ample. Estimation explicite la fonction de Hilbert-Samuel {#Sec:HS} ================================================== Le but de ce paragraphe est de démontrer le théorème \[Thm:HMgeometryique\]. On fixe un corps commutatif $k$ de caractéristique zéro. Soit $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$. On suppose donné un tour de fibrations sur courbes $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^d$ de $X$. Pour tout système linéaire gradué $V_\sbullet$ d’un faisceau inversible gros $L$ sur $X$, on introduit un invariant birationnel $\varepsilon^{\Theta}(V_\sbullet)$, construit dans la suite. Si $d=0$, alors on définit $$\varepsilon^{\Theta}(V_\sbullet):=\max(g(\Theta)-1,1).$$ Lorsque $d\geqslant 1$, on désigne par $W_\sbullet$ la fibre générique de $p_{0*}(V_\sbullet)$, qui est un système linéaire gradué de $L|_{X_1}$ contenant un diviseur ample. Soit en outre $\Theta':=(p_i:X_i\rightarrow C_i)_{i=1}^d$, qui est un tour de fibrations sur courbes de $X_1$. Alors l’invariant $\varepsilon^{\Theta}(V_\sbullet)$ est définie de façon récursive comme $$\label{Equ:termeerreur}\varepsilon^{\Theta}(V_\sbullet)=\mu_0\varepsilon^{\Theta'}(W_\sbullet)+\Big(\frac{\operatorname{vol}(W_\sbullet)}{d!}+\varepsilon^{\Theta'}(W_\sbullet)\Big)\max(g_0-1,1),$$ où $\mu_0$ et $g_0$ sont respectivement les premières coordonnées des vecteurs $\mu_{\max}^{\Theta}(V_\sbullet)$ et $g(\Theta)$. On voit aussitôt de la définition que, si $V_\sbullet$ est contenu dans un autre système linéaire gradué $V'_\sbullet$, alors on a (cf. la remarque \[Rem:comparaisonmu\]) $$\label{Equ:compaeps}\varepsilon^{\Theta}(V_\sbullet)\leqslant\varepsilon^{\Theta}(V_\sbullet').$$ Pour tout entier $p\geqslant 1$, soit $V^{(p)}_{\sbullet}$ le système linéaire gradué $\bigoplus_{n\geqslant 0}V_{np}$ de $L^{\otimes p}$. La fibre générique de $p_{0*}(V_\sbullet^{(p)})$ s’identifie à $W^{(p)}_\sbullet$. On a $\operatorname{vol}(W^{(p)}_\sbullet)=p^d\operatorname{vol}(W^{(p)})$. En outre, on a $\mu_{\max}^{\Theta}(V_\sbullet^{(p)})=p\mu_{\max}^{\Theta}(V_\sbullet)$. On obtient alors de la formule récursive que $$\varepsilon^\Theta(V_\sbullet^{(p)})\leqslant p^d\varepsilon^{\Theta}(V_\sbullet)$$ \[Thm:majorationdeHS\] Soit $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$ muni d’un tour de fibrations sur courbes $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^d$, et $L$ un faisceau inversible sur $X$, où $k$ est un corps de caractéristique zéro. Si $V_\sbullet=\bigoplus_{n\geqslant 0}V_n$ est un système linéaire gradué de $L$, qui contient un diviseur ample, alors on a $$\label{Equ:majoration}\operatorname{rg}_k(V_1)\leqslant\mathrm{vol}(V_\sbullet)+\varepsilon^{\Theta}(V_\sbullet).$$ Les invariants figurant à droite de l’inégalité sont des invariants birationnels. Quitte à passer à la normalisation de $X$, on peut supposer que $X$ est un schéma normal. On raisonne par récurrence sur $d$. Dans le cas où $d=0$, le schéma $X$ est une courbe régulière de genre $g(\Theta)$. Pour tout entier $n\geqslant 0$, soit $L_n$ le sous-$\mathcal O_X$-module de $L^{\otimes n}$ engendré par $V_n$. D’après le théorème \[Thm:passageauxfibres\], $L_\sbullet=\bigoplus_{n\geqslant0}L_n$ est un système gradué en fibrés vectoriels de $L$, qui contient un diviseur ample et vérifie $\operatorname{vol}(L_\sbullet)=\operatorname{vol}(V_\sbullet)$. En outre, l’homomorphisme $L_n\otimes L_m\rightarrow L_{n+m}$ est non-nul dès que $L_n$ et $L_m$ sont tous non-nuls. On obtient alors (cf. la remarque \[Rem:systemegradue\]) $$\operatorname{vol}(V_\sbullet)=\operatorname{vol}(L_\sbullet)=\lim_{n\rightarrow+\infty}\frac{\deg(L_n)}{n}=\sup_{n\geqslant 1}\frac{\deg(L_n)}{n}$$ En particulier, si $V_1$ est non-nul, alors $$\operatorname{rg}_k(V_1)\leqslant h^0(L_1)\leqslant\deg(L_1)+\max(g(X)-1,1)\leqslant\operatorname{vol}(V_\sbullet)+\max(g(X)-1,1).$$ Cette inégalité est aussi vraie lorsque $\operatorname{rg}_k(V_1)=\{0\}$. Le théorème est donc démontré pour le cas où $d=0$. Traitons maintenant le cas général. Soient $$(v_0,\ldots,v_d)=\operatorname{vol}^\Theta(V_\sbullet),\;(\mu_0,\ldots,\mu_d)=\mu_{\max}^{\Theta}(V_\sbullet)\;\text{ et }\;(g_0,\ldots,g_d)=g(\Theta).$$ Soient $E_\sbullet:=p_{0*}(V_\sbullet)$ et $W_\sbullet$ la fibre générique de $E_\sbullet$. D’après le théorème \[Thm:passageauxfibres\], $E_\sbullet=\bigoplus_{n\geqslant 0}E_n$ est un système gradué en fibrés vectoriels de $L$ qui contient un diviseur ample et vérifie la relation $\operatorname{vol}(E_\sbullet)=\operatorname{vol}(V_\sbullet)$. Soit $\eta$ la fibre générique de $C_0$. Pour tout nombre réel $t$ et tout entier $n\in\mathbb N$, soit $W^{t}_n:=(\mathcal F^{nt}E_n)_\eta$, où $\mathcal F$ est la $\mathbb R$-filtration de Harder-Narasimhan sur $E_n$. D’après le corollaire \[Cor:filtrationdeHNenalg\], $W^t_\sbullet=\bigoplus_{n\geqslant 0}W_n^t$ est un système linéaire gradué de $L_\eta$. En outre, $W_\sbullet^t$ contient un diviseur ample lorsque $t<\mu_0$, et devient trivial lorsque $t>\mu_0$ (cf. [@Boucksom_Chen lemme 1.6]), et on a (cf. [@Boucksom_Chen corollaire 1.13]) $$\label{Equ:volcommeinte} \operatorname{vol}(E_\sbullet)=(d+1)\int_0^{\mu_0}\operatorname{vol}(W_\sbullet^t)\,\mathrm{d}t.$$ Si $E_1$ est un fibré vectoriel non-nul, d’après la formule on obtient[^7] $$\deg_+(E_1)=\int_0^{\mu_0}\operatorname{rg}(W_1^t)\,\mathrm{d}t.$$ On applique l’hypothèse de récurrence à $W_\sbullet^t$ pour chaque $t$ et obtient $$\label{Equ:majoratiodegplus}\deg_+(E_1)\leqslant\int_0^{\mu_0}\Big(\frac{\operatorname{vol}(W_1^t)}{d!}+\varepsilon^{\Theta'}(W_\sbullet^t)\Big)\,\mathrm{d}t\leqslant\frac{\operatorname{vol}(E_\sbullet)}{(d+1)!}+\mu_0\varepsilon^{\Theta'}(W_\sbullet),$$ où $\Theta'=(p_i:X_i\rightarrow C_i)_{i=1}^d$. Dans la deuxième inégalité, on a utilisé les relations et . Enfin, le théorème \[Thm:h0etdegplus\] montre que $$\begin{split}h^0(E_1)&\leqslant\deg_+(E_1)+\operatorname{rg}(W_1)\max(g_0-1,1)\\ &\leqslant\frac{\operatorname{vol}(E_\sbullet)}{(d+1)!}+\mu_0\varepsilon^{\Theta'}(W_\sbullet)+\Big(\frac{v_1}{d!}+\varepsilon^{\Theta'}(W_\sbullet^{\Theta'})\Big)\max(g_0-1,1), \end{split}$$ où on a appliqué l’hypothèse de récurrence à $W_\sbullet$ dans la deuxième inégalité. Comme $\operatorname{rg}(V_1)\leqslant h^0(E_1)$ et comme $\operatorname{vol}(E_\sbullet)=\operatorname{vol}(V_\sbullet)$, on obtient $$\operatorname{rg}(V_1)\leqslant\frac{\operatorname{vol}(V_\sbullet)}{(d+1)!}+\mu_0\varepsilon^{\Theta'}(W_\sbullet)+\Big(\frac{v_1}{d!}+\varepsilon^{\Theta'}(W_\sbullet)\Big)\max(g_0-1,1).$$ La démonstration est donc achevée, compte tenu de la formule récursive définissant $\varepsilon^{\Theta}(V_\sbullet)$. Dans la démonstration de l’inégalité , on a seulement utilisé le fait que l’algèbre graduée $W_\sbullet$ est filtrée par la $\mathbb R$-filtration de Harder-Narasimhan. Ainsi le théorème se généralise naturellement dans le cadre de système linéaire gradué filtré et conduit au corollaire suivant qui sera utile plus loin dans l’étude des systèmes linéaires gradués arithmétiques. \[Cor:majHS\] Soit $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$ muni d’un tour de fibrations sur courbes $\Theta$. Soient $L$ un faisceau inversible gros sur $X$ et $V_\sbullet$ un système linéaire gradué de $L$ qui contient un diviseur ample. On suppose que chaque espace vectoriel $V_n$ est muni d’une $\mathbb R$-filtration $\mathcal F$ de sorte que $$\label{Equ:sousmult}(\mathcal F^aV_n)\cdot(\mathcal F^bV_m)\subset \mathcal F^{a+b}V_{n+m}$$ pour tout $(a,b)\in\mathbb R^2$ et tout $(n,m)\in\mathbb N^2$. Soit en outre $$\lambda_{\max}^{\mathrm{asy}}(V_\sbullet)=\sup_{n\geqslant 1}\frac{\sup\{t\,|\,\mathcal F^tV_n\neq\{0\}\}}{n}.$$ Alors on a $$\int_0^{+\infty}\operatorname{rg}_k(\mathcal F^tV_1)\,\mathrm{d}t\leqslant\int_0^{+\infty}\frac{\operatorname{vol}(V^t_\sbullet)}{d!}\,\mathrm{d}t+\lambda_{\max}^{\mathrm{asy}}(V_\sbullet)\varepsilon^{\Theta}(V_\sbullet),$$ où $d=\dim(X)$ et $V_\sbullet^t=\bigoplus_{n\geqslant 0}\mathcal F^{nt}V_n$. Pour tout entier $n\geqslant 1$, on note $$\lambda_{\max}(V_n)=\sup\{t\,|\,\mathcal F^tV_n\neq\{0\}\}.$$ La condition montre que la suite $(\lambda_{\max}(V_n))_{n\geqslant 1}$ est sur-additive, d’où $\lambda_{\max}(V_n)\leqslant n\lambda_{\max}^{\mathrm{asy}}(V_\sbullet)$. Par conséquent, pour tout entier $n\geqslant 1$ et tout $t>\lambda_{\max}^{\mathrm{asy}}(V_\sbullet)$, on a $V_n^t=\mathcal F^{nt}V_n=\{0\}$. En outre, d’après [@Boucksom_Chen lemme 1.6], pour tout $t<\lambda_{\max}^{\mathrm{asy}}(V_\sbullet)$, le système linéaire gradué $V_\sbullet^t$ contient un diviseur ample. On applique le théorème \[Thm:majorationdeHS\] à $V_\sbullet^t$ et obtient $$\operatorname{rg}_k(\mathcal F^tV_1)\leqslant\frac{\operatorname{vol}(V_\sbullet^t)}{d!}\,\mathrm{d}t+\varepsilon^{\Theta}(V_\sbullet^t)\leqslant \frac{\operatorname{vol}(V_\sbullet^t)}{d!}\,\mathrm{d}t+\varepsilon^{\Theta}(V_\sbullet),\quad t<\lambda_{\max}^{\mathrm{asy}}(V_\cdot).$$ L’intégration de cette inégalité pour $t\in [0,\lambda_{\max}^{\mathrm{asy}}(V_\sbullet)[$ conduit à l’inégalité souhaitée. Fibrés vectoriels adéliques =========================== Dans ce paragraphe, l’expression $K$ désigne un corps de nombres. Soit $\mathcal O_K$ la fermeture intégrale de $\mathbb Z$ dans $K$. On entend par *place* de $K$ toute classe d’équivalence de valeurs absolues non-triviales sur $K$, où deux valeurs absolues sont dites équivalentes si elles définissent la même topologie sur $K$. On désigne par $M_K$ l’ensemble de toutes les places de $K$. Pour toute place $v$ de $K$, on désigne par $|.|_v$ une valeur absolue dans la place $v$ qui prolonge soit la valeur absolue usuelle sur $\mathbb Q$, soit l’une des valeurs absolues $p$-adiques. On désigne par $K_v$ le complété de $K$ par rapport à la valeur absolue $|.|_v$. Il s’avère que $|.|_v$ s’étend de façon unique sur la clôture algébrique $\overline K_v$. On désigne par $\mathbb C_v$ le complété de $\overline K_v$, qui est à la fois algébriquement clos et complet. On rappelle que la famille $(|.|_v)_{v\in M_K}$ de valeurs absolues vérifie la formule du produit $$\label{Equ:formuleduproduit} \forall\, a\in K^{\times},\; \sum_{v\in M_K}\frac{[K_v:\mathbb Q_{v}]}{[K:\mathbb Q]}\ln |a|_v=0.$$ Il s’avère que l’ensemble des places de $K$ qui ne prolongent pas $\infty\in M_{\mathbb Q}$ (représentant la valeur absolue usuelle de $\mathbb Q$) correspond biunivoquement à l’ensemble des points fermés de $\operatorname{Spec}\mathcal O_K$. Soit $E$ un espace vectoriel de rang fini sur $K$. Si $\boldsymbol{e}=(e_1,\ldots,e_n)$ est une base de $E$, alors elle définit pour chaque place $v\in M_K$ une norme $\|.\|_{\boldsymbol{e},v}$ sur $E\otimes_K\mathbb C_v$ telle que $$\|\lambda_1e_1+\cdots+\lambda_ne_n\|_{\boldsymbol{e},v}=\begin{cases} \max\{|\lambda_1|_v,\ldots,|\lambda_n|_v\},&\text{si $v$ est non-archim\'edienne},\\ (|\lambda_1|_v^2+\cdots+|\lambda_n|_v^2)^{1/2},&\text{si $v$ est archim\'edienne}. \end{cases}$$ Cette norme est invariante sous l’action du groupe de Galois $\mathrm{Gal}(\mathbb C_v/K_v)$. On appelle *fibré vectoriel adélique* (cf. [@Gaudron08]) sur $\operatorname{Spec}K$ toute donnée $\overline E=(E,(\|.\|_{v}))$ d’un espace vectoriel de rang fini $E$ sur $K$ et une famille de normes, où $\|.\|_v$ est une norme sur $E\otimes_K\mathbb C_v$, qui est invariante sous l’action du groupe de Galois $\mathrm{Gal}(\mathbb C_v/K_v)$, et est ultramétrique lorsque $v$ est une place non-archimédienne. On demande en plus l’existence d’une base $\boldsymbol{e}$ de $E$ telle que $\|.\|_v=\|.\|_{\boldsymbol{e},v}$ pour toute sauf un nombre fini de places $v\in M_K$. Si $E$ est de rang $1$ sur $K$, on dit que $\overline E$ est un *fibré inversible adélique*. Si, pour toute place archimédienne $v$, la norme $\|.\|_v$ est hermitienne, on dit que $\overline E$ est *hermitien*. Soit $\overline L$ un fibré inversible adélique sur $\operatorname{Spec}K$. On définit le *degré d’Arakelov* de $\overline L$ comme $$\operatorname{\widehat{\deg}}(\overline L):=-\sum_{v\in M_K}[K_v:\mathbb Q_{v}]\ln\|s\|_v,$$ où $s$ est un élément non-nul de $L$. D’après la formule du produit , cette définition ne dépend pas du choix de $s$. On introduit aussi la version normalisée du degré d’Arakelov comme $$\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline L):=\frac{\deg(L)}{[K:\mathbb Q]}.$$ La notion de *fibré vectoriel normé* (ou *hermitien*) sur $\operatorname{Spec}\mathcal O_K$ est aussi largement utilisée dans la littérature. Rappelons qu’un fibré vectoriel normé sur $\operatorname{Spec}\mathcal O_K$ est la donnée $\overline{\mathcal E}$ d’un $\mathcal O_K$-module projectif et de type fini muni d’une famille de norme $(\|.\|_v)_{v\in M_{K,\infty}}$ indexée par l’ensemble des places archimédiennes de $K$, où chaque $\|.\|_{v}$ est une norme sur $\mathcal E\otimes_{\mathcal O_K}\mathbb C_v$, invariante sous l’action du groupe de Galois $\mathrm{Gal}(\mathbb C/K_v)$. Le fibré vectoriel normé $\overline{\mathcal E}$ est dit hermitien si chaque norme $\|.\|_v$ est hermitienne. Étant donné un fibré vectoriel normé $\overline{\mathcal E}$ sur $\operatorname{Spec}\mathcal O_K$, on obtient naturellement une structure de fibré vectoriel adélique pour $E=\mathcal E_K$ (qui est hermitienne lorsque $\overline{\mathcal E}$ est hermitien), où la norme en une place non-archimédienne $\mathfrak p$ est induite par la structure de $\mathcal O_K$-module de $\mathcal E$ : la boule unité fermée de $\|.\|_{\mathfrak p}$ est $\mathcal E\otimes_{\mathcal O_K}\mathcal O_{\mathfrak p}$, où $\mathcal O_{\mathfrak p}$ est l’anneau de valuation de $\mathbb C_{\mathfrak p}$. Réciproquement, un fibré vectoriel adélique sur $\operatorname{Spec}K$ provient nécessairement d’un fibré vectoriel normé sur $\operatorname{Spec}\mathcal O_K$ pourvu que toutes ses normes indexées par les places non-archimédiennes sont pures[^8]. On renvoie les lecteurs dans [@Gaudron09 proposition 3.10] pour une démonstration. Filtration de Harder-Narasimhan ------------------------------- Soit $\overline E=(E,(\|.\|_v)_{v\in M_K})$ un fibré vectoriel adélique sur $\operatorname{Spec}K$ qui est hermitien. Pour toute place $v\in M_K$, la norme $\|.\|_v$ sur $E_{\mathbb C_v}$ induit une norme sur ${\mathrm{d\acute{e}t}}(E_{\mathbb C_v})$ qui est une ultranorme (resp. une norme hermitienne) lorsque $v$ est une place ultramétrique (resp. une place archimédienne), et telle que $$\forall\,(s_1,\ldots,s_r)\in E^r,\quad\|s_1\wedge\cdots\wedge s_r\|_v=\prod_{i=1}^r\|s_i\|_v,$$ où $r$ est le rang de $E$. Ainsi ${\mathrm{d\acute{e}t}}(\overline E):=({\mathrm{d\acute{e}t}}E,(\|.\|_v)_{v\in M_K})$ devient un fibré inversible adélique sur $\operatorname{Spec}K$. On définit le *degré d’Arakelov normalisé* de $\overline E$ comme celui de ${\mathrm{d\acute{e}t}}(\overline E)$, noté comme $\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline E)$. Si de plus $E$ est non-nul, on désigne par $\operatorname{\widehat{\mu}}(E)$ le quotient $\operatorname{\widehat{\deg}}_n(\overline E)/\operatorname{rg}(E)$, appelé la *pente* de $\overline E$. Le formalisme de la théorie de Harder-Narasimhan est encore valable dans le cadre des fibrés vectoriels adéliques hermitiens. En particulier, il existe un sous-espace vectoriel $E_{\operatorname{des}}$ de $E$ tel que $$\operatorname{\widehat{\mu}}(\overline E_{\operatorname{des}})=\operatorname{\widehat{\mu}}_{\max}(\overline E):=\sup_{0\neq F\subset E}\operatorname{\widehat{\mu}}(\overline F)$$ et que $E_{\operatorname{des}}$ contient tous les sous-espaces vectoriels non-nuls $F$ de $E$ tel que $\overline F$ admet $\operatorname{\widehat{\mu}}_{\max}(\overline E)$ comme pente. On renvoie les lecteurs dans [@Gaudron08 §5.1] pour les détails. Similairement au cas de fibrés vectoriels sur une courbe, on peut construire de façon récursive un drapeau $$0=E_0\subsetneq E_1\subsetneq\ldots\subsetneq E_n=E$$ de sous-espaces vectoriels de $E$ de sorte que $E_i/E_{i-1}=(E/E_{i-1})_{\operatorname{des}}$ pour tout $i\in\{1,\ldots,n\}$, où on a considéré des normes quotients sur $E/E_{i-1}$. Si on désigne par $\alpha_i$ la pente de $\overline E_i/\overline E_{i-1}$, alors on a $$\alpha_1>\alpha_2>\ldots>\alpha_n.$$ La dernière pente $\alpha_n$ est appelée la *pente minimale* de $\overline E$, notée comme $\operatorname{\widehat{\mu}}_{\min}(\overline E)$. On désigne par $P_{\overline E}$ la fonction concave et affine par morceau définie sur l’intervalle $[0,\operatorname{rg}(E)]$, qui est affine sur chaque intervalle $[\operatorname{rg}(E_{i-1}),\operatorname{rg}(E_i)]$ et de pente $\alpha_i$. Cette fonction est appelée le *polygone de Harder-Narasimhan* de $\overline E$. On désigne par $\mathcal F_{\mathrm{HN}}$ la $\mathbb R$-filtration décroissante sur $E$ telle que $$\mathcal F_{\mathrm{HN}}^t(E)=\sum_{\begin{subarray}{c}1\leqslant i\leqslant n\\ t\leqslant\alpha_i\end{subarray}}E_i.$$ Cette filtration est appelée la *filtration de Harder-Narasimhan* de $\overline E$. \[Def:degplus\] Soit $\overline E$ un fibré vectoriel adélique hermitien sur $\operatorname{Spec}K$. Soit $r$ le rang de $E$. Pour tout $i\in\{1,\ldots,r\}$, on désigne par $\operatorname{\widehat{\mu}}_i(\overline E)$ la pente du polygone de Harder-Narasimhan $P_{\overline E}$ sur l’intervalle $[i-1,i]$, appelée la *$i^{\text{\`eme}}$ pente* de $\overline E$. En outre, on désigne par $\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)$ la valeur maximale du polygone $P_{\overline E}$ sur l’intervalle $[0,r]$. Avec les notation de la définition précédente, on a $$\label{Equ:deg+}\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)=\sum_{\begin{subarray}{c} 1\leqslant i\leqslant r\\ \widehat{\mu}_i(\overline E)\geqslant 0 \end{subarray}}\operatorname{\widehat{\mu}}_i(\overline E)=\int_0^{+\infty}\operatorname{rg}(\mathcal F^t_{\mathrm{HN}}(E))\,\mathrm{d}t.$$ Comme la fonction $P_{\overline E}$ est affine sur chaque intervalle $[i-1,i]$, on obtient que $$\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)=\max_{i\in\{0,\ldots,r\}}P_{\overline E}(i)=\max_{i\in\{0,\ldots,r\}}\sum_{1\leqslant j\leqslant i}\operatorname{\widehat{\mu}}_j(\overline E).$$ En outre, comme la fonction $P_{\overline E}$ est concave, on a $\operatorname{\widehat{\mu}}_1(\overline E)\geqslant\ldots\geqslant\operatorname{\widehat{\mu}}_r(\overline E)$. La première égalité de est donc démontrée. Soit $0=E_0\subsetneq E_1\subsetneq\ldots\subsetneq E_n=E$ le drapeau de Harder-Narasimhan de $\overline E$. Pour tout $j\in\{1,\ldots,n\}$, soit $\alpha_j$ la pente de la fonction $P_{\overline E}$ sur l’intervalle $[\operatorname{rg}(E_{j-1}),\operatorname{rg}(E_j)]$. Par la définition de $\mathcal F^t_{\mathrm{HN}}$, on a $$\mathrm{d}\operatorname{rg}(\mathcal F_{\mathrm{HN}}^t(E))=-\sum_{j=1}^n\big(\operatorname{rg}(E_j)-\operatorname{rg}(E_{j-1})\big)\delta_{\alpha_j}=-\sum_{i=1}^r\delta_{\operatorname{\widehat{\mu}}_i(\overline E)}$$ comme mesures boréliennes sur $\mathbb R$, où $\delta_x$ désigne la mesure de Dirac en $x$. On obtient alors $$\int_0^{+\infty}\operatorname{rg}(\mathcal F_{\mathrm{HN}}^t(E))\,\mathrm{d}t=-\int_{[0,+\infty[}t\,\mathrm{d}\operatorname{rg}(\mathcal F_{\mathrm{HN}}^t(E))=\sum_{\begin{subarray}{c} 1\leqslant i\leqslant r\\ \widehat{\mu}_i(\overline E)\geqslant 0 \end{subarray}}\operatorname{\widehat{\mu}}_i(\overline E).$$ Le résultat suivant relie la fonction $\operatorname{\widehat{\deg}_{\mathrm{n+}}}(.)$ au nombre d’éléments effectifs dans un fibré vectoriel adélique. Pour tout fibré vectoriel adélique $\overline E$ sur $\operatorname{Spec}K$, on désigne par $\widehat{H}^0(\overline E)$ l’ensemble des éléments $s\in E$ tels que $\sup_{v\in M_K}\|s\|_v\leqslant 1$ (un tel élément est appelé une *section effective* de $\overline E$). C’est un ensemble fini. On désigne par $\widehat{h}^0(\overline E)$ le nombre réel $\ln(\#\widehat{H}^0(\overline E))$. Rappelons d’abord un résultat de Gillet et Soulé [@Gillet-Soule91 théorème 2]. Pour tout entier $n\geqslant 1$, on introduit une constante $C(K,n)$ comme la suite $$nd_K\ln(3)+n(r_1+r_2)\ln(2)+\frac{n}{2}\ln|\Delta_K|-r_1\ln(V(B_n)n!)-r_2\ln(V(B_{2n})(2n)!)+\ln((d_Kn)!),$$ où $d_K=[K:\mathbb Q]$, $\Delta_K$ est le discriminant du corps $K$, $B_n$ désigne la boule unité dans $\mathbb R^n$, $V(.)$ est la mesure de Lebesgue, et $r_1$ et $r_2$ sont respectivement le nombre des places réelles et complexes de $K$. Rappelons que la formule de Sterling $$n!\sim \sqrt{2\pi n}(n/e)^n\quad (n\rightarrow+\infty)$$ et la relation $$V(B_n)=\frac{\pi^{n/2}}{\Gamma(n/2+1)}$$ montrent que $$C(K,n)=\frac 12[K:\mathbb Q]n\ln(n)+O(n),\qquad n\rightarrow+\infty.$$ Soit $\overline E$ un fibré vectoriel adélique sur $\operatorname{Spec}K$ qui provient d’un fibré vectoriel hermitien sur $\operatorname{Spec}\mathcal O_K$. Soit $n$ le rang de $E$ sur $K$. Alors on a $$\label{Equ:Gillet-Soule}\Big|\widehat{h}^0(\overline E)-\widehat{h}^0(\overline\omega_{K/\mathbb Q}\otimes \overline E^\vee)- \operatorname{\widehat{\deg}}(\overline E)\Big|\leqslant C(K,n),$$ où $\overline\omega_{K/\mathbb Q}$ est le fibré inversible adélique associé à $\omega_{\mathcal O_K}=\operatorname{Hom}_{\mathbb Z}(\mathcal O_K,\mathbb Z)$ muni de la famille de normes $(\|.\|_v)_{v\in M_{K,\infty}}$ telles que $\|\mathrm{tr}_{K/\mathbb Q}\|_v=1$ pour tout $v\in M_{K,\infty}$. Le fibré inversible adélique $\overline{\omega}_{K/\mathbb Q}$ devrait être considéré comme le faisceau dualisant relative arithmétique de $\operatorname{Spec}K\rightarrow\operatorname{Spec}\mathbb Q$. Son degré d’Arakelov est $\ln|\Delta_K|$. \[Lem:estdeh0\] Soit $\overline E$ un fibré vectoriel adélique non-nul sur $\operatorname{Spec}K$ qui provient d’un fibré vectoriel hermitien sur $\operatorname{Spec}\mathcal O_K$. (a) Si $\operatorname{\widehat{\mu}}_{\max}(\overline E)<0$, alors $\widehat{h}^0(\overline E)=0$. (b) Si $\operatorname{\widehat{\mu}}_{\min}(\overline E)>[K:\mathbb Q]^{-1}\ln|\Delta_K|$, alors $|\widehat{h}^0(\overline E)-\operatorname{\widehat{\deg}}(\overline E)|\leqslant C(K,\operatorname{rg}(E))$. (c) Si $\operatorname{\widehat{\mu}}_{\min}(\overline E)\geqslant 0$, alors $|\widehat{h}^0(\overline E)-\operatorname{\widehat{\deg}}(\overline E)|\leqslant\ln|\Delta_K|\operatorname{rg}(E)+C(K,\operatorname{rg}(E))$. Pour tout nombre réel $t$, on désigne par $\mathcal O(t)$ le fibré inversible adélique sur $\operatorname{Spec}K$ dont l’espace vectoriel sous-jacent s’identifie à $K$ et telle que $\|1\|_{\mathfrak p}=1$ pour toute place non-archimédienne $\mathfrak p$ et que $\|1\|_{\sigma}=\mathrm{e}^{-1}$ pour toute place archimédienne $\sigma$. Le degré d’Arakelov du fibré inversible adélique $\mathcal O(t)$ est alors $[K:\mathbb Q]t$. \(a) On suppose que $\overline E$ possède une section effective non-nulle, qui définit un homomorphisme injectif de $\mathcal O(0)$ vers $\overline E$. D’après l’inégalité de pentes (cf. [@Gaudron08 lemme 6.4]), on obtient $\operatorname{\widehat{\mu}}_{\max}(\overline E)\geqslant 0$. \(b) Comme $\operatorname{\widehat{\mu}}_{\min}(\overline E)>[K:\mathbb Q]^{-1}\ln|\Delta_K|$, on a $$\operatorname{\widehat{\mu}}_{\max}(\overline{\omega}_{K/\mathbb Q}\otimes\overline E{}^\vee)=\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline{\omega}_{K/\mathbb Q})+\operatorname{\widehat{\mu}}_{\max}(\overline E{}^\vee)=[K:\mathbb Q]^{-1}\ln|\Delta_K|-\operatorname{\widehat{\mu}}_{\min}(\overline E)<0.$$ Par le résultat obtenu dans (a), on obtient $\widehat{h}^0(\overline{\omega}_{K/\mathbb Q}\otimes\overline E^\vee)=0$. L’inégalité annoncée provient donc de la formule . \(c) Soient $\varepsilon>0$ un nombre réel et $t=[K:\mathbb Q]^{-1}\ln|\Delta_K|+\varepsilon$. On a alors $\operatorname{\widehat{\mu}}_{\min}(\overline E\otimes\mathcal O(t))>[K:\mathbb Q]^{-1}\ln|\Delta_K|$. D’après (b), on obtient $$\widehat{h}^0(\overline E\otimes\mathcal O(t))\leqslant\operatorname{\widehat{\deg}}(\overline E\otimes\mathcal O(t))+C(K,\operatorname{rg}(E))=\operatorname{\widehat{\deg}}(\overline E)+t\operatorname{rg}(E)+C(K,\operatorname{rg}(E)).$$ En outre, comme $t>0$, on a $\widehat{h}^0(\overline E)\leqslant\widehat{h}^0(\overline E\otimes\mathcal O(t))$. On obtient donc $$\widehat{h}^0(\overline E)-\operatorname{\widehat{\deg}}(\overline E)\leqslant t\operatorname{rg}(E)+C(K,\operatorname{rg}(E)).$$ De plus, l’inégalité implique que $$\widehat{h}^0(\overline E)-\operatorname{\widehat{\deg}}(\overline E)\geqslant\widehat{h}^0(\overline{\omega}_{K/\mathbb Q}\otimes\overline E^\vee)-C(K,\operatorname{rg}(E))\geqslant -C(K,\operatorname{rg}(E)).$$ On obtient alors $$|\widehat{h}^0(\overline E)-\operatorname{\widehat{\deg}}(\overline E)|\leqslant t\operatorname{rg}(E)+C(K,\operatorname{rg}(E)).$$ Comme $\varepsilon$ est arbitraire, on obtient le résultat souhaité. Soit $\overline E$ un fibré vectoriel adélique non-nul sur $\operatorname{Spec}K$ qui provient d’un fibré vectoriel hermitien sur $\operatorname{Spec}\mathcal O_K$. Alors on a $$\label{Equ:h0degset} \big|\widehat{h}^0(\overline E)-[K:\mathbb Q]\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)\big|\leqslant\operatorname{rg}(E)\ln|\Delta_K|+C(K,\operatorname{rg}(E)).$$ Soient $0=E_0\subsetneq E_1\subsetneq\ldots\subsetneq E_n=E$ le drapeau de Harder-Narasimhan de $E$, et $\alpha_i=\operatorname{\widehat{\mu}}(\overline E_i/\overline E_{i-1})$. Soit $j$ le dernier indice dans $\{1,\ldots,n\}$ tel que $\alpha_j\geqslant 0$. Si un tel indice n’existe pas, on prend $j=0$ par convention. Par définition on a $\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)=\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline E_j)$. Si $j=0$, alors $\operatorname{\widehat{\mu}}_{\max}(\overline E)=0$, et donc $\widehat{h}^0(\overline E)=0$. En outre, par convention on a $\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)=0$. L’inégalité est donc triviale. Si $j>0$, alors $\operatorname{\widehat{\mu}}_{\max}(\overline E/\overline E_j)=\alpha_{j-1}<0$. Par conséquent, on a $\widehat{h}^0(\overline E/\overline E_j)=0$ et donc $\widehat{h}^0(\overline E)=\widehat{h}^0(\overline E_j)$. En outre, on a $\operatorname{\widehat{\mu}}_{\min}(\overline E_j)=\alpha_j\geqslant 0$, et par le lemme \[Lem:estdeh0\].(c), on obtient $$\begin{split}&\quad\;\big|\widehat{h}^0(\overline E)-[K:\mathbb Q]\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E)\big|=\big|\widehat{h}^0(\overline E_j)-[K:\mathbb Q]\operatorname{\widehat{\deg}_{\mathrm{n+}}}(\overline E_j)\big|\\ &\leqslant\operatorname{rg}(E_j)\ln|\Delta_K|+C(K,\operatorname{rg}(E_j))\leqslant\operatorname{rg}(E)\ln|\Delta_K|+C(K,\operatorname{rg}(E)). \end{split}$$ Le résultat est donc démontré. Filtration par hauteur ---------------------- Soit $\overline E=(E,(\|.\|_v)_{v\in M_K})$ un fibré vectoriel adélique sur $\operatorname{Spec}K$ (qui n’est pas nécessairement hermitien). Pour tout élément non-nul $s$ de $\overline E$, on désigne par $h_{\overline E}(s)$ le nombre $-\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline{Ks})$, appelé la *hauteur normalisée* de $s$. Plus généralement, si $K'$ est une extension finie de $K$, on peut construire un fibré vectoriel adélique $\overline E\otimes_KK'$ sur $\operatorname{Spec}K'$ dont l’espace vectoriel sous-jacent est $E_{K'}:=E\otimes_KK'$ et dont la norme en $v'\in M_{K'}$ s’identifie à $\|.\|_v$ avec $v\in M_K$, où $v'$ prolonge $v$. Ainsi on peut définir la hauteur normalisée pour tout élément non-nul de $E_{K'}$. En outre, pour toute extension finie $K''$ de $K'$, la hauteur normalisée de $s\in E_{K'}$ s’identifie à celle de son image canonique dans $E_{K''}$. Cette observation permet d’étendre $h_{\overline E}$ en une fonction sur l’ensemble des vecteurs non-nuls dans $E_{K^{\mathrm{a}}}$, où $K^{\mathrm{a}}$ désigne la clôture algébrique de $K$. En outre, d’après la formule du produit, la fonction $h_{\overline E}$ est invariante sous la multiplication par un scalaire non-nul dans $K^{\mathrm{a}}$. Ainsi on peut la considérer comme une fonction définie sur l’ensemble des points algébriques de $\mathbb P(E^\vee)$. Il s’avère que cette fonction s’identifie à la hauteur absolue par rapport au faisceau inversible universel $\mathcal O_{E^\vee}(1)$ muni des métriques de Fubini-Study. Pour tout nombre réel $t$, on désigne par $\mathcal F_{\mathrm{ht}}^t(E_{K^{\mathrm{a}}})$ le sous-$K^{\mathrm{a}}$-espace vectoriel de $E_{K^{\mathrm{a}}}$ engendré par tous les vecteurs non-nuls $s$ vérifiant $h_{\overline E}(s)\leqslant -t$. La famille $(\mathcal F_{\mathrm{ht}}^t(E_{K^{\mathrm{a}}}))_{t\in\mathbb R}$ définit une $\mathbb R$-filtration décroissante de l’espace vectoriel $E_{K^{\mathrm{a}}}$. Ses points de saut successifs (où on compte les multiplicités) sont la version logarithmique des minima absolus définis par Roy et Thunder [@Roy_Thunder96] ([voir aussi [@Gaudron_Remond13] pour une présentation détaillée de différentes notions de minima]{}) dans le cadre de la géométrie des nombres adéliques, et par Soulé dans le cadre de la géométrie d’Arakelov[^9]. Pour tout entier $i\in\{1,\ldots,\operatorname{rg}(E)\}$, on désigne par $\Lambda_i(\overline E)$ le nombre $$\sup\{t\in\mathbb R\,:\,\operatorname{rg}_{K^{\mathrm{a}}}(\mathcal F^t_{\mathrm{ht}}(E_{K^{\mathrm{a}}}))\geqslant i\},$$ appelé le $i^{\text{\`eme}}$ minimum absolu logarithmique de $\overline E$. La comparaison entre les minima successifs et les pentes successives d’un fibré vectoriel adélique hermitien est un problème naturel. Les inégalités $\Lambda_i(.)\leqslant\widehat{\mu}_i(.)$ sont relativement standards et résultent de l’inégalité de pentes. Cependant la comparaison au sens inverse est beaucoup plus délicate. Conjecturalement [on a $$\label{Equ:compainv}\mu_i(\overline E)\leqslant\Lambda_i(\overline E)+\frac 12\ln(\operatorname{rg}(E)),\quad i\in\{1,\ldots,\operatorname{rg}(E)\}$$]{} pour tout fibré vectoriel adélique hermitien $\overline E$ sur $\operatorname{Spec}K$. Une approche possible pour attaquer ce problème est d’établir une version absolue du théorème de transférence à la Banaszczyk [@Banaszczyk95], qui n’est malheureusement pas encore disponible. Dans la suite, on établit une version plus faible de l’inégalité . Il s’agit d’une comparaison explicite entre $$\sum_{i=1}^{\operatorname{rg}(E)}\max(\operatorname{\widehat{\mu}}_i(\overline E),0)\quad\text{et}\quad\sum_{i=1}^{\operatorname{rg}(E)}\max(\Lambda_i(\overline E),0)$$ qui provient du lemme de Siegel absolu dû à Bombieri-Vaaler [@Bombieri_Vaaler83] et Zhang [@Zhang95]. Si $\overline E$ est un fibré vectoriel adélique hermitien sur $\operatorname{Spec}K$, alors on a $$\label{Equ:Siegeltronque}\sum_{i=1}^{\operatorname{rg}(E)}\max(\operatorname{\widehat{\mu}}_i(\overline E),0)\leqslant\sum_{i=1}^{\operatorname{rg}(E)}\max(\Lambda_i(\overline E),0)+\frac 12\operatorname{rg}(E)\ln(\operatorname{rg}(E)).$$ Rappelons que le lemme de Siegel absolu montre que, pour tout fibré vectoriel adélique hermitien $\overline E$ sur $\operatorname{Spec}K$, on a (cf. [@Gaudron08 théorème 4.14] et [@Gaudron09 §2.1.3] [^10]) $$\label{Equ:sigeltronque2}\sum_{i=1}^{\operatorname{rg}(E)}\Lambda_i(\overline E)\geqslant \operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline E)-\frac{1}{2}\operatorname{rg}(E)\ln(\operatorname{rg}(E)).$$ L’inégalité est donc vraie lorsque $\operatorname{\widehat{\mu}}_{\min}(\overline E)\geqslant 0$. Dans le cas général, il existe un sous-espace vectoriel $F$ de $E$ tel que $\widehat{\mu}_{\min}(\overline F)\geqslant 0$ et que $$\label{Equ:mufetmue}\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline F)=\sum_{i=1}^{\operatorname{rg}(F)}\widehat{\mu}_i(\overline F)=\sum_{i=1}^{\operatorname{rg}(E)}\max(\widehat{\mu}_i(\overline E),0).$$ On peut choisir $F$ comme le dernier sous-espace vectoriel dans le drapeau de Harder-Narasimhan de $E$ vérifiant $\widehat{\mu}_{\min}(\overline F)\geqslant 0$. Comme $F$ est un sous-espace vectoriel de $E$, on a $\Lambda_i(\overline F)\leqslant\Lambda_i(\overline E)$ pour tout $i\in\{1,\ldots,\operatorname{rg}(F)\}$. L’inégalité appliquée à $\overline F$ montre alors que $$\operatorname{\widehat{\deg}_{\mathrm{n}}}(F) \leqslant\sum_{i=1}^{\operatorname{rg}(F)}\Lambda_i(\overline F)+\frac 12\operatorname{rg}(F)\ln(\operatorname{rg}(F))\leqslant\sum_{i=1}^{\operatorname{rg}(F)}\Lambda_i(\overline E)+\frac 12\operatorname{rg}(E)\ln(\operatorname{rg}(E)).$$ D’après la formule , on en déduit l’inégalité . [L’inégalité est une conséquence immédiate de . En outre,]{} en utilisant le lemme de Siegel absolu, on peut montrer que, si $\overline E$ est un fibré vectoriel adélique sur $\operatorname{Spec}K$ qui est hermitien, alors on a $$\label{Equ:mu1etlambda1} \widehat{\mu}_1(\overline E)\leqslant\Lambda_1(\overline E)+\frac12\ln(\operatorname{rg}(E)).$$ On renvoie les lecteurs dans [@Gaudron_Remond §3.2] pour une démonstration. Le cas non-hermitien -------------------- Soit $\overline E$ un fibré vectoriel adélique général. Gaudron a défini dans [@Gaudron08] le degré d’Arakelov[^11] de $\overline E$ comme la différence entre la caractéristique d’Euler-Poincaré de $\overline E$ et celle du fibré vectorial adélique trivial dont le rang est $\operatorname{rg}(E)$. Cela lui permet de généraliser la notion des pentes successives dans un cadre plus général des fibrés vectoriels adéliques non nécessairement hermitiens. Comme dans le cas hermitien, les pentes successives de $\overline E$ sont définies comme les pentes du polygone de Harder-Narasimhan de $\overline E$, dont le graphe est le bord supérieur de l’enveloppe convexe des points dans $\mathbb R^2$ de coordonnées $(\operatorname{rg}(F),\operatorname{\widehat{\deg}_{\mathrm{n}}}(\overline F))$, où $F$ parcourt l’ensemble des sous-espaces vectoriels de $E$. À l’aide de la méthode d’ellipsoïde de John-Löwner[^12], on peut associer à $E$ une structure de fibré vectoriel adélique hermitien $(\|.\|_{v}')_{v\in M_K}$ de sorte que $\|.\|_{v}'=\|.\|_v$ si $v$ est une place non-archimédienne, et $$\|.\|_v'\leqslant \|.\|_v\leqslant (\operatorname{rg}(E))^{1/2}\|.\|_{v}'$$ lorsque $v$ est archimédienne. Si on note $\overline E{}'$ le fibré vectoriel adélique $(E,(\|.\|_v')_{v\in M_K})$, on a $\operatorname{\widehat{\mu}}_{i}(\overline E)\leqslant\operatorname{\widehat{\mu}}_i(\overline E{}')$ et $\Lambda_i(\overline E{}')\leqslant\Lambda_i(\overline E)+\frac 12\ln(\operatorname{rg}(E))$ pour tout $i\in\{1,\ldots,\operatorname{rg}(E)\}$. On obtient à partir de l’inégalité (appliquée à $\overline E{}'$) la relation suivante $$\label{Equ:Siegeltronquenonhem}\sum_{i=1}^{\operatorname{rg}(E)}\max(\operatorname{\widehat{\mu}}_i(\overline E),0)\leqslant\sum_{i=1}^{\operatorname{rg}(E)}\max(\Lambda_i(\overline E),0)+\operatorname{rg}(E)\ln(\operatorname{rg}(E)).$$ Majoration de la fonction de Hilbert-Samuel arithmétique {#Sec:majroationarith} ======================================================== Soit $K$ un corps de nombres. Dans ce paragraphe, on établit un analogue arithmétique du corollaire \[Cor:majHS\] pour un système gradué en fibrés vectoriels adéliques sur $\operatorname{Spec}K$. Soit $X$ un schéma projectif et intègre de dimension $d\geqslant 1$ sur $\operatorname{Spec}K$, $L$ un faisceau inversible gros sur $X$. On entend par *système linéaire gradué* de $L$ *en fibrés vectoriels adéliques* tout système linéaire gradué $E_\sbullet=\bigoplus_{n\geqslant 0}E_n$ de $L$ dont chaque composante homogène $E_n$ est muni d’une structure de fibré vectoriel adélique sur $\operatorname{Spec}K$ de telle sorte que $$\label{Equ:suradditive}\|s\cdot s'\|_{v}\leqslant\|s\|_v\cdot\|s'\|_v$$ pour tout couple $(n,m)\in\mathbb N^2$ et tous $s\in E_{n,\mathbb C_v}$, $s'\in E_{m,\mathbb C_v}$. Cette inégalité montre que la suite $(\Lambda_1(\overline E_n))_{n\geqslant 1}$ est sur-additive. Donc la suite $(\Lambda_1(\overline E_n)/n)_{n\geqslant 1}$ converge vers un élément dans $\mathbb R\cup\{+\infty\}$ pourvu que $E_n\neq\{0\}$ pour tout entier $n$ suffisamment positif. On désigne par $\widehat{\mu}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)$ cette limite[^13]. En outre, si $E_\sbullet$ contient un diviseur ample et si $\widehat{\mu}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)<+\infty$, il est démontré dans [@Boucksom_Chen théorème 2.8] que la suite $$\frac{(d+1)!}{n^{d+1}}\sum_{i=1}^{\operatorname{rg}(E_n)}\max(\Lambda_i(\overline E_n),0),\qquad n\geqslant 1.$$ converge vers un nombre réel que l’on notera comme $\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\sbullet)$. \[Thm:HMarithmetique\] Soit $X$ un schéma projectif et géométriquement intègre sur $\operatorname{Spec}K$, $L$ un faisceau inversible gros sur $X$ et $\overline E_\sbullet=\bigoplus_{n\geqslant 0}\overline E_n$ un système linéaire gradué de $L$ en fibrés vectoriels adéliques. On suppose que le système linéaire gradué $E_\sbullet$ contient un diviseur ample et que $X_{K^{\mathrm{a}}}$ possède un tour de fibrations sur courbes $\Theta$. Alors on a $$\sum_{i=1}^{\operatorname{rg}(E_1)}\max(\Lambda_i(\overline E_1),0)\leqslant \operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\sbullet)+\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}}).$$ Pour tout entier $n\in\mathbb N$ et tout nombre réel $t$, on désigne par $\mathcal F^tE_{n,K^{\mathrm{a}}}$ le sous-$K^{\mathrm{a}}$-espace vectoriel engendré par les vecteurs non-nuls $s\in E_{n,K^{\mathrm{a}}}$ tels que $h_{\overline E_n}(s)\leqslant -t$. Il s’avère que $(\mathcal F^tE_{n,K^{\mathrm{a}}})_{t\in\mathbb R}$ est une $\mathbb R$-filtration décroissante de $E_{n,K^{\mathrm{a}}}$. En outre, la relation montre que $$(\mathcal F^{t_1}E_{n_1,K^{\mathrm{a}}})\cdot(\mathcal F^{t_2}E_{n_2,K^{\mathrm{a}}})\subset\mathcal F^{t_1+t_2}E_{n_1+n_2,K^{\mathrm{a}}}$$ pour tous $(n_1,n_2)\in\mathbb N^2$ et $(t_1,t_2)\in\mathbb R^2$. Pour tout $t\in\mathbb R$, $E^t_{\sbullet}:=\bigoplus_{n\geqslant 0}\mathcal F^{nt}E_{n,K^{\mathrm{a}}}$ est alors un système linéaire gradué de $L_{K^{\mathrm{a}}}$. D’après [@Boucksom_Chen lemme 1.6], ce système linéaire gradué contient un diviseur ample dès que $t<\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)$. En outre, on a (d’après le corollaire 1.13 du *loc. cit.*) $$\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\sbullet)=(d+1)\int_0^{+\infty}\operatorname{vol}(E^t_\sbullet)\,\mathrm{d}t=(d+1)\int_0^{\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)}\operatorname{vol}(E^t_\sbullet)\,\mathrm{d}t,$$ où $d$ est la dimension de $X$. Le théorème \[Thm:majorationdeHS\] appliqué à $E_\sbullet^t$ montre que $$\operatorname{rg}(\mathcal F^tE_{1,K^{\mathrm{a}}})\leqslant\operatorname{vol}(E^t_{\sbullet,K^{\mathrm{a}}})+\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}}^t)\leqslant\operatorname{vol}(E^t_{\sbullet,K^{\mathrm{a}}})+\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}}).$$ On en déduit $$\begin{split}&\quad\;\sum_{i=1}^{\operatorname{rg}(E_1)}\max(\Lambda_i(\overline E_1),0)=\int_0^{\Lambda_1(\overline E_1)}\operatorname{rg}(\mathcal F^tE_{1,K^{\mathrm{a}}})\,\mathrm{d}t=\int_0^{\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)}\operatorname{rg}(\mathcal F^tE_{1,K^{\mathrm{a}}})\,\mathrm{d}t\\ &\leqslant \int_0^{\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)}\frac{\operatorname{vol}(E_\sbullet^t)}{d!}\,\mathrm{d}t+\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}})=\frac{\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\bullet)}{(d+1)!}+\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}}), \end{split}$$ où la deuxième égalité provient du fait que $\Lambda_1(\overline E_1)\leqslant\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)$. La démonstration est donc achevée. On déduit du théorème précédent et l’inégalité le résultat suivant. \[Cor:majrationdemu\] Avec les notations du théorème précédent, on a $$\sum_{i=1}^{\operatorname{rg}(E_1)}\max(\operatorname{\widehat{\mu}}_i(\overline E_1),0)\leqslant\frac{\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\sbullet)}{(d+1)!}+\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}})+\operatorname{rg}(E_1)\ln(\operatorname{rg}(E_1)).$$ Soit $\pi:\mathscr X\rightarrow\operatorname{Spec}\mathcal O_K$ un morphisme projectif et plat d’un schéma intègre $\mathscr X$ vers $\operatorname{Spec}\mathcal O_K$. Par faisceau inversible hermitien sur $\mathscr X$, on entend un $\mathcal O_{\mathscr X}$-module inversible $\mathscr L$ dont le tire en arrière sur $\mathscr X^{\mathrm{an}}$ est muni d’une métrique continue qui est invariante par la conjugaison complexe, où $\mathscr X^{\mathrm{an}}$ désigne l’espace analytique complexe associé à $\mathscr X\otimes_{\mathbb Z}\mathbb C$. Étant donné un faisceau inversible hermitien $\overline{\mathscr L}$ sur $\mathscr X$, on peut construire une structure de fibré vectoriel adélique sur $H^0(X,\mathscr L_K)$. En une place finie $\mathfrak p$, la norme $\|.\|_{\mathfrak p}$ sur $H^0(X,\mathscr L_K)\otimes_K{\mathbb C_{\mathfrak p}}$ provient de la structure de $\mathcal O_K$-module de $\pi_*(\mathscr L)$ : la boule unité fermé pour la norme $\|.\|_{\mathfrak p}$ s’identifie à $\pi_*(\mathscr L)\otimes_{\mathcal O_K}\mathcal O_{\mathfrak p}$, où $\mathcal O_{\mathfrak p}$ désigne l’anneau de valuation de $\mathbb C_{\mathfrak p}$. En une place infinie $\sigma:K\rightarrow\mathbb C$, la norme $\|.\|_{\sigma}$ est la norme sup : pour tout élément $s\in H^0(X,\mathscr L_K)\otimes_{K,\sigma}\mathbb C$, on a $$\|s\|_\sigma:=\sup_{x\in\mathscr X_\sigma(\mathbb C)}\|s(x)\|.$$ On utilise l’expression $\pi_*(\overline{\mathscr L})$ pour désigner ce fibré vectoriel adélique. Ainsi $$\overline E_\sbullet=\bigoplus_{n\geqslant 0}\pi_*(\overline{\mathscr L}{}^{\otimes n})$$ devient un système linéaire gradué en fibrés vectoriels adéliques sur $\operatorname{Spec}K$. Il s’avère que le nombre $\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\sbullet)$ décrit le comportement asymptotique du nombre de sections effectives de $\overline{\mathscr L}{}^{\otimes n}$ lorsque $n\rightarrow+\infty$. En effet, en utilisant l’inégalité et la méthode d’ellipsoïde de John-Löwner, on peut montrer que $$\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline{E}_\sbullet)=\frac{1}{[K:\mathbb Q]}\lim_{n\rightarrow+\infty}\frac{\widehat{h}^0(\overline E_n)}{n^{d+1}/(d+1)!},$$ où $d$ est la dimension relative de $\pi$. Rappelons que la limite figurant dans le terme de droite de la formule est appelé le *volume arithmétique*, noté comme $\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})$. On renvoie les lecteurs dans [@Moriwaki07] où cette notion a été proposée. Dans le cas où le faisceau inversible hermitien $\overline{\mathscr L}$ est arithmétiquement nef[^14], le volume arithmétique de $\overline{\mathscr L}$ s’identifie au nombre d’auto-intersection arithmétique $\widehat{c}_1(\overline{\mathscr L})^{d+1}$. On déduit alors du corollaire \[Cor:majrationdemu\] le résultat suivant. \[Thm:HSarithmetic\] Soit $\pi:\mathscr X\rightarrow\operatorname{Spec}\mathcal O_K$ un morphisme projectif et plat. On suppose que la fibre générique géométrique $\mathscr X_{K^{\mathrm{a}}}$ est géométriquement intègre et admet un tour de fibrations sur courbes $\Theta$. Alors, pour tout faisceau inversible hermitien $\overline{\mathscr L}$ sur $\mathscr X$ qui est arithmétiquement nef et génériquement gros, il existe une fonction explicite $F_{\overline{\mathscr L}}:\mathbb N\rightarrow [0,+\infty[$ telle que $$\label{Equ:estimationdeF}F_{\overline{\mathscr L}}(n)=[K:\mathbb Q]\frac{c_1(\mathscr L_K)^d}{(d-1)!}n^d\ln(n)+O(n^d),\quad n\rightarrow+\infty,$$ et que $$\label{Equ:estdeg}\operatorname{\widehat{\deg}}(\pi_*(\overline{\mathscr L}{}^{\otimes n}))\leqslant\frac{\widehat{c}_1(\overline{\mathscr L})^{d+1}}{(d+1)!}n^{d+1}+F_{\overline{\mathscr L}}(n),$$ où $d$ est la dimension relative de $\pi$. Pour tout entier $n\geqslant 1$, on note $r_n=\operatorname{rg}(\pi^*(\mathscr L^{\otimes n}))$. Comme $\mathscr L_K$ est nef et gros, on a (cf. [@LazarsfeldI corollaire 1.4.38]) $$\label{Equ:estrn}r_n=\frac{c_1(\mathscr L_K)^d}{d!}n^d+O(n^{d-1}).$$ Soit $\overline E_{\sbullet}=\bigoplus_{n\geqslant 0}\pi_*(\overline{\mathscr L}{}^{\otimes n})$. Compte tenu de [@Boucksom_Chen lemme 2.6], la relation $\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)<+\infty$ est satisfaite. D’après le corollaire \[Cor:majrationdemu\], on obtient $$\operatorname{\widehat{\deg}}(\pi_*(\overline{\mathscr L}{}^{\otimes n}))\leqslant\sum_{i=1}^{r_n}\max(\operatorname{\widehat{\mu}}_i(\overline E_n),0)\leqslant [K:\mathbb Q]\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline E_\sbullet)\frac{n^{d+1}}{(d+1)!}+F_{\overline{\mathscr L}}(n)$$ avec $$F_{\overline{\mathscr L}}(n)=[K:\mathbb Q]\big(n\widehat{ \mu}_{\max}^{\mathrm{asy}}(\overline{E}_\sbullet)\varepsilon^{\Theta}(E^{(n)}_{\sbullet,K^{\mathrm{a}}})+r_n\ln(r_n)\big)$$ Comme $\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}}^{(n)})\leqslant n^{d-1}\varepsilon^{\Theta}(E_{\sbullet,K^{\mathrm{a}}})$, on déduit de la relation . Enfin, comme $\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})=[K:\mathbb Q]\operatorname{\widehat{\operatorname{vol}}_{\mathrm{n}}}(\overline{E}_\sbullet)$, l’inégalité est démontrée. Si on remplace $\operatorname{\widehat{\deg}}(.)$ par la somme des minima absolus successifs dans le théorème précédent, on peut obtenir une majoration asymptotique où le terme d’erreur est $O(n^d)$. Plus précisément, il existe une fonction explicite $\widetilde F_{\overline{\mathscr L}}:\mathbb N\rightarrow[0,+\infty[$ telle que $\widetilde F_{\overline{\mathscr L}}(n)=O(n^d)$ et que $$\sum_{i=1}^{r_n}\Lambda_i(\pi_*(\overline{\mathscr L}{}^{\otimes n}))\leqslant \frac{\widehat{c}_1(\overline{\mathscr L})^{d+1}}{(d+1)!}n^{d+1}+\widetilde F_{\overline{\mathscr L}}(n).$$ Cela suggère que le terme sous-dominant (de l’ordre $n^d\ln(n)$) dans le théorème de Riemann-Roch arithmétique (cf. [@Soule92 §2.2]) peut provenir du choix de métrique et de la comparaison entre certains invariants arithmétiques de fibré vectoriel hermitien sur $\operatorname{Spec}\mathcal O_K$. Des phénomènes similaires existent aussi dans l’étude des surfaces arithmétiques, comme par exemple [@Faltings84 théorème 3] (voir aussi [@Abbes-Bouche §5.1]). Dans la suite, on établit un analogue du théorème \[Thm:HSarithmetic\] pour la fonction $\widehat{h}^0$. En utilisant le deuxième théorème de Minkowski et la filtration par les minima (usuels), on peut majorer $\widehat{h}^0(\pi_*(\overline{\mathscr L}{}^{\otimes n}))$ par une fonction explicite. Soit $\overline M$ un fibré vectoriel normé sur $\operatorname{Spec}\mathbb Z$ (i.e. un réseau dans un espace vectoriel normé de dimension fini sur $\mathbb R$). Pour tout entier $i\in\{1,\ldots,n\}$, on définit le $i^{\text{\`eme}}$ *minimum logarithmique* de $\overline M$ comme $$\lambda_i(\overline M):=\sup\Big\{t\in\mathbb R\,\Big|\,\operatorname{rg}_{\mathbb Q}\big(\mathrm{Vect}_{\mathbb Q}\{s\in M\,|\,\|s\|\leqslant\mathrm{e}^{-t}\}\big)\geqslant i\Big\}.$$ Rappelons que la caractéristique d’Euler-Poincaré de $\overline M$ est défini comme $$\chi(\overline M)=\ln\frac{\operatorname{vol}(B(M_{\mathbb R},\|.\|))}{\mathrm{covol}(M)},$$ où $B(M_{\mathbb R},\|.\|)$ désigne la boule unité fermée dans $M_{\mathbb R}$, $\operatorname{vol}$ est une mesure de Haar sur $M_{\mathbb R}$ et $\mathrm{covol}(M)$ est la mesure de $M_{\mathbb R}/M$ par rapport à la mesure induite par $\operatorname{vol}$. Soit $\overline M$ un réseau de rang $r>0$ dans un espace vectoriel normé. On a $$\widehat{h}^0(\overline M)\leqslant \sum_{i=1}^r\max(\lambda_i(\overline M),0)+r\ln(2)+\ln(2r!).$$ Quitte à remplacer $M$ par le sous-réseau engendré par les éléments $s\in M$ vérifiant $\|s\|\leqslant 1$, on peut supposer que $\lambda_i(\overline M)\geqslant 0$ pour tout $i\in\{1,\ldots,r\}$. Rappelons que le deuxième théorème de Minkowski montre que $$r\ln(2)-\ln(r!)\leqslant\chi(\overline M)-\sum_{i=1}^r\lambda_i(\overline M)\leqslant r\ln(2).$$ En outre, si on fixe une base de $M$ sur $\mathbb Z$ et identifie $M_{\mathbb R}$ à $\mathbb R^r$ via cette base, alors $B(M_{\mathbb R},\|.\|)$ est un corps convexe symétrique dans $\mathbb R^r$ dont le volume est $\chi(\overline{M})$ (où on a considéré la mesure de Haar standard sur $\mathbb R^r$). D’après un résultat de Blichfeldt (cf. [@Henze13 page 372]), on a $$\label{Equ:Minkowski2}\widehat{h}^0(\overline M)\leqslant \ln(r!\exp(\chi(\overline M))+r)\leqslant \ln(2r!)+\chi(\overline M),$$ où la deuxième inégalité provient de l’hypothèse $\forall\,i,\;\lambda_i(\overline M)\geqslant 0$. En effet, sous cette hypothèse on a $r!\exp(\chi(\overline M))\geqslant 2^r> r$. D’après la deuxième inégalité de , on obtient le résultat. \[Thm:majorationh0\] Soit $\pi:\mathscr X\rightarrow\operatorname{Spec}\mathcal O_K$ un morphisme projectif et plat. On suppose que la fibre générique $\mathscr X_K$ est intègre. Alors, pour tout faisceau inversible hermitien $\overline{\mathscr L}$ sur $\mathscr X$ tel que $\mathscr L_K$ soit gros, il existe une fonction explicite $G_{\overline{\mathscr L}}:\mathbb N\rightarrow [0,+\infty[$ telle que $$\label{Equ:estimationdeG}G_{\overline{\mathscr L}}(n)=[K:\mathbb Q]\frac{\operatorname{vol}(\mathscr L_K)}{(d-1)!}n^d\ln(n)+O(n^d),\quad n\rightarrow+\infty,$$ et que $$\label{Equ:estdegG}\widehat{h}^0(\pi_*(\overline{\mathscr L}{}^{\otimes n}))\leqslant\frac{\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})}{(d+1)!}n^{d+1}+G_{\overline{\mathscr L}}(n),$$ où $d$ est la dimension relative de $\pi$. Pour tout entier $n\geqslant 1$, on désigne par $R_n$ le rang de $\pi_*(\mathscr L^{\otimes n})$ sur $\mathbb Z$. Comme $\mathscr L_K$ est gros, d’après le théorème \[Thm:majorationdeHS\] on a $$\label{Equ:estRn} R_n=[K:\mathbb Q]\frac{\operatorname{vol}(\mathscr L_K)}{d!}n^d+o(n^{d-1}).$$ Soit $\overline E_\sbullet=\bigoplus_{n\geqslant 0}{\pi_*(\overline{\mathscr L}{}^{\otimes n})}$. On considère chaque $\overline E_n$ comme un réseau dans l’espace vectoriel normé $E_n\otimes_{\mathbb Q}\mathbb R$, qui s’identifie à $$\bigoplus_{v\in M_{K,\infty}}E_n\otimes_{K}K_v.$$ Si $\boldsymbol{s}=(s_v)_{v\in M_{K,\infty}}$ est un élément de $E_n\otimes_{\mathbb Z}\mathbb R$, la norme de $\boldsymbol{s}$ est définie comme $$\max_{v\in M_{K,\infty}}\|s_v\|_v.$$ On considère $\mathscr X_K$ comme un schéma projectif sur $\operatorname{Spec}\mathbb Q$ (que l’on notera comme $\mathscr X_{\mathbb Q}$ dans la suite). De même, on considère $\mathscr L_K$ comme un faisceau inversible sur $\mathscr X_{\mathbb Q}$ (que l’on notera comme $\mathscr L_{\mathbb Q}$). Ainsi $E_\sbullet$ devient le système linéaire gradué total de $\mathscr L_{\mathbb Q}$. Pour chaque entier $n\geqslant 0$, on munit $E_n$ (comme espace vectoriel sur $\mathbb Q$) de la $\mathbb R$-filtration $\mathcal F$ par les minima : $$\mathcal F^t(E_n)=\mathrm{Vect}_{\mathbb Q}\{s\in\pi_*(\mathscr L^{\otimes n})\,:\,\|s\|\leqslant \mathrm{e}^{-t}\}.$$ Quitte à passer à une modification birationnelle (la quantité $\widehat{h}^0(\pi_*(\overline{\mathscr L}{}^{\otimes n}))$ augment éventuellement, tandis que $\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})$ reste inchangé), on peut supposer que $\mathscr X_{\mathbb Q}$ admet un tour de fibrations sur courbes $\Theta$. D’après le corollaire \[Cor:majHS\] (appliqué à $\overline{\mathscr L}{}^{\otimes n}$), on obtient $$\sum_{i=1}^{R_n}\max(\lambda_i(\overline E_n),0)\leqslant\frac{\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})}{(d+1)!}n^{d+1}+n^d\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)\varepsilon^{\Theta}(E_\sbullet),$$ où on a utilisé les relations $\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet^{(n)})=n\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)$ et $\varepsilon^{\Theta}(E_\sbullet^{(n)})\leqslant n^{d-1}\varepsilon^{\Theta}(E_\sbullet)$. D’après le lemme précédent, on obtient $$\sum_{i=1}^{R_n}\max(\lambda_i(\overline E_n),0)\leqslant\frac{\operatorname{\widehat{\operatorname{vol}}}(\overline{\mathscr L})}{(d+1)!}n^{d+1}+G_{\overline{\mathscr L}}(n)$$ avec $$G_{\overline{\mathscr L}}(n)=n^d\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)\varepsilon^{\Theta}(E_\sbullet)+(R_n+1)\ln(2)+R_n\ln(R_n).$$ Enfin, la relation montre aussitôt que $$G_{\overline{\mathscr L}}(n)=[K:\mathbb Q]\frac{\operatorname{vol}(\mathscr L_K)}{(d-1)!}n^d\ln(n)+O(n^d).$$ Le théorème est donc achevé. \[Rem:comparaisonYuanZhang\] Dans le cas où $\mathscr X$ est une surface arithmétique, il est intéressant de comparer le théorème précédent à la majoration obtenu dans [@Yuan_Zhang13] (voir aussi l’analogue de ce travail dans le cadre de corps de fonctions [@Yuan_Zhang13b]). Asymptotiquement le terme $G_{\overline{\mathscr L}}(n)$ est meilleur que le terme d’erreur $$\label{Equ:yuan-zhang}4n[K:\mathbb Q]\deg(\mathscr L_K)\ln(n[K:\mathbb Q]\deg(\mathscr L_K))$$ dans le théorème A du *loc. cit.*. Cependant le terme ne dépend que de l’information géométrique de la fibre générique $\mathscr L_K$. On se demande si une combinaison des deux méthodes ne donne pas une majoration effective de la somme des minima logarithmiques positifs de $\pi_*(\overline{\mathscr L}{}^n)$ dont le terme d’erreur est d’ordre $n^d$ et ne dépend que de la géométrie de $\mathscr L_K$ dans le cas où $\mathscr X$ est une surface arithmétique. Le cas de caractéristique positif ================================= Dans ce paragraphe, on établit l’analogue du théorème \[Thm:majorationdeHS\] dans le cas où la caractéristique du corps de base est strictement positif. Soit $k$ un corps de caractéristique quelconque. La conclusion de la proposition \[Pro:produitdefiltrationHN\], qui est équivalente à la semi-stabilité du produit tensoriel de tout couple de fibrés vectoriels semi-stables sur la courbe projective régulière définie sur $k$, n’est cependant pas vrai en général. On renvoie les lecteurs dans [@Gieseker73] pour un contre-exemple. La méthode de $\mathbb R$-filtration de Harder-Narasimhan que l’on a développée dans §\[Sec:HS\] n’est plus valable dans ce cadre-là. On propose d’utiliser les minima successifs dans le cadre de corps de fonction pour surmonter cette difficulté. Soient $C$ une courbe projective et régulière sur $\operatorname{Spec}k$. On désigne par $K$ le corps des fonctions rationnelles sur $C$. Si $E$ est un fibré vectoriel sur $C$, on désigne par $\mathcal O_E(1)$ le faisceau inversible universel du schéma $\mathbb P(E)$. On peut définir une fonction de hauteur sur l’ensemble des $k$-points de $\mathbb P(E)$ à valeurs dans $K$ comme la suite. Si $x$ est un point dans $\mathbb P(E)_k(K)$, il se prolonge en une section $\mathscr P_x:C\rightarrow\mathbb P(E)$ de $\mathbb P(E)$. On définit la hauteur $x$ comme $$h_E(x):=\deg(\mathscr P_x^*\mathcal O_E(1)).$$ La fonction de hauteur nous permet de définir une filtration $\mathcal F$ sur l’espace $K$-vectoriel $E_{K}$ (la fibre générique de $E$) comme la suite[^15] $$\label{Equ:filtrationparminima}\mathcal F^t(E_{K}):=\mathrm{Vect}_{K}\big\{x\in\mathbb P(E^\vee)_k(K)\,|\,h_{E^{\vee}}(x)\leqslant -t\big\}.$$ Pour tout entier $i\in\{1,\ldots,\operatorname{rg}(E)\}$, on désigne par $\lambda_i(E)$ le plus grand nombre réel $t$ tel que $\operatorname{rg}(\mathcal F^t(E_{K}))\geqslant i$. Les nombres $\lambda_i(E)$ et la filtration $\mathcal F$ sont reliées par la formule suivante : $$\label{Equ:lambdaint} \sum_{i=1}^{\operatorname{rg}(E)}\max(\lambda_i(E),0)=\int_0^{+\infty}\operatorname{rg}_K(\mathcal F^t(E_K))\,\mathrm{d}t=\int_0^{\lambda_1(E)}\operatorname{rg}_K(\mathcal F^t(E_K))\,\mathrm{d}t.$$ Les quantités $\lambda_i(E) $ devraient être considérées comme l’analogue des minima successifs dans le cadre de corps de fonction. On désigne par $M_K$ l’ensemble des points fermés dans la courbe $C$, considéré comme l’ensemble des places du corps de fonctions $K$. Pour tout point $v\in M_K$, on désigne par $|.|_v$ la valeur absolue sur $K$ définie comme $$\forall\,f\in K^{\times}\,\quad |f|_v=\exp(-[k(v):k]\mathrm{ord}_v(f)),$$ où $k(v)$ est le corps résiduel de $v$ et $\mathrm{ord}_v(f)$ désigne l’ordre d’annulation de $f$ en $v$. Comme dans le cas de corps de nombres, on désigne par $K_v$ le complété de $K$ par rapport à cette valeur absolue et $\mathbb C_v$ le complété d’une clôture algébrique de $K_v$, sur lequel la valeur absolue $|.|_v$ s’étend de façon unique. On désigne par $\mathcal O_v$ l’anneau de valuation de $\mathbb C_v$ par rapport à cette valeur absolue (qui est non-archimédienne). Si $E$ est un fibré vectoriel sur $C$, alors sa structure de $\mathcal O_C$-module définit, pour chaque place $v\in M_K$, une norme $\|.\|_{E,v}$ sur $E\otimes_{\mathcal O_C}\mathbb C_v$ dont la boule unité fermée est $E\otimes_{\mathcal O_C}\mathcal O_v$. Cette norme est invariante par l’action du groupe de Galois $\mathrm{Gal}(\mathbb C_v/K_v)$. On obtient alors un fibré vectoriel adélique sur $\operatorname{Spec}K$ au sens de [@Gaudron08 §3], et il s’avère que l’on peut exprimer le degré de $E$ sous la forme $$\deg(E)=-\sum_{v\in M_K}\ln\|s_1\wedge\cdots\wedge s_r\|_{E,v},$$ où $(s_1,\ldots,s_r)$ est une base quelconque de $E_K$. En outre, les nombres $\lambda_i(E)$ sont précisément les minima successifs logarithmiques suivant Thunder [@Thunder96] dans le cadre de corps de fonctions. D’après un résultat de Roy et Thunder [@Roy_Thunder96 théorème 2.1] : on a $$\label{Equ:Siegelfun}\sum_{i=1}^{\operatorname{rg}(E)}\lambda_i(E)\geqslant\deg(E)-\operatorname{rg}(E)\ell(g(C)),$$ où $g(C)$ le genre de $C$ et $\ell$ est une fonction affine qui ne dépend que du degré effectif du corps de fonction $K$. On renvoie les lecteur dans [@Roy_Thunder96 page 5] pour la forme explicite de cette fonction. On en déduit le résultat suivant. Soit $C$ une courbe projective régulière définie sur un corps $k$. Si $E$ est un fibré vectoriel sur $C$, on a $$\label{Equ:majdedeg+l} \deg_+(E)=\sum_{i=1}^{\operatorname{rg}(E)}\max({\mu}_i(E),0)\leqslant\sum_{i=1}^{\operatorname{rg}(E)}\max(\lambda_i(E),0)+\operatorname{rg}(E)\ell(g(C)).$$ Quitte à remplacer $E$ par le dernier sous-fibré vectoriel de pente minimale positive dans le drapeau de Harder-Narasimhan de $E$, on peut supposer que $\mu_{\min}(E)\geqslant 0$. Dans ce cas-là on a $$\deg_+(E)=\deg(E)=\sum_{i=1}^{\operatorname{rg}(E)}\lambda_i(E)+\operatorname{rg}(E)\ell(g(C))\leqslant \sum_{i=1}^{\operatorname{rg}(E)}\max(\lambda_i(E),0)+\operatorname{rg}(E)\ell(g(C)),$$ où la deuxième égalité provient de . Le résultat est donc démontré. Soient $E$ et $F$ deux fibrés vectoriels sur $C$. Si $x$ et $y$ sont respectivement deux $k$-points de $\mathbb P(E^\vee)$ et $\mathbb P(F^\vee)$ à valeurs dans $K$, alors $x\otimes y$ (vu comme un sous-espace vectoriel de rang un de $E_K\otimes F_K$) est un $k$-point de $\mathbb P(E^\vee\otimes F^\vee)$ à valeurs dans $K$ qui vérifie la relation suivante $$h_{E^\vee\otimes F^\vee}(x\otimes y)=h_{E^\vee}(x)+h_{F^\vee}(y).$$ On obtient donc le résultat suivant : Soit $E_\sbullet=\bigoplus_{n\geqslant 0}E_n$ un $\mathcal O_C$-algèbre graduée. On suppose que chaque $E_n$ est un fibré vectoriel sur $C$. Pour tout $(a,b)\in\mathbb R^2$ et tout $(n,m)\in\mathbb N^2$, la relation suivante est vérifiée : $$\label{Equ:filtree} \mathcal F^a(E_{n,K})\mathcal F^b(E_{m,K})\subset\mathcal F^{a+b}(E_{n+m,K}),$$ où $\mathcal F$ désigne la $\mathbb R$-filtration par minima définie dans . \[Rem:comparaisonlambda1\] On fixe un faisceau inversible ample $M$ sur $C$. Soit $a$ le degré de $M$. Si $E$ est un fibré vectoriel non-nul sur $C$, les inégalités $$\lambda_1(E)\leqslant\mu_{\max}(E)\leqslant\lambda_1(E)+g-1+a$$ sont toujours vérifiées. La première inégalité est triviale. Pour la deuxième inégalité, on peut utiliser l’invariance de la quantité $\mu_{\max}(E)-\lambda_1(E)$ par le produit tensoriel d’un $\mathcal O_C$-module inversible. Quitte à remplacer $E$ par le produit tensoriel de $E$ avec une puissance tensorielle (éventuellement d’exposant négatif) du faisceau inversible $M$, on peut supposer $g-1<\mu_{\max}(E)\leqslant g-1+a$. D’après le théorème de Riemann-Roch, on obtient que $\lambda_1(E)\geqslant 0$ (cf. [@Chen_pm lemme 2.1]). On obtient donc $\mu_{\max}(E)-\lambda_1(E)\leqslant g-1+a$. Cette inégalité montre que, si $E_\sbullet$ est une $\mathcal O_C$-algèbre graduée en fibrés vectoriels sur $C$ telle que $E_n$ soit non-nul pour $n$ assez grand, alors on a $$\lim_{n\rightarrow+\infty}\frac{\mu_{\max}(E_n)}{n}=\lim_{n\rightarrow+\infty}\frac{\lambda_1(E_n)}{n}\in\mathbb R\cup\{+\infty\}.$$ En outre, la relation montre que la suite $\lambda_1(E_n)$ est sur-additive. On obtient donc $$\forall\,p\geqslant 1,\quad \lambda_1(E_p)\leqslant p\lim_{n\rightarrow+\infty}\frac{\mu_{\max}(E_n)}{n}.$$ La proposition précédente nous permet de retrouver les résultats présentés dans §\[Sec:HS\], quitte à remplacer la filtration de Harder-Narasimhan par la filtration par minima. Soit $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$, muni d’un tour de fibrations sur courbes $\Theta$. On suppose que la dimension de Krull de $X$ est $d+1$. Soient $L$ un faisceau inversible gros sur $X$ et $V_\sbullet$ un système linéaire gradué de $L$, qui contient un diviseur ample. Si $d=0$, on définit $$\widetilde{\varepsilon}^{\Theta}(V_\sbullet):=\max(g(\Theta)-1,1),$$ où $g(\Theta)$ est le genre du tour $\Theta$ défini dans §\[Sec:tourdefibration\]. Si $d\geqslant 1$, on définit $\widetilde{\varepsilon}^{\Theta}(V_\sbullet)$ de façon récursive comme la suite. On suppose que $\Theta$ est de la forme $(p_0:X\rightarrow C_0,\Theta')$, où $\Theta'$ est un tour de fibrations sur courbes de la fibre générique de $p_0$. Soient en outre $g_0$ le genre de la courbe $C_0$, $\mu_0:=\mu^{p_0}_{\max}(V_\sbullet)$ (cf. la définition \[Def:imagedirecte\]), et $W_\sbullet$ la fibre générique de $p_{0*}(V_\sbullet)$. On définit $$\widetilde{\varepsilon}^{\Theta}(V_\sbullet)=\mu_0\widetilde{\varepsilon}^{\Theta'}(W_\sbullet)+\Big(\frac{\operatorname{vol}(W_\sbullet)}{d!}+\widetilde\varepsilon^{\Theta'}(W_\sbullet)\Big)(\max(g_0-1,1)+\ell(g_0)).$$ On voit aussitôt que $\widetilde{\varepsilon}^{\Theta}(V_\sbullet)\leqslant\widetilde{\varepsilon}^{\Theta}(V_\sbullet')$ si $V_\sbullet$ est contenu dans un autre système linéaire gradué $V_\sbullet'$ d’un autre faisceau inversible gros $L'$. En outre, par récurrence sur $d$ on peut vérifier que $\widetilde{\varepsilon}^{\Theta}(V_\sbullet^{(p)})\leqslant p^d\widetilde{\varepsilon}^{\Theta}(V_\sbullet)$ pour tout entier $p\geqslant 1$. \[Thm:majorationdeHSbis\] Soit $X$ un schéma projectif et intègre sur $\operatorname{Spec}k$ muni d’un tour de fibrations sur courbes $\Theta=(p_i:X_i\rightarrow C_i)_{i=0}^d$, et $L$ un faisceau inversible sur $X$. Si $V_\sbullet=\bigoplus_{n\geqslant 0}V_n$ est un système linéaire gradué de $L$, qui contient un diviseur ample, alors on a $$\label{Equ:majorationbis}\operatorname{rg}_k(V_1)\leqslant\frac{\mathrm{vol}(V_\sbullet)}{(d+1)!}+\widetilde{\varepsilon}^{\Theta}(V_\sbullet).$$ La démonstration est presque identique à celle du théorème \[Thm:majorationdeHS\]. Il suffit de remplacer les filtrations de Harder-Narasimhan par les filtrations par minima. Le cas où $d=0$ utilise notamment les estimés démontrées dans le théorème \[Thm:h0etdegplus\] qui sont valables pour tout corps $k$, et la démonstration reste donc inchangée. Dans la suite, on suppose $d\geqslant 1$. On suppose en outre que $\Theta$ est de la forme $(p_0:X\rightarrow C_0,\Theta')$, où $\Theta'$ est un tour de fibrations sur courbes de la fibre générique de $p_0$. Soient $g_0$ le genre de la courbe $C_0$, $\mu_0:=\mu^{p_0}_{\max}(V_\sbullet)$, et $W_\sbullet$ la fibre générique de $E_\sbullet=p_{0*}(V_\sbullet)$. On munit $W_\sbullet$ de la filtration par minima $\mathcal F$ et on note $W_\sbullet^t=\bigoplus_{n\geqslant 0}\mathcal F^{nt}W_n$. On a encore $$\operatorname{vol}(V_\sbullet)=\operatorname{vol}(E_\sbullet)=(d+1)\int_0^{\mu_0}\operatorname{rg}(W_\sbullet^t)\,\mathrm{d}t,$$ où la première égalité provient du théorème \[Thm:passageauxfibres\] et on a utilisé la relation $\lambda_1(E_1)\leqslant\mu_0$ (cf. la remarque \[Rem:comparaisonlambda1\]) dans la deuxième égalité. En outre, les relations et montre que $$\deg_+(E_1)\leqslant\int_0^{\mu_0}\operatorname{rg}(W_1^t)\,\mathrm{d}t+\operatorname{rg}(W_1)\ell(g_0).$$ On applique l’hypothèse de récurrence à $W_\sbullet^t$ et obtient $$\begin{split}\deg_+(E_1)&\leqslant\int_0^{\mu_0}\Big( \frac{\operatorname{vol}(W_1^t)}{d!}+\widetilde{\varepsilon}^{\Theta'}(W_\sbullet^t)\Big)\,\mathrm{d}t+\operatorname{rg}(W_1)\ell(g_0)\\ &\leqslant\frac{\operatorname{vol}(V_\sbullet)}{(d+1)!}+\mu_0\widetilde{\varepsilon}^{\Theta'}(W_\sbullet^t)+\operatorname{rg}(W_1)\ell(g_0). \end{split}$$ On en déduit $$\begin{split}\operatorname{rg}(V_1)&\leqslant h^0(E_1)\leqslant\deg_+(E_1)+\operatorname{rg}(W_1)\max(g_0-1,1)\\ &\leqslant\frac{\operatorname{vol}(V_\sbullet)}{(d+1)!}+\mu_0\widetilde{\varepsilon}^{\Theta'}(W_\sbullet^t)+\operatorname{rg}(W_1)(\ell(g_0)+\max(g_0-1,1)). \end{split}$$ On applique alors l’hypothèse de récurrence à $W_\sbullet$ et obtient le résultat souhaité. [^1]: La convergence de cette suite a été démontrée dans [@Chen08]. On peut aussi la déduire du théorème de Fujita arithmétique [@Chen10; @Yuan09]. [^2]: En effet, l’homomorphisme naturel de $E_{n,\operatorname{des}}\otimes E_{m,\operatorname{des}}$ vers $E_{n+m}$ est non-nul. En outre, le corollaire \[Cor:filtrationdeHNenalg\] montre que $E_{n,\operatorname{des}}\otimes E_{m,\operatorname{des}}$ est semi-stable de pente $\mu_{\max}(E_n)+\mu_{\max}(E_m)$. On obtient donc l’inégalité souhaitée. [^3]: C’est-à-dire que $E_n$ est l’image de l’homomorphisme $\varphi^*(V_n)\rightarrow\pi_*(L^{\otimes n})$ induit par l’inclusion $V_n\rightarrow H^0(X,L^{\otimes n})=\varphi_*(\pi_*(L^{\otimes n}))$ via l’adjonction entre les foncteurs $\varphi_*$ et $\varphi^*$, où $\varphi:C\rightarrow\operatorname{Spec}k$ désigne le morphisme structurel. [^4]: Cela provient d’un analogue dans le cadre de corps de fonction du corollaire 4.8 de [@Zhang95]. On peut suivre la stratégie de *loc. cit.*. La démonstration est plus simple car les places archimédienne ne se manifestent pas dans le problème. [^5]: Ici la platitude est équivalente à la surjectivité du morphisme, cf. [@LiuQing proposition 4.3.9]. [^6]: cf. la définition \[Def:imagedirecte\] pour la construction de $\mu_{\max}^{p_0}(V_\sbullet)$. [^7]: D’après la remarque \[Rem:systemegradue\], on a $\mu_{\max}(E_1)\leqslant\mu_{\max}^{\mathrm{asy}}(E_\sbullet)=\mu_0$. Donc $W_1^t=\{0\}$ lorsuqe $t>\mu_0$. [^8]: Soit $\overline E=(E,(\|.\|_v)_{v\in M_K})$ un fibré vectoriel adélique sur $\operatorname{Spec}K$. Pour toute place non-archimédienne $\mathfrak p$, la norme $\|.\|_{\mathfrak p}$ est dite pure si l’image de sa restriction à $E$ s’identifie à l’image de la valeur absolue $|.|_{\mathfrak p}:K\rightarrow\mathbb R$. [^9]: Dans son exposé au colloque “*Arakelov theory and its arithmetic applications.*” le 22 février 2010 à Regensbourg [^10]: Comme on considère les minima absolus, le défaut de pureté est anodin ici. [^11]: Si on fixe un isomorphisme d’espaces vectoriels $\phi:E\rightarrow K^n$, où $n=\operatorname{rg}_K(E)$, alors le degré d’Arakelov de $\overline E$ est défini comme $$\operatorname{\widehat{\deg}}(\overline E)=\ln\frac{\mathrm{vol}(\phi(\mathbb B(\overline E)))}{\mathrm{vol}(\mathbb B(\overline K^n))},$$ où $\mathbb B(.)$ désigne la boule unité adélique, et $\mathrm{vol}$ désigne une mesure de Haar sur l’espace adélique $\mathbb A_K^n$. Cette définition ne dépend pas du choix de $\phi$ et $\mathrm{vol}$. [^12]: On renvoie les lecteurs dans [@Gaudron08 §4] pour les détails. [^13]: Pour tout fibré vectoriel adélique non-nul (non nécessairement hermitien) $\overline F$ sur $\operatorname{Spec}K$, on a $\Lambda_1(\overline F)\leqslant\operatorname{\widehat{\mu}}_{1}(\overline F)\leqslant\Lambda_1(\overline F)+ \ln(\operatorname{rg}(F))$. Par conséquent, $\operatorname{\widehat{\mu}}_{\max}^{\mathrm{asy}}(\overline E_\sbullet)$ est égal à $\displaystyle\lim_{n\rightarrow+\infty}\operatorname{\widehat{\mu}}_{\max}(\overline E_n)/n.$ [^14]: C’est-à-dire que $\mathscr L$ est nef relativement à $\pi$, la métrique de $\overline{\mathscr L}$ est pluri-sous-harmonique, et la fonction hauteur sur l’ensemble des points algébriques de $\mathscr X_K$ définie par $\overline{\mathscr L}$ est à valeurs positives. [^15]: Rappelons qu’un élément dans $\mathbb P(E^\vee)_k(K)$ correspond à un sous-espace $K$-vectoriel de rang $1$ de $E_{K}$.

ArXiv

--- abstract: | The [Aladin]{} interactive sky atlas, developed at CDS, is a service providing simultaneous access to digitized images of the sky, astronomical catalogues, and databases. The driving motivation is to facilitate direct, visual comparison of observational data at any wavelength with images of the optical sky, and with reference catalogues. The set of available sky images consists of the STScI Digitized Sky Surveys, completed with high resolution images of crowded regions scanned at the MAMA facility in Paris. A Java WWW interface to the system is available at: [http://aladin.u-strasbg.fr/]{}. author: - François Bonnarel - Pierre Fernique - Olivier Bienaymé - Daniel Egret - Françoise Genova - 'Mireille Louys[^1]' - François Ochsenbein - Marc Wenger - 'James G. Bartlett[^2]' date: 'Received 6 December 1999 / Accepted 16 December 1999' subtitle: A Reference Tool for Identification of Astronomical Sources title: The ALADIN Interactive Sky Atlas --- Introduction ============ The CDS ------- The Centre de Données astronomiques de Strasbourg (CDS) defines, develops, and maintains services to help astronomers find the information they need from the very rapidly increasing wealth of astronomical information, particularly on-line information. In modern astronomy, cross-matching data acquired at different wavelengths is often the key to the understanding of astronomical phenomena, which means that astronomers have to use data and information produced in fields in which they are not specialists. The development of tools for cross-identification of objects is of particular importance in this context of multi-wavelength astronomy. A detailed description of the CDS on-line services can be found, e.g., in Egret et al. ([@cds-amp2]) and in Genova et al. ([@cds-hub], [@cds], [@cds2000]), or at the CDS web site[^3]. The ALADIN Project ------------------ Several sites currently provide on-line access to digitized sky surveys at different wavelengths: this is, for instance, the case of Digitized Sky Survey (DSS) at STScI (Morrison [@1995adass...4..179M]), and of similar implementations at other sites, providing quick access to cutouts of the compressed DSS images. [SkyView]{} at HEASARC (McGlynn et al. [@skyview]) can generate images of any portion of the sky at wavelengths in all regimes from radio to gamma-ray. Some of these services provide simultaneous access to images and to catalogue data. The [SkyCat]{} tool, recently developed at ESO (Albrecht et al. [@skycat]), addressed this concern in the context of the European Southern Observatory scientific environment (in view of supporting future users of the Very Large Telescope); [SkyCat]{} uses a standardized syntax to access heterogeneous astronomical data sources on the network. [Aladin]{} has been developed independently by the CDS since 1993 as a dedicated tool for identification of astronomical sources – a tool that can fully benefit from the whole environment of CDS databases and services, and that is designed in view of being released as a multi-purpose service to the general astronomical community. [Aladin]{} is an interactive sky atlas, allowing the user to visualize a part of the sky, extracted from a database of images from digitized surveys and observational archives, and to overlay objects from the CDS catalogues and tables, and from reference databases ([Simbad]{} and NED), upon the digitized image of the sky. It is intended to become a major cross-identification tool, since it allows recognition of astronomical sources on the images at optical wavelength, and at other wavelengths through the catalogue data. Expected usage scenarios include multi-spectral approaches such as searching for counterparts of sources detected at various wavelengths, and applications related to careful identification of astronomical objects. [Aladin]{} is also heavily used for the CDS needs of catalogue and database quality control. In the case of extensive undertakings (such as checking the astrometric quality for a whole catalogue), it is expected that [Aladin]{} will be useful for understanding the characteristics of the catalogue or survey, and for setting up the parameters to be adjusted while fine tuning the cross-matching or classification algorithms, by studying a sample section of objects or fields. A discussion of the usage of such a tool for cross-identification can be found in Bartlett & Egret ([@xid-179]), where it is shown how *training sets* are used to build likelihood ratio tests. The [Aladin]{} interactive atlas is available in three modes: a simple previewer, a Java interface, and an X-Window interface. We describe here mostly the Java interface which is publicly accessible on the World-Wide Web. Access modes ============ After a long phase of development, (see e.g., Paillou et al. [@paillou]), [Aladin]{} has been first distributed to a limited number of astronomy laboratories in 1997, as an X-Window client program, to be installed on a Unix machine on the user side. The client program interacts with the servers running on Unix workstations at CDS (image server, catalogue server, [Simbad]{} server) and manages image handling and plane overlays. The strategy of having a client program on the user side is difficult to maintain on the long run. The World-Wide Web offers, with the development of Java applications (or *applets*), a way to solve this difficulty. Actually, there is still a *client* program: this is the Java applet itself, that the user receives from the WWW server. Most current Internet browsers are able to make it run properly, so that the user does not have to install anything special other than an Internet browser. As a consequence, [Aladin]{} is currently available in the three following modes: [Aladin]{} previewer: : a pre-formatted image server provides a compressed image of fixed size ($14.1\arcmin \times 14.1\arcmin$ for the DSS-I) around a given object or position. When an object name is given, its position is resolved through the [Simbad]{} name resolver. Anchors pointing to the previewer are integral part of the World-Wide Web interfaces to the [Simbad]{} database[^4] and to the CDS bibliographic service[^5]. The result page also gives access to the full resolution FITS image for download. [Aladin]{} Java: : this is the primary public interface, supporting queries to the image database and overlays from any catalogue or table available at CDS, as well as from [Simbad]{} and NED databases. Access to personal files is not possible (due to security restrictions of the Java language). These restrictions do not apply to the *stand-alone* version, which can be installed and run on a local *Java virtual machine*. [Aladin X]{}: : The X-Window [Aladin]{} client provides most of the functionalities of the [Aladin]{} Java interface, plus more advanced functions, as described below (section \[AladinX\]). The image database ================== Database summary ---------------- The [Aladin]{} image dataset consists of: - The whole sky image database from the first Digitized Sky Survey (DSS-I) digitized from photographic plates and distributed by the Space Telescope Science Institute (STScI) as a set of slightly compressed FITS images (with a resolution of $1.8\arcsec$); DSS-II is also currently being integrated into the database (see below); - Images of *crowded* fields (Galactic Plane, Magellanic Clouds) at the full resolution of $0.67\arcsec$, scanned at the [*Centre d’Analyse des Images*]{} (MAMA machine) in Paris; - Global plate views ($5\degr \times 5\degr$ or $6\degr \times 6\degr$ according to the survey) are also available for all the plates contributing to the image dataset: these are built at CDS by averaging blocks of pixels from the original scans; - Other images sets, or user-provided images, in FITS format, having suitable World Coordinate System information in the header (see e.g. Greisen & Calabretta [@wcs]); this functionality is currently available only for the Java stand-alone version. Building the database contents ------------------------------ The [Aladin]{} project has set up collaborations with the major groups providing digitizations of sky surveys. The original surveys are made of photographic Schmidt plates obtained at Palomar in the North, and ESO/SERC in the South, and covering the whole sky at different epochs and colours (see e.g., MacGillivray [@potsdam]). The database currently includes the first Digitized Sky Survey (DSS-I) produced by the Space Telescope Institute (Lasker [@dss]), for the needs of the Hubble Space Telescope. To create these images, the STScI team scanned the first epoch (1950/1955) Palomar $E$ Red and United Kingdom Schmidt $J$ Blue plates (including the SERC J Equatorial Extension and some short V-band plates at low galactic latitude) with a pixel size of $1.7\arcsec$ ($25{\mu}m$). The low resolution and a light data compression (factor of 10) permit storage of images covering the full sky on a set of 102 CD-ROMs. DSS-II images in the R-band (from Palomar POSS-II F and UK Schmidt SES, AAO-R, and SERC-ER), scanned with a $1\arcsec$ ($15{\mu}m$) sampling interval (see Lasker [@dss-ii]) are gradually being included into the system, and will soon be followed by DSS-II images in the B-band (POSS-II J). In addition, high resolution digitalization of POSS-II, SERC-J, SERC-SR, SERC-I, or ESO-R plates featuring crowded regions of the sky (Galactic Plane and Magellanic Clouds) have been provided by the MAMA facility at the [*Centre d’Analyse des Images*]{} (CAI), Observatoire de Paris (Guibert [@MAMA]). Sampling is $0.67\arcsec$ per pixel ($10{\mu}m$). Currently, these high resolution images cover about 15% of the sky, and are stored in a juke-box of optical disks, with a capacity of 500 Gigabytes. The image server ---------------- The image server for [Aladin]{} had to be able to deal with various survey data, in heterogeneous formats (uncompressed FITS, compressed JPEG or PMT – see Section \[compression\], etc.). For that, an object-oriented design was chosen, allowing an easy manipulation of image calibrations and headers, through the use of object classes. Image compression or decompression, image reconstruction, and in a near future, part of the recalibration, are seen as class methods. Images are currently divided into subimages of $500 \times 500$ pixels (DSS-I), $768 \times 768$ pixels (DSS-II), or $1024 \times 1024$ pixels (MAMA). The 1.5 million subimages are described by records stored in a relational database, encapsulated by several classes of the image management software. When an image of the sky is requested, the original subimages containing the corresponding sky area are retrieved through SQL commands, and the resulting image is built on the fly. Usage scenarios =============== In this section we will focus on describing the usage of the [Aladin]{} Java interface, as it is available now (November 1999). Access ------ The [Aladin]{} home page is available through the CDS Web server at the following address: http://aladin.u-strasbg.fr/ This site provides access to [Aladin]{} documentation, including scientific reports, recent publications, etc. Query modes ----------- The typical usage scenario starts with a request of the digitized image for an area of the sky defined by its central position or name of central object (to be resolved by [Simbad]{}). The size of the sky field is determined by the photographic survey used: it is $14.1\arcmin$ in the case of the DSS-I. Astrometric information comes from the FITS header of the DSS image, and is generally accurate to the arcsecond (with deviations up to several arcsec. in exceptional cases, on plate edges). In a subsequent step, the interface, illustrated by Figs. \[query-panel\] and \[sample-image\], allows the user to stack several information planes related to the same sky field, to superimpose the corresponding data from catalogues and databases, and to obtain interactive access to the original data. The possible information planes are the following: - Image pixels from the [Aladin]{} database of digitized photographic plates (DSS-I, MAMA, DSS-II); functionalities include zooming capabilities, inverse video, modification of the color table; - Information from the [Simbad]{} database (Wenger et al. [@simbad]); objects referenced in [Simbad]{} are visualized by color symbols overlaid on top of the black and white image; the shape and color of the symbols can be modified on request, and written labels can be added for explicit identification of the objects; these features are also available for all the other information planes; - Records from the CDS library of catalogues or tables ([VizieR]{}[^6], Ochsenbein et al. [@vizier]); the user can select the desired catalogue from a preselected list including the major reference catalogues such as the Tycho Reference Catalogue (ESA [@tyc]; H[ø]{}g et al. [@trc]), GSC (Lasker et al. [@gsc]), IRAS Point Source Catalog, or USNO A2.0 (Monet [@usno]); the user can alternatively select the catalogues for which entries may be available in the corresponding sky field, using the [VizieR]{} query mechanism by position (see \[catserver\]), catalogue name or keyword; - Information from the NED database: objects referenced in the NASA/IPAC Extragalactic Database[^7] (Helou et al. [@ned]) can also be visualized through queries submitted to the NED server at IPAC; - Archive images will gradually become available through the corresponding mission logs: Hubble Space Telescope images are currently available (see Fig. \[M16\] for an example), and more archives will follow. - Local, user data files can also be overlaid, but, because of current limitations of the Java applications, this feature is only available in the stand-alone version, or in [Aladin X]{}. The stack of images and graphics is made visible to the user (under the eye icon, on the right of Fig. 2) so that each plane can be turned on and off. The status of queries is also easily visualized. For all information planes ([Simbad]{}, [VizieR]{}, NED) links are provided to the original data. This is done in the following way: when selecting an object on the image, with mouse and cursor, it is possible to call for the corresponding information which will appear in a separate window on the Internet browser. It is also possible to select with the mouse and cursor all objects in a rectangular area: the corresponding objects are listed in a panel on the bottom of the [Aladin]{} window; this list includes basic information (name, position and, when applicable, number of bibliographical references) and anchors pointing to the original catalogue or database. At any moment the position of the cursor is translated in terms of right ascension and declination on the sky and visualized in the top panel of the [Aladin]{} window. Additional features are available, such as a tool for computing angular distance between marked objects. The *standalone* version gives access to additional facilities, not available through the Web, such as printing and saving the images and data. The catalogue server {#catserver} -------------------- The ability to access all [VizieR]{} catalogues and tables directly from [Aladin]{} is a unique feature which makes it an extremely powerful tool for any cross-identification or classification work. The “*Select around target*” request relies on a special feature – the genie of the lamp: this is the ability to decide which catalogues, among the database of (currently) over 2,600 catalogues or tables, contain data records for astronomical objects lying in the selected sky area. In order to do that, an index map of [VizieR]{} catalogues is produced (and kept up-to-date), on the basis of about ten pixels per square degree: for each such ‘pixel’ the index gives the list of all catalogues and tables which have entries in the field. When a user hits the button “*Select around target*”, this index is queried and the list of useful catalogues is returned. It is possible, at this stage, either to list all catalogues, or to produce a subset selected on the basis of keywords. Note that, as the index “pixels” generally match an area larger than the current sky field, there is simply a good chance, but not 100%, to actually obtain entries in the field when querying one of the selected tables. Cache ----- The images of the 30,000 most cited objects in [Simbad]{} are pre-computed and available on a cache on magnetic disk. For these objects, the image is served much faster than for other objects where the image has to be extracted from the Digitized Sky Survey. Usage statistics ---------------- As the newest service developed by CDS, [Aladin]{} has not yet been widely publicized, and its usage is in a steeply growing phase. Currently about 10,000 queries are processed monthly, generating the extraction of more than 5,000 images. Image compression {#compression} ================= Astronomical image compression in the context of [Aladin]{} has been discussed in detail by Louys et al. ([@louys]). For the [Aladin]{} Java interface and for the [Aladin]{} previewer, the current choice has been to deliver to the user an image in JPEG 8-bit format, constructed from the original FITS images. JPEG is a general purpose standard which is supported by all current Internet browsers. The size of such an image does not exceed 30 kBytes, and thus the corresponding network load is very small. In the near future, the Pyramidal Median Transform (PMT) algorithm, implemented in the MR-1 package (Starck et al. [@pmt]), will be used within [Aladin]{} for storing or transferring new image datasets, such as additional high resolution images (see again Louys et al. [@louys] for details). The corresponding decompression package is being written in Java code, and could be downloaded on request for use within the Java interface. Aladin X {#AladinX} ======== The [Aladin]{} X-Window interface is the testbed for further developments. It is currently only distributed for the Unix Solaris operating system. Interested potential users should contact CDS for details. Source extraction ----------------- [Aladin X]{} includes a procedure for source extraction. The current mechanism will soon be replaced by SExtractor (Bertin & Arnouts ). Plate calibrations ------------------ While the first level astrometric calibrations are given by the digitizing machines, a second level is being developed that will allow the user to *recalibrate* the image with a new set of standards taken, for example, from the Tycho Reference Catalogue. The photometric calibrations (surface and stellar) will eventually also be performed within [Aladin]{}, by using the Guide Star Photometric Catalogs (GSPC I and II; Ferrari et al. [@gspc2]; Lasker et al. [@gspc1]). Users will thus be able to work on the details of local astrometric and photometric plate calibrations in order to extract the full information from the digitized plates. Integration of distributed services =================================== While the CDS databases have followed different development paths, the need to build a transparent access to the *whole set* of CDS services has become more and more obvious with the easy navigation permitted by hypertext tools. [Aladin]{} has become the prototype of such a development, by giving comprehensive simultaneous access to [Simbad]{}, the [VizieR]{} Catalogue service, and to external databases such as NED, using a client/server approach and, when possible, standardized query syntax and formats. In order to be able to go further, the CDS has built a general data exchange model, taking into account all types of information available at the Data Center, known under the acronym of GLU for Générateur de Liens Uniformes – Uniform Link Generator (Fernique et al. [@glu]). More generally, with the development of the Internet, and with an increasing number of on-line astronomical services giving access to data or information, it has become critical to develop new tools providing access to distributed services. This is, for instance, the concern expressed by NASA through the AstroBrowse project (Heikkila et al. [@astrobrowse]). A local implementation of this concept is available at CDS (AstroGlu: Egret et al. [@astroglu]). Future developments =================== An important direction of development in the near future is the possibility of providing access to images from other sky surveys or deep field observations: obvious candidates are the DENIS (Epchtein [@denis]) and 2MASS (Skrutskie [@2mass]) near-infrared surveys. The first public point source catalogues resulting from these surveys are already available through [Aladin]{}, since they are included in the [VizieR]{} service. This has already proved useful for validating survey data in preliminary versions of the DENIS catalogue (Epchtein et al. [@denis-psc]). The CDS team will also continue to enrich the system functionality. The users play an important role in that respect, by giving feedback on the desired features and user-friendliness of the interfaces. New developments are currently considered as additional modules which will be incorporated to the general release only when needed, possibly as optional downloads, in order to keep the default version simple and efficient enough for most of the Web applications. On a longer term, the CDS is studying the possibility of designing *data mining* tools that will help to make a fruitful use of forthcoming very large surveys, and will be used for cross-matching several surveys obtained, for instance, at different wavelengths. A first prototype, resulting from a collaboration between ESO and CDS, in the framework of the VLT scientific environment is currently being implemented (Ortiz et al. [@ortiz]). CDS acknowledges the support of INSU-CNRS, the Centre National d’Etudes Spatiales (CNES), and Université Louis Pasteur. We are indebted to Michel Crézé who initiated the project while being Director of the CDS, and to all the early contributors to the [Aladin]{} project: Philippe Paillou, Joseph Florsch, Houri Ziaeepour, Eric Divetain, Vincent Raclot. Collaboration with STScI, and especially with the late Barry Lasker, and with Brian McLean, is gratefully acknowledged. We thank Jean Guibert and René Chesnel from CAI/MAMA for their continuous support to the project. The Digitized Sky Survey was produced at the Space Telescope Science Institute under U.S. Government grant NAG W-2166. The images of these surveys are based on photographic data obtained using the Oschin Schmidt Telescope on Palomar Mountain and the UK Schmidt Telescope. Java is a registered trademark of Sun Microsystems. Albrecht, M. A., Brighton, A., Herlin, T., et al., 1997, in [*Astronomical Data Analysis Software and Systems VI*]{}, ASP Conf. Ser. 125, p. 333 Bartlett, J.G., Egret, D., 1997, in [*New Horizons from Multi-Wavelength Sky Surveys*]{}, IAU Symp. 179, Kluwer Academic Publ., p. 437 Bertin, E., Arnouts, S. 1996, A&AS 117, 393 Egret, D., Crézé, M. , Bonnarel, F., et al., 1995, in [*Information & On-line Data in Astronomy*]{}, Egret & Albrecht (Eds.), Kluwer Academic Publ., p. 163 Egret, D., Fernique, P., Genova, F., 1998, in [*Astronomical Data Analysis Software and Systems VII*]{}, ASP Conf. Ser. 145, 416 Epchtein, N., 1998, in *The impact of near-infrared surveys on galactic and extragalactic astronomy*, Proc. 3rd. Euroconf., Kluwer Academic Publ., ASSL 230, p. 3 Epchtein, N., Deul, E., Derriere, S., et al., 1999, A&A 349, 236 ESA, 1997, The Hipparcos and Tycho Catalogues, ESA SP–1200 Fernique, P., Ochsenbein, F., Wenger, M., 1998, in [*Astronomical Data Analysis Software and Systems VII*]{}, ASP Conf. Ser. 145, p. 466 Ferrari, A., Siciliano, E.D., Pizzuti, A., et al., 1994, in [*Astronomy from Wide-Field Imaging*]{}, IAU symposium 161, H.T. MacGillivray & E.B. Thomson (Eds.), Kluwer Academic Publ., p. 301 Genova, F., Bartlett, J.G., Bienaymé, O., et al., 1996, Vistas in Astronomy 40, 429 Genova, F., Bartlett, J.G., Bonnarel, F., et al., 1998, in [*Astronomical Data Analysis Software and Systems VII*]{}, ASP Conf. Ser. 145, p. 470 Genova, F., et al., 2000, A&AS, *in press* (CDS) Greisen, E.W., Calabretta, M., 1995, in [*Astronomical Data Analysis Software and Systems IV*]{}, ASP Conf. Ser. 77, p. 233 Guibert, J. 1992, in [*Digitized Optical Sky Surveys*]{}, H.T. MacGillivray & E.B. Thomson (Eds.), Kluwer Academic Publ., p. 103 Heikkila, C.W., McGlynn, T.A., White, N.E., 1999, in [*Astronomical Data Analysis Software and Systems VIII*]{}, ASP Conf. Ser. 172, p. 221 Helou, G., Madore, B.F., Schmitz, M., et al., 2000, A&AS, *in press* (NED) H[ø]{}g, E., Kuzmin, A., Bastian, U., et al., 1998, A&A 335, 65 Lasker, B.M., Sturch, C.R., Lopez, C., et al., 1988, ApJS 68, 1 Lasker, B.M., Sturch, C.R., McLean, B.J., et al., 1990, AJ 99, 2019 Lasker, B.M., 1992, in [*Digitized Optical Sky Surveys*]{}, H.T. MacGillivray & E.B. Thomson (Eds.), Kluwer Academic Publ., p. 87 Lasker, B.M., 1994, in [*Astronomy from Wide-Field Imaging*]{}, IAU symposium 161, H.T. MacGillivray & E.B. Thomson (Eds.), Kluwer Academic Publ., p.167 Louys, M., Starck, J.-L., Mei, S., Bonnarel, F., Murtagh, F., 1999, A&AS 136, 579 MacGillivray, H.T., et al., Editors, 1994, [*Astronomy from Wide-Field Imaging*]{}, Postdam, Germany, Kluwer Academic Publ., pp. 1-760 McGlynn, T., Scollick, K., White, N., 1997, in [*New Horizons from Multi-Wavelength Sky Surveys*]{}, IAU Symp. 179, Kluwer Academic Publ., p. 465 Monet, D., et al., 1998, [*The PMM USNO A2.0 Catalogue*]{}, US Naval Observatory Flagstaff Station Morrison, J. E., 1995, in [*Astronomical Data Analysis Software and Systems IV*]{}, ASP Conf. Ser. 77, p. 179 Ochsenbein, F., Bauer, P., Genova, F., 2000, A&AS, *in press* (VizieR) Ortiz, P., Ochsenbein, F., Wicenec, A., Albrecht, M., 1999, in [*Astronomical Data Analysis Software and Systems VIII*]{}, ASP Conf. Ser., 172, 379 Paillou, Ph., Bonnarel, F., Ochsenbein, F., Crézé, M., 1994, in [*Astronomy from Wide-Field Imaging*]{}, IAU symposium 161, H.T. MacGillivray and E.B. Thomson (Eds.), Kluwer Academic Publ., p. 347 Skrutskie, M., 1998, in “The impact of near-infrared surveys on galactic and extragalactic astronomy”, Proc. 3rd. Euroconf., Kluwer Academic Publ., ASSL 230, 11 Starck, J.-L., Murtagh, F., Pirenne, B., Albrecht, M., 1996, PASP 108, 446 Wenger, M., Ochsenbein, F., Egret, D., et al., 2000, A&AS, *in press* (Simbad) [^1]: Laboratoire des Sciences de l’Informatique, de l’Image et de la Télédétection, ENSPS, Université Louis Pasteur, F-67000 Strasbourg, France [^2]: *Present address:* Observatoire Midi-Pyrénées, Toulouse, France [^3]: *Internet address:* http://cdsweb.u-strasbg.fr/ [^4]: *Internet address*: http://simbad.u-strasbg.fr/Simbad [^5]: *Internet address*: http://simbad.u-strasbg.fr/biblio.html [^6]: *Internet address:* http://vizier.u-strasbg.fr/ [^7]: *Internet address:* http://nedwww.ipac.caltech.edu/

ArXiv

--- abstract: 'Soft solids, such as elastomers, have an instability toward the nucleation and growth of sharply creased surface folds. We develop an analogy between this instability and a first order phase transition by constructing an isolated fold as a domain wall-like structure separating two, coexisting scale invariant deformations of an incompressible elastomer: an affine deformation, and a crease. We relate the existence of this surface folding instability and the striking non-existence of a phase boundary separating the two scale invariant deformations to certain convexity properties of the elastomer free energy.' author: - Evan Hohlfeld title: Coexistence of scale invariant states in a scale free system --- Liquefaction of argon, nematic ordering of the liquid crystal 5CB, and allotropy in iron are well studied examples of first order phase transitions in materials. These phenomena can be understood at a macroscopic scale as coexistences between different scale free, homogenous states—phases. Recently, soft solids such as a hydrogels, elastomers, and tissues, were discovered to possess a novel kind of scale free instability with an uncanny, though imperfect resemblance to a first order phase transition [@Hohlfeld-Mahadevan; @Hohlfeld-Mahadevan2012; @Chen]. When a soft solid surface is sufficiently compressed, infinitesimal, sharply creased folds can nucleate and grow [@Hohlfeld-Mahadevan; @Chen]. This critical compression is independent of the sample shape and marks the coexistence of localized folds, called *sulci*, and a smooth free surface. With further quasi-static compression, work flows in from the displaced boundaries of the sample to the individual sulci as these grow. During growth, the compression close to each sulcus remains at the coexistence value [@Hohlfeld-Mahadevan2012]. The amount of work is proportional to the folded *volume*, and hence can be interpreted as the work of transformation. Beyond a second, larger critical compression the flat free surface loses metastability and sulci spontaneously form [@Biot; @Hohlfeld-Mahadevan]. This second critical compression can be understood as a kind of spinodal point for the flat surface [@Onuki89; @Hohlfeld-Mahadevan2012]. The analogy with a phase transition fails, however, when we attempt identify a sulcus with a “droplet” of a new phase, since, for example, no clear phase boundary divides the sulcus from the surrounding elastomer [@Hohlfeld-Mahadevan2012]. In this Letter we propose an explanation for the partial analogy between the sulcification (or creasing) instability of an elastomer surface and a phase transition. Whereas phase transitions are transformations between distinct *homogenous*, scale invariant states, instabilities like sulcification occur when a system transitions between scale invariant, though potentially *inhomogeneous* states. We explain how to construct these states, the transitions between them, and the striking absence of phase boundaries. Recent interest in sulcification has focused on the effects of swelling [@Tanaka; @Hayward; @Hayward2010; @Zalachas], growth [@Benamar; @Suo; @Bayly], applied fields [@Wang], mechanical confinement[@Gent; @Ghatak; @P.-M.-Reis:2009fk; @Hohlfeld-Mahadevan; @Mora; @Suo1], and imperfections [@Hutchinson], on how patterns form[@Hohlfeld-Mahadevan2012; @Chen; @Tallinen], and on connections to other phenomena such as plastic folding [@Sundaram]. In previous work we calculated the critical strain for sulcificiation [@Hohlfeld-Mahadevan] and developed a bifurcation and pattern formation theory based on treating sulcification point as novel kind of nonlinear critical point [@Hohlfeld-Mahadevan2012]. The later work was built on the supposition that an isolated sulcus in a critically compressed rubber half-space is a local minimum of energy up to translations and changes of scale. At the critical point, the scaling mode changes stability, allowing a nucleated sulcus to grow from infinitesimal size. Linearization fails to detect this instability because scaling changes the size of the sulcus without changing the magnitude of its deformation gradient. The possibility of such an undetectable small amplitude instability involving a continuous deformation was actually first anticipated by Weierstrass when he distinguished so-called weak (i.e. merely linearly stable) local minima from strong local minima. The detection of Weierstrass instabilities has remained an open problem, however [@Ball-Marsden]. Our exposition here adds to these ongoing discussions by showing how to construct an isolated sulcus solution in a simple elastomer, and how this construction might generalize to more complex systems. Our plan is to systematically develop the analogy between sulcification and a phase transition. To frame this discussion, we consider a sample of incompressible neo-Hookean elastomer. This is a model material for many rubbery solids which are dramatically softer in simple shear than in volumetric compression. It is compactly specified by the Lagrangian density $$\mathcal{L}\left(\mathbf{A}\right)=\frac{\mu}{2} A^i_\alpha A_i^\alpha - p\left(\det [A^i_\alpha] -1 \right) \label{eq:NHM}$$ where the matrix $\mathbf{A}=\partial \mathbf{x}/\partial \mathbf{X}$ is the gradient of the deformation $\mathbf{x}(\mathbf{X})$ of a reference volume. The first term in (\[eq:NHM\]) governs the entropic elasticity of the network chains with shear modulus $\mu$, and the second term, involving the Lagrange multiplier $p$, enforces the nonlinear incompressibility constraint $\det \partial \mathbf{x}/\partial \mathbf{X}=1$. When restricted to isochoric deformations so that the constraint term in (\[eq:NHM\]) is identically zero, $\mathcal{L}=\mathcal{F}$, the free energy density of the elastomer. In terms of $\mathcal{F}$, we impose the volumetric constraint by setting $\mathcal{F}(\mathbf{A})=+\infty$ for $\det(\mathbf{A})\ne1$. The Lagrangian density $\mathcal{L}$ is invariant under rigid translations and rotations of $\mathbf{x}$ and $\mathbf{X}$, separately, i.e. $\mathbf{x}\to\mathcal{O}\mathbf{x} + \mathbf{c}$ or $\mathbf{X}\to\mathcal{O}\mathbf{X}+\mathbf{c}$ where $\mathcal{O}$ is an orthogonal matrix and $\mathbf{c}$ is a vector. The first of these symmetries is the so-called objectivity of the model while the second expresses that the elastomer is a homogenous and isotropic material. Our focus is on another symmetry of $\mathcal{L}$. Because the Lagrangian density $\mathcal{L}$ is a function of the dimensionless deformation gradient $\mathbf{A}$, it is also invariant under the scale transformation $\{\mathbf{x},\mathbf{X}\}\to \{L\mathbf{x},L\mathbf{X}\}$ for any scaling factor $L>0$. Unlike rigid translations and rotations, however, the measure $d\mathbf{X}$ is not invariant under scale transformations. Thus while a scale transformation is not a symmetry of the Lagrangian $L=\int_{\Omega} \mathcal{L}d\mathbf{X}$—and so not associated with a Noetherian conservation law—it is nevertheless a symmetry of the Euler-Lagrange equation. Many material models possess scale symmetry as an intermediate asymptotic property between the molecular length scale and the system scale. The configurations of small enough patches of material, which are still larger than the small-scale cut-off, must be independent of both the macroscopic and microscopic details. In particular such configurations must become invariant under rescaling as the system size is taken to infinity. To compare the elastomer with other materials, it is helpful to view the deformation gradient $\mathbf{A}$ as a tensor order parameter. Then, what distinguishes the elastomer from other materials with tensor order parameters, say nematic liquid crystals, is the *compatibility constraint* that $\mathbf{A}=\partial \mathbf{x}/\partial \mathbf{X}$. A similar constraint holds for the scalar phase variables that appear in theories of smectic liquid crystals and superconductors. What distinguishes the elastomer from smectics and superconductors is the vector nature of the “phase” variable $\mathbf{x}$. Because of the compatibility constraint, an external field conjugate to $\mathbf{A}$ is equivalent to the application of tractions to the boundary of a sample. The compatibility constraint plays a fundamental role in our analysis below. Even though rubbers are stable enough materials to use in car tires, let us for the moment forget this stability and compute the “phase diagram” corresponding to (\[eq:NHM\]), viewed as an abstract material. For simplicity, we consider only plane strain deformations so that $\mathbf{x},\mathbf{X}\in\mathbb{R}^2$. In this exercise, we would identify different phases with homogenous values of $\mathbf{A}$ corresponding to distinct local minima of the free energy density $\mathcal{F}$. This is, for example, how the austenite and martensite phases of shape-memory alloys are described. If $\mathcal{F}$ has multiple local minima, then we can consider the coexistence of any two phases. Because of the compatibility constraint, phase boundaries between such phases can involve at most a rank-one jump in $\mathbf{A}$. That is if $\mathbf{A}_1$ and $\mathbf{A}_2$ are the two phases then $\mathbf{A}_1-\mathbf{A}_2=\mathbf{a}\otimes \mathbf{N}$ where $\mathbf{N}$ is the interface normal and $\mathbf{a}$ is some vector. (Otherwise the elastomer material would tear along the phase boundary.) One can check that $\mathcal{F}$ is rank-one convex[^1]. Invoking Maxwell’s famous argument, we conclude that no two phases can coexist for any value of applied field (i.e. imposed stress). Hence the neo-Hookean material has no phase transitions. Since Maxwell’s argument is based on convexity, it is perhaps troubling that the neo-Hookean free energy density is not actually convex in $\mathbf{A}$. This is easy to see if one imagines rotating a homogeneous volume by $180^\circ$ so that $\mathbf{A}\to-\mathbf{A}$. The chord joining these states passes though $\mathbf{A}=\mathbf{0}$, which is a point of infinite energy because of the isochoric constraint. To see the consequence of this non-convexity of $\mathcal{L}$, we reframe our exercise to examine an infinite elastomer-vacuum interface. We are now looking for interfacial phase transitions. First let us fix a Cartesian coordinate systems $(X^1,\,X^2)$ in the material frame and $(x^1,\, x^2)$ in the laboratory frame so that the elastomer occupies the material space $X^2<0$. We can freely assume the deformation $\mathbf{x}(\mathbf{0})=\mathbf{0}$ so that the deformed elastomer also occupies and laboratory space $x^2<0$. Because of the free boundary condition on $X^2=0$ we must have $\mathbf{A}=\mathrm{diag}(\lambda,1/\lambda)$, where $1-\lambda$ is the lateral compression of the elastomer interface. We can check the stability of the interface by computing the speeds of shear and Rayleigh waves, which must be real in a stable material. The shear wave speed is always real (because of rank-one convexity), but the Rayleigh wave speed is imaginary for compressions exceeding $1-\lambda_B\approx0.45$ [@Biot]. This surface instability is known as Biot’s instability. Note that if the free energy density $\mathcal{F}$ were convex, Biot’s instability would not exist. In many materials, the vanishing of a wave speed indicates a phase transition. For instance, the speed of sound vanishes at the liquid-vapor critical point. In our present model, the shear wave speed vanishes as $\mu\to0$, and the $\mu=0$ limit of Lagrangian density (\[eq:NHM\]) describes a perfect, incompressible fluid. Let’s, then, conduct the following thought experiment. Suppose a thermal fluctuation (or other perturbation) triggers Biot’s instability. Recent post-buckling analysis of Biot’s instability has shown that Biot’s instability is sub-critical, which suggests the hypothetical surface phase transition is first order [@Hutchinson2]. Like most first order phase transitions, we expect that the fluctuation nucleates an initially localized, growing “bubble” of some highly deformed configuration. Because the neo-Hookean material has no bulk phase transitions, we might guess that the emerging phase in this bubble involves some kind of frozen surface wave or roughness. If the amplitude of this roughness were finite, we could use the scale symmetry of $\mathcal{L}$ to rescale the amplitude to zero and recover the initial, flat interface at large scales. But as the flat interface is unstable, it must be that the alternative phase has infinite amplitude fluctuations. The signature of such a fluctuation-dominated state in a finite size system would be the non-attainment of the energy minimum. That is, a minimizing sequence of deformations starting from a super critically compressed flat interface would develop wilder and wilder surface undulations at finer and finer scales. A precise criterion for the non-attainment of an energy minimum was given by C. B. Morrey who proved that so-called quasiconvex energy densities always attain their energy minimum [@Morrey][^2]. The free energy density $\mathcal{F}$ satisfies a stronger condition called polyconvexity [@Ball][^3], so its form rules out any roughening of the interface. Because of the quasiconvex form of $\mathcal{F}$, at long times, any finite volume of elastomer surrounding the initial disturbance must settle into a new equilibrium, which in an infinite system must be scale invariant. So we now propose to identify all deformations which are invariant under scale transformation. ![\[fig:fig1\] A sulcus in a unit-diamater half-disc compressed to a width $\lambda$ (blowups on right) is mapped to a domain wall in a strip (representative configurations $A-C$) by conformal change of variable and coordinates \[double headed arrow, also Eqs. (\[eq:COV\])\]. Color (blue to red) indicates a linear blow-up of pressure as $s\to -\infty$. The size of the sulcus $L$ is dual to the position of the domain wall on the $s$-axis in the strip. The bifurcation diagram (blue line) relates $L^2\propto \lambda-\lambda^*$ as was explained in [@Hohlfeld-Mahadevan2012]; the critical compression $1-\lambda^*\approx0.35$](fig1.pdf){width="\columnwidth"} Scale invariant deformations are most easily identified by picking an origin and changing to logarithmic-polar coordinates (which is a conformal coordinate transformation from the standard Cartesian coordinates on $\mathbb{R}^2$). We define new coordinates systems $(T,\Theta)$ and $(s,\phi)$ by the relations $$\begin{aligned} X^1&=\cos(\Theta)e^T &\quad X^2&=\sin(\Theta)e^T\\ x^1&=\cos(\phi)e^s &\quad x^2&=\sin(\phi)e^s.\label{eq:COV}\end{aligned}$$ (I.e. $T=\log R,$ the radial coordinate in the reference body, and likewise for $s$.) This coordinate transformation maps an annulus in the the Cartesian coordinates to a finite strip in the logarithmic-polar coordinates as illustrated in Fig. \[fig:fig1\]. In particular, the $X^2<0$ half-plane is mapped to the infinite strip $(T,\Theta)\in [-\infty,\infty]\times[-\frac{\pi}{2},\frac{\pi}{2}]$. More importantly, this coordinate transformation maps rescaling in the Cartesian coordinates to simultaneous translation in $s$ and $T$. The coordinate transformation also maps power-law type singularities in the deformation gradient to $T$-translation invariant deformation gradients on the infinite strip. In the original Cartesian coordinates, the Euler-Lagrange equation for (\[eq:NHM\]) is $\partial S^\alpha_i/\partial X^\alpha=0,$ where the nominal stress $S^\alpha_i=\partial \mathcal{L}/\partial A^i_\alpha$. The divergence form of this equations is direct consequence of the compatibility constraint on $\mathbf{A}$. In the new coordinates, the equations of equilibrium take the form (in units where $\mu=1$) $$\begin{aligned} -\left(\frac{\partial^2 s}{\partial T^2} + \frac{\partial^2 s}{\partial \Theta^2}\right) + \left(\frac{\partial \phi}{\partial T}\right)^2 + \left(\frac{\partial \phi}{\partial \Theta}\right)^2 - \left(\frac{\partial s}{\partial T}\right)^2 - \left(\frac{\partial s}{\partial \Theta}\right)^2 - \frac{\partial p}{\partial T}\frac{\partial \phi}{\partial \Theta} + \frac{\partial p}{\partial \Theta}\frac{\partial \phi}{\partial T}&=0\\ -\left(\frac{\partial^2 \phi}{\partial T^2} + \frac{\partial^2 \phi}{\partial \Theta^2}\right) + 2 \left(\frac{\partial \phi}{\partial T}\frac{\partial s}{\partial T} + \frac{\partial \phi}{\partial \Theta}\frac{\partial s}{\partial \Theta}\right)+ \frac{\partial p}{\partial T}\frac{\partial s}{\partial \Theta} - \frac{\partial p}{\partial \Theta}\frac{\partial s}{\partial T}&=0,\label{eq:xformed-1}\end{aligned}$$ the isochoric constraint becomes $$\frac{\partial s}{\partial T}\frac{\partial \phi}{\partial \Theta} - \frac{\partial s}{\partial \Theta}\frac{\partial \phi}{\partial T} = e^{2(T-s)},\label{eq:xformed-2}$$ and boundary conditions on the surfaces $\Theta=\pm \frac{\pi}{2}$ (i.e. the erstwhile free surface $X^2=0$) are $$\frac{\partial s}{\partial \Theta} - p\frac{\partial \phi}{\partial T}=0,\quad \frac{\partial \phi}{\partial \Theta} + p\frac{\partial s}{\partial T}=0.\label{eq:BCs}$$ Scale invariant deformations, that is $T$-homogenous solutions to Eqs. (\[eq:xformed-1\]-\[eq:BCs\]), can be parametrized as $\partial s/ \partial \Theta=f(\Theta)$, $\partial \phi/\partial \Theta = g(\Theta)$, $\partial s/\partial T=1$, and $\partial \phi/\partial T =0$. Inspection of Eqs. (\[eq:xformed-1\]-\[eq:BCs\]) shows that such a form requires $p=CT + h(\Theta)$. Solutions with $C=0$ correspond to affine deformations. In the logarithmic-polar coordinates these are $s=T + \frac{1}{2}\log\left(\lambda^2\sin^2\Theta + \frac{1}{\lambda^2}\cos^2\Theta\right)$, $\lambda^2 \tan \phi = \tan \Theta$. Solutions with $C\ne 0$ correspond to creased deformations with different opening angles $\beta$ and with different amounts of shear applied to the boundaries $\Theta=\pm \pi/2$. All of these deformations develop infinite compressive stresses normal to these boundaries as $T\to -\infty$, which would violate the second of Eqs. (\[eq:BCs\]) unless $\beta=0$ so that the crease closes to self-contact. The only solution with $\beta=0$ and no surface shear is $s=T-\frac{1}{2}\log 2$, $\phi=2 \Theta$, $C=\frac{3}{2}$, and $h=const.$[@Silling]. Note that the creased deformation spontaneously breaks the $x^1$-translation symmetry of the interface. The creased ($C\ne0$) deformation and affine ($C=0$) deformations are all linearly stable, but for different reasons. The affine deformations are linearly stable (for $\lambda>\lambda_B$) because the local curvature of the free energy density function provides a restoring force for small deformations. The crease is stable because it generates its own confining pressure. Formally, this existence of this pressure can be traced the terms in Eqs. (\[eq:xformed-1\]) which are quadratic in $s$ and $\phi$ and to the incompressibility constraint. These terms arise in the change of variable to logarithmic-polar coordinates because of the tensor character of the nominal stress $S^\alpha_i=\partial\mathcal{L}/\partial A^i_\alpha$, and the divergence form of the Euler-Lagrange equation $\partial S^\alpha_i/\partial X^\alpha$. Thus the same features of the elastomer that yielded the non-convexity of $\mathcal{F}$ (incompressibility) and the stability of the homogenous phase in spite of this non-convexity (the compatibility constraint) also stabilize the inhomogeneous scale invariant deformation (i.e. the crease). The crease and the affine deformations are distinct manifolds of linearly stable, $T$-translation invariant configurations. So just as for two distinct stable phases, there must be a domain wall type solution which interpolates from the creased manifold at $T=-\infty$ to the affine manifold at $T=+\infty$. Coexistence between the two scale invariant states occurs when the domain wall can freely traverse the strip, see Fig. \[fig:fig1\]. Transforming this mobile domain wall back to Cartesian coordinates yields the sulcus deformation of the critically compressed half-space. The location of the domain wall in the strip gives the size of the sulcus in the half-space, hence the mobility of the domain wall is equivalent to the soft growth mode of a sulcus in a critically strained half-space. While the width of most domain walls is set by microscopic physics, the rank-one convexity of $\mathcal{F}$ precludes the formation of sharp phase boundary. Hence the width of our domain wall must be fixed by the width of the strip. The domain wall turns out to have an exponential profile, which translates to power law tail of the sulcus in the Cartesian coordinates and gives rise to the energy of transformation [@Hohlfeld-Mahadevan2012]. Practically, we find the domain wall solution by solving Eqs. (\[eq:xformed-1\]-\[eq:BCs\]) numerically using finite element and continuation methods as in Ref. [@Hohlfeld-Mahadevan]. We impose the creased and affine deformations as boundary conditions on opposite ends of a long strip in the $T-\Theta$ plane. The two imposed deformations can have a relative shift in the $x^2$ direction. The value of this shift, which is the depth of the crease within the sulcus, is fixed (relative to the size of the sulcus) by requiring that the integral of the traction in the $x^2$ direction vanishes on either constrained surface. We find coexistence between the crease and an affine deformation at the critical compression of $1-\lambda^*\approx 0.35$, which is similar to what has been reported elsewhere [@Hohlfeld-Mahadevan; @Suo1]. The principle implied our discussion of the incompressible neo-Hookean model is that while ordinary phase stability follows from rank-one convexity of the free energy, this property is not strong enough to preclude all processes which resemble phase transitions. In general, scale free material models can have multiple scale invariant potential wells, and some of these may correspond to *inhomogeneous* states. However, a local analysis of the free energy density will only detect the homogenous states, i.e. phases. Rather, one should look for all scale invariant configurations, whether or not these are homogenous. Whenever a model has multiple scale invariant potential wells, whether or not they correspond to distinct phases, there are associated coexistence conditions between every pair. When coexistence occurs, the mode in which one state transforms into the other (i.e. the growth mode) is an emergent soft mode of the whole system which can dramatically change its response to external stimuli. It is interesting to note that our construction doesn’t directly use the free energy density $\mathcal{F}$, so it makes sense even for systems which are out of equilibrium. We also remark that integrability of the strain energy density close to the origin $\mathbf{X}=\mathbf{0}$ implies the coordinate-transformed deformation gradient must in any case be bounded as $T\to -\infty$. So a potential alternative to a $T$-translation homogenous state, such as a crease, is a $T$-periodic state. Such configurations might be relevant in the finely twinned microstructure that sometimes forms during a martensitic phase transition. Our method of constructing a sulcus potentially has an immediate application in liquid crystal elastomers. A recent experiment suggests that sulcification or creasing in these materials may induce topological changes in the nematic director field [@Islam]. The existence of a director field in these materials potentially enriches the space of inhomogeneous, scale invariant deformations beyond the crease studied here. [10]{} and [L. Mahadevan]{}, Phys. Rev. Lett. [**106**]{}, 105702 (2011). E. Hohlfeld and L. Mahadevan, Phys. Rev. Lett. [**109**]{}, 025701 (2012). D. Chen, S. Cai, Z. Suo, and R. C. Hayward, Phys. Rev. Lett. [**109**]{}, 038001 (2012). . (John Wiley and Sons, New York, 1965). , Phys. Rev. A [**39**]{}, 5932–5948 (1989). in [*Molecular Conformation and Dynamics of Macromolecules in Condensed Systems*]{}, edited by M. Nagasawa (Elsevier, New York, 1988). , [J. Kim]{}, and [R. C. Hayward]{}, Soft Matter [**4**]{}, 564–569 (2008). and [J. Yoon]{} and [R. C. Hayward]{}, Nature Materials [**9**]{},159-164 (2010). N. Zalachas, S. Cai, Z. Suo, Y. Lapusta, Int. J. Solids and Struc. [**50**]{}, 920–927 (2013). P. V. Baly, R. J. Okamoto, G. Xu, Y. Shi, and L. A. Taber, Phys. Biol. [**10**]{}, 016005 (2013). , [S. Cai]{} and [Z. Suo]{}, Euro. Phys. Lett. [**95**]{}, 64002 (2011). , [Y. Couder]{} and [M. A. Guedeau-Boudeville]{} and [M. Ben Amar]{}, Phys. Rev. Lett. [**107**]{}, 018103 (2011). Q. Wang, L. Zhang, and X. Zhao, Phys. Rev. Lett. [**106**]{}, 118301 (2011). S. Mora, M. Abkarian, H. Tabuteau, and Y. Pomeau, Soft Matter [**7**]{}, 10612-10619 (2011). W. Hong, X. Zhao, Z. Suo, Appl. Phys. Lett. [**95**]{}, 111901 (2009). and [I. S. Cho]{}, Surface instabilitiesin compressed or bent rubber blocks. Rubber Chem. and Tech. [**72**]{}, 253–262 (1999). and [A. L. Das]{}, Phys. Rev. Lett. [**99**]{}, 076101 (2007). , [F. Corson]{}, [A. Boudaoud]{}, and [B. Roman]{}, Phys. Rev. Lett., [**103**]{}, 045501 (2009) and [J. W. Hutchinson]{}, Proc. R. Soc. A [**468**]{}, 94-115 (2011). T. Tallinen, J. S. Biggins, and L. Mahadevan, Phys. Rev. Lett. [**110**]{}, 024302 (2013). N. K. Sundaram, Y. Guo, and S. Chandrasekar, Phys. Rev. Lett [**109**]{}, 106001 (2012). and [J. E. Marsden]{}, Arch. Rat. Mech. Anal. [**86**]{}, 251–277 (1984). Y. Grabovsky and T. Mengsha, Trans. Amer. Math. Soc. [**361**]{}, 1495–1541 (2009). J. Hutchinson, Phil. Trans. R. Soc. A, (2013) C. B. Morrey, Jr., Pacific J. Math. [**2**]{}, 25-53 (1952). E. Acerbi and N. Fusco, Arch. Rat. Mech. Anal. [**21**]{}, 152–145 (1984). J. M. Ball, Arch. Rat. Mech. Anal. [**63**]{}, 337–403 (1977). , J. App. Mech. [**58**]{}, 70–74 (1991). M. F. Islam, M. Nobili, F. Ye, T. C. Lubensky, and A. G. Yodh, Phys. Rev. Lett [**95**]{}, 148301 (2005). [^1]: A function $F$ is rank-one convex if $F(\mathbf{A} + (1-t) \mathbf{a}\otimes \mathbf{b})\le tF(\mathbf{A}) + (1-t)F(\mathbf{A}+\mathbf{a}\otimes \mathbf{b})$ for every number $t\in [0,1]$, matrix $\mathbf{A}$, and pair of vectors $\mathbf{a}$, $\mathbf{b}$. In our case this condition is equivalent to the Lagrange-Hadamard inequality, or a finite shear wave speed at all states of deformation. [^2]: A function $F$ is quasiconvex if $\int_Q F(\mathbf{A}+\mathbf{B}(\mathbf{X}))dV\ge F(\mathbf{A})\mathrm{vol}(Q)$ where $Q$ is the unit cube in $\mathbb{R}^d$, for every matrix $\mathbf{A}$ and *function* $\mathbf{B}$ which is the gradient of a $Q$-periodic function. Quasiconvexity can be compared to Jensen’s inequality, which automatically holds for convex functions. [^3]: A function $F$ is polyconvex if it is a convex function of null Lagrangians. For example if $F(\mathbf{A},b)$ is a jointly convex function of the matrix $\mathbf{A}$ and number $b$, then $F(\mathbf{A}, \det\mathbf{A})$ is polyconvex.

ArXiv

--- abstract: 'We derive a rigorous scaling law for minimizers in a natural version of the regularized Cross-Newell model for pattern formation far from threshold. These energy-minimizing solutions support defects having the same character as what is seen in experimental studies of the corresponding physical systems and in numerical simulations of the microscopic equations that describe these systems.' address: - 'Dept. of Math., Univ. of Arizona, 617 N. Santa Rita Ave., Tucson, AZ 85721, USA' - 'Dept. of Math., Univ. of Arizona, 617 N. Santa Rita Ave., Tucson, AZ 85721, USA' author: - 'N. M. Ercolani' - 'S.C. Venkataramani' title: A variational theory for point defects in patterns --- [^1] Introduction ============ This paper reports on some recent progress that has been made in the analytical modeling of defect formation, far from threshold, in pattern forming physical systems. We will take a moment here to very briefly sketch the physical and mathematical background that motivates what is done in this paper. The relevant class of pattern-forming physical systems to consider are those in which the spatial physical field can be described as planar and the first bifurcation from a homogeneous state, having arbitrary translational symmetry in the plane, produces a striped pattern which has only a discrete periodic symmetry in one direction. This *symmetry-breaking* occurs at a critical threshold; above this threshold the pattern can deform and, further away, *defects* can form. It is the desire to understand and model this process of defect formation that motivates our study. A good particular example of these kinds of physical systems is a high Prandtl number Rayleigh-Bénard convection experiment. The critical threshold in this example is the critical Rayleigh number at which fluid convection is initiated from the sub-threshold homogeneous conducting state. The “striped pattern” here can be taken to be the horizontal cross-section of the temperature field at the vertical midpoint of the experimental cell in which *convection rolls* have formed. Because of its periodic structure, the striped pattern can be described in terms of a periodic form function of a *phase*, $\theta = \vec{k}\cdot\vec{x}$, where the magnitude of $\vec{k}$ is the wavenumber of the pattern and the orientation of $\vec{k}$ is perpendicular to the stripes. Here $\vec{x}=(x,y)$ is a physical point in the plane. Even though the striped pattern will deform far from threshold, over most of the field (and in particular away from defects) it can be locally approximated as a function of a well-defined phase, $\theta(\vec{x})$, for which a local wavevector can be defined as $\vec{k} = \nabla\theta$ which differs little from a constant vector unless one varies over distances on the order of many stripes in the pattern. This slowly-varying feature of pattern formation far from threshold motivates the introduction of a *modulational* ansatz in the microscopic equations describing these physical systems from which an order parameter equation for the behavior of the phase can be formally derived. This was originally done by Cross and Newell [@CN]. These equations are variational and from our perspective it is advantageous to study their solutions by studying the behavior of the minimizers of the variational problem. The version of the variational problem that we study corresponds to the following energy functional on a given domain $\Omega$ with specified Dirichlet boundary values. $$\label{eq:gl} {\mathcal{E}}^\mu(\Theta) = \mu \int_\Omega \left(\Delta_{\vec{X}} \Theta \right)^2 d\vec{X} + \frac{1}{\mu} \int_\Omega (1 - |\nabla_{\vec{X}} \Theta|^2)^2\, d\vec{X}\ ,$$ which is expressed in terms of *slow* variables stemming from the modulational ansatz mentioned above: $\vec{X} = (X,Y) = \left(\mu x, \mu y\right); \Theta = \frac{\theta}{\mu}$. We refer to this functional as the *regularized Cross-Newell* (RCN) *Energy*. It consists of two parts: a non-convex functional of the gradient (the CN part) plus a quadratic functional of the Hessian matrix of $\Theta$, which is the regularizing singular perturbation. Without this regularization, the CN variational equations admit non-physical caustic formation. Instead, by studying the limit of minimizers of ${\mathcal{E}}^\mu$ as $\mu \to 0$, one may be able to identify the formation of a physical defect as a limiting jump discontinuity or other kind of singularity in the wavevector field associated to the $\mu$-indexed family of minimizing phase fields. For more details on what has been rather tersely outlined above, we refer the reader to [@EINP] where analytical results on the asymptotic limit of minimizers for RCN and their defects in certain geometries are also derived. See also [@ET] where further refinements and generalizations are developed. We further mention that the variational problem associated to (\[eq:gl\]) also arises in other physical contexts (unrelated to pattern formation) where it is known as the *Aviles-Giga energy* [@AG]. We now turn to the focus of this paper. The kind of defects that are seen to arise far from threshold are not supported by asymptotic minimizers of (\[eq:gl\]) if the class of functions over which one is varying is restricted to be single-valued phases. In particular, one can see for purely topological reasons that this restriction rules out *disclinations* [@EINP]. In [@EINP2], physical, numerical and experimental arguments are developed which make a strong case in support of the hypothesis that the correct order parameter model for the phase in pattern forming systems far from threshold should come from a variational problem admitting test functions which are *multi-valued* and in particular *two-valued*. In physical parlance this is often expressed by saying that the wavefield $\vec{k}$ should be allowed to be a *director field*; i.e. an unoriented vector field. One figure (see Fig. \[fig:sh-zip\] below) from [@EINP2] will help to crystallize the issue and the focus of this paper. This figure shows seven numerical simulations, each done in a horizontal strip, of a solution to the *Swift-Hohenberg equation* which is a generic model of microscopic equations for a pattern forming system. Each of these is run far from theshold but with differing boundary conditions imposed at the edges. In each case the boundary conditions impose a constant orientation of the stripe at the edges such that the normal to the stripe is $(\cos(\alpha), \sin(\alpha))$ along the top edge and $(\cos(\alpha), - \sin(\alpha))$ along the bottom edge. The only thing that changes from one simulation to the next is the value of $\alpha$ which in the figure is recorded on the left in each respective cell. The results of [@EINP] together with symmetry considerations establish that for an analogous domain and boundary values, the asymptotic minimizers of (\[eq:gl\]), within the class of single-valued phases, should have the form shown in the bottom-most cell of Figure (\[fig:sh-zip\]). That is, they should have wavevectors very close to $(\cos(\alpha), \sin(\alpha))$ in the upper region of the cell and very close to $(\cos(\alpha), -\sin(\alpha))$ in the lower region of the cell with a boundary layer around the mid-line in which the wavevector transitions smoothly but rapidly from one state to the other. These minimizers are dubbed *knee solutions* in [@EINP] and in the limit as $\mu \to 0$, they tend to a configuration in which there is a sharp jump in the wavevector along the mid-line. This kind of defect is called a *grain boundary*. In other words, the theory for (\[eq:gl\]) with single-valued phases predicts that the grain boundary should be the limiting defect independent of the value of $\alpha$. The different result appearing in Figure (\[fig:sh-zip\]) was one of the pieces of evidence sited in [@EINP2] to argue the necessity for the larger variational class of multi-valued phases, even in such simple geometries as those of Figure (\[fig:sh-zip\]). In this paper we are going to carry out a careful analytical study of the RCN variational problem in exactly this geometry but within a larger class of two-valued phases. We will firmly establish that the form of the asymptotic minimizers in this more general model does in fact depend non-trivially on $\alpha$. In addition, the construction of test functions in section \[sec:u\_bound\] and the numerical simulations in section \[sec:results\] gives some intuitive and experimental support to the belief that the stable solutions of the RCN equations qualitatively resemble what is seen in the Swift-Hohenberg simulations. In [@EINP2], the term Swift-Hohenberg “zippers” was coined to refer to the problem studied in Figure (\[fig:sh-zip\]). In this paper we will be studying *Cross-Newell zippers*. Setup {#sec:prelim} ===== We are given an angle $\alpha$ that determines the boundary conditions on the pattern as $y \rightarrow \pm \infty$ by $$\nabla \theta \rightarrow (\cos(\alpha), \pm \sin(\alpha)) \text{ as } y \rightarrow \pm \infty.$$ Note that this differs from the setup underlying the Swift-Hohenberg zippers in that the boundary conditions are placed at $\pm \infty$ in the $y$-direction rather than at finite values of $y$. This simplifies our technical considerations in that we don’t need to worry about adjusting the location of these boundaries as $\alpha$ changes. Also, all of the patterns we want to consider here are *shift-periodic* in the $x$-direction. This allows us to reduce our study to domains that are periodic in $x$. We introduce the (small) parameter $\epsilon = \cos(\alpha)$ and we define the period $l = \pi/\epsilon$. We consider the following variational problem on the strip ${\mathcal}{S}^\epsilon \equiv \{(x,y) | 0 \leq x < l, y\geq 0\}$: Minimize ${\mathcal}{F}^\epsilon[\theta;a,\delta]$ given by $${\mathcal}{F}^\epsilon[\theta;a,\delta] = \iint_{{\mathcal}{S}^\epsilon} \left\{[ \Delta \theta]^2 + (1 - |\nabla \theta|^2)^2 \right\} dx dy$$ over all $a \in [0,1], \delta \in \mathbb{R}$ and $\theta$ satisfying the boundary conditions $$\begin{gathered} \theta(x,0) = 0 \quad \text{ for } 0 \leq x < a l; \label{eq:bc} \\ \theta_y(x,0) = 0 \quad \text{ for } a l \leq x < l; \nonumber \\ \theta(x,y) - \epsilon x \text{ is periodic in $x$ with period $l$ for each } y \geq 0; \nonumber \\ \theta(x,y) - \left[\epsilon x + \sqrt{1 - \epsilon^2} y + \delta\right] \in H^2({\mathcal}{S}^\epsilon). \nonumber\end{gathered}$$ We take a moment here to explain the considerations that have motivated the mixed Dirichlet-Neumann boundary conditions here, the first two boundary conditions in (\[eq:bc\]) above. We argued in the introduction that in order to capture the physically relevant minimizers, the RCN variational problem needed to allow for multi-valued phases in its admissible class of test functions. However, the numerical results on the Swift-Hohenberg zippers suggest that in certain symmetrical geometries the appropriate multi-valuedness can be introduced in a tractable fashion. Indeed in the case of the SH zippers we see that the symmetry of the boundary conditions between the upper and lower edges of the domain is preserved in the symmetry of all of the exhibited solutions about the middle horizontal axis; i.e., the reflection in $y$ about the $y=0$ axis. This suggests that a single-valued phase could describe the solution in the upper half-plane with the solution in the lower half-plane given as a symmetric reflection of that in the upper half-plane about $y=0$. ![An illustration of the appropriate boundary conditions.[]{data-label="fig:bcs"}](figs/boundaryconds.eps) Figure \[fig:bcs\] illustrates two instances of the form that we expect these zippers to take in the infinite (in $y$) geometry. The figure on the left illustrates level curves (*stripes* in the parlance of the introduction) of what we will shortly define to be a *self-dual knee solution*. This is indeed symmetric about the mid-axis, which we will take to be the $y=0$ axis; moreover, one can see that its gradient field along $y=0$ is tangential to this axis. Thus the gradient field in the upper half-plane is completely symmetrical to that in the lower half-plane under reflection about $y=0$. However, for the striped pattern on the right in figure \[fig:bcs\], this is not the case. There are regions, illustrated for example by the darkened interval along $y=0$, where the gradient field is tangential to this axis; but, there are other regions, illustrated for example by the lightened interval along $y=0$, where the gradient field needs to be perpendicular to this axis. By reflection symmetry this field will point upwards in the upper half-plane and downward in the lower half-plane. This cannot be supported by a vector field but it is allowable for a director field. This indicates that in this region a two-valued phase is required. To get at the conditions on the phase itself we observe that patterns of the type illustrated here are analytically given in terms of a form function $F$ of the phase $\theta = \theta(x,y)$ such that $F$ is locally periodic of period $2\pi$ in $\theta$ and such that $F(\theta(x,y))$ is even in $y$ and smooth in $(x,y)$. In order to allow $\theta$ to be two-valued we also require $F$ to be even in $\theta$. (An example of a global form function having these properties is $F = \cos$.) It follows from these requirements that either $\theta(x,y)$ is even in $y$, in which case $\theta_y(x,0) = 0$, a Neumann boundary condition; or, $\theta(x,y)$ is an odd function of $y$ modulo $\pi$, in which case $\theta_y(x,0) = n\pi$ for some integer $\pi$, a Dirichlet boundary condition. Thus to realize the pattern on the right in figure \[fig:bcs\] in terms of a single-valued phase in the upper half-plane, we would need to take the Neumann boundary condition on the darkened interval and the Dirichlet boundary condition on the lightened interval. This is what we have done in (\[eq:bc\]). For the self-dual knee pattern on the left we would take the entire boundary condition to be Neumann. The functional ${\mathcal}{F}^\epsilon$ is the RCN energy functional but with the scaling $\mu$ removed. It is appropriate to do this because the demonstration that the nature of the RCN minimizers depends on $\alpha$ is independent of this scaling. The first and the second conditions impose a mixed Dirichlet/Neumann boundary condition at $y = 0$, the third condition imposes (shifted-) periodicity in $x$ and the last condition ensures that the test functions $\theta$ approach the straight parallel roll patterns $\epsilon x + \sqrt{1 - \epsilon^2} y + \delta$ as $y \rightarrow \infty$. Note that the dependence of the functional ${\mathcal}{F}^\epsilon$ on the parameter $\epsilon$ is through the dependence of the domain ${\mathcal}{S}^\epsilon$ and the boundary conditions on $\epsilon$. The parameters $a$ and $\delta$ are determined by minimization. The parameter $a$ is a measure of the fraction of the boundary at $y = 0$ that has a Dirichlet boundary condition, and $\delta$ represents the [*asymptotic phase shift*]{}, that is the difference in phases between the test function $\theta$ and the roll pattern $\hat{\theta}(x,y) = \epsilon x + \sqrt{1 - \epsilon^2} y$ which satisfies $\hat{\theta}(0,0) = \theta(0,0) = 0$. The case where $a$ is set to zero is considered in earlier references [@EINP]. The test functions $\theta(x,y)$ satisfy a pure Neumann boundary condition at $y = 0$ and the minimizers in this case are the self-dual knee solutions $$\theta_{neu}(x,y) = \epsilon x + \log(\cosh(\sqrt{1-\epsilon^2} y)).$$ These solutions have an asymptotic phase shift of $-\log(2)$ and the energy of the minimizers in the strip ${\mathcal}{S}^\epsilon$ is given by $${\mathcal}{F}^\epsilon[\theta_{neu};0,-\log(2)] = \frac{ 4 \pi \sqrt{1-\epsilon^2}}{3 \epsilon}. \label{eq:chevrons}$$ The existence of $(\theta^\epsilon,a^\epsilon,\delta^\epsilon)$ minimizing ${\mathcal}{F}^\epsilon$ can be shown from the direct method in the calculus of variations. We also prove the following results about the minimizers, and their energy – [*Upper bound*]{} There is a constant $E_0$ such that ${\mathcal}{F}^\epsilon[\theta^\epsilon;a^\epsilon,\delta^\epsilon] \leq E_0$ for all $\epsilon \in (0,1]$. \[thm:u\_bound\] We prove this result in sec. \[sec:u\_bound\] by exhibiting an explicit test function satisfying this bound. Note the implication that the minimizers for sufficiently small $\epsilon$ cannot be the self-dual solutions, since the energy in Eq. (\[eq:chevrons\]) diverges as $\epsilon \rightarrow 0$. Consequently, $a^{\epsilon} > 0$ for sufficiently small $\epsilon$. [*Lower bound*]{} There are constants $E_1 > 0$ and $\epsilon_0 > 0$ such that, even for the optimal test function $\theta^\epsilon$ and the optimal parameter values $a^\epsilon$ and $\delta^\epsilon$, we have ${\mathcal}{F}^\epsilon[\theta^\epsilon;a^\epsilon, \delta^\epsilon] \geq E_1$ for all $\epsilon \leq \epsilon_0$. Further, there are constants $0 < \alpha_1 < \alpha_2$ such that $1 - \alpha_2 \epsilon < a^\epsilon < 1 - \alpha_1 \epsilon$ for sufficiently small $\epsilon$. \[thm:l\_bound\] We prove this result in sec. \[sec:l\_bound\]. Combining this result with the preceding theorem, we obtain a rigorous scaling law for the energy of the minimizer, and for the quantity $(1-a)$ as $\epsilon \rightarrow 0$. As a corollary to Theorem \[thm:l\_bound\], we find that an $O(1)$ part of the energy of the minimizer concentrates on the set, $a l \leq x \leq l, 0 \leq y \leq 1$. This can be interpreted as saying that a nontrivial part of the energy of the minimizer lives in the region of the convex-concave disclination pair [@EINP2]. Upper bound {#sec:u_bound} =========== We will first show an upper bound for the enrgy functional ${\mathcal}{F}^{\epsilon}$, uniform in $\epsilon$, by constructing a family of explicit test function whose energy is uniformly bounded. The idea for the construction of these test functions comes from the self-dual ansatz [@EINP] which requires that the energy density of the functional $\mathcal{F}$ should be *equi-partitioned* between its two terms. Functions satisfying this ansatz solve the self-dual (resp., anti-self-dual) equation: $$\label{eq:selfdual} \Delta\theta =\pm (1 - |\nabla\theta|^2).$$ Solutions of this equation can be constructed via the logarithmic transform $$\theta = \pm \log u$$ which reduces (\[eq:selfdual\]) to the linear Helmholtz equation (\[selfdual\]). We refer the reader to [@EINP; @ET] for more background on self-dual reduction. Self-dual test functions for the CN-Zipper problem {#zipper} -------------------------------------------------- ### Existence We consider the Helmholtz equation in the upper half-plane, $$\label{selfdual} \Delta u - u = 0$$ subject to the mixed boundary conditions $$\begin{aligned} u(x,0) &=& e^{-n\pi} \,\,\, n\ell < x < (n+a)\ell \label{2} \\ u_y(x,0) &=& 0 \,\,\,\,\,\, (n+a)\ell \leq x \leq (n+1)\ell \label{3}\end{aligned}$$ and with asymptotic behavior for large $y$ given by const. $\exp(-\epsilon x - \sqrt{1-\epsilon^2} y)$ where $\ell = \pi/\epsilon$ and $a \in (0,1)$. We seek a shift-periodic solution, meaning that we change variables to $w = e^{\epsilon x} u(x,y)$ and look for periodic solutions of $$\label{shiftper} Lw = \Delta w -2\epsilon\partial_x w -(1-\epsilon^2)w=0,$$ with boundary conditions of periodicity in $x$ of period $\ell$; mixed boundary conditions at $y=0$, $$\begin{aligned} w(x,0) &=& e^{\epsilon x} \,\,\,\,\,\, 0 < x < a\ell \label{5}\\ w_y(x,0) &=& 0 \,\,\,\,\,\, a\ell \leq x \leq \ell \label{6};\end{aligned}$$ and with asymptotic behavior for large $y$ given by const. $\exp(- \sqrt{1-\epsilon^2} y)$. Given such a $u$, $\theta = -\log u$ would satisfy the boundary conditions (\[eq:bc\]). (However, for notational simplicity, in the remainder of this section we will set $\theta = \log u$.) We now let ${\mathcal}{S}^\epsilon$ denote the half-cylindrical domain, $\ell$-periodic in $x$ and with $y>0$. The existence of a weak solution to (\[shiftper\]) satisfying the above boundary conditions can be established via the Lax-Milgram theorem with appropriate energy estimates. However, in order to derive uniform asymptotic energy estimates (as $\epsilon \to 0$) for the CN Zipper problem we need to go beyond existence results and try to construct a more explicit representation of the solution to (\[selfdual\]). Unfortunately, at present, the solutions one can construct using Greens function methods and the like do not yield sufficient a priori boundary regularity near $y=0$ to control the asymptotic behavior of the energy in this *finite* part of ${\mathcal}{S}^\epsilon$. We will therefore instead study solutions of a self-dual problem with modified boundary conditions (more precisely, with pure Dirichlet boundary conditions). Subsequently we will make a local modification of these solutions near the boundary to produce functions (no longer global self-dual solutions) whose asymptotic energy we can control *and* which are valid test functions for the Cross-Newell Zipper problem. The modified boundary value problem we consider is (\[shiftper\]) with (\[5\]-\[6\]) replaced by the pure Dirichlet boundary condition $$w(x,0) = \left\{\begin{array}{c} e^{\epsilon x} \,\,\,\,\,\, 0 < x < a\ell \\ q_a(x) \,\,\,\,\,\, a\ell \leq x \leq \ell \\ \end{array}\right. \leqno{\begin{array}{c} (\ref{5}^\prime)\\ (\ref{6}^\prime)\\ \end{array}}$$ where $q_a(x)$ is a function which smoothly interpolates, up through second derivatives, between $e^{\epsilon x}$ at $x=a\ell$ on the left and $e^{\epsilon x - \pi}$ at $x=\ell$ on the right. There are clearly many choices for such a function; the precise choice for our purposes will be made later at the end of subsection \[energy\]. By elliptic regularity [@Evans], the solution to this boundary value problem satisfies $w(x,y) \in H^2\left({\mathcal}{S}^\epsilon\right)$. In the following sections we will construct the solutions to this problem and study its asymptotics relative to the RCN energy $\mathcal{F}^\epsilon$. ### Explicit Construction The whole plane Green’s function for the Helmholtz equation (\[selfdual\]) is explicitly given in terms of the Bessel potential [@Evans]: $$\label{Bessel} {G}(x,y;\xi,\eta)=\frac{1}{4\pi} \int_0^\infty e^{-t} \frac{dt}{t} \exp\left( -\frac{1}{4t}\{(x-\xi)^2 + (y-\eta)^2\}\right).$$ In terms of this Green’s function we can then represent a solution to (\[selfdual\]), with asymptotic behavior for large $y$ given by const. $\exp(-\epsilon x - \sqrt{1-\epsilon^2} |y|)$, as $$\begin{aligned} \label{soln} u^\epsilon(x,y) &=& \int_{-\infty}^\infty \rho^\epsilon(\xi){G}(x,y;\xi,0) d\xi.\end{aligned}$$ Note that $$\begin{aligned} \label{dirbase} u_y^\epsilon(x,y) &=& \int_{-\infty}^\infty \rho^\epsilon(\xi){G}_y(x,y;\xi,0) d\xi\\ \nonumber &=& -\int_{-\infty}^\infty \rho^\epsilon(\xi){G}_\eta(x,y;\xi,0) d\xi\end{aligned}$$ solves (\[selfdual\]) with respect to the standard Dirichlet boundary condition which equals minus the jump of $u_y^\epsilon$ along the $x$-axis. One may check directly (see (\[FT\])) that in fact $\rho^\epsilon(\xi) = -2 u_y^\epsilon(\xi,0)$ almost everywhere. Integrating (\[dirbase\]) with respect to $y$ gives $$\begin{aligned} \label{intdirbase} u^\epsilon(x,y) + f(x) &=& \int_{-\infty}^\infty \rho^\epsilon(\xi){G}(x,y;\xi,0) d\xi.\end{aligned}$$ Since both $u^\epsilon(x,y)$ and the RHS of (\[intdirbase\]) decay as $y\uparrow \infty$, it follows that $f(x)\equiv 0$. This is consistent with the ansatz (\[soln\]), taking $\rho^\epsilon(\xi)$ to be the jump in the normal derivative of $u^\epsilon$ along $y=0$. We make the following shift-periodic ansatz for $\rho^\epsilon$, $$\rho^\epsilon(\xi+\ell)e^{\epsilon(\xi+\ell)} = \rho^\epsilon(\xi)e^{\epsilon\xi}.$$ With this one can expand out (\[soln\]) more explicitly as $u^\epsilon(x,y)=$ $$\begin{aligned} &=& \frac{1}{4\pi}\sum_{n \in \textbf{Z}} \int_0^\infty\frac{dt}{t} e^{-(t+\frac{y^2}{4t})} \int_{n\ell}^{(n+1)\ell} \rho^\epsilon(\xi) \exp\left(\frac{(x-\xi)^2}{-4t}\right) d\xi \label{11}\\ &=& \frac{1}{4\pi}\sum_{n \in \textbf{Z}} \int_0^\infty\frac{dt}{t} e^{-(t+\frac{y^2}{4t})} \int_{0}^{\ell} e^{-n\pi}\rho^\epsilon(\xi) \exp\left(\frac{(x-(\xi + n\ell))^2}{-4t}\right) d\xi \label{12}\\ &=& \frac{1}{4\pi} \int_0^\infty\frac{dt}{t} e^{-(t+\frac{y^2}{4t})} \int_{0}^{\ell}\rho^\epsilon(\xi) \sum_{n \in \textbf{Z}}e^{-n\pi} \exp\left(\frac{(x-(\xi + n\ell))^2}{-4t}\right) d\xi \label{13}\\ &=& \frac{e^{-\epsilon x}}{4\pi} \int_0^\infty\frac{dt}{t} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \int_{0}^{\ell}d\xi\rho^\epsilon(\xi)e^{\epsilon\xi}\sum_{n \in \textbf{Z}} \exp\left(\frac{((x-2\epsilon t)-(\xi + n\ell))^2}{-4t}\right) \label{14}\\ &=& \frac{e^{-\epsilon x}}{\sqrt{4\pi}} \int_0^\infty\frac{dt}{t^\frac{1}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \frac{1}{\ell}\int_{0}^{\ell} \rho^\epsilon(\xi) e^{\epsilon \xi} \,\,\vartheta_3\left(\frac{-(x - \xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right) d\xi \label{15}\end{aligned}$$ In (\[11\]) we have interchanged the order of integration which is justified by Tonelli’s Theorem; in (\[12\]) we’ve made the substitution $\xi = \xi_n + n\ell$ and in (\[13\]) we’ve commuted the sum past the integrals which is justified by monotone convergence–all terms in the series are positive and hence the partial sums are monotonic. In (\[14\]) we write each summand as a single exponential and then appropriately complete the square in each exponent. Finally in (\[15\]) we apply Jacobi’s identity [@WW]. Here $\vartheta_3$ is one of the Jacobi theta functions, in this setting explicitly given as $$\begin{aligned} \label{Jacobi} \vartheta_3\left(\frac{- x + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right) &=& 1 + 2\sum_{n=1}^\infty e^{-\left(\frac{2\pi}{\ell}\right)^2 n^2 t} \cos\left(\frac{2\pi n}{\ell}\left(x-\frac{2\pi t}{\ell}\right) \right)\end{aligned}$$ Finally, from (\[15\]) we can express our candidate for the solution to (\[shiftper\]), ($\ref{5}^\prime - \ref{6}^\prime$) as $$\begin{aligned} {w^\epsilon}(x,y) = \frac{1}{\sqrt{4\pi}} \int_0^\infty\frac{dt}{t^\frac{1}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \frac{1}{\ell}\int_{0}^{\ell} p^\epsilon(\xi) \,\,\vartheta_3\left(\frac{-(x - \xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right) d\xi,\label{shiftpersoln}\end{aligned}$$ where $p^\epsilon(\xi) = \rho^\epsilon(\xi) e^{\epsilon \xi}$. ### Data Characterization, periodized and in Fourier Space From the previous sections we have that $p^\epsilon(\xi)$ is periodic of period $\ell$; also ${w^\epsilon}(x,y)$ is periodic in $x$ of period $\ell$ and $=e^{\epsilon x}$ along $(0, a\ell)$ when $y=0$. Moreover, taking the Fourier transform of (\[shiftpersoln\]) one finds that the Fourier coefficients, in $x$, must satisfy $$\begin{aligned} \label{FT} \{\widehat{w^\epsilon(x,y)}\}(n,y) &=& \frac{1}{2}\frac{1}{\sqrt{1+\epsilon^2(2n+i)^2}}{\{\widehat{p^\epsilon(\xi)}\}}(n)e^{-\sqrt{1+\epsilon^2(2n+i)^2}y}\end{aligned}$$ for each value of $y$. Taking the limit as $y\to 0$ on both sides of (\[FT\]) gives $$\begin{aligned} \label{FTsoln} \{\widehat{w^\epsilon(x,0)}\}(n) &=& \frac{1}{2}\frac{1}{\sqrt{1+\epsilon^2(2n+i)^2}}{\{\widehat{p^\epsilon(\xi)}\}}(n).\end{aligned}$$ This is a determining conditions for $p^\epsilon(\xi)$. We note that differentiating (\[FT\]) with respect to $y$ and setting $y = 0$ demonstrates that $p^\epsilon(x) = - 2 w^\epsilon_y(x,0)$, at least in the $L^2$ sense. Since $w^\epsilon(x,0) \in H^{2}(S^1)$, it follows, by comparison, that $2\sqrt{1+ \epsilon^2(2n+i)^2}\widehat{w^\epsilon}(n)\in h^1(\mathbb{Z})$. Given this we can now define $$\label{Besspot} p^\epsilon(x) \doteq \left\{2\sqrt{1+ \epsilon^2(2n+i)^2}\widehat{w^\epsilon}(n)\right\}^\vee(x)$$ which characterizes $p^\epsilon$ as an element of $H^1(S^1)$. It follows from Sobolev’s lemma [@Evans] that $p^\epsilon$ can be taken to be continuous. This last observation also justifies the existence of the Fourier coefficients $\{\widehat{p^\epsilon}\}(n)$ that were formally introduced in (\[FT\]). ### Large $y$ asymptotics {#largey} We now determine the large $y$ asymptotics of (\[soln\]). By (\[FT\]), $w^\epsilon$ has a Fourier representation given by $$\begin{aligned} {w^\epsilon}(x,y) &=& \sum_{n \in \mathbb{Z}} \widehat{w^\epsilon}(n) e^{-\sqrt{1+\epsilon^2(2n+i)^2}y} e^{\frac{2\pi i nx}{\ell}}\\ &=& \widehat{w^\epsilon}(0) e^{-\sqrt{1-\epsilon^2}y} + \mathcal{O}\left(e^{-2 \sqrt{1 + 3\epsilon^2} y}\right).\end{aligned}$$ (We note that for large $y$ this series converges uniformly to a smooth, in fact real-analytic, function of $x$.) Moreover, $w^\epsilon(0) = \frac{1}{\ell}\int_{0}^{\ell}w^\epsilon(x,0) dx$ is non-zero since by the maximum principle [@Evans] applied to the elliptic PDE (\[shiftper\]) on the cylinder $[0,\ell] \times \left(-\infty, \infty\right)$, the integrand, $w^\epsilon(x,0)$, is non-negative and in fact, by ($\ref{5}^\prime-\ref{6}^\prime$), non-vanishing on $[0,\ell]$ (the definition of $q_a$ which we give later will insure that this is so). Energy Estimates {#energy} ---------------- We will now try to show that the regularized Cross-Newell energy of $\theta(x,y) = \log u(x,y)$ is uniformly bounded in $\epsilon$. This would establish a uniform (in $\epsilon$) upper bound for the energy minimizers. Recall that the energy is calculated by integrating the energy density over the domain ${\mathcal}{S}^\epsilon$. Making this estimate breaks naturally into the consideration of two regions: $[0, \ell] \times \{y\geq M_\epsilon\}$ and $[0, \ell] \times \{y < M_\epsilon\}$ where $M_\epsilon$ is to be determined. We remark that the so-called “knee solution” of the self-dual equation provides an upper bound for the energy for values of $\epsilon$ bounded away from zero. So we only need to be concerned with small values of $\epsilon$. Since $u^\epsilon(x,y) = e^{-\epsilon x}w^\epsilon(x,y)$ solves the Helmholtz equation, it will suffice to bound the density $(1-|\nabla\theta^\epsilon|^2)^2$ (since the integral of this density equals that of $(\Delta \theta^\epsilon)^2$ for self-dual solutions). ### Estimates in $[0, \ell] \times \{y\geq M_\epsilon\}$ We begin by considering the domain for large $y$. Since $$\begin{aligned} &&\nabla\theta^\epsilon(x,y) = \frac{\nabla u^\epsilon}{u^\epsilon}(x,y) = \left( \begin{array}{c} -\epsilon \\ 0 \\ \end{array} \right) + \frac{\nabla w^\epsilon}{w^\epsilon}(x,y),\end{aligned}$$ we may reduce our considerations to studying the asymptotics of $w^\epsilon$ and its first derivatives. It will be convenient to replace the convolution integral in (\[shiftpersoln\]) by the Fourier series whose coefficients are the product of the Fourier coefficients of $p^\epsilon$ and the $\vartheta_3$ series. This results in the following alternative representation of $w^\epsilon$: $$\begin{aligned} {w^\epsilon}(x,y) &=& \frac{1}{\sqrt{4\pi}} \sum_{n\in \mathbb{Z}}\int_0^\infty\frac{dt}{t^\frac{1}{2}} e^{- \left((1+\epsilon^2(2n+i)^2)t+\frac{y^2}{4t}\right)}\widehat{p^\epsilon} (n)e^{\frac{2\pi i n x}{\ell}}.\end{aligned}$$ With the change of variables, $$s = \frac{t}{y}$$ this representation takes the form $$\begin{aligned} \label{asympw} {w^\epsilon}(x,y) &=& \sqrt{\frac{y}{4\pi}} \sum_{n\in \mathbb{Z}}\int_0^\infty\frac{ds}{s^\frac{1}{2}} e^{-\frac{y}{4}\left(\frac{s}{s_n^2} + \frac{1}{s}\right)}\widehat{p^\epsilon}(n)e^{\frac{2\pi i n x}{\ell}},\end{aligned}$$ where $s_n = \frac{1}{2\sqrt{1 + \epsilon^2 (2n+i)^2}}$. The critical point of the exponent is $s=s_n$ and the expansion of the exponent in the $nth$ term of the series near this critical point has the form $$\begin{aligned} \frac{s}{s_n^2} + \frac{1}{s} &=& \frac{2}{s_n}\left(1 + \frac{(s-s_n)^2}{s_n^2}+ \mathcal{O}\left(\frac{(s-s_n)^3}{s_n^3}\right)\right).\end{aligned}$$ An asymptotic expansion in large $y$ may be developed for the integral in each term of the series (\[asympw\]) by the method of Laplace. By the uniform convergence of the series (for large $y$), the asymptotic expansion of the series is equivalent to the sum of the asymptotic expansions from each term. We implement this strategy to find the leading order, large $y$ behavior, and next corrections, for $w^\epsilon$, $w_x^\epsilon$ and $w_y^\epsilon$: $$\begin{aligned} w^\epsilon(x,y) &=& \sqrt{\frac{y}{4\pi}} \sum_{n\in \mathbb{Z}}s_n^\frac{1}{2} e^{-\frac{y}{s_n}}\int_{-1}^\infty \frac{dz}{(1+z)^\frac{1}{2}} e^{- \frac{y}{2s_n}\left(z^2+\mathcal{O}(z^3)\right)}\widehat{p^\epsilon}(n)e ^{\frac{2\pi i n x}{\ell}}, \\ w_x^\epsilon(x,y) &=& -2 i \epsilon \sqrt{\frac{y}{4\pi}} \sum_{n\ne 0} s_n^\frac{1}{2} e^{-\frac{y}{s_n}}\int_{-1}^\infty \frac{dz}{(1+z)^\frac{1}{2}} e^{-\frac{y}{2s_n}\left(z^2+\mathcal{O}(z^3)\right)}n \widehat{p^\epsilon}(n)e^{\frac{2\pi i n x}{\ell}}, \\ w_y^\epsilon(x,y) &=& \frac{1}{2y}w^\epsilon - \frac{1}{2}\sqrt{\frac{y}{4\pi}} \sum_{n\in \mathbb{Z}}s_n^{-\frac{1}{2}} e^{-\frac{y}{s_n}}\int_{-1}^\infty \frac{dz}{(1+z)^\frac{1}{2}} e^{- \frac{y}{2s_n}\left(z^2+\mathcal{O}(z^3)\right)}\widehat{p^\epsilon}(n)e ^{\frac{2\pi i n x}{\ell}},\end{aligned}$$ where in the $n^{th}$ term of each series, $z=\frac{s-s_n}{s_n}$, respectively. We can now apply Laplace’s method to each term and then observe that the dominant contributions for large $y$ come from the $0, +1, -1$ Fourier modes. Retaining just these we derive the following asymptotic behavior for $\nabla \log w^\epsilon$: $$\begin{aligned} \frac{w_x^\epsilon}{w^\epsilon}(x,y) &=& \frac{-4\epsilon\sqrt{1-\epsilon^2}}{\widehat{p^\epsilon}(0)}e^{-2\epsilon^2y}\Im\left(\widehat{p^\epsilon}(1) e^{-2i \epsilon^2 y}\right) = \mathcal{O}\left(\epsilon e^{-2\epsilon^2 y}\right)\\ \frac{w_y^\epsilon}{w^\epsilon}(x,y) &=& \frac{1}{2y} -\sqrt{1-\epsilon^2}\frac{1 + \Re\left(\frac{\widehat{p^\epsilon}(1)}{\widehat{p^\epsilon}(0)} e^{2i(\epsilon x - \epsilon^2 y)}\right)e^{-2\epsilon^2 y} +\mathcal{O}\left(\epsilon^2e^{-2\epsilon^2 y}\right)}{1 + \Re\left(\frac{\widehat{p^\epsilon}(1)}{\widehat{p^\epsilon}(0)}e^{2i(\epsilon x - \epsilon^2 y)}\right)e^{-2\epsilon^2 y} +\mathcal{O}\left(\epsilon^2e^{-2\epsilon^2 y}\right)}\\ &=& -\sqrt{1-\epsilon^2} + \frac{1}{2y} +\mathcal{O}\left(\epsilon^2 e^{-2\epsilon^2 y}\right).\end{aligned}$$ Based on these asymptotics we can now estimate the energy in the large $y$ domain. $$\begin{aligned} \nabla \theta^\epsilon &=& \left( \begin{array}{c} -\epsilon \\ -\sqrt{1-\epsilon^2} \\ \end{array} \right) + \left( \begin{array}{c} \mathcal{O}\left(\epsilon e^{-2\epsilon^2 y} \right) \\ \mathcal{O}\left(\frac{1}{y} + \epsilon^2 e^{-2\epsilon^2 y} \right) \\ \end{array} \right)\end{aligned}$$ from which it follows that $$\begin{aligned} |\nabla \theta^\epsilon|^2 &=& 1 + \mathcal{O}\left(\epsilon^2 e^{-2\epsilon^2 y} \right) + \mathcal{O}\left(\frac{1}{y}\right)\end{aligned}$$ and so the energy density $$\begin{aligned} \left(1- |\nabla \theta^\epsilon|^2\right)^2 &=& \mathcal{O}\left(\epsilon^4 e^{-4\epsilon^2 y} \right) + \mathcal{O}\left(\frac{\epsilon^2}{y} e^{-2\epsilon^2 y} \right) + \mathcal{O}\left(\frac{1}{y^2}\right).\end{aligned}$$ &gt;From this it follows that the “large $y$” part of the total energy is bounded as $$\mathcal{F}^\epsilon_{y\geq M_\epsilon} \lesssim \frac{1}{\epsilon M_\epsilon}.$$ Thus, if we take $M_\epsilon = c/\epsilon$, this part of the total energy will remain finite as $\epsilon \to 0$. ### Estimates in $[0,\ell] \times \left\{y <M_\epsilon\right\}$ We next turn to consideration of the energy density in the *finite* part of the domain. To facilitate this consideration we will sometimes make the uniformizing change of variables $z = \epsilon \xi$ and $h = \epsilon x$ in the Jacobi theta function (\[Jacobi\]): $$\begin{aligned} \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)=\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) \label{perJacobi}\end{aligned}$$ In what follows we will assume that $a$ is chosen to depend on $\epsilon$ in such a way that $1-a^\epsilon = \mathcal{O}(\epsilon)$. We will make use here of the following single-layer potential counterpart of the double-layer potential representation (\[shiftpersoln\]), which in fact can be deduced directly from a change of variables in (\[asympw\]): $\noindent {w^\epsilon}(x,y)=$ $$\begin{aligned} &=& \frac{-y}{\sqrt{4\pi}} \int_0^\infty\frac{dt}{t^\frac{3}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \frac{1}{\ell}\int_{0}^{\ell} w^\epsilon(\xi,0) \,\,\vartheta_3\left(\frac{-(x - \xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right) d\xi,\label{shiftpersoln2}\end{aligned}$$ We study the asymptotic behavior of the convolution integral in (\[shiftpersoln2\]) for $x \in (0, a\ell)$ and for *times* $t$ of order less than $1/\epsilon$: $$\begin{aligned} \label{convoest} &&\frac{1}{\ell}\int_{0}^{\ell} w^\epsilon(\xi) \,\, \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)d\xi \\ \nonumber &=&\frac{1}{\ell}\int_{0}^{\ell} e^{\epsilon \xi} \,\, \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)d\xi\\ \nonumber &+& \frac{1}{\ell}\int_{a\ell}^{\ell} \left(w^\epsilon(\xi) - e^{\epsilon \xi}\right) \,\, \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)d\xi \\ \nonumber &=& \frac{1}{\pi}\int_{0}^{\pi} e^{z} \,\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz\\ \nonumber &+& \frac{1}{\pi}\int_{0}^{\pi} \left(q_a(\frac{z}{\epsilon}) - e^{z}\right) \,\, \vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi},\frac{-4\epsilon^2 t}{\pi}\right) dz \\ \nonumber &=& e^{\epsilon x} + o(\epsilon),\end{aligned}$$ where in the third line above, the form of the integrals follows from making the change of variables as in (\[perJacobi\]). In the second integral we smoothly extend $q_a(\frac{z}{\epsilon}) - e^z$ to be zero on $(0,a\pi)$. The final line follows for $t$ of order less than $1/\epsilon$ because in this regime the Jacobi theta function inside the convolution behaves as a *Dirac comb* as $\epsilon \to 0$. The second term has this asymptotic behavior because $h \in (0,a\pi)$ and the support of $q_a(\frac{z}{\epsilon}) - e^z$ is complementary to this interval, so that this integral decays exponentially to zero with $\epsilon$, as with a Dirac sequence away form its support. Based on (\[convoest\]) we can estimate $w^\epsilon$ as $$\begin{aligned} \label{w-limit} {w}^\epsilon(h,y) &=& \nonumber \frac{-y}{\sqrt{4\pi}} \left[\int_0^{1/\epsilon}\frac{dt}{t^\frac{3}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \left(e^{\epsilon x} + {o}(\epsilon)\right)\right] + \mathcal{O}\left(e^{-1/\epsilon}\right)\\ &=& e^{\epsilon x} e^{-\sqrt{1-\epsilon^2}y}+ o(\epsilon).\end{aligned}$$ The evaluation of the previous integral may be deduced from a basic Bessel identity (see [@AS] 9.6.23). In order to estimate $\nabla\theta^\epsilon$, we also need to estimate the $x$ and $y$ derivatives of $w^\epsilon(x,y)$. To this end we first consider the $x$-derivative of the internal convolution integral which equals $$\begin{aligned} &&\partial_x \frac{1}{\pi}\int_{0}^{\pi} \left( q_a(\frac{z}{\epsilon}) - e^z \right) \,\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz \\ &=& \epsilon \frac{1}{\pi}\int_{0}^{\pi} \left( q_a(\frac{z}{\epsilon}) - e^z \right) \,\,\partial_h \vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz \\ &=& \nonumber - \epsilon\frac{1}{\pi}\int_{0}^{\pi} \left( q_a(\frac{z}{\epsilon}) - e^z \right) \,\,\partial_z\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz.\end{aligned}$$ Integrating by parts, the above derivative may be rewritten as $$\begin{aligned} \label{xderiv}&& \frac{1}{\pi}\int_{0}^{\pi} \epsilon\partial_z \left( q_a(\frac{z}{\epsilon}) - e^z \right)\, \,\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz\\ &=& \nonumber o(\epsilon)\,\, \mbox{for}\,\,\, h \in (0,a\pi),\end{aligned}$$ as for the second integral in the last line of (\[convoest\]). Thus, $$\begin{aligned} \nonumber \frac{u^\epsilon_x}{u^\epsilon} &=& -\epsilon + \frac{\frac{y}{\sqrt{4\pi}} \int_0^\infty\frac{dt}{t^\frac{3}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \frac{1}{\pi}\int_{0}^{\pi} \epsilon\partial_z \left( q_a(\frac{z}{\epsilon}) - e^z \right)\,\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz}{\frac{y}{\sqrt{4\pi}} \int_0^\infty\frac{dt}{t^\frac{3}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \frac{1}{\pi}\int_{0}^{\pi} \left( q_a(\frac{z}{\epsilon}) - e^z \right)\,\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz}\\ \label{xlogd}&=& -\epsilon + \frac{o(\epsilon)}{e^{\epsilon x} + o(\epsilon)} = \mathcal{O}(\epsilon)\end{aligned}$$ For the $y$ logarithmic derivative we have $$\begin{aligned} \nonumber \frac{{u}^\epsilon_{y}}{{u}^\epsilon} &=& \frac{1}{y} - \frac{\frac{y}{2}\int_0^\infty\frac{dt}{t^\frac{5}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \frac{1}{\ell}\int_{0}^{\ell} w^{\epsilon}(\xi,0) \vartheta_3\left(\frac{-(x - \xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right) d\xi}{\int_0^\infty\frac{dt}{t^\frac{3}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})}\frac{1}{\ell}\int_{0}^\ell w^{\epsilon}(\xi,0) \vartheta_3\left(\frac{-(x- \xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right) d\xi } \\ \nonumber &=& \frac{1}{y} - \frac{\frac{y}{2}\int_0^\infty\frac{dt}{t^\frac{5}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})} \left(e^{\epsilon x} + o(\epsilon)\right)}{\int_0^\infty\frac{dt}{t^\frac{3}{2}} e^{-((1-\epsilon^2)t+\frac{y^2}{4t})}\left(e^{\epsilon x} + o(\epsilon)\right)}\\ \label{ylogd} &=& \frac{1}{y} - \frac{K_{-\frac{3}{2}}(y)}{K_{-\frac{1}{2}}(y)} + o(\epsilon) = -1 + o(\epsilon).\end{aligned}$$ The last equivalence follows from a Bessel recurrence identity [@AS], formula 9.6.26, together with formula 9.6.6. Thus we finally have $$\begin{aligned} \nabla \theta^\epsilon = \left( \begin{array}{c} 0 \\ -1 \\ \end{array} \right) + \mathcal{O}(\epsilon)\end{aligned}$$ and hence $(1-|\nabla\theta^\epsilon|^2)^2 = \mathcal{O}(\epsilon^2)$. Since the domain $[0,a\ell] \times \left\{y < M_\epsilon\right\}$ has dimensions $1/\epsilon \times 1/\epsilon$, the total energy in this region is also asymptotically finite. ### Modification of the Self-dual Test Funciton It remains to estimate the energy in the region $[a\ell, \ell] \times \left\{y < M_\epsilon\right\}$ which has dimensions $\mathcal{O}(1) \times 1/\epsilon$. The question of the finiteness of the energy of the $\theta^\epsilon$ we have been considering in this region is beside the point for general purpose of but, this self-dual solution does not satisfy the boundary condition (\[6\]) in this region. As stated earlier we are going to modify the self-dual test function in this region so that the boundary condition (\[6\]) is satisfied. To that end we fix a small value of $\delta$ and let $\mathbf{B}(\delta)$ denote the $\delta$-neighborhood of $[a\ell, \ell]$ in $\mathcal{S}^\epsilon$. We modify $w^\epsilon$ in this neighborhood as follows. Define $$\begin{aligned} \label{finalsoln} \widetilde{w^\epsilon}(x,y) &=& \phi_1(x,y) w^\epsilon(x,y) + \phi_2(x,y) w_2(x,y)\end{aligned}$$ where $\{\phi_1,\phi_2\}$ is a partition of unity subordinate to the cover of $\mathcal{S}^\epsilon$ given by $$\begin{aligned} U_1 &=& \mathcal{S}^\epsilon\backslash \mathbf{B}(\delta/2) \\ U_2 &=& \mathbf{B}(\delta)\end{aligned}$$ and $w_2(x,y) = w^\epsilon(x,0)\cosh\left( y\right)$, where $w^\epsilon(x,0)$ here is defined as in $(\ref{5}^\prime - \ref{6}^\prime)$. One has $$\begin{aligned} \phi_1 &=& \left\{\begin{array}{cc} 1 & U_1\backslash \mathbf{B}(\delta)\\ 0 & \mathbf{B}(\delta/2) \\ \end{array}\right. \\ \phi_2 &=& \left\{\begin{array}{cc} 1 & \mathbf{B}(\delta/2) \\ 0 & U_1\backslash \mathbf{B}(\delta)\\ \end{array} \right.\end{aligned}$$ and $\phi_1 + \phi_2 \equiv 1$. It is straightforward to check that $\widetilde{w^\epsilon}(x,y)$ satisfies the boundary conditions (\[5\]) and (\[6\]): $$\begin{aligned} \lim_{y\to 0} \widetilde{w^\epsilon}(x,y) &=& \phi_1(x,0) w^\epsilon(x,0) + \phi_2(x,0) w^\epsilon(x,0)\\ &=& \left(\phi_1(x,0)+ \phi_2(x,0)\right) w^\epsilon(x,0)\\ &=& w^\epsilon(x,0)\\ &=& e^{\epsilon x}\end{aligned}$$ for $x\in [0,a\ell]$. For $x \in [a\ell, \ell]$, $$\begin{aligned} \lim_{y\to 0} \widetilde{w^\epsilon}_y(x,y) &=& \left(\phi_{1y}(x,0)+ \phi_{2y}(x,0)\right) w^\epsilon(x,0) + \phi_2(x,0) w^\epsilon(x,0)\sinh(0) \\ &=& \left(\phi_1 + \phi_2\right)_y (x,0) w^\epsilon(x,0) + 0 \\ &=& 0.\end{aligned}$$ Thus, $\log \widetilde{w^\epsilon}$ is an admissible test function for the regularized Cross-Newell variational problem. We can now estimate the energy of this test function in $\mathbf{B}(\delta)$. The energy density in this region is bounded and therefore the energy in $\mathbf{B}(\delta)$ is finite. ### Estimates for the “outer” solution in $\left( [a\ell,\ell] \times \left\{y < M_\epsilon\right\}\right)$ It remains to estimate the energy in $\left( [a\ell,\ell] \times \left\{y < M_\epsilon\right\}\right) \backslash \mathbf{B}(\delta) $. To proceed with this we will need a more specific definition of $q_a$ which we now give. Note first that by our assumption that $1 - a^\epsilon = \mathcal{O}(\epsilon)$, the interval $[a\ell,\ell]$ remains of size $\mathcal{O}(1)$ for arbitrarily small values of $\epsilon$. We will now further pin this down by setting $1-a^\epsilon = c \epsilon$ for a value of $c$ that is fixed, independent of $\epsilon$. Consequently, $[a\ell,\ell]$ is now an interval of fixed length $c\pi$ which can therefore also be represented as $[\ell - c\pi,\ell]$. Recall that $q_a$ needs to be built so that on this interval it matches, through second order, to $e^{\epsilon x}$ at the left endpoint and similarly to $e^{\epsilon x - \pi}$ at the right endpoint. Toward this end we observe that the required leading order value on the right is $1$, independent of $\epsilon$ while on the left the leading order value limits to the stable value of $e^\pi$ as $\epsilon \to 0$. Choosing a value $\nu > 0$ that is small with respect to $c\pi$, we define a compressed *tanh*-profile that interpolates between the point $ (x_0, y_0) = (\ell - c\pi + \nu, e^\pi + \gamma)$ and the point $ (x_1, y_1) = (\ell - \nu, 1 - \gamma)$ and where $\gamma > 0$ is another chosen value required to be smaller than $1$. (This last requirement will insure that the positivity claim made at the end of subsection \[zipper\] holds.) Explicitly this tanh-profile is given by $$\begin{aligned} T(x) &=& \frac{e^\pi + 1}{2} + \left(\frac{e^\pi - 1}{2} + \gamma\right) \tanh \left( \frac{x - (\ell - c\frac{\pi}{2})}{\left(x-(\ell-\nu)\right)\left(x-(\ell - c\pi + \nu)\right)}\right).\end{aligned}$$ Note that the profile of $T(x)$ is independent of $\epsilon$. The only way in which $T$ depends on $\epsilon$ is that this profile translates uniformly with $\ell$ as $\epsilon$ changes. We will define $q_a(x) = T(x)$ on the subinterval $[x_0, x_1] = [\ell - c\pi + \nu,\ell - \nu]$ of $[\ell - c\pi,\ell]$. Next we will define the piece of $q_a(x)$ on the left that interpolates between the point $(\ell - c\pi, e^{\pi - \epsilon c \pi})$ and the point$(x_0, y_0)$. Choose a value $\sigma > 0$ that is small with respect to $\nu$. Consider the covering of $[\ell - c\pi,\ell - c\pi + \nu]$ by the two sets $V_1 = [\ell - c\pi,x_0 - \sigma)$ and $V_2 = (\ell - c\pi + \sigma, x_0]$ and let $\{\psi_1(x), \psi_2(x)\}$ be a partition of unity subordinate to this cover which means, in particular, that $$\begin{aligned} \psi_1 &=& \left\{\begin{array}{cc} 1 & [\ell - c\pi, \ell - c\pi + \sigma)\\ 0 & (x_0 - \sigma, x_0] \\ \end{array}\right. \\ \psi_2 &=& \left\{\begin{array}{cc} 1 & (x_0 - \sigma, x_0] \\ 0 & [\ell - c\pi, \ell - c\pi + \sigma). \\ \end{array} \right.\end{aligned}$$ On $[\ell - c\pi, x_0]$ we define $$\begin{aligned} \label{qa} q_a(x) &=& \psi_1(x) e^{\epsilon x} + \psi_2(x) (e^\pi + \gamma).\end{aligned}$$ It is straightforward to see that with these choices $q_a(x)$ is smooth throughout $[\ell - c\pi, \ell - \nu)$ and satisfies the smooth matching conditions on the left. Moreover, it is clear from the functions comprising (\[qa\]) that $q_a$ and its derivatives remain bounded on $[\ell - c\pi, \ell - \nu)$ as $\epsilon \to 0$. A similar construction may be made on the right; i.e., on $(\ell - c\pi + \nu, \ell]$. This completes our description of $q_a(x)$. The study of the convolution integral in (\[shiftpersoln2\]) in the region where $x \in [a\ell, \ell]$ now proceeds similarly to what was done in (\[convoest\]) and subsequent formulae. In particular, the analogous result to (\[convoest\]) is that for $x \in (a\ell, \ell)$ and for *times* $t$ of order less than $1/\epsilon$: $$\begin{aligned} \label{convoest2} &&\frac{1}{\ell}\int_{0}^{\ell} w^\epsilon(\xi) \,\, \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)d\xi \\ \nonumber &=&\frac{1}{\ell}\int_{0}^{\ell} q_a(\xi) \,\, \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)d\xi\\ \nonumber &+& \frac{1}{\ell}\int_{a\ell}^{\ell} \left(w^\epsilon(\xi) - q_a(\xi)\right) \,\, \vartheta_3\left(\frac{- (x-\xi) + 2\epsilon t}{\ell}, \frac{-4\pi t}{\ell^2}\right)d\xi \\ \nonumber &=& \frac{1}{\pi}\int_{0}^{\pi} q_a(\frac{z}{\epsilon}) \,\,\vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi}, \frac{-4\epsilon^2 t}{\pi}\right) dz\\ \nonumber &+& \frac{1}{\pi}\int_{0}^{\pi} \left(w^\epsilon(\frac{z}{\epsilon}) - q_a(\frac{z}{\epsilon}) \right) \,\, \vartheta_3\left(\frac{-(h - z) + 2\epsilon^2 t}{\pi},\frac{-4\epsilon^2 t}{\pi}\right) dz \\ \nonumber &=& q_a(x) + o(\epsilon),\end{aligned}$$ with $q_a(x)$ here bounded away from zero, independent of $\epsilon$, by our earlier choice of $\gamma$. Hence the denominators in the estimates analogous to (\[xlogd\]) and (\[ylogd\]) are under control. In the subsequent formulae the roles of $e^{\epsilon x}$ and $q_a(x)$ are effectively interchanged as above and all proceeds as before. The result is that the energy in $\left( [a\ell,\ell] \times \left\{y < M_\epsilon\right\}\right)\backslash \mathbf{B}(\delta)$ is asymptotically bounded like $\mathcal{O}(\epsilon)$. It thus follows that the total energy of our family of test functions is uniformly bounded in $\epsilon$. Lower bound {#sec:l_bound} =========== Following the ideas of Jin and Kohn [@JK], we will prove [*ansatz-free*]{} lower bounds for the functional ${\mathcal}{F}^{\epsilon}$ by identifying vector fields $\Sigma(\nabla\theta)$ such that $${\mathcal}{F}^{\epsilon}[\theta;a,\delta] \geq C^{-1} \left|\iint_{{\mathcal}{S}^\epsilon} \nabla \cdot \Sigma(\nabla \theta) dx dy \right|.$$ This allows us to obtain information about the energy ${\mathcal}{F}^\epsilon$ purely in terms of the boundary conditions on $\theta$. To avoid the proliferation of symbols, here and henceforth, $C, C', C_1$, [*etc*]{} denote (finite) constants whose precise value is unimportant, and different occurrences of the same symbol might denote different values of the constants. $e_1,e_2,K,K_1$, [*etc*]{} denote constants that have the same value in all their occurrences. A smooth vector function $\Sigma(p,q) = (\Sigma_1, \Sigma_2)$ is [*subordinate to the energy*]{} if $$\begin{gathered} \left|\frac{\partial\Sigma_1(p,q)}{\partial p}\right| + \left|\frac{\partial\Sigma_1(p,q)}{\partial q} + \frac{\partial\Sigma_2(p,q)}{\partial p}\right| + \left| \frac{\partial\Sigma_2(p,q)}{\partial q}\right| \leq C |1 - p^2 - q^2| \label{eq:subordinate}\end{gathered}$$ for some $C < \infty$. If $\Sigma$ is subordinate to the energy, it follows that $$\begin{aligned} |\nabla \cdot \Sigma(\nabla \theta)| & \leq \left|\frac{\partial\Sigma_1(p,q)}{\partial p}\right| |\theta_{xx}| + \left|\frac{\partial\Sigma_1(p,q)}{\partial q} + \frac{\partial\Sigma_2(p,q)}{\partial p}\right||\theta_{xy}| + \left| \frac{\partial\Sigma_2(p,q)}{\partial q}\right||\theta_{yy}| \\ & \leq C|1-\theta_x^2 - \theta_y^2| |\nabla \nabla \theta|,\end{aligned}$$ where we use the identification $p = \theta_x, q = \theta_y$ and $|\nabla \nabla \theta|^2 = \theta_{xx}^2 + 2 \theta_{xy}^2 + \theta_{yy}^2$. Consequently, $$\begin{aligned} {\mathcal}{F}^{\epsilon}[\theta;a,\delta] & = \iint_{{\mathcal}{S}^\epsilon} \left\{[\nabla \nabla \theta]^2 + (1 - |\nabla \theta|^2)^2\right\} dx dy - 2 \int_{\partial {\mathcal}{S}^\epsilon} \theta_x d \theta_y \\ & \geq 2 \iint_{{\mathcal}{S}^\epsilon} |\nabla \nabla \theta| |1 - |\nabla \theta|^2| dx dy - 2 \int_{\partial {\mathcal}{S}^\epsilon} \theta_x d \theta_y \\ & \geq \frac{2}{C} \left|\iint_{{\mathcal}{S}^\epsilon} \nabla \cdot \Sigma(\nabla \theta) dx dy \right| - 2 \int_{\partial {\mathcal}{S}^\epsilon} \theta_x d \theta_y \\ & \geq C^{-1} \left|\iint_{{\mathcal}{S}^\epsilon} \nabla \cdot \Sigma(\nabla \theta) dx dy \right|.\end{aligned}$$ In obtaining the last equation, we use the fact that $$\int_{\partial {\mathcal}{S}^\epsilon} \theta_x d \theta_y = 0$$ for the boundary conditions in (\[eq:bc\]). There are constants $e_1,K_1 > 0$ such that for all $\epsilon \in (0,1]$,$a \in [0,1]$ and $\delta \in \mathbb{R}$, we have $${\mathcal}{F}^\epsilon[\theta;a,\delta] \geq \frac{e_1 \epsilon^2}{(1-a)^2} - K_1\epsilon^2.$$ \[lem:squeeze\] Let $\phi \geq 0$ be a smooth, compactly supported function such that $$\begin{aligned} & \phi(0) = 1, \\ & \phi(1) < \phi(0),\\ & f(p) = p \phi(p^2) \text{ has a single maximum at } p=1.\end{aligned}$$ An explicit example of a function $\phi$ whith these properties is $$\phi(p) = \begin{cases} \exp\left[\frac{1}{2} - \frac{1}{(2-p)(p+1)} - \frac{p}{4} \right] & p \in (-1,2) \\ 0 & \text{ otherwise } \end{cases}$$ Let $b = (1-a)/\epsilon$. Define the vector field $\Sigma(p,q)$ by $$\begin{aligned} \Sigma_2(p,q) & = p \phi(b^2 p^2) \\ \Sigma_1(p,q) & = - \int_0^q \left[\phi(b^2 (1-\eta^2)) + 2(b^2(1-\eta^2)\phi'(b^2(1-\eta^2)) \right] d\eta.\end{aligned}$$ Since $\phi$ has compact support, it follows that $\Sigma$ is bounded on $\mathbb{R}^2$. An explicit calculation shows that the quantities $\Sigma_{1,p}$ and $\Sigma_{2,q}$ are zero. Also, $$\begin{aligned} \left|\frac{\partial\Sigma_1(p,q)}{\partial q} + \frac{\partial\Sigma_2(p,q)}{\partial p}\right| & = \left|\phi(b^2 p^2) + 2 b^2 p^2 \phi'(b^2 p^2) - \phi(b^2 (1-q^2)) \right.\\ &\left. - 2b^2(1-q^2)\phi'(b^2(1-q^2)\right| \\ & \leq C b^2 |(1-p^2 - q^2)|\end{aligned}$$ where $$C = \sup_{x,y}\left|\frac{\phi(x) + 2 x \phi'(x) - \phi(y) - 2 y \phi'(y)}{x-y}\right| \leq \sup_z |3 \phi'(z) + 2 z \phi''(z)|$$ is clearly finite since $\phi$ is compactly supported and twice differentiable. This proves that $\Sigma$ is subordinate to the energy. Since $\nabla \cdot \Sigma(\nabla \theta) = (\Sigma_{2,p} + \Sigma_{1,q}) \theta_{xy}$ we obtain $$\begin{aligned} \left|\iint \nabla \cdot \Sigma(\nabla \theta)\, dx dy\right| & \leq Cb^2 \iint| (1 - \theta_x^2 - \theta_y^2) \theta_{xy}| \, dx dy \nonumber \\ & \leq \frac{C b^2}{2} \left[ \iint (1 - \theta_x^2 - \theta_y^2)^2\, dx dy + \iint [\nabla \nabla \theta]^2\, dx dy \right] \nonumber \\ & = \frac{Cb^2}{2} {\mathcal}{F}^\epsilon[\theta;a,\delta] \label{eq:lbnd1}\end{aligned}$$ Integrating by parts, we have $$\begin{aligned} \iint \nabla \cdot \Sigma(\nabla \theta)\, dx dy = & \Sigma_2(\epsilon,\sqrt{1-\epsilon^2}) \frac{\pi}{\epsilon} - \int_0^{a \pi/\epsilon} \Sigma_2(0,\theta_y(x,0)) dx \\ & - \int_{a \pi/\epsilon}^{\pi/\epsilon} \Sigma_2(\theta_x(x,0),0)dx,\end{aligned}$$ where the contributions from the boundaries at $x = 0$ and $x = \pi/\epsilon$ cancel due to the periodicity. By construction, $\Sigma_2(p,0) = 0$ and $\Sigma_2(p,q)$ has a maximum value at $p = 1/b$. Consequently, $$\begin{aligned} \int_0^{a \pi/\epsilon} \Sigma_2(0,\theta_y(x,0)) dx & = 0 \\ \int_{a \pi/\epsilon}^{\pi/\epsilon} \Sigma_2(\theta_x(x,0),0) dx & \leq \frac{(1 - a) \pi}{\epsilon} \frac{\phi(1)}{b} = \pi \phi(1).\end{aligned}$$ Also, $\phi(0) = 1$ and $\phi$ is Lipschitz so that $$\Sigma_2(\epsilon,\sqrt{1-\epsilon^2}) = \epsilon \phi(b^2 \epsilon^2) = \epsilon \phi((1-a)^2) \geq \epsilon(1 - C' (1-a)^2),$$ for some finite $C'$. Combining these estimates with (\[eq:lbnd1\]), we obtain $${\mathcal}{F}^\epsilon[\theta;a,\delta] \geq \frac{2 \pi}{C b^2}\left[\phi(0) - \phi(1) - C'(1-a)^2 \right],$$ and rewriting $b$ in terms of $a$ and $\epsilon$ yields the desired conclusion. The above lemma shows that the energy grows without bound as the quantity $(1-a)$ becomes small. However, we do not have [*a priori*]{} control on the size of $(1-a)$. Consequently, to obtain a lower bound for the energy, we need a complementary estimate which shows that the energy grows as the quantity $(1-a)$ becomes large. To prove this result, we first construct a vector field $\Sigma$ subordinate to the energy functional as follows: Let $\psi \geq 0$ be a smooth, compactly supported function such that $$\begin{aligned} & \psi(0) = 1 \\ & \int_0^{\infty} (1-\xi^2)\psi(\xi^2) d\xi = 0\end{aligned}$$ We can always construct such a function, given $\chi \geq 0$, a compactly supported function with $\chi(0) = 1$. Observe that $$\int_0^\infty (1 - \xi^2) \chi\left(\frac{\xi^2}{\eta^2}\right) d \xi = \eta(A_0 - \eta^2 A_1),$$ where $A_0,A_1 > 0$. Consequently, by an appropriate choice of $\eta$, we get $\psi(x) = \chi(x/\eta^2)$ with the required properties. We define the functions $\zeta(q^2)$ and $\sigma(q^2)$ by $$\begin{aligned} \zeta(q^2) & = \int_0^q (q - \eta) \psi(\eta^2) d \eta \nonumber \\ \sigma(q^2) & = \int_0^q (q - \eta) (1 - \eta^2) \psi(\eta^2) d \eta \label{eq:zeta_sigma}\end{aligned}$$ Note that the functions $\zeta$ and $\sigma$ are well defined for positive values of their arguments, that is the expressions on the right hand sides of the above equations are even functions of $q$. From these expressions, we have $$\begin{aligned} \frac{\partial^2}{\partial q^2} \zeta(q^2) = \psi(q^2); & \quad \zeta(0) = 0 \\ \frac{\partial^2}{\partial q^2} \sigma(q^2) = (1-q^2)\psi(q^2); & \quad \sigma(0) = 0\end{aligned}$$ We will also use the same letters $\zeta$ and $\sigma$ to denote smooth extensions of the functions defined above to all of $\mathbb{R}$. We will pick extensions such that the supports of $\zeta$ and $\sigma$ are contained in $[-1,\infty)$. Let $\varphi \geq 0$ be a compactly supported function and set $$\begin{aligned} V(p,q) = & \, \, \varphi(p^2)\left[\sigma(q^2) - p^2\zeta(q^2)\right] \nonumber \\ & - \int_0^p (p-\xi)\left\{ \sigma(1-\xi^2) \left[ \frac{\partial^2}{\partial \xi^2} \varphi(\xi^2) \right] - \zeta(1-\xi^2) \left[ \frac{\partial^2}{\partial \xi^2} (\xi^2 \varphi(\xi^2)) \right] \right\}\, d\xi. \label{eq:defnv}\end{aligned}$$ $V$ is now an even function of $p$ and $q$. Define the vector field $\Sigma$ by $$\Sigma(p,q) = \left( -\frac{\partial}{\partial p} V,\frac{\partial}{\partial q} V \right) \label{eq:def}$$ From (\[eq:def\]) it follows that $$\frac{\partial\Sigma_1(p,q)}{\partial q} + \frac{\partial\Sigma_2(p,q)}{\partial p} = 0.$$ Also, $$\left|\frac{\partial\Sigma_2(p,q)}{\partial q}\right| = \left|\frac{\partial^2 V(p,q)}{\partial q^2}\right| = |\varphi(p^2)\psi(q^2)||1-p^2-q^2| \leq C_1 |1-p^2-q^2|.$$ With $C_1 = \sup|\varphi(p^2) \psi(q^2)| < \infty$. Finally, an explicit calculation yields $$\begin{aligned} \left|\frac{\partial\Sigma_1(p,q)}{\partial p}\right| & = \left|\left[ \frac{\partial^2}{\partial p^2} \varphi(p^2) \right]\left( \sigma(q^2) - \sigma(1-p^2) \right) - \left[ \frac{\partial^2}{\partial p^2} p^2 \varphi(p^2) \right]\left( \zeta(q^2) - \zeta(1-p^2) \right)\right| \nonumber \\ & \leq C_2 |1 - p^2 - q^2|\end{aligned}$$ where $C_2$ can be bounded in terms of the support of $\varphi$ and the maximum values of $|\varphi|, |\varphi'|, |\varphi''|, |\zeta'|$ and $|\sigma'|$. Clearly $\varphi$ and all it’s derivatives are uniformly bounded since it is smooth and compactly supported. From (\[eq:zeta\_sigma\]), we have $$\begin{aligned} \zeta'(q^2) & = \frac{1}{2q}\int_0^q \psi(\eta^2) d \eta \\ \sigma'(q^2) & = \frac{1}{2q}\int_0^q (1 - \eta^2) \psi(\eta^2) d \eta\end{aligned}$$ Since $\psi$ is compactly supported, these derivatives vanish as $q^2\rightarrow \infty$, implying that $\zeta'$ and $\sigma'$ are bounded for all positive values of the argument. Since $\sigma$ and $\zeta$ are smooth and are identically zero if their arguments are sufficiently negative, it follows that $C_2 < \infty$. It thus follows that the vector field $\Sigma$ is subordinate to the energy functional. For future use, let us record a few observations that follow directly from the construction: $\Sigma_2(p,0) = 0$ since $\Sigma_2$ is an odd function of $q$. $$\Sigma_{2,q}(0,q) = V_{qq}(0,q) = \psi(q^2)(1-q^2).$$ Consequently, the non-degenerate critical points are at $ q = \pm 1$. Differentiating in $q$, we get $$\Sigma_{2,qq}(0,\pm 1) = \mp 2 \psi(1),$$ so that $\Sigma_2(0,q)$ has a maximum at $q = 1$ and a minimum at $q = -1$. Finally, $$\Sigma_2(0,q) = 2 q \sigma'(q^2) = \int_0^q (1-\xi^2) \psi(\xi^2) d \xi,$$ so that $\Sigma_2(0,1) > 0$ and $\Sigma_2(0,q) \rightarrow 0$ as $q \rightarrow \infty$. $M = \int_0^1 (1-\xi^2) \psi(\xi^2) d \xi$ will denote the maximum value of $\Sigma_2(0,q)$. $$\begin{aligned} \Sigma_2(\epsilon,\sqrt{1-\epsilon^2}) & = \varphi(\epsilon^2) \int_0^{\sqrt{1-\epsilon^2}} (1- \epsilon^2 -\xi^2) \psi(\xi^2) d \xi \\ & \geq M - K \epsilon^2\end{aligned}$$ for a constant $K < \infty$. In obtaining the last line, we use $$\begin{aligned} |1-\varphi(\epsilon^2)| & \leq C_1 \epsilon^2 \\ \left|\int_0^{\sqrt{1-\epsilon^2}} \psi(\xi^2) d \xi \right| & \leq C_2 \\ \left|\int_{\sqrt{1-\epsilon^2}}^1 (1-\xi^2) \psi(\xi^2) d \xi \right| & \leq C_3 \epsilon^2\end{aligned}$$ for some bounded constants $C_1,C_2,C_3$. \[obs:estimate\] There are constants $e_2,K_2 > 0$ such that for all $\epsilon \in (0,1]$, $a \in [0,1]$ and $\delta \in \mathbb{R}$, we have $${\mathcal}{F}^\epsilon[\theta;a,\delta] \geq \frac{e_2(1-a)}{\epsilon} - K_2 \epsilon.$$ \[lem:extend\] For the case $a = 0$, corresponding to the self-dual minimizers, this estimate captures the right scaling of the minimum energy as $\epsilon \rightarrow 0$. The proof of the lemma follows from estimating a lower bound for the functional ${\mathcal}{F}^\epsilon[\theta;a,\delta]$ using the vector field $\Sigma$ that we constructed above. For the vector field $\Sigma$ we have $$\begin{aligned} \iint \nabla \cdot \Sigma(\nabla \theta)\, dx dy = & \Sigma_2(\epsilon,\sqrt{1-\epsilon^2}) \frac{\pi}{\epsilon} - \int_0^{a \pi/\epsilon} \Sigma_2(0,\theta_y(x,0)) dx \\ & - \int_{a \pi/\epsilon}^{\pi/\epsilon} \Sigma_2(\theta_x(x,0),0) dx.\end{aligned}$$ As before, the contributions from the boundaries at $x = 0$ and $x = \pi/\epsilon$ cancel due to the periodicity. By construction, $\Sigma_2(p,0) =0$ and $\Sigma_2(0,q)$ has a maximum value $M$ at $q = 1$. Consequently, $$\int_0^{a \pi/\epsilon} \Sigma_2(0,\theta_y(x,0)) dx \leq \frac{M a \pi}{\epsilon}$$ From observation \[obs:estimate\], we obtain $$\Sigma_2(\epsilon,\sqrt{1-\epsilon^2}) \frac{\pi}{\epsilon} \geq \frac{M \pi}{\epsilon} - K \pi \epsilon,$$ Since $\Sigma$ is subordinate to the energy, $${\mathcal}{F}^\epsilon[\theta;a,\delta] \geq \frac{M \pi}{C \epsilon}\left[1 - a - K \epsilon^2 \right],$$ which yields the desired conclusion. We can now prove theorem \[thm:l\_bound\] using lemma \[lem:squeeze\] and lemma \[lem:extend\]. Let $b$ denote the quantity $(1-a)/\epsilon$. From lemma \[lem:squeeze\] we get $${\mathcal}{F}^\epsilon[\theta;a,\delta] \geq \frac{e_1}{b^2} - K_1 \epsilon^2 \geq 3 \left(e_1 e_2^2\right)^{1/3} - 2 e_2 b - K_1 \epsilon^2$$ where the last inequality comes from linearizing the convex function $e_1 b^{-2}$ at $b = (e_1/e_2)^{1/3}$. Combining this estimate with the conclusion of lemma \[lem:extend\], we get $${\mathcal}{F}^\epsilon[\theta;a,\delta] \geq \max\left(3 \left(e_1 e_2^2\right)^{1/3} - 2 e_2 b - K_1 \epsilon^2, e_2 b - K_2 \epsilon \right) \geq (e_1 e_2^2)^{1/3} - \frac{2 K_2 \epsilon +K_1 \epsilon^2 }{3}.$$ If we set $$\epsilon_* = \min\left( \frac{(e_1 e_2^2)^{1/6}}{\sqrt{K_1}}, \frac{(e_1 e_2^2)^{1/3}}{2 K_2}\right),$$ for all $\epsilon < \epsilon_*$, all $a \in [0,1]$ and all $\theta$ satisfying the boundary conditions in (\[eq:bc\]), we have $${\mathcal}{F}^{\epsilon}[\theta;a,\delta] \geq \frac{(e_1 e_2^2)^{1/3}}{3} \equiv E_1.$$ Combining the upper bound ${\mathcal}{F}^{\epsilon}[\theta^\epsilon;a^\epsilon,\delta^\epsilon] \leq E_0$ in theorem \[thm:u\_bound\] with the lower bounds for ${\mathcal}{F}^\epsilon$ in lemma \[lem:squeeze\] and lemma \[lem:extend\], it follows that for $$\epsilon < \min\left(\sqrt{\frac{E_0}{K_1}},\frac{E_0}{K_2}\right),$$ we have $$\sqrt{\frac{e_1}{2E_0}} < \frac{1-a^\epsilon}{\epsilon} < \frac{2 E_0}{e_2}.$$ Consequently, $$1 - \alpha_2 \epsilon < a^\epsilon < 1 - \alpha_1 \epsilon,$$ for sufficiently small $\epsilon$ with $\alpha_1 = \sqrt{e_1/(2E_0)}$ and $\alpha_2 = 2 E_0/e_2$. Numerical Results {#sec:results} ================= In this section, we will present the results of numerical simulations that illustrate and clarify our analysis of the energy and also the structure of the minimizers for the regularized Cross-Newell energy $\mathcal{F}^\epsilon$ within the class of functions given by (\[eq:bc\]). For our numerical simulations, we restrict ourself to the finite domain, ${\mathcal}{R}^{\epsilon} = \{(x,y) \, |\, 0 \leq x \leq l = \pi/\epsilon, 0 \leq y \leq L\}$, where $L \gg 1$ is a length scale much larger than the typical wavelength of the pattern. The boundary conditions in (\[eq:bc\]) which are appropriate for the semi-infinite strip ${\mathcal}{S}^\epsilon$ are modified for the finite domain as follows – $$\begin{gathered} \theta(x,0) = 0 \quad \text{ for } 0 \leq x < a l; \nonumber \\ \theta_y(x,0) = 0 \quad \text{ for } a l \leq x < l; \nonumber \\ \theta(x,y) - \epsilon x \text{ is periodic in $x$ with period $l$ for each } y \in [0,L]; \nonumber \\ \theta(x,L) = \left[\epsilon x + \sqrt{1 - \epsilon^2} L + \delta\right] \label{eq:num_bc}\end{gathered}$$ It is rather straightforward to show that there exist $\theta^\epsilon \in H^2({\mathcal}{R}^\epsilon)$ satisfying (\[eq:num\_bc\]) for an $a^\epsilon \in [0,1)$ and $\delta^\epsilon \in \mathbb{R}$ minimizing the functional $${\mathcal}{F}^\epsilon[\theta;a,\delta] = \iint_{{\mathcal}{R}^\epsilon} \left\{[ \Delta \theta]^2 + (1 - |\nabla \theta|^2)^2 \right\} dx dy.$$ The existence of a minimizer is immediate from the following lemma: Let $0 \leq a < 1$, and $\delta \in \mathbb{R}$ be given. $\rho_j \in L^2({\mathcal}{R}^\epsilon)$ is a sequence of functions that converges weakly to zero. $H^2_{per}$ denotes the completion of periodic (in $x$) functions on ${\mathcal}{R}^\epsilon$ with respect to the $H^2$ norm. If $\theta_j \in H^2_{per}({\mathcal}{R}^\epsilon)$ is a sequence satisfying (in the sense of trace) $$\begin{gathered} \Delta \theta_j = \rho_j \nonumber \\ \theta_j(x,0) = 0 \quad \text{ for } 0 \leq x < a l; \nonumber \\ \partial_y \theta_j(x,0) = 0 \quad \text{ for } a l \leq x < l; \nonumber \\ \theta(x,L) = 0\end{gathered}$$ it follows that, up to extraction of a subsequence and relabelling, we have $\nabla \theta_j \rightarrow 0$ in $L^4({\mathcal}{R}^\epsilon,\mathbb{R}^2)$. Elliptic regularity along with the given boundary conditions implies that the sequence $\theta_j$ is bounded in $H^2({\mathcal}{R}^\epsilon)$. The compactness of the embedding $H^2({\mathcal}{R}^\epsilon) \hookrightarrow W^{1,4}({\mathcal}{R}^\epsilon)$ [@Evans] proves the lemma. If $\tilde{\theta}_j$ is an infimizing sequence for ${\mathcal}{F}^\epsilon[\theta;a,\delta]$ subject to the boundary conditions in (\[eq:num\_bc\]), then let $\theta_j = \tilde{\theta}_j -\varphi$, where $\varphi$ is a smooth function on ${\mathcal}{R}^\epsilon$ satisfying the boundary conditions in (\[eq:num\_bc\]). It then follows from the form of ${\mathcal}{F}^\epsilon$ and the fact that $\tilde{\theta}_j$ is infimizing that $\Delta \theta_j$ is a bounded sequence in $L^2$, and so converges weakly to a limit $\rho^*$. Applying the compactness result of the preceding lemma with reference to the sequence $\rho_j = \Delta \theta_j - \rho^*$, we obtain the existence of a minimizier for the functional ${\mathcal}{F}^\epsilon[\theta;a,\delta]$ for a fixed $a$ and $\delta$. Note that, for a given $a$ it is easy to construct smooth transformations $\psi_t : {\mathcal}{R}^\epsilon \rightarrow {\mathcal}{R}^\epsilon$ such that $\psi_0$ is the identity, if $\theta$ satisfies the boundary conditions in (\[eq:num\_bc\]), then $\theta \circ \psi_t$ satisfies the same boundary conditions with the fraction of the boundary with a Dirichlet boundary condition equaling $a(1+t)$. Further, the energy ${\mathcal}{F}^\epsilon[\theta \circ \psi_t;a(1+t),\delta]$ is a smooth function of $t$ for sufficiently small $t$. A standard argument now implies that, for a given $\delta$ the map $$a \mapsto \inf_{\theta} {\mathcal}{F}^\epsilon[\theta;a,\delta]$$ is continuous for $a \in (0,1)$. A similar argument shows that the map is also continuous at $a = 0$. In Lemma \[lem:squeeze\] we showed that $\liminf_{a \rightarrow 1} {\mathcal}{F}^\epsilon[\theta;a,\delta] = \infty$. Combining these results, we see that $$\inf_{a,\theta} {\mathcal}{F}^\epsilon[\theta;a,\delta] = \inf_a \left[\inf_{\theta} {\mathcal}{F}^\epsilon[\theta;a,\delta]\right].$$ We now consider variations $\theta \rightarrow \theta_t = \theta + t \chi(y/L)$, where $\chi$ is a smooth function vanishing identically on $[0,1/3]$ and equal to 1 on $[2/3,1]$. The functions $\theta_t$ satisfy the boundary conditions in (\[eq:num\_bc\]), except the asymptotic phase shift is given by $\delta + t$. A similar argument as above shows that the map $$\delta \mapsto \inf_{\theta} {\mathcal}{F}^\epsilon[\theta;a,\delta]$$ is continuous and it is easy to see that $ {\mathcal}{F}^\epsilon[\theta;a,\delta] \rightarrow \infty$ as $\delta \rightarrow \pm \infty$. In particular, this proves the existence of an optimal $\delta$, and combining with the results from above, we see that the minimizer $\theta^{\epsilon}, a^\epsilon, \delta^{\epsilon}$, can be obtained by successive minimization in each of the factors. This suggests the following discretization for the functional ${\mathcal}{F}^\epsilon$, which should converge as the grid spacings $\eta,\zeta \rightarrow 0$. We define a grid by $x_i = i \eta, i = 0,1,2,\ldots,m-1, y_j = j \zeta, j = 0,1,2,\ldots,n$, where $\eta = l/m, \zeta = L/n$. The discretization of the test function $\theta(x,y)$ is $$\theta_{i,j} = \theta(i \eta,j \zeta)$$ We define the difference operator $\delta^{\pm}_x$ by $$(\delta^{\pm}_x \theta)_{i,j} = \pm\frac{\theta_{i \pm 1,j} - \theta_{i,j}}{\eta}$$ with similar definitions for $\delta^{\pm}_y$. In terms of the discretization, the boundary conditions are $$\begin{gathered} \theta_{i,0} = 0 \quad \text{ for } 0 \leq i < k; \nonumber \\ \delta_y^+\theta_{i,0} = 0 \quad \text{ for } k \leq i < m; \nonumber \\ \theta_{m,j} = \theta_{0,j} + \pi \quad j = 0,1,2,\ldots,n \\ \theta_{i,n} = \frac{\pi i}{m} + \sqrt{1-\epsilon^2} L + \delta \quad i = 0,1,2,\ldots m-1 \label{eq:discr_bc}\end{gathered}$$ and the Energy functional is discretized as $${\mathcal}{F}^\epsilon \approx \eta \zeta \sum_{i = 0}^{m-1} \sum_{j = 0}^{n} \left[ (\delta_x^+ \delta_x^- + \delta_y^+ \delta_y^-) \theta_{i,j}\right]^2 + \left[ \frac{(\delta_x^+ \theta)_{i,j}^2 +(\delta_x^- \theta)_{i,j}^2 +(\delta_y^+ \theta)_{i,j}^2 +(\delta_y^- \theta)_{i,j}^2 }{2} - 1 \right]^2$$ Computing this functional requires assigning values for $\theta_{i,j}$ with $i = -1, j = -1$ and $j = n+1$. The values for $i = -1$ are obtained form the shift-periodicity of $\theta$ by $\theta_{-1,j} = \theta_{m-1,j} - \pi$. The values for T $j = n+1$ are assigned using the Dirichlet boundary condition $\theta_{i,n+1} = \frac{\pi i}{m} + \sqrt{1-\epsilon^2} (L+\zeta) + \delta$. This functional is minimized using MATLAB’s conjugate-gradient minimization. Fig. \[fig:hyster\] shows the results from minimizing the RCN energy over the pattern $\theta$ and also the phase shift $\delta$, for different values of $\epsilon$, and for a range of values of $a \approx k/m$. The results do indeed suggest that the (partial) minimization with respect to the pattern and the asymptotic phase yields a functional that [*depends continuously*]{} on $a$. Further, this functional has first-order (discontinuous) phase transition at a bifurcation value $\epsilon^*$, below which the global minimizer has $a \neq 0$. Fig. \[fig:energy\] shows the energy of the minimizer (minimizing over the pattern, asymptotic phase and the parameter $a$) as a function of $\epsilon$. Note that the minimum energy is a non-differentiable function of $\epsilon$, as one would expect for a first-order phase transition. Figure \[fig:patterns\] show the numerically obtained minimizing patterns at various values of $\epsilon$. Note that, for sufficiently large $\epsilon$, the minimizers are the knee-solutions (\[eq:chevrons\]) with $a = 0$, whereas for sufficiently small $\epsilon$, the minimizers have convex-concave disclination pairs, and have $a \neq 0$.  \[fig:patterns\] [^2] [7]{} M. Abramowitz and I. Stegun. Handbook of Mathematical Functions, U.S. Govt. Printing Office, Washington, D.C., 1972. P. Aviles, Y. Giga, On Lower Semicontinuity of a Defect Energy obtained by a Singular Limit of the Ginzburg-Landau Type Energy for Gradient Fields, Proc. Roy. Soc. Edinburgh, **129A** (1999) 1-17. M.C. Cross, A.C. Newell, Convection Patterns in Large Aspect Ratio Systems, Physica D **10** (1984) 299-328. N.M. Ercolani, R. Indik, A.C. Newell, T. Passot, The Geometry of the Phase Diffusion Equation, J. Nonlinear Sci. **10** (2000) 223-274. N.M. Ercolani, R. Indik, A.C. Newell, T. Passot, Global Description of Patterns Far from Onset: A Case Study, Physica D **184** (2003) 127-140. N.M. Ercolani, M. Taylor, The Dirichlet-to-Neumann Map, Viscosity Solutions to Eikonal Equations, and the Self-Dual Equations of Pattern Formation, Physica D **196** (2004) 205-223. L. C. Evans. Partial Differential Equations, American Mathematical Society, Providence, RI, 2002. W. Jin, R. Kohn, Singular Perturbation and the Energy of Folds, J. Nonlinear Sci. **10** (2000) 355-390. E.T. Whittaker and G.N. Watson. A Course in Modern Analysis, Cambridge University Press, Cambridge, England, 1990. [^1]: *A Variational Theory for Point Defects in Patterns* [^2]: **Acknowledgements:** N. M. Ercolani was supported in part by NSF grant DMS-0073087; S.C. Venkataramani was supported in part by an NSF CAREER Award DMS–0135078.

ArXiv

--- abstract: 'We develop a first-principles theoretical approach to doping in field-effect devices. The method allows for calculation of the electronic structure as well as complete structural relaxation in field-effect configuration using density-functional theory. We apply our approach to ionic-liquid-based field-effect doping of monolayer, bilayer, and trilayer ZrNCl and analyze in detail the structural changes induced by the electric field. We show that, contrary to what is assumed in previous experimental works, only one ZrNCl layer is electrochemically doped and that this induces large structural changes within the layer. Surprisingly, despite these structural and electronic changes, the density of states at the Fermi energy is independent of the doping. Our findings imply a substantial revision of the phase diagram of electrochemically doped ZrNCl and elucidate crucial differences with superconductivity in Li intercalated bulk ZrNCl.' author: - Thomas Brumme - Matteo Calandra - Francesco Mauri bibliography: - 'ZrNCl.bib' title: | Electrochemical doping of few layer ZrNCl from first-principles:\ electronic and structural properties in field-effect configuration --- Introduction ============ In recent years, materials with reduced dimensionality have attracted a lot of attention because of their interesting physical properties and proposed applications range from electronics to sensing. In addition to graphene, which is probably the most studied two-dimensional (2D) material, monolayers or few-layer systems (nanolayers) of layered materials such as transition metal dichalcogenides[@radisavljevic2011; @ye2012; @yuan2013; @radisavljevic2013; @wu2013; @das2013_1; @das2013_2] or chloronitrides[@ye2010; @kasahara2010; @schurz2011; @zhangs2012; @ekino2013] are gaining importance because of their intrinsic band gap. The possibility of doping these few layer systems with field-effect transistors (FETs) is particularly appealing. In ionic-liquid-based FETs the charging of the nanolayers is substantial and thus allows for electrochemical doping of few-layer materials[@yuan2009; @ye2010; @ye2012; @zhang2012; @taniguchi2012; @yuan2013; @perera2013] with surface charges of the order of $10^{14}\:\mathrm{cm}^{-2}$. In the field of condensed matter physics, this prospect is very appealing as it becomes possible to dope ultrathin films of layered material and explore their phase diagram to detect charge-density waves, spin-density waves, or superconducting instabilities in reduced dimensionality. Furthermore, it was recently shown that the device characteristics of FETs using an ionic-liquid are much better than those of back-gated devices[@perera2013]. An eminent example is the electrochemical doping of the semiconducting chloronitride ZrNCl[@ye2010]. It has been shown by Ye [*et al.*]{} that $10-30\:\mathrm{nm}$ ZrNCl flakes can be made first metallic and then superconducting by electrochemical doping. A similar behavior has also been detected in MoS$_2$ flakes[@ye2012]. Despite these challenging experimental perspectives, the understanding of electronic and structural properties at high electric field in FET configuration is still limited. For example, the effects of the large electric field (in the order of $1\:\mathrm{V/nm}$[@ye2010; @ye2012]) on the structural properties of the nanolayer remains an open question. Moreover, in the case of ZrNCl, the surface charge, and consequently the effective doping, changes by an order of magnitude depending on the assumed thickness of the charge layer[@ye2010]. It is then crucial to address these issues to determine the phase diagram of electrochemically doped ZrNCl. We develop a first-principles theoretical approach to doping in field-effect devices. The method allows for calculation of the electronic structure as well as complete structural relaxation in field-effect configuration using density-functional theory. We apply our approach to ionic-liquid-based field-effect doping of monolayer, bilayer, and trilayer ZrNCl. The paper is organized as follows: in Sec. II we give a brief introduction to field-effect doping and a description of the model we use in order to investigate such systems from first-principles, along with the computational details. Results for the specific example of the electrochemical doping of ZrNCl are presented in Sec. III. First, we show the structural changes in field-effect configuration and then investigate the distribution of the induced charge in more detail. Afterwards we use a simple model to describe the modification of the band structure and show that, despite the structural and electronic changes, the density of states at the Fermi energy is independent of the doping. Finally, conclusions are presented in Sec. IV. Theoretical background ====================== Field-effect configuration -------------------------- In Fig. \[fig:fet\] the typical setup for a field-effect measurement is shown. The system is separated from the gate electrode by a dielectric such as silicon oxide. Applying a gate voltage between the system and the gate results in a change in the Fermi level of the metal ($E_F^G$) with respect to the one in the system ($E_F^S$). Hence, the dielectric is polarized due to the opposite charges at the dielectric/metal and dielectric/system interfaces. The distribution of charges in the system/dielectric/ metal-gate system generates the potential profile shown in Fig. \[fig:fet\]. Inside the dielectric, a constant and finite electric field (linear potential) occurs due only to the opposite signs of the charges on the different sides of the dielectric. At the system/dielectric interface (on the system side), the potential varies very quickly because of the strongly inhomogeneous charge distribution in the first few layers of the system. Finally, inside the system and far from the dielectric, the potential felt by the electrons oscillates with a periodicity that is related to the crystal periodicity and is determined by the ionic positions. In this region, the spatial average of the electric field is zero. The same situation occurs inside the metal gate far from the dielectric. In order to describe such a setup in theory, different methods have been used in the past. In perhaps the most advanced method one takes into account the connection of the system to the gate electrode and the back contact, effectively simulating an open-boundary system. This can be achieved by coupling the calculation of the electronic structure of the system—often done within density-functional theory (DFT)—with the Keldysh nonequilibrium Green’s function (NEGF) formalism[@liu2012; @gong2013]. However, these calculations are very expensive since the electrostatic potential of the system has to be determined self-consistently. This is achieved by solving the Poisson equation with the bulk-like Hartree potentials of the electrodes as boundary conditions on the interfaces between the electrodes and the system. In fact, by using NEGF, one is often restricted to use the structure which was relaxed without electric field, effectively neglecting the interplay of electric field, charging, electronic structure, and the geometry of the system[@liu2012; @gong2013]. Furthermore, these simulations of the full system/dielectric/metal-gate system are not only difficult and time consuming but also unnecessary, as we are mainly interested in what happens at the system/dielectric interface. We thus need to simplify the modeling of the FET measurements. The first simplification is to include the external electric field and eliminate the metallic gate. This amounts to replacing the metallic gate with a charged plate that simulates the charge occurring on the gated metal surface in contact with the dielectric[@gava2009; @uchida2009]. It is important to remark that, if the charged plate is neglected and only an external field is added, no charging can occur on the system side. After eliminating the gate, the next step is trying to reduce the size of the dielectric or to eliminate it completely. Since we expect that the relevant physics occurs inside the system in close contact with the dielectric, it is possible to replace the several nanometer thick-dielectric with a much thinner one. This is not an approximation, but is what is found in modern nanoelectronics, where, in order to increase the FET capacity, thinner and thinner dielectrics are used. Finally, the ultimate approximation is the complete removal of the dielectric, and including the charged plate close to the position of the system/dielectric interface. The total removal of the dielectric can be achieved by following the steps mentioned above, as long as the ions of the system are fixed. However, we also want to simulate the structural relaxation of the system in FET configuration, which means that we need to allow the ions to move. In this case an additional problem occurs. Having totally eliminated the dielectric, the ions can move too close to the charged plate which simulates the gate electrode (opposite charge on the plate and the system). This can be avoided by representing the dielectric with a constant potential barrier which effectively mimics the physical repulsion exerted by the dielectric. This is the approximation adopted in the present work. Up to now, we have introduced a simpler model of the system/dielectric/gate system that works with open boundary conditions. An additional conceptual step is needed in order to simulate the FET using periodic boundary conditions (PBC). With PBC, the charged plate interacts with the periodic image of the system generating a spurious electric field. This is unphysical, as in experiments far from the dielectric and inside the system or the metal gate the average electric field must be zero. One way to circumvent this issue is to include a mirror image of the system that forces the electric field to be zero between the periodic images[@uchida2009]. However this implies calculating a system with twice the number of atoms and results in an higher computational load. A smarter solution comes from electrostatics. It is indeed sufficient to add an additional dipole plate (two planes of opposite charge[@bengtsson1999]) generating an electric field that exactly cancels the one on the left side of the gate in Fig. \[fig:fet\]. In this case no additional computational load is needed and the condition of zero electric field far from the system/dielectric interface is enforced. Model {#monopole} ----- The general setup we use to model a system in FET configuration using PBC in the three spatial directions is shown in Fig. \[fig:pots\]. The nanolayers are placed in front of a charged plane[@gava2009; @uchida2009] which models the gate as already mentioned in the last section (henceforth referred to as “monopole”). The layers are then charged with the same amount of opposite charge $-n_\mathrm{dop}\,A$ ($n_\mathrm{dop}$ and $A$ are the number of doped electrons per unit area and the area of the unit cell parallel to the surface respectively) which will lead to a finite electric field in the region between the gate and the system. In the vacuum region between the repeated images (PBC) the electric field has to be zero in order to correctly determine the changes in the electronic structure and the geometry for such a field effect setup. However, the dipole of the charged system and the gate leads to an artificial electric field between the different slabs of the repeated unit cell since in PBC the electric field averaged over the unit cell is zero. We thus include an electric dipole generated by two planes of opposite charge[@neugebauer1992; @bengtsson1999; @meyer2001] which are in the vacuum region next to the monopole (Fig. \[fig:pots\]). We furthermore include a potential barrier to avoid the direct interaction between the charge density of the system and the monopole/dipole. The resulting total potential together with the definition of the different variables used in the following discussion is shown in Fig. \[fig:pots\]. The total energy of the system in this model-FET setup is given as $$\begin{aligned} \label{eq:etot} E_\mathrm{tot} &= E^\mathrm{per}_\mathrm{tot} + E_\mathrm{mono} + E_\mathrm{dip} + E_b.\end{aligned}$$ The first term appearing on the right-hand side of Eq. (\[eq:etot\]) is the total energy calculated within PBC using the periodic, bare ion-electron potential $V^i_\mathrm{per}\left(\mathbf{r}\right)$ $$\begin{aligned} \label{eq:eper} E^\mathrm{per}_\mathrm{tot} &= \sum_i\langle\psi_i\left|T\right|\psi_i\rangle+E_H+E_\mathrm{XC}\notag\\ &\quad+E_\mathrm{i-i}+\int_\Omega n\left(\mathbf{r}\right)V^i_\mathrm{per}\left(\mathbf{r}\right)d^3r.\end{aligned}$$ Here the first four terms represent the kinetic energy, Hartree energy, exchange-correlation energy, and Madelung energy, respectively, while the last term is the bare ion-electron energy with $n\left(\mathbf{r}\right)$ being the electron density. $E_\mathrm{mono}$ is the energy associated with the monopole while $E_\mathrm{dip}$ is the energy due to the dipole. The interaction with the potential barrier which is used to mimic the dielectric is given by the last term in Eq. (\[eq:etot\]), $E_b$. The total charge density in the cell $\Omega$ is the sum of the electronic part, the ionic part, and the monopole charge $$\begin{aligned} \label{eq:rhotot} \rho^\mathrm{tot}\left(\mathbf{r}\right)&=\rho^e\left(\mathbf{r}\right)+\rho^i\left(\mathbf{r}\right)+\rho^\mathrm{mono}\left(\mathbf{r}\right),\notag\\ &=-n\left(\mathbf{r}\right)+\sum_j Z_j\delta\left(\mathbf{r}-\mathbf{R}_j\right)\notag\\ &\quad-n_\mathrm{dop}\,\delta(z-z_\mathrm{mono}),\end{aligned}$$ where $\rho^e\left(\mathbf{r}\right)$, $\rho^i\left(\mathbf{r}\right)$, and $\rho^\mathrm{mono}\left(\mathbf{r}\right)$ are the electron, ion, and monopole charge densities, respectively. Furthermore, $Z_j$ is the (pseudo) atomic charge of atom $j$ at position $\mathbf{R}_j$, and the monopole with a total charge of $-n_\mathrm{dop}\,A$ per unit cell is located at $z=z_\mathrm{mono}$. Within PBC, the interaction energy associated to the presence of the monopole is given by $$\begin{aligned} \label{eq:emono} E_\mathrm{mono} &= -\int_\Omega\rho^\mathrm{tot}\left(\mathbf{r}\right)\,V_\mathrm{mono}\left(\mathbf{r}\right)d^3r,\end{aligned}$$ where $V_\mathrm{mono}$ can be written as $$\begin{aligned} \label{eq:potential} V_\mathrm{mono}\left(\mathbf{r}\right) &= -2\pi\,n_\mathrm{dop}\,\left(-\left|\overline{\mathrm{z}}\right|+\frac{\overline{\mathrm{z}}^2}{L}+\frac{L}{6}\right).\end{aligned}$$ Here $\overline{\mathrm{z}}=z-z_\mathrm{mono}$ with $\overline{\mathrm{z}}\in\left[-L/2;L/2\right]$ measures the distance to the monopole which is at $z_\mathrm{mono}$ and $L$ is the length of the periodically repeated unit cell along $\mathbf{\hat{z}}$ (see Fig. \[fig:pots\]). Notice that $V_\mathrm{mono}\left(\mathbf{r}\right)$ represents the potential generated in PBC by the monopole plane of total charge $-n_\mathrm{dop}$ placed at $z_\mathrm{mono}$ (the linear term) and by a uniform jellium density of opposite total charge $+n_\mathrm{dop}$ (the quadratic term). This jellium charge cancels the uniform background charge which is used in standard plane-wave *ab initio* codes in order to have a neutral system[@gava2009]. The last term in Eq. (\[eq:potential\]) is constant and it is conventionally chosen in order to have $\int V_\mathrm{mono}\left(\mathbf{r}\right) d^3r=0$. To eliminate the electrostatic interactions between repeated cells in the $z$ direction, we introduce a dipole formed by two opposite-charged planes at $z_\mathrm{dip}-d_\mathrm{dip}/2$ and $z_\mathrm{dip}+d_\mathrm{dip}/2$. The dipole should be placed between the monopole and the vacuum. This condition is realized by choosing $z_\mathrm{dip}+d_\mathrm{dip}/2=z_\mathrm{mono}-\epsilon$, where $\epsilon$ is a small positive number. In order to calculate the dipole energy $E_\mathrm{dip}$ we need to determine the electric dipole moment per unit surface $m$ of the full system ([*i.e.*]{}, using the sum of ionic, electronic, and monopole charge density)[@bengtsson1999; @meyer2001] $$\begin{aligned} \label{eq:dipole} m &= \int_\Omega\left[\frac{f\left(\tilde{z}\right)}{A}\,\rho^\mathrm{tot}\left(\mathbf{r}\right)\right]\,d^3r.\end{aligned}$$ Here $f\left(\tilde{z}\right)$ is a periodic function having different definitions for values of $z$ that are inbetween the two charged planes of the dipole or elsewhere, namely: $$\begin{aligned} \label{eq:f} f\left(\tilde{z}\right) &= \left\{ \begin{matrix} \tilde{z}-L/2 &\forall \tilde{z}\geq d_\mathrm{dip}/2\\ -\tilde{z}\left(L-d_\mathrm{dip}\right)/d_\mathrm{dip} &\forall \tilde{z}<d_\mathrm{dip}/2 \end{matrix} \right.\end{aligned}$$ with $\tilde{z}=z-z_\mathrm{dip}$, $\tilde{z}\in\left[-d_\mathrm{dip}/2;L-d_\mathrm{dip}/2\right]$ (and periodically repeated). The dipole energy is given by[@notedipc] $$\begin{aligned} \label{eq:Edip} E_\mathrm{dip} &= -\frac{1}{2}\,\int_\Omega\rho^\mathrm{tot}\left(\mathbf{r}\right)\,V_\mathrm{dip}\left(\mathbf{r}\right)d^3r,\end{aligned}$$ while the potential generated by the two planes of opposite charge can be written as $$\begin{aligned} \label{eq:dip_pot} V_\mathrm{dip}\left(\mathbf{r}\right) &= -\frac{4\pi m}{L}\,f\left(\tilde{z}\right).\end{aligned}$$ Note that in a self-consistent calculation the charge density changes with each iteration. Thus, the dipole moment per unit surface $m$ and the corresponding potential which represents the dipole have to be recalculated on each iteration until self-consistency is achieved. Finally, we place a barrier, mimicking the repulsion by the dielectric, to separate the system from the monopole plane. Such barrier, of height $V_0$ and thickness $d_b$, starts at $z_1=z_\mathrm{dip}-d_\mathrm{dip}/2$, in correspondence of the first plane of the dipole, and ends at $z_2=z_\mathrm{dip}-d_\mathrm{dip}/2+d_b$. The associated energy is $$\begin{aligned} \label{eq:ebarrier} E_b &= -\int_\Omega V_b\left(\mathbf{r}\right)\rho^\mathrm{tot}\left(\mathbf{r}\right)d^3r,\end{aligned}$$ where $V_b\left(\mathbf{r}\right)$ is a periodic function of $z$ defined on the interval $z\in[0,L]$ as $$\begin{aligned} \label{eq:vbarrier} V_b\left(\mathbf{r}\right) &= \left\{ \begin{matrix} V_0,&\forall z\in[z_1,z_2]\\ 0, &\forall z\notin[z_1,z_2] \end{matrix} \right..\end{aligned}$$ In the present implementation, we have chosen a linear transition from $V_b=0$ to $V_b=V_0$ within $d_\mathrm{dip}$ in order to minimize the fluctuation of the potential in $\mathbf{k}$ space. The final, full Kohn-Sham potential of the gated system can be written as $$\begin{aligned} \label{eq:vks} V_\mathrm{KS}\left(\mathbf{r}\right)&=\frac{\delta\left[E_\mathrm{tot}-\sum_i\langle\psi_i\left|T\right|\psi_i\rangle\right]}{\delta n\left(\mathbf{r}\right)}\notag\\ &= V^i_\mathrm{per}\left(\mathbf{r}\right) + V_H\left(\mathbf{r}\right) + V_\mathrm{XC}\left(\mathbf{r}\right)\notag\\ &\quad + V_\mathrm{mono}\left(\mathbf{r}\right) + V_\mathrm{dip}\left(\mathbf{r}\right) + V_b\left(\mathbf{r}\right),\end{aligned}$$ where $V^i_\mathrm{per}\left(\mathbf{r}\right)$ is the bare, periodic electron-ion potential, $V_H\left(\mathbf{r}\right)$ is the normal Hartree potential $$\begin{aligned} \label{eq:vh} V_H\left(\mathbf{r}\right) &= \int_\Omega\frac{n\left(\mathbf{r}'\right)}{\left|\mathbf{r}-\mathbf{r}'\right|} d^3r',\end{aligned}$$ $V_\mathrm{XC}\left(\mathbf{r}\right)$ is the exchange-correlation potential, and $V_\mathrm{mono}\left(\mathbf{r}\right)$, $V_\mathrm{dip}\left(\mathbf{r}\right)$, $V_b\left(\mathbf{r}\right)$, are the monopole, dipole, and potential barrier, given in Eqs. (\[eq:potential\]), (\[eq:dip\_pot\]), and (\[eq:vbarrier\]), respectively. The different parts of the Kohn-Sham potential are also shown in Fig. \[fig:pots\]. For the structural optimization we need to calculate the total force $\mathbf{F}^j$ on atom $j$. Using the Hellman-Feynman theorem and Eq. (\[eq:etot\]) the force can be written as[^1] $$\begin{aligned} \label{eq:ftotal} \mathbf{F}^j &= -\int\rho^e\left(\mathbf{r}\right)\nabla_{\mathbf{R}_j}V^i_\mathrm{per}\left(\mathbf{r}\right)d^3r-\nabla_{\mathbf{R}_j}E_{i-i}\notag\\ &\quad +Z_j\left(\frac{\partial V_\mathrm{mono}\left(\mathbf{R}_j\right)}{\partial\mathbf{r}}+\frac{\partial V_\mathrm{dip}\left(\mathbf{R}_j\right)}{\partial\mathbf{r}}\right.\notag\\ &\qquad\qquad\left.+\frac{\partial V_b\left(\mathbf{R}_j\right)}{\partial\mathbf{r}}\right).\end{aligned}$$ Using this equation we can now relax the system in an FET setup. We explicitly verified the force calculation by finite differences. Computational details {#DFT} --------------------- All calculations were performed using DFT with a modified version of the [PWscf]{} code of the [Quantum ESPRESSO]{} package[@quantumespresso]. We used the Perdew-Zunger local density approximation (LDA) exchange-correlation functional[@pz81] together with ultrasoft pseudopotentials including core corrections. We cross-checked our results with the Perdew-Burke-Ernzerhof (PBE) functional[@pbe96] including dispersion corrections[@grimme2006] (“+D2”). A plane-wave basis set was used to describe the valence electron wave function and density up to a kinetic energy cutoff of $45$ and $500\:\mathrm{Ry}$ ($1\:\mathrm{Ry}\approx13.6\:\mathrm{eV}$), respectively. The electronic eigenstates have been occupied with a Fermi-Dirac distribution, using an electronic temperature of $300\:\mathrm{K}$. The BZ integration has been performed with a Monkhorst-Pack grid[@monkhorst1976] of $64\times64\times1$ for the charged systems and $16\times16\times1$ $\mathbf{k}$ points for the neutral one. In order to correctly determine the Fermi energy in the charged system, we performed a non-self-consistent calculation on a denser $\mathbf{k}$-point grid of $90\times90\times1$ points starting from the converged charge density. The self-consistent solution of the Kohn-Sham equations was obtained when the total energy changed by less than $10^{-9}\:\mathrm{Ry}$ and the maximum force on all atoms was less than $5\cdot10^{-4}\:\mathrm{Ry}\:a_0^{-1}$ ($a_0\approx0.529177\:{\textup{\AA}}$ is the Bohr radius). A tight force convergence threshold was necessary since the force on the atoms due to the charged plane representing the gate can be as small as $10^{-2}\:\mathrm{Ry}\:a_0^{-1}$ for the lighter nitrogen atoms. Using the experimental lattice parameters of bulk $\beta$-ZrNCl as starting geometry[@shamoto1998], we relaxed the unit cell for the bulk system until the total stress on the system was less than $0.1\:\mathrm{kbar}$. The lattice parameters thus determined agree well with the experimental values: for the in-plane lattice constant we found a value of $a=3.5622\:{\textup{\AA}}$ (Exp.: $a=3.5974\:{\textup{\AA}}$) and for the perpendicular constant we calculated $c=27.035\:{\textup{\AA}}$ (Exp.: $a=27.548\:{\textup{\AA}}$). The deviation of about 1% is in the range of variations typically found for LDA calculations. Using PBE+D2 the deviations can be decreased to less than 1%, but the convergence for the layered-2D systems is slower, especially if ZrNCl is close to the position $z_0$ of the monopole potential. Since we are mainly interested in the general behavior of the system under the influence of a gate voltage we continue to use the LDA. We cross-checked our results for selected gate voltages using PBE+D2 and got the same general behavior with only small changes. The final geometry of the bulk system was used as the starting geometry for the calculations of the layered-2D systems fixing the size of the unit cell in plane and increasing the perpendicular size such that the vacuum region between the repeated images was at least $20\:{\textup{\AA}}$. We investigated single-, double-, and triple-layer systems with different doping levels ranging from $0.05$ to $0.35$ electrons per unit cell ([*i.e.*]{}, $0.025$ to $0.175$ electrons per Zr atom since we are using the hexagonal unit cell) which corresponds to a doping charge per unit surface ranging from $n_\mathrm{dop}\approx4.55\cdot10^{13}\:\mathrm{cm^{-2}}$ to $n_\mathrm{dop}\approx3.18\cdot10^{14}\:\mathrm{cm^{-2}}$. The dipole was placed at $z_\mathrm{dip}=d_\mathrm{dip}/2$ with $d_\mathrm{dip}=0.01\:L$ and $L\approx35\:{\textup{\AA}}$ for single- and double-layer ZrNCl, and $L\approx46\:{\textup{\AA}}$ in the trilayer case. A potential barrier with an height of $V_0=1.5\:\mathrm{Ry}$ and a width of $d_b=0.1\:L$ was used in order to prevent the ions from moving too close to the monopole. The final results were found to be independent of the separation of the dipole planes, as well as the barrier height and width as long as it is high or thick enough to ensure that $\rho^e\left(z_\mathrm{mono}\right)=\rho^e\left(z_\mathrm{dip}\right)=0$. Results ======= Structural changes under electrochemical doping {#geometry} ----------------------------------------------- In the relevant experiment of Ref.  on ZrNCl an ionic liquid was used as dielectric and the gate voltage $V_G$ was applied between the ionic liquid and ZrNCl. The mobile cation and anion of the ionic liquid move towards the oppositely charged electrode which in turn leads to the formation of an electric double layer (EDL) at the interface to ZrNCl. The EDL at the ZrNCl/ionic liquid interface acts as a capacitor with a huge capacitance (“supercapacitor”) and can accumulate charges in the transport channel between source and drain. Using such an electric double-layer transistor one can achieve large carrier densities up to $n_\mathrm{dop}=8\times10^{14}\,\mathrm{cm^{-2}}$ (see Ref. ) thus allowing the investigation of the transition from a band insulator to a metal and, eventually, to a superconductor[@ye2010]. The thickness of the EDL is mainly determined by the size of the ions, the microscopic structure of the surface, and the interaction between both. In this work we simulate the gate with the monopole potential (Eq. (\[eq:potential\])) as shown in Fig. \[fig:pots\], which has a fixed position $z_\mathrm{mono}$. We thus need to relax the nanolayers towards the charged plane of the monopole potential until the attractive force between the two oppositely charged systems is compensated by the repulsive force due to the electron density within the potential barrier, $V_b$. In this section we analyze the final geometry with increasing doping ([*i.e.*]{}, increasing strength of the monopole potential) in detail. Figure \[fig:distdef\] shows the atomic structure of a double-layer ZrNCl together with the definition of the atoms which are used in the following discussion and their position with respect to the interface between system and ionic liquid. The bonding between the neutral layers of ZrNCl is very weak[@heid2005; @akashi2012] which allows for the intercalation with alkaline elements or even molecules[@kasahara2010]. Thus, the interaction between ZrNCl and the monopole potential representing the gate could in principle strongly change the interlayer distance depending on the distribution of the doping charge. In our calculations however, we find that the interlayer distance is nearly unchanged. Instead the thickness of the nanolayers system is increased, due mainly to the change in the Cl-Zr distance for the chlorine closest to the gate, while the other distances are nearly constant as can be seen in Fig. \[fig:Cldist\]. The interlayer distance in the bilayer case changes by less than 1% compared to the value of bulk ZrNCl. Triple-layer ZrNCl shows the same general behavior—we find an increase of about 1% in the distance between the middle layer and the one next to the charged plane from $3.537\:{\textup{\AA}}$ for $n_\mathrm{dop}\approx9\cdot10^{13}\:\mathrm{cm^{-2}}$ to $3.576\:{\textup{\AA}}$ for $n_\mathrm{dop}\approx27\cdot10^{13}\:\mathrm{cm^{-2}}$ (bulk value $3.543\:{\textup{\AA}}$). The absolute value of these interlayer distances, in principle, depends on the correct description of the van der Waals interactions between the layers and, thus, might be questionable. The most important result is that for all investigated systems the major changes in the geometry occur in the layer next to the monopole (see Fig. \[fig:distdef\]). This is also the case for the different Zr-N bond lengths, and angles of the Zr-N rhomboid. In the double- and triple-layer system relevant changes occur only in the layer next to the monopole. As the relaxed geometry of the ZrNCl layer closer to the monopole are very similar for the single-, double-, and triple-layer case, we only present the results of the structural minimization of a single layer (see Fig. \[fig:ZrNdist\_angles\]). Due to the large difference between the electronegativity of nitrogen and zirconium, the nitrogen is partially negative. Accordingly, N is attracted by the positively charged plane representing the gate. This leads to an increase of the bond length labeled “Zr 2 - N 1” which is the bond with the nitrogen closer to the monopole as can be seen in Fig. \[fig:ZrNdist\_angles\](a). On the other hand, the bond length in which the zirconium is closer to the monopole is decreased (“Zr 1 - N 2”). Since the in-plane bonds are nearly unchanged the N-Zr-N angles are modified considerably as can be seen in Fig. \[fig:ZrNdist\_angles\](b). These changes in the geometry of the layer next to the monopole are due to the fact that the doping charge is mainly localized within this layer as we will see in the next section. Electronic structure -------------------- As shown in the last section, the structure of the layer closest to the monopole is strongly affected by the electric field. This is due to the interplay between the electric field, the doping charge, and the specific band structure of ZrNCl. The conduction band minimum is located at the $\mathrm{K}$ and $\mathrm{K}'$ point[@akashi2012; @sugimoto2004; @heid2005] and has a nearly quadratic in-plane dispersion $E(\mathbf{k})\propto k^2_x+k^2_y$. Furthermore, the band consists mainly of Zr $d$-states which together with the weak interlayer interaction leads to a negligible dispersion along $k_z$. Thus, it can be expected that the doping charge is mainly localized within the center of the layers closest to the ionic liquid. Figure \[fig:charge\_ildos\](a) shows the distribution of the conduction band electrons in bilayer ZrNCl averaged over the directions parallel to the layers for three doping levels ranging from $n_\mathrm{dop}\approx9\cdot10^{13}\:\mathrm{cm^{-2}}$ to $n_\mathrm{dop}\approx27\cdot10^{13}\:\mathrm{cm^{-2}}$. The doping charge occupying the conduction band is mainly localized in the center of the layer closest to the monopole and has a strong asymmetry between the two Zr-N planes on either side. This asymmetry is due to the electric field and the induced structural changes. The field induce not only an asymmetry in the geometry and the distribution of the conduction band electrons but also an increased surface dipole $m$. In Fig. \[fig:charge\_ildos\](b) the difference between the full charge density of the doped system $\rho^n_{||}$ and the undoped system $\rho^0_{||}$ is shown, where $\rho^0_{||}$ was calculated with the relaxed geometry of the charged one. This charge difference $\left(\rho^n_{||}-\rho^0_{||}\right)$ represents the screening charge of the gate electric field, including both the metallic screening of the conduction band and the dielectric screening of the lower, occupied valence bands which rearrange under the applied electric field. The field has two main effects as can be seen in Fig. \[fig:charge\_ildos\]: firstly, it leads to an asymmetry in the charge distribution and consequently in the geometry as discussed in the last section, and secondly, it increases the surface dipole—not only due to the structural changes but also due to the rearrangement of the charge density of the undoped system $\rho^0_{||}$. In fact, the induced dipole is largest close to the last chlorine (Fig. \[fig:charge\_ildos\](b)), even if the doping charge in the conduction band is localized within the center of the layer (Fig. \[fig:charge\_ildos\](a)). The strong localization of the doping charge within the first layer also leads to a complete screening of the electric field. Figure \[fig:pot\] shows the difference of the planar averaged potential $U\left(\mathbf{r}\right)=V_\mathrm{KS}\left(\mathbf{r}\right)-V_\mathrm{XC}\left(\mathbf{r}\right)$ of the charged and the uncharged system $\Delta U\left(\mathbf{r}\right)=U^n\left(\mathbf{r}\right)-U^0\left(\mathbf{r}\right)$, the derivative of which being proportional to the effective electric field $\partial\Delta U\left(\mathbf{r}\right)/\partial z\propto E_z$. The slope of the potential close to the potential barrier at $z\approx3\:{\textup{\AA}}$ increases with increasing doping and the total potential changes mainly within the layer of ZrNCl which is closer to the monopole. The largest change and, thus, the largest electric field can be found close to the outermost chlorine atom. The band structure for different number of layers is shown in Fig. \[fig:bands\]. Since the doping charge is localized within the layer close to the ionic liquid/system interface, only the electronic structure of this layer is influenced and, accordingly, the degeneracy of the conduction band at the $\mathrm{K}$ and $\mathrm{K}'$ point is lifted. The bands localized on the doped layer are shifted down in energy while those of the undoped layers are unchanged as can be seen in Figs. \[fig:bands\](b) and (c) for bilayer and trilayer ZrNCl, respectively. Also the former valence band is nearly unchanged despite the geometrical changes ([*cf.*]{}, Figs. \[fig:Cldist\] and \[fig:ZrNdist\_angles\]). Aligning the band structure for different number of layers to the minimum of the conduction band $E^c_0$ as in Fig. \[fig:bands\_close\] demonstrates again that there is effectively no difference between the monolayer/bilayer/trilayer systems. Thus, in this section we focus in the following on the single-layer system and analyze the variations of the partially occupied band on doping in more detail. Charging of the monolayer does not simply act as a rigid doping as it also affects the shape of the band. In order to describe the changes in the band structure quantitatively we model the dispersion relation $E(\mathbf{k})$ near the $\mathrm{K}$ and $\mathrm{K}'$ points by a 4^th^ order Taylor expansion in $\mathbf{k}$, compatible with the symmetry of the bands: $$\begin{aligned} \label{eq:dispersion} E\left(\mathbf{k}\right) &= E^c_0 + \frac{\hbar^2}{2\,m^*}\,\left(k_x^2 + k_y^2\right) + B\,\left(k_x^2 + k_y^2\right)^2\notag\\ &\quad+ C\,\left(k_x^3 - 3\,k_x\,k_y^2\right).\end{aligned}$$ Here $k_x$ is aligned with the $\Gamma\rightarrow\mathrm{K}$ direction and $|\mathbf{k}|=k$ measures the distance from the $\mathrm{K}$, $\mathrm{K}'$ special points. The $k^3$ term is needed to describe the trigonal warping of the conduction band while both the 3^rd^ the 4^th^ order term are necessary to allow for the deviations of the density of states from the constant energy dependence of a perfect 2D electron gas. The parameters $m^*$, $B$, and $C$ obtained by fitting Eq. (\[eq:dispersion\]) to the band structure of single-layer ZrNCl for three different doping levels are depicted in Fig. \[fig:parameters\]. Most interestingly, the effective mass $m^*$ changes by 20% in the doping regime investigated here, which would lead to the same change in the density of states (DOS) for a perfect 2D electron gas which can be modeled using just the first two terms in Eq. (\[eq:dispersion\]). Yet the parameters $B$ and $C$, which describe the deviations from a quadratic dispersion for higher values of $k$, change considerably when the doping is increased. The decrease from about $-22\:\mathrm{eV\,a_0^4}$ to $-69\:\mathrm{eV\,a_0^4}$ for $B$ and from about $-17\:\mathrm{eV\,a_0^3}$ to $-24\:\mathrm{eV\,a_0^3}$ for $C$ compensate the change in the effective mass $m^*$. Accordingly the DOS at the Fermi energy stays more or less constant as can be seen in Fig. \[fig:dos\]. In fact, specific heat measurements on Li intercalated ZrNCl suggested that the DOS at the Fermi energy is not increased with increasing doping[@kasahara2009], which cannot be explained assuming a rigid doping since the DOS is not flat ([*cf.*]{}, Refs.  and Fig. \[fig:dos\]). In our simulated electric double-layer transistor (EDLT) the DOS at the Fermi energy is constant since the band structure is changed upon doping. However, does this picture hold for bulk Li$_x$ZrNCl? As a first attempt to simulate the Li intercalation in the bulk system with our model, we substitute the Li$_x$ plane in bulk Li$_x$ZrNCl ([*i.e.*]{}, periodic in all three dimension) with a charged plane as described in Eq. (\[eq:potential\])—a sketch of the model is shown in Fig. \[fig:symmetricmodel\]. Similarly to the EDLT setup we find a decreasing effective mass even if the decrease is smaller as can be seen as green (light gray) line in Fig. \[fig:mass\](a) labeled “symmetric”. Since a rigid structure in field-effect configuration also shows this behavior (red (grey) curve in Fig. \[fig:mass\](a) labeled “EDLT without relaxation”), it can neither be related to the EDLT setup nor to the field-induced structural changes. Instead, it might be a system-specific property. However, the decreasing effective mass is again compensated by the change in the parameters $B$ and $C$ as can be seen in Fig. \[fig:mass\](b). Further increasing the doping would ultimately lead to an increasing DOS at the Fermi energy for all investigated systems. Finally, we also note that the maximum possible doping is smaller if the nanolayer system is not relaxed in the presence of the electric field generated by the potential in Eq. (\[eq:potential\]). This shows again the importance of the structural relaxation for the correct theoretical description of ZrNCl in an FET setup. Implications for superconductivity ================================== Electrochemical doping of ZrNCl is appealing as an insulator-superconductor transition occurs as a function of charging[@ye2010], as shown in Fig. \[fig:phase\_diagram\]. The critical doping of the insulator-superconductor transition and the behavior of the superconducting critical temperature as a function of doping depends crucially on the depth of the induced charge layer in the sample. Indeed, if the induced charge layer is deeper, then for a given gate voltage the charge density per layer (and thus the doping) is smaller. Thus knowledge of the charge distribution in the ZrNCl flakes is a crucial parameter to determine the phase diagram. While the gate voltage and the total induced charge (Hall effect) are accessible quantities, the depth of the induced charge layer is not. Thus experimentalist are forced to use simplified screening models to access this quantity. In the case of electrochemically doped ZrNCl, Ye [*et al.*]{} in Ref.  assumed that the induced charge is localized on two ZrNCl layers. Under this assumption, they obtained the phase diagram in Fig. \[fig:phase\_diagram\], showing a similar behavior for Li-intercalated and electrochemically doped ZrNCl. However, we have shown in the last section that this assumption is not well grounded, as the depth of the induced charge layer is only one ZrNCl layer, and not two. Furthermore we have shown that the charge distribution inside the outermost ZrNCl layer is strongly inhomogeneous. These findings imply a redrawing of the ZrNCl phase diagram, as shown in Fig. \[fig:phase\_diagram\]. In particular, in the revised phase diagram of electrochemically doped samples, superconductivity occurs at doping larger then $0.12$ and T$_c$ [*increases*]{} monotonically from a minimum value of $14\:\mathrm{K}$ up to a maximum of $\approx 15\:\mathrm{K}$ at a doping of $\approx 0.15$. This is in striking contrast with chemically intercalated bulk ZrNCl samples, in which superconductivity occurs at a doping of $0.05$ with T$_c=15.5\:\mathrm{K}$. Moreover, in the bulk case, T$_c$ [*decreases*]{} monotonically as a function of doping, as show in Fig. \[fig:phase\_diagram\]. Thus, our work points out crucial differences in superconductivity in Li intercalated bulk ZrNCl. Further calculations for, [*e.g.*]{}, the transition temperature are beyond the scope of this paper and will be addressed in a separate work. Conclusions =========== In this work we developed a first principles approach to field-effect doping. The method allows for the calculation of electronic structure and structural optimization under an applied electric field in an FET configuration. We have applied the method to the transition-metal chloronitride ZrNCl and have shown that the electric field induces substantial deformation of bond lengths and angles in the outermost ZrNCl layer in contact with the ionic liquid. The most evident deformation generated by the electric field is the large change of the the N-Zr-N angles. The electronic structure is also affected by the large electric field as the effective mass of the electron is substantially reduced as a function of increasing charge. The density of states at the Fermi level is, however, essentially independent on doping. Furthermore, we have shown that the depth of the induced charge layer is only one ZrNCl layer and that the charge distribution inside the outermost ZrNCl layer is strongly inhomogeneous. These findings imply a substantial revision of the phase diagram of electrochemically doped ZrNCl and elucidate crucial differences with superconductivity in Li intercalated bulk ZrNCl. The authors acknowledge financial support of the Graphene Flagship and of the French National ANR funds within the *Investissements d’Avenir programme* under reference ANR-11-IDEX-0004-02, ANR-11-BS04-0019 and ANR-13-IS10-0003-01. Computer facilities were provided by CINES, CCRT and IDRIS (project no. x2014091202). [^1]: Notice that if the barrier potential $V_b\left(\mathbf{r}\right)$ is correctly chosen, the nuclei never enter the barrier region and the last term of Eq.(\[eq:ftotal\]) is always equal to zero. The repulsive force on the system is mediated by the tail of the electron distribution within the barrier region. Such repulsion is included in the terms of Eq.(\[eq:ftotal\]) which implicitly or explicitly depend on the electron charge density $\rho^e\left(\mathbf{r}\right)$.

ArXiv

--- abstract: 'We report the first detection of a water megamaser in a radio-loud galaxy, 3C403, and present a follow-up study using the VLA. 3C403 has been observed as a part of a small sample of FRII galaxies with evidence of nuclear obscuration. The isotropic luminosity of the maser is $\sim$ 1200 . With a recessional velocity of c[*[z]{}*]{} $\sim$ 17680 kms$^{-1}$ it is the most distant water maser so far reported. The line arises from the densest ($>$ 10$^{8}$ cm$^{-3}$) interstellar gas component ever observed in a radio-loud galaxy. Two spectral features are identified, likely bracketing the systemic velocity of the galaxy. Our interferometric data clearly indicate that these arise from a location within 0.1$^{\prime\prime}$ ($\approx$ 110 pc) from the active galactic nucleus. We conclude that the maser spots are most likely associated with the tangentially seen parts of a nuclear accretion disk, while an association with dense warm gas interacting with the radio jets cannot yet be ruled out entirely.' author: - 'A.' - 'C.' - 'M.' - 'K. M.' - 'A.' - 'L.' title: 'Probing the obscuring medium around active nuclei using masers: The case of 3C403' --- INAF-Osservatorio Astronomico\ Loc. Poggio dei Pini, Strada 54\ 09012 Capoterra (CA)\ Italy Introduction ============ So far, water megamasers have been detected in radio-quiet active galactic nuclei (AGN; for a definition of radio-quiet AGN, see e.g. ), mostly in Seyfert2 and LINER galaxies. Under the assumption that all megamasers are associated with molecular material orbiting around the central engine (e. g. ) or interacting with the nuclear jet(s) of the host galaxy (e. g. ) and that their amplification is unsaturated (i.e. the maser intensity grows linearly with the background radio continuum), one should expect a much higher detection rate in radio-loud AGN than is actually observed. This fact should be especially true for the radio-loud AGN classified as narrow-lined FRIIs for which the AGN unification scheme (e.g. ) requires the presence of geometrically and optically thick obscuring structure. Samples of radio galaxies belonging to the broad-lined and narrow-lined FRII class have been recently observed with the Effelsberg telescope (Tarchi et al., in prep; Lara et al., priv. comm.). As in the case of FRIs [@henkel98], no maser detections have been obtained. Sample selection ================ Our sample comprises all nearby ($z<0.1$) 3 FRIIs spectrally classified as High Excitation Galaxies (HEGs, ) with nuclear equivalent widths of the \[O[iii]{}\]$\lambda$5007 emission line EW(\[O[iii]{}\])$>10^{4}$ [Å]{} (Fig.\[chiab\]). A high value for the nuclear EW(\[O[iii]{}\]) in HEGs has been interpreted as a hint for the obscuration of the central ionizing continuum source [@chiaberge02]. In these sources the nuclear ionizing continuum would be obscured to our line-of-sight and only a small fraction of the emission is seen through scattered light. Therefore, our selection criteria provide us with a sample of galaxies with both high radio flux densities and nuclear obscuration of the central ionizing source. ![Equivalent width of the \[O[iii]{}\] emission line, measured with respect to the central compact core emission, is plotted vs. the ratio between the optical central compact core to radio core flux (for details, see also Chiaberge et al., 2002). LEG: Low excitation galaxy; HEG: High excitation galaxy. The targets of our sample with their respective names are those within circles.[]{data-label="chiab"}](heg_past.eps){width="20pc"} Observations and data reduction =============================== Observations of the $6_{16} - 5_{23}$ transition of H$_2$O (rest frequency: 22.235 GHz) were carried out with the 100-m telescope of the MPIfR at Effelsberg[^1] in January and March 2003. The beam width (HPBW) was 40$^{\prime\prime}$. Flux calibration was obtained by measuring W3(OH) (see ). Gain variations of the telescope as a function of elevation were taken into account (Eq.1 of ). The pointing accuracy was better than 10$^{\prime\prime}$. The follow-up VLA[^2] A-array observations of the detected maser in 3C403 were performed in July 2003 with two IFs and a bandwith of 12.5 MHz each centered on one of the two maser features. Using 32 spectral channel a velocity resolution of $\sim$ 6kms$^{-1}$ was reached. The beam width (HPBM) was $\sim$ 0.1$^{\prime\prime}$ and the total on-source observation time was about $\sim$ 8 hours. All data were reduced using standard procedures belonging either to the GILDAS or the AIPS software packages. Results ======= ![Maser lines in 3C403. [*Upper panel*]{} Spectral resolution of 78 kHz = 1.15kms$^{-1}$. [*Lower panel*]{} Spectral resolution of 1.25 MHz = 17.8kms$^{-1}$. The arrow marks the nominal systemic velocity of the galaxy = 17688kms$^{-1}$.[]{data-label="2spec"}](2spec_mod.eps){width="20pc"} The maser spectra of 3C403, taken in January with the Effelsberg telescope, are shown in Fig.\[2spec\]. The profile is composed of two main components (asymmetrically) bracketing the nominal systemic velocity of the galaxy (c$z$ = 17688kms$^{-1}$, see next section): the stronger one has a velocity of c$z$ = 17827$\pm$1kms$^{-1}$, a width of 31$\pm$2kms$^{-1}$, and a flux density peak of 23$\pm$3mJy; the weaker one has a velocity of c$z$ = 17644$\pm$5kms$^{-1}$, a width of 53$\pm$8kms$^{-1}$, and a peak flux density of 4.0$\pm$0.5mJy. Using a distance of 235Mpc ($H_{\rm 0}$ = 75kms$^{-1}$Mpc$^{-1}$), the total isotropic luminosities are 950$\pm$140 and 280$\pm$55 for the main components, respectively. Hence, the maser in 3C403 is the most distant and one of the most luminous water maser so far reported (the most luminous maser, with 6000, and, before this discovery also the most distant one, is in TXFS2226-184; ) The outcome of the VLA observations (Fig.\[vla\]; right-hand side) confirms the presence of the two main maser features, and indicates that the maser emission is unresolved and arise from a location within 0.1$^{\prime\prime}$ ($\approx$ 110 pc) from the active galactic nucleus. The strength of the weaker line is consistent with the single-dish result, while the stronger line is much weaker than that obtained with Effelsberg. Because of the unresolved nature of the emission it is unlikely that the loss of flux density in the feature is due to extended emission resolved-out because of the higher resolution. More probably, the cause is a flare-down of the component due to strong variability, a phenomenon already well known to exist in water masers. ![[*Left-hand side*]{} VLA 8.4 GHz maps in grey-scale (top) and contours (bottom) of 3C403 at a resolution of $\sim$ 2.5$^{\prime\prime}$ (FWHM). For details, see Black et al. (1992). [*Right-hand side*]{} The two main water maser features in 3C403 observed with the VLA A-array. Both lines arise from a location within 0.1$^{\prime\prime}$ ($\approx$ 110 pc) from the active galactic nucleus (RA${\rm _{B1950}}$ = 19$^{\rm h}$ 52$^{\rm m}$ 15.8$^{\rm s}$; Dec${\rm _{B1950}}$ = 02$^{\rm \circ}$ 30$^{\rm \prime}$ 24$^{\rm \prime\prime}$).[]{data-label="vla"}](perzwolle.eps){width="30pc"} Discussion and conclusions ========================== The standard unified scheme of AGN requires an obscuring region, possibly containing molecular gas that surrounds the central engine and that effectively shields the inner few parsecs from view, if the radio axis lies close to the plane of the sky [@antonucci93]. In the innermost part, at radii up to some tenths of a parsec, this material is likely to form a rapidly rotating accretion disk around a central supermassive black hole. At larger distances (up to about 50-100 pc) the atomic and molecular gas is possibly distributed in a toroidal structure providing obscuration of the central regions to particular lines-of-sight. Indeed, very few direct detections of molecular gas in radio-loud galaxies have been reported so far ($\rm H_{2}$ in CygnusA: ; CO in 3C293: ). The detection discussed here strongly favors the presence of molecular material near the central engines of at least some FRIIs. Past negative results in molecular line surveys may also be a consequence of observational sensitivity limits (such a possibility was also mentioned by ). Furthermore, because of the high gas densities required for H$_{2}$O masers to operate ($>$10$^{8}$cm$^{-3}$; e. g. ), our detection represents the densest interstellar gas component ever observed in a radio-loud galaxy (typical values for molecular gas densities range between $\sim 10^{3}$ and $ \sim 10^{5}$ cm$^{-3}$; e. g. ). Accretion disk or jet interaction? ---------------------------------- Because of the large uncertainty of the systemic velocity[^3] ($V_{\rm sys}$) of 3C403, the following discussion and (preliminary) conclusions are based on the assumption that $V_{\rm sys}$ is placed (as indicated in Fig.\[2spec\]) between the two main H$_{2}$O maser components. As mentioned in Sect.1, interferometric studies of H$_2$O megamasers have shown that the emission is either associated with a nuclear accretion disk (for NGC4258, see e.g. ) or with the radio-jets interacting with dense molecular material near the center (for Mrk348, see ). The interferometric observation performed with the VLA in its A configuration confirms that also the megamaser in 3C403 has a nuclear origin. Only VLBI interferometric observations at milliarcsecond resolution (spatial scales of $\approx$ 1 parsec) will allow us to determine (or to provide an upper limit to) the extent of the emission and to pinpoint the exact location of the H$_2$O emitting region(s) in order to see if the maser is associated with an accretion disk or radio jets. Nevertheless, a qualitative discussion is possible on the basis of the single-dish spectra of Fig. \[2spec\] and the new VLA observation. [*[Accretion disk]{}*]{}: this scenario is particularly supported by the expected almost edge-on orientation of the nuclear obscuring layer (see ). As in the case of the Seyfert galaxy NGC 4258 (e.g. ), the spectrum should then show three distinct groups of features: one centered at the systemic velocity of the galaxy (originating along the line of sight to the nucleus) and two groups symmetrically offset from the systemic velocity, arising from those parts of the disk that are viewed tangentially. If the latter are the two lines we are observing, the rotational velocity of the disk is $\sim$100kms$^{-1}$ (which is, we have to point out, very small when compared with that derived for NGC4258). In 3C403 the systemic lines seem instead to be missing. To explain this fact, partly following the hypothesis proposed by , we could argue that the circumnuclear disk is actually a [*[thin rotating ring]{}*]{} of the type modeled by @ponomarev94 (1994; their Fig.2), where masing gives a double-peaked profile only. [*[Jet interaction]{}*]{}: The handful of bright knots visible in the radio images shown in Fig.\[vla\] (left-hand side) hints at the presence of jets interacting in several regions with the interstellar medium. The profile of the spectrum does not contradict a ‘jet-origin’ of the detected maser emission in 3C403. If we assume a symmetric molecular distribution, the two observed features could be interpreted as the red-shifted and blue-shifted counterparts of the maser line, arising from opposite jets close to the core. Antonucci, R. 1993, [*ARA&A*]{}, [**31**]{}, pp. 473 Baum, S.A., Heckman, T. and van Breugel, W. 1990, [*ApJS*]{}, [**74**]{}, pp. 389 Black, A.R.S., Baum, S.A., Leahy, J.P., et al. 1992, [*MNRAS*]{}, [**256**]{}, pp. 186 Chiaberge, M., Capetti, A. and Celotti, A. 2002, [*A&A*]{}, [**394**]{}, pp. 791 Claussen, M.J., Diamond, P.J., Braatz, J.A., Wilson, A.S. and Henkel, C. 1998, [*ApJ*]{}, [**500**]{}, pp. L129 Elitzur, M., Hollenbach, D.J. and McKee, C.F. 1989, [*ApJ*]{}, [**346**]{}, pp. 983 Evans, A.S., Sanders, D.B., Surace, J.A. and Mazzarella, J.M. 1999, [*ApJ*]{}, [**511**]{}, pp. 730 Gallimore, J.F., Henkel, C., Baum, S.A., et al. 2001, [*ApJ*]{}, [**556**]{}, pp. 694 Henkel, C., Wang, Y.P, Falcke, H., Wilson, A.S. and Braatz, J. A. 1998, [*A&A*]{}, [**335**]{}, pp. 463 Jackson, N. and Rawlings, S. 1997, [*MNRAS*]{}, [**286**]{}, pp. 241 Kellermann, K.I., Sramek, R., Schmidt, M., Shaffer, D.B. and Green, R. 1989, [*AJ*]{}, [**98**]{}, pp. 1195 Koekemoer, A.M., Henkel, C., Greenhill, L.J., et al. 1995, [*Nature*]{}, [**378**]{}, pp. 697 Mauersberger, R., Wilson, T. L. and Henkel, C. 1988, [*A&A*]{}, [**201**]{}, pp. 123 Miyoshi, M., Moran, J., Herrnstein, J., et al. 1995, [*Nature*]{}, [**373**]{}, pp. 127 Morganti, R., Oosterloo, T.A., Tadhunter, C.N., et al. 2001, [*MNRAS*]{}, [**323**]{}, pp. 331 Peck, A.B., Henkel, C., Ulvestad, J.S., et al. 2003, [*ApJ*]{}, accepted, \[astro-ph/0303423\] Ponomarev, V.O., Smith, H.A., Strelnitski, V.S. 1994, [*ApJ*]{}, [**424**]{}, pp. 976 Urry, C.M. and Padovani, P. 1995, [*PASP*]{}, [**107**]{}, pp. 803 Wilman, R.J., Edge, A.C., Johnstone, R.M., Crawford, C.S. and Fabian, A.C. 2000, [*MNRAS*]{}, [**318**]{}, pp. 1232 [^1]: The 100-m telescope at Effelsberg is operated by the Max-Planck-Institut f[ü]{}r Radioastronomie (MPIfR) on behalf of the Max-Planck-Gesellschaft (MPG). [^2]: The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. [^3]: Velocities derived from optical emission lines may be uncertain or biased by motions of the emitting gas (e.g. ). From the rotation curve measured by , we deduce that these uncertainties in 3C403 are $<$ 100kms$^{-1}$.

ArXiv

--- abstract: 'We study the tolerance to congestion failures in communication networks with scale-free topology. The traffic load carried by each damaged element in the network must be partly or totally redistributed among the remaining elements. Overloaded elements might fail on their turn, triggering the occurrence of failure cascades able to isolate large parts of the network. We find a critical traffic load above which the probability of massive traffic congestions destroying the network communication capabilities is finite.' author: - Yamir Moreno - 'Romualdo Pastor-Satorras' - 'Alexei V[á]{}zquez' - Alessandro Vespignani title: 'Critical load and congestion instabilities in scale-free networks' --- Complex heterogeneous connectivity patterns have been recently identified in several natural and technological networks [@strogatz; @bara02; @doro]. The Internet and the World-Wide-Web (WWW) networks, where nodes represent routers or web pages and edges physical connections or hyper-links, appear to have a topology characterized by the presence of “hubs” with many connections to peripherical nodes. Empirical evidence recently collected shows that this distinctive feature finds its statistical characterization in the presence of heavy-tailed degree distributions [@bara00; @fal; @broder00; @gov00; @gcalda; @gov01; @caida; @romu01]. In the Internet, for instance, the statistical analysis reveals that the degree distribution $P(k)$, defined as the probability that any node has $k$ links to other nodes, is well approximated by a power-law behavior $P(k)\sim k^{-\gamma}$, with $\gamma\approx 2.2$ [@fal; @gov00; @romu01]. This makes the Internet a capital example of the recently identified class of scale-free (SF) networks [@bara02]. The statistical physics approach has been proved to be a valuable tool for the study of complex networks, and several interesting results concerning dynamical processes taking place on complex networks have been recently reported. In particular, the absence of the percolation [@newman00; @havlin01] and epidemic [@pv01a; @virusreview] thresholds in SF networks has a large impact because of its potential practical implications. The absence of the percolating threshold, indeed, prompts to an exceptional tolerance to random damages [@barabasi00]. This is a property that assumes a great importance in communication networks, guaranteeing the connectivity capabilities of the system. Percolation properties of SF networks refer only to the static topological connectivity properties [@newman00; @havlin01]. On the other hand, in the Internet and other communications networks, many instabilities are due to traffic load congestions [@labo1; @labo2]. The traffic load carried on the failing nodes or connections is automatically diverted to alternative paths on the networks and instabilities can spread from node to node by an avalanche of traffic congestions and overloads. For instance, route flaps have led to the transient loss of connectivity for large portions of the Internet. These instabilities are thus of a dynamical nature and depend on how information is routed and distributed in the network. The models proposed so far, however, deal with regular structures [@jap1; @jap2; @watts02] and do not take into account the complex topology of SF networks. In this paper, we propose a simple model aimed at the study of failure cascades generated by the redistribution of traffic load by congested links or nodes in SF networks. We find that the system behavior depends on the average traffic load imposed to the network. Above a critical value of the average traffic load, a single failure has a finite probability of triggering a congestion avalanche affecting a macroscopic part of the network. The present analysis thus reveals the existence of a transition from a free phase to a congested one as a function of the amount of traffic carried by the network. Contrary to what happens for the static percolation transition [@havlin01; @newman00], loaded SF networks exhibit a finite threshold above which the system can develop macroscopic instabilities with respect to small damages if we consider the dynamics of the traffic carried on top of them. The results provided here represent a first step towards a more complete modeling of traffic instabilities in real communication networks. In order to include the degree fluctuations of SF networks we shall use in the following the Barabási-Albert model [@bar99]. This is a stochastic growth model in which one starts from a small number $m_0$ of nodes and at each time step a new node is introduced. The new node is connected preferentially to $m$ old ones (for the simulations we used $m=3$) with a probability $\Pi(k_i)=k_i/\sum_j k_j$ proportional to the degrees $k_i$ of the nodes. The repeated iteration of this scheme gives as a result a complex network with a topological structure characterized by a power-law degree distribution $P(k)=2m^2k^{-\gamma}$ with $\gamma=3$ and average degree $\langle k \rangle=2m$. In principle, one might also consider a more general class of complex networks with variable power-law degree distributions [@bara02; @doro]. To simulate the flow of data packets on SF networks, taking into account the load redistribution in case of damages, we need to specify the initial state of the network; i.e. the load of traffic flowing through each link. An estimate of such load, assuming that the routing takes place following the minimum path, is given by the total number of shortest paths between any two nodes in the network that pass through the node $i$. This magnitude is called betweenness or load [@new01; @goh01] and has been recently studied in SF networks. This property of the network can also be defined in terms of links. In real systems, however, the amount of traffic carried by each link is a fluctuating quantity that depends on many variables such as number of users, routing agreements, and available bandwidth. For this reason, we associate to each link connecting the nodes $i$ and $j$ of the network a load $\ell_{i,j}$ drawn from a probability distribution that specifies the initial traffic load of the system. For simplicity, we have considered a uniform distribution $U(\ell)$ for $0< \ell < 1$, taking the form $$U(\ell) = \left\{ \begin{array}{cll} \frac{1}{2 \langle \ell \rangle }, & \ell \in [0, 2 \langle \ell \rangle ] & \mathrm{if} \, \langle \ell \rangle \leq 0.5\\ &&\\ \frac{1}{2(1- \langle \ell \rangle) }, & \ell \in [2 \langle \ell \rangle -1 ,1 ] & \mathrm{if} \,\langle \ell \rangle \geq 0.5 \end{array} \right..$$ This uniform distribution implies that the minimum initial load carried by a link is bounded by a nonzero value for an average load $\langle l\rangle > 0.5$, which means that there will be no links with load smaller than this lower bound. The results reported in this paper were obtained with the initial distribution $U(\ell)$. In order to test the universality of the critical behavior, we have also considered the distribution $F(\ell)= (\langle \ell \rangle^{-1} -1) (1-\ell)^{(\langle \ell \rangle^{-1} -2)}$, $\ell \in [0, 1 ]$, which allows the existence of links with a very small load, irrespective of the average load $\langle\ell\rangle$ flowing through the system. Both distributions $U(\ell)$ and $F(\ell)$ yield the same qualitative results. Along with the load, we associate to each link the same capacity $C$ that, without loss of generality, we fix equal to one. This choice can be considered as a first approximation since the actual difference between line bandwidths in communication networks can be large. In this perspective, we consider as the most important source of heterogeneity the flow of different amounts of load through the network. The dynamics of the model is defined by a simple threshold process. A link is selected at random and overloaded by raising its traffic. When the load carried by a link is $\ell_{i,j}>C$, i.e. when it exceeds the link’s capacity, the link is considered congested and the load it carries is diverted among its (not overloaded) neighboring links. This amounts to consider that the time scale of the local congestion is greater than the time scale characterizing the reorganization of the routing procedure. The redistribution of the load on its turn might provoke that other links become overloaded, thus triggering a cascade of failures. We have explored two physically different settings of the load redistribution rule. The first consists on equally distributing the load of a congested link among the non-congested neighboring links. We refer to it as the deterministic redistribution rule, respectively. The second case will be called random redistribution because when a link is overloaded, a random amount of load is redistributed to each of the remaining working links in its neighborhood. Finally we note that in the rare event in which the congested link has no active neighbors, its load can be equally shared among all the remaining working lines of the network or just be considered as lost from the network. This amounts to a conserved or dissipative redistribution rule. Many physical systems display criticality only when energy is conserved [@jen98; @marro]. In distributed networks such as the Internet, however, it is common to discard packets if there is not a route available at the moment. As we shall see in the following, the results do not depend qualitatively on the conserved nature of the traffic load. We have performed large-scale numerical simulations by applying repeatedly the rules stated above on BA networks. The sizes of the networks used in the simulations range from $N=5 \times 10^3$ nodes ($15 \times 10^3$ links) to $N=10^5$ nodes ($3 \times 10^5$ links). All numerical results have been obtained by averaging over 10 different networks and, at least, 100 different realizations of the initial load distribution. In order to inspect the occurrence of dynamic instabilities, we construct the phase diagram of the system. The order parameter can be identified as the probability $P_\mathcal{G}$ of having a giant component $\mathcal{G}$ of connected nodes with size of the order of the network size. The giant component is defined as the largest component of the network made by nodes connected by active links, after the system has reached a stable state (when $\ell_{i,j} < C$ for all $i$ and $j$). The existence of a giant component implies that a macroscopic part of the network is still functional. If the giant component of the network is zero, the communication capabilities of the network are destroyed and a congestion of the order of the system size builds up. It is worth noticing that although we have determined the giant component size in terms of nodes and the dynamical rules of the model are expressed in terms of links, the results are completely equivalent since a connected node is defined as a node with at least one active link. In Fig. \[figure1\] we plot the order parameter $P_\mathcal{G}$ as a function of the average load $\langle\ell\rangle$. At low values of the average load, the network always reaches a stable state in which the number of isolated nodes is very small and with probability $P_\mathcal{G}=1$ the network has a giant component of connected nodes of the order of the system size. When increasing the load imposed on the network, the system starts to develop instabilities. In particular, above a critical load $\langle\ell\rangle_c^I$, whose value depends on the model considered, with a finite probability the system evolves to a congested state without giant component of connected nodes; i.e. the largest set of connected active nodes has a density of order $N^{-1}$. This implies a probability of having a giant component $P_\mathcal{G}<1$, which is decreasing as the load is progressively increased. At an average load $\langle\ell\rangle_c^{II}\simeq 0.82$ we get that $P_\mathcal{G}=0$, signalling that, with probability one, any instability will propagate until the complete fragmentation of the network. It is worth remarking that this scenario is rather different from the percolation one in which the probability of having a giant component is abruptly dropping from one to zero at the transition point. Here, the probability decays continuously to zero and we have a wide region of $\langle\ell\rangle$ where the initial instability can trigger a destructive congestion with probability $1-P_\mathcal{G}$. Fig. \[figure2\] illustrates the probability $p(S)$ that the isolated network has a size $S$ in the case that no giant component of connected nodes has survived. The distribution is rather peaked also at relatively small values of $\langle \ell \rangle$, almost affecting the totality of the network. The phase diagram obtained in Fig. \[figure1\] points out that the value of the average load at which $P_\mathcal{G}=0$ is relatively high. On the other hand, the value at which $P_\mathcal{G}$ is appreciably smaller than one is well below the theoretical capacity of the network measured as the capacity $C=1$ of the individual links ($\langle\ell\rangle_c^I\simeq 0.15 - 0.3$, see Fig. \[figure1\]). This evidence defines a wide region of load values in which there is a small but finite probability that a small instability propagates through the system and could explain why in real communication networks congestions of diverse degree can be encountered from time to time. In Fig. \[figure1\] we also report curves obtained for the different dynamical rules defined previously and it is interesting to remark that the highest level of stability has been obtained for the deterministic and conservative version of the model ($\langle\ell\rangle_c^I\simeq 0.3$). Moreover, the dynamics of the transition does not depend on the distribution of initial loads and the network size $N$. Another way to shed light on the congestion dynamics is to inspect the process of generation of the congestion instability. We define the size $s$ of a congestion burst or avalanche as the total number of simultaneously overloaded links. The cumulative distribution $P(s)$ of avalanches of size larger than $s$ for several values of the average load imposed on the network and four different system sizes have been plotted in Fig. \[figure3\] for the random and dissipative definition of the present model. The main plot is in log-linear scale, so that a straight line corresponds to a power law of the form $p(s) \sim s^{-1}$ for the probability of observing an avalanche of size $s$. Power-laws with exponent $-1$ have been found for several characteristic features of Internet traffic such as latency times, queue lengths, and congestion lengths [@jap1; @mag; @will]. In the figure we focus on region close to the stable region $\langle \ell \rangle\simeq 0.20$, which means that the power-law behavior extends to values far from the instability transition. This fact confirms that it is not necessary that the network operates very close to a critical point in order to observe power-laws in the distribution of several quantities. The inset in Fig. \[figure3\] shows that the cumulative size of overloaded links also scales with the system size, the scaling dynamics, however, remaining the same. This may help understand why power-law distributions observed in real communication networks have been measured for different network sizes, i.e. both for local networks and for networks that extend to a very large scale. In summary, we have introduced a simple threshold model aimed at the description of instabilities due to load congestion that takes into account the topological properties of SF networks. The results obtained point out that the network can freely handle traffic up to some critical average load $\langle\ell\rangle_c^I$. Above this level the network faces partial congestions that start to build up local bottlenecks in various places and small instabilities might trigger macroscopic outages with a finite probability. Above a critical load value $\langle\ell\rangle_c^{II}\simeq 0.82$ any small instability leads to the whole network collapse. In the intermediate region of network load, the number of simultaneous line casualties follows a power-law resembling what has been observed in experimental studies of the Internet. We hope that our work will provide hints for accurate modeling of the Internet and the WWW large-scale traffic behavior. This work has been partially supported by the European Commission - Fet Open project COSIN IST-2001-33555. R.P.-S. acknowledges financial support from the Ministerio de Ciencia y Tecnología (Spain). [99]{} S. H. Strogatz, Nature (London) [**410**]{}, 268 (2001). R. Albert and A.-L. Barabási, Rev. Mod. Phys. [**74**]{}, 47 (2002). S. N. Dorogovtsev and J. F. F. Mendes, Adv. Phys. [**51**]{}, 1079 (2002). A.-L. Barabási, R. Albert and H. Jeong, Physica A [**281**]{}, 69 (2000). M. Faloutsos, P. Faloutsos, and C. Faloutsos, Comp. Com. Rev. [**29**]{}, 251 (2000). A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins and J. Wiener, Comp. Networks [**33**]{}, 309 (2000). R. Govindan and H. Tangmunarunkit, in [*Proceedings of IEEE INFOCOM 2000*]{}, Tel Aviv, Israel (Ieee, Piscataway, N.J. 2000). G. Caldarelli, R. Marchetti and L. Pietronero, Europhys. Lett. [**52**]{}, 386 (2000). Q. Chen, H. Chang, R. Govindan, S. Jamin, S.J. Shenker and W. Willinger, INFOCOM 2002. Twenty-First Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE , Volume: 2 , 2002 A. Broido and K. C. Claffy, in SPIE International symposium on Convergence of IT and Communication, (Denver, CO, 2001). R.Pastor-Satorras, A. Vázquez, and A. Vespignani, Phys. Rev. Lett. [**87**]{}, 258701 (2001); A. Vázquez, R.Pastor-Satorras, and A. Vespignani, Phys. Rev. E [**65**]{}, 066130 (2002). R. Cohen, K. Erez, D. ben-Avraham, and S. Havlin, [*Phys. Rev. Lett.*]{} [**85**]{}, 4626 (2000); D. S. Callaway, M. E. J. Newman, S. H. Strogatz, and D. J. Watts, [*Phys. Rev. Lett.*]{} [**85**]{}, 5468 (2000). R. Pastor-Satorras, and A. Vespignani, Phys. Rev. Lett. [**86**]{}, 3200 (2001). R. Pastor-Satorras, and A. Vespignani, [*Handbook of Graphs and Networks: From the Genome to the Internet*]{}, eds. S. Bornholdt and H.G. Schuster (Wiley-VCH, Berlin, 2002). R. Albert, H. Jeong, and A.-L. Barabási, Nature [**406**]{}, 378 (2000). C. Labovitz, G.R. Malan and F. Jahanian, INFOCOM ’99. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, Volume: 1, (1999). C. Labovitz, A. Ahuja, F. Jahanian, Fault-Tolerant Computing, 1999. Digest of Papers. Twenty-Ninth Annual International Symposium on , 278, (1999) M. Takayasu, H. Takayasu, and T. Sato, Physica A [**233**]{}, 924 (1996); A. Y. Tretyakov, H. Takayasu, and M. Takayasu, Physica A [**253**]{}, 315 (1998). T. Ohira, R. Sawatari, Phys. Rev. E [**58**]{}, 193 (1998); S. Valverde, and R. V. Solé, Physica A, to appear (2002). D.J. Watts, Proc. Nat. Acad. Sci. USA,[**99**]{}, 5766 (2002). A.-L. Barabási, and R. Albert, Science [**286**]{}, 509 (1999); A.-L. Barabási, R. Albert, and H. Jeong, Physica A [**272**]{}, 173 (1999). M. J. Newman, Phys. Rev. E. [**64**]{}, 016132 (2001). K.-I. Goh, B. Kahng, and D. Kim, Phys. Rev. Lett. [**87**]{}, 278701 (2001). H. J. Jensen, [*Self-Organized Criticality*]{} (Cambridge University Press, Cambridge, England, 1998) and references therein. J. Marro and R. Dickman, [*Nonequilibrium phase transitions in lattice models*]{} (Cambridge University Press, Cambridge, 1999). M. O. Magnasco, [*The thunder of distant Net storms*]{}, preprint nlin.AO/0010051 (2000). W. Willinger, R. Govindan, S. Jamin, V. Paxson and S. Shenker, Proc. Nat. Acad. Sci. USA [**99**]{} 2573 (2002).

ArXiv

--- abstract: 'It is well known that many biochemical processes in the cell such as gene regulation, growth signals and activation of ion channels, rely on mechanical stimuli. However, the mechanism by which mechanical signals propagate through cells is not as well understood. In this review we focus on stress propagation in a minimal model for cell elasticity, actomyosin networks, which are comprised of a sub-family of cytoskeleton proteins. After giving an overview of th actomyosin network components, structure and evolution we review stress propagation in these materials as measured through the correlated motion of tracer beads. We also discuss the possibility to extract structural features of these networks from the same experiments. We show that stress transmission through these networks has two pathways, a quickly dissipative one through the bulk, and a long ranged weakly dissipative one through the pre-stressed actin network.' address: - '$^1$Raymond & Beverly Sackler School of Chemistry, Tel Aviv University, Tel Aviv 6997801, Israel' - '$^2$Department of Chemical Engineering, Ilse Kats Institute for Nanoscale Science and Technology, Ben Gurion University of the Negev, Beer-Sheva 84105, Israel' author: - 'Adar Sonn-Segev,$^{1}$ Anne Bernheim-Groswasser,$^{2}$ Yael Roichman$^{1*}$' title: 'Dynamics in Steady State In Vitro Acto-Myosin Networks' --- July 2014 [*Keywords*]{}: cytoskeleton, stress transmission, microrheology Structure and components of In vitro acto-mysoin networks ========================================================= Two of the central functions of the cytoskeleton are to support the cell’s shape and to control its motion. These roles are done predominantly by variants of actomyosin networks consisting mainly of a structural protein actin, and a molecular motor, myosin. Additional actin binding proteins may modify the structure of actin networks and their polymerization kinetics. Network building blocks ----------------------- ### Actin and actin polymerization {#actin-and-actin-polymerization .unnumbered} Actin is an abundant and highly conserved protein in most eukaryotic cells[@Dominguez2011]. Actin networks are responsible for many cellular processes such as cell motility, cell division, and the maintenance of cellular integrity[@Dominguez2011; @Pollard2009; @Cooper2003; @Blanchoin2014]. To accommodate all of these roles actin networks adopt a variety of structures within cells ranging from finely branched networks in the lamellipodia to thick bundles in stress fibers[@Blanchoin2014]. It is becoming increasingly clear that cells control the structure and kinetics of their actin networks by numerous actin binding proteins (ABP). These ABP are involved in many processes including nucleation, elongation, and branching of actin filaments, in filament disassembly via severing and depolymerization, and in bundling and crosslinking of filaments into fibers and unorganized networks[@Dominguez2011; @Blanchoin2014]. ![G-actin monomers (PDB:1J6Z) polymerize into actin filaments with a pointed $(-)$ end and a barbed $(+)$ end. Within the filament ATP is hydrolyzed to ADP. During polymerization and treadmilling monomers are added preferentially on the barbed end and released preferentially from the pointed end after hydrolyzation. (b) Myosin monomers (PDB:2MYS) are comprised of a head and tail domains. The head domain consists of an actin binding site and a lever arm, which takes part in the power stroke of the motor. The tail domains in myosin II coil to form dimers which then self-assemble into minifilaments.[]{data-label="fig:actin"}](structure) The actin monomer (G-actin) is a globular protein with a molecular mass of approximately 42 kDa[@Otterbein2001]. G-actin is asymmetric having a pointed end $(-)$ and a barbed end $(+)$ (see Fig. \[fig:actin\](a)). In physiological conditions G-actin is found preferentially in an ATP (adenosine triphosphate) bound state. The kinetics of the actin filament growth is thermodynamically limited by the slow nucleation process; after a stable nucleus is created (with three or four monomers) filament elongation proceeds rapidly. The double-stranded helical actin filament retains the asymmetry of G-actin supporting a barbed and pointed end. The elongation of F-actin filaments can occur from both ends, however the critical concentration above which growth is advantageous is over 12 folds higher at the pointed end[@Wegner1983]. Starting from high monomer concentrations, the elongation process reaches a steady-state in which G-actin in ATP state is added preferentially to the barbed end and G-actin in ADP state is released from the pointed end. In cells this treadmilling process is mediated by actin binding proteins allowing for much faster turnover rates and in consequence, faster reorganization of the cytoskeleton. As a result, local control over actin binding proteins concentration allows for local changes of the turnover rate and hence to spatial heterogeneity within the cell [@Dominguez2011; @Blanchoin2014]. Also in-vitro kinetics of polymerization of actin can be mediated through addition of actin binding proteins. For example, capping protein binds to the fast growing barbed end and is used, in proper stoichiometry, to control filaments length[@Lodish2000; @Schafer1996; @TDPollard2003]. Single actin filaments are semi-flexible polymers with a persistence length of approximately $10$ [$\mu$m]{}, a rupture force of $110$ pN, a bending energy of $4\cdot10^{21}$ J at room temperature, a buckling force of $0.4$ pN for a $1$ [$\mu$m]{} long filament, and an estimated Young modulus of approximately $2$ GPa [@Blanchoin2014; @Janmey1999]. The viscoelastic properties of an actin network are determined not only by those of the single filaments, but also by the network structure. A semi-dilute suspension of actin filaments will form a physical gel, in which entanglements supply some resistance to shear. An addition of cross linking proteins is essential for such networks to resist flow and maintain their integrity once myosin motors are introduced[@Blanchoin2014; @Backouche2006; @Reymann2012; @Haviv2008; @Ideses2013]. Networks made of actin filaments support a large variety of viscoelastic properties highly dependent on the type of actin binding proteins present[@Lieleg2010] and on the salt concentrations in the polymerization buffer[@Strelnikova2016]. For example, fascin organizes actin into parallel filament bundles, while $\alpha$-actinin arranges it into antiparallel bundles. High concentrations of bundling proteins produce networks of thick bundles with much higher stiffness than networks made of crosslinked single actin filaments[@Lieleg2010]. ### Myosin minifilaments self assembly {#myosin-minifilaments-self-assembly .unnumbered} One of the most important families of actin binding proteins are myosins, the molecular motors associated with actin filaments. They are mechano-enzymes generating force by hydrolysis of ATP. This superfamily of motor proteins includes at least twenty four classes performing different roles within cells and muscles [@Hartman2012; @Lodish2000; @Syamaladevi2012]. For example, organelle transport involves myosin V [@Desnos2007], muscle contraction and cytokinesis involve myosin II, which was the first to be discovered [@Kuhne1864] and the most studied of all myosins [@Lodish2000]. All Myosins are approximately $1000-2000$ residues long, and are composed of a heavy chain consisting of three domains: a conserved globular head domain (motor domain) containing an actin binding site and an ATP binding site, a flexible $\alpha$ helix neck domain associated with regulatory proteins called light chains, and a tail domain (Fig. \[fig:actin\](b)). The tail domain is tailored to the function of the specific myosin, such as supporting a cargo binding site. In many myosins, including myosin II, the tail domain mediates the formation of two headed dimers by forming a long rod-like coiled coil structure binding the two monomers[@Hartman2012; @Lodish2000; @Syamaladevi2012; @Reisler1980]. In-vitro at high salt (KCl) concentrations these dimers are stable, however at low salt concentrations multiple myosin II dimers self-assemble in an antiparallel manner to form thicker filaments termed minifilaments[@Reisler1980], reminiscent of the contractile apparatus in muscle (Fig. \[fig:actin\](b)) [@Kaminer1966; @Koretz1979; @Davis1988; @Shutova2012]. The resulting structure of the minifilaments includes a bare zone at the center and head domains positioned at both ends of the filament. The prevailing model for minifilament self-assembly is that myosin dimers nucleate small bipolar structures[@Davis1982; @Koretz1982; @Pollard1982] that continue to grow by addition of monomers[@Katsura1973]. This bipolar structure is essential for myosin function in sliding two actin filaments relative to each other, and allows myosin II to function as a crosslinker between actin filaments. Control of minifilament length and number of dimers in each myosin minifilament, in vitro, is achieved by bringing the solution to a desire ionic strength. The rate at which this process is carried out may affect the assembly process. [@Kaminer1966; @Koretz1979; @Ideses2013]. Myosin II motors use the chemical energy released in ATP hydrolysis to perform mechanical work via conformational changes in the head domain where both the ATP and actin binding sites are located. The motion is then amplified by the neck domains [@Hartman2012; @Syamaladevi2012; @Sweeney2010]. Myosin II motors move with discrete steps of 5-15 nm reducing to 4-5 nm under load, and generating 1-9 pN forces[@Lodish2000; @Sweeney2010; @Finer1994; @Ruegg2002; @Kaya2013]. Due to the multiple actin binding sites on actin filaments, a myosin minifilament can move stochastically or continuously (processively) along the actin filament. The degree of processivity of the motor motion depends on the average binding time of the myosin heads to actin (which depends on ATP concentration) and on the number of heads in each minifilament. Self organization of actomyosin networks {#self} ---------------------------------------- The minimal combination of actin monomers, myosin minifilaments and a crosslinking protein at the right salt and ATP concentration is used to create active self-organized gels in-vitro. Such polymer networks organize through a universal process of initiation, coarsening, and failure, i.e., rupture or global compression[@Backouche2006]. The myosin motors play an active part in this evolution process, mostly in the stages of coarsening and failure. The extent of network remodeling as well as the mode of failure may vary extensively with the concentration and relative amounts of actin, myosin, and crosslinker protein. It also depends on the type of crosslinking protein and on ATP and salt concentrations [@Ideses2013; @Lieleg2009]. For example, in Fig. \[network\] the three stages of network evolution are depicted from initiation Fig. \[network\](a) through coarsening Fig. \[network\](b) and failure through a 10 fold compression Fig. \[network\](c). ![Actomyosin network evolution. Actin is polymerized in the presence of fascin and large mysin minifilaments (150 dimers per minifilament) [@Ideses2013]. Fluorescent labeling allows direct observation of the actin bundles (red) and myosin minifilaments (green). (a) The network close to the initiation of polymerization, (b) after significant coarsening, (c) after 10 fold compression.[]{data-label="network"}](network) In cells actomyosin networks appear in a large variety of morphologies well suited for their different functions. By choice of appropriate crosslinking proteins, ATP concentrations, and myosin minifilaments’ concentration and size [@Ideses2013; @Backouche2006], many such morphologies can be obtained in-vitro. ### Initiation {#initiation .unnumbered} Little is known of the initial steps of network formation in in-vitro actomyosin gels due to the sub-diffraction-limit size of the actin filaments at this stage. It is known, however, that polymerization is limited by the nucleation of actin filaments, and that an addition of fascin will accelerate network formation[@Haviv2008a].In the presence of fascin, myosin can further shorten the time it takes the networks to visibly form at intermediate concentrations approaching those of fascin [@Ideses2013]. ### Coarsening {#coarsening .unnumbered} Once an initial network is formed the actin network goes through a motor driven reorganization process. During this stage the thickness of actin bundles increases as does the average mesh size [@Bendix2008; @Ideses2013]. The length scale of reorganization depends on network connectivity and on the force generation process [@Kohler2011]. The former depends mainly on the crosslinking protein (concentration and type), and the latter on minifilament size and the average attachment time of a single myosin head to actin. The actin architecture which is influenced by nucleation sites [@Reymann2012] and by crosslinker type and density affects significantly the myosin induced reorganization [@Smith2007]. For example, a simple system of actin and myosin can reach nematic ordering, but a trace amount of biotin/neutravidin crosslinking sites results in disordered cluster formation [@Smith2007]. The reorganization of the actin network is a result of the activity of myosin minifilaments. Myosin minifilaments anchored on two strands of actin can either slide the actin filaments apart or bring them closer. The long time result of the filament sliding and buckling is the formation of actin bundles and asters, and the pulling of excess slack in the actin network [@Murrell2012; @Backouche2006; @SoareseSilva2011]. In general, myosin activity depends both on the size of the myosin minifilaments and on ATP concentration. The processivity of the motor increases with the number of heads on a minifilament, but decreases with a decrease in the binding time of a myosin head to actin via an increase in ATP concentration. Therefore, in networks containing small minifilaments the motors can be either active, but not attached to the actin, active and attached, or attached but immobile, the probability of which depends on ATP concentration [@Smith2007; @Vogel2013]. In addition, the total force applied by the myosin motors depends linearly on minifilament size. Finally, larger minifilaments promote larger contractile units that increase the bundling rate of the acting filaments [@Thoresen2013; @Vogel2013]. The time it takes an active network to reorganize before it reaches a steady state or fails via collapse or rupture processes can range from a few minutes to several hours. During this time the network undergoes a stiffening process due to increase in actin bundles thickness and the reduction of filament bending fluctuation entropy. This gradual stiffening occurs for low and high motor concentrations as was demonstrated by using single particle microrheology [@Stuhrmann2012; @SoareseSilva2011]. In addition, the stiffening due to internal forces applied by myosin motors was found to be equivalent to the stiffening of actin gels stretched externally [@Koenderink2009]. ### Arrest, collapse, and rupture {#arrest-collapse-and-rupture .unnumbered} In most conditions the built up of tension due to myosin activity during the reorganization of active gels results in macroscopic compression of the gels. A recent theory identifies four competing compaction mechanisms: sarcomerelike contraction due to motors stalling at the barbed end[@Liverpool2005], motion of a finite sized motor crosslinking to filaments from the intersection towards the barbed end causing contraction[@Dasanayake2011], flexible minifilaments zipping filaments together[@Lenz2012], and deformable actin filaments[@MacKintosh2008; @Mizuno2007]. This theoretical work predicts that the main mechanism for compression is the latter, namely, a local symmetry breaking in which the actin scaffold’s deformation results in mesoscopic compression regardless of the sliding direction of two actin filaments generated by a myosin motor. Actin buckling due to this mechanism results from perpendicular forces rather than longitudinal buckling [@Lenz2014]. If the gels are held at the boundary or at high myosin concentrations, rupture occurs instead of compression. From monitoring the rupture process of such gels it was discovered that the final state of these gels before failure is a critically connected state [@Alvarado2013]. A range of network connectivity and motor activity is required to reach such a critically connected state which can then develop global compression [@Wang2012; @Bendix2008]. Measuring stress propagation in viscoelastic materials {#stress} ====================================================== One way to measure the stress propagation through a given medium is to look at its mechanical response, e.g., at the displacement field resulting from a point perturbation. This concept is used, for example, in traction force microscopy[@Style2014]. Small tracers particles embedded in the medium have proved to be good markers for monitoring the deformation field caused by a perturbation[@Ladam2003]. Alternatively, stress propagations can be extracted from the correlation in displacement of two such embedded tracer particles. When one particle is perturbed a stress field is created in the medium resulting in a displacement field entailing the other particle. Since the second particle is moving in response to the motion of the first bead their movement is correlated. The perturbing force on the first particle may be externally or internally applied (e.g., by an external agent such as an optical or magnetic tweezers, or by a nearby molecular motor), or it can be induced by thermal fluctuations. If the beads are subjected to stochastic motion due to thermal or active fluctuations, the material’s mechanical response, which is a deterministic quantity, may be distorted by the fluctuation induced noise. In such cases averaging over time and ensemble is required to characterize properly this response. We define ${\Delta r_\alpha (t, \tau) = r_\alpha(t+\tau)-r_\alpha(t)}$ as the vector displacement of individual tracers, where $t$ is the absolute time and $\tau$ is the lag time. The time and ensemble-averaged tensor product of the vector displacements is a measure of stress propagation, and is a function of distance and lag time: $$D_{\alpha, \beta}(r, \tau) = \langle \Delta r_\alpha ^i(t, \tau)\Delta r_\beta ^j(t, \tau)\delta[r - R^{ij}(t)]\rangle_{i \not= j, t} , \label{dis_2p_vec}$$ where $i$ and $j$ label different particles, $\alpha$ and $\beta$ label different coordinates, and $R^{ij}(t)$ is the distance between particles $i$ and $j$ at time $t$. Here the average is taken over the distinct terms $(i \not= j)$; the self term yields $ \langle \Delta r^2(\tau) \rangle \times \delta(r)$, the one-particle mean-squared displacement (MSD$^{\textup{1P}}$). The two-point correlation for particles in an incompressible continuum is calculated by treating each thermal particle as a point stress source and mapping its expected strain field [@Landau1986]. ![Schematic of two-point displacement components. In this sketch, the longitudinal component $D_\parallel=\langle \Delta r^1_\parallel(\tau)\Delta r^2_\parallel(\tau) \rangle$ is the product of the displacement component projected along the line separating the tracers by distance $r$. The transverse component $D_\perp=\langle \Delta r^1_\perp(\tau)\Delta r^2_\perp(\tau) \rangle$ is the product of the displacement component projected perpendicular to the line connecting the pair. []{data-label="fig:2P"}](2P) Spatially, $D_{\alpha,\beta}(r, \tau)$ can be decomposed into longitudinal $D_{\parallel}$ and transverse $D_{\perp}$ components, where the former is the component of the motion along the center-to-center separation vector of the two tracers (see Fig. \[fig:2P\]), while the latter is the component orthogonal to the separation vector. In an isotropic medium the off-diagonal component vanishes by symmetry. For an incompressible medium, to lowest order in $a/r$, where $a$ is the tracer particle radius, the amplitudes of the two components are related via $D_\perp=\frac{1}{2}D_\parallel. \label{Eq:perp_para}$ Typically, $D_\parallel$ is the stronger component and hence easiest to measure in experiments from a signal-to-noise perspective. Correlated motion measurements are used in thermodynamic equilibrium to measure the complex shear modulus of viscoelastic media. This technique is called two-point (2P) microrheology [@Crocker2000], and was developed in 2000 as an improvement of one-point (1P) microrheology. In 1P microrheology the MSD$^\textup{1P}$ of a single tracer particle is used to extract the shear modulus of the material it is embedded in through the generalized Stokes-Einstein relation (GSER)[@Mason1995; @Mason1997; @Squires2010]. 2P microrheology takes advantage of the interparticle mechanical coupling, characterized by $D_{\alpha,\beta}(r, \tau)$ to robustly extract bulk material properties. Complex fluids contain structural elements, such as particles in a colloidal suspension or a polymer network in a gel. For these systems we may define $r_{GC}$ as a measure of the largest structural element in a complex fluid, e.g. the diameter of a colloidal particle or the mesh size of a polymer network. In a medium that is homogeneous (and isotropic) at long length-scales ($r>r_{CG}$), the strain field resulting from thermal motion of a particle is proportional to the tracer’s motion and decays as $a/r$, where $r$ is the distance from the tracer. On these scales the functional form of the decay in motion correlation is the same as in a simple incompressible fluid. This is a manifestation of momentum conservation on such scales[@Sonn-Segev2014]. The correlated motion of two particles with separation $r$ is driven only by modes with wavelengths greater than the separation distance. Therefore, two tracers that are separated by more than the coarse-grained length-scale $r_{CG}$ will depend on the coarse-grained, macroscopic complex modulus. At this range of separations the material is treated as homogeneous, $ D_{\parallel}(r, \tau), D_{\perp}(r, \tau)\sim r^{-1}$ within this range, and the shear modulus of the material can be determined using the relation [@Crocker2000] $$\tilde{D}_{\parallel}(r, s) = \frac{k_BT}{2 \pi rs \tilde{G}(s)}, \label{GSER_2PM}$$ where $ \tilde{D}_{\parallel}(r, s)$ is the temporal Laplace transform of $D_{\parallel}(r, \tau)$ and $\tilde{G}(s)$ is the temporal Laplace transform of the complex shear modulus. Comparing the longitudinal two-point correlation to the generalized Stokes-Einstein equation used in 1P microrheology, $\langle \Delta \tilde{r}^2 (s) \rangle = d{k_{\rm B}T}/3\pi a s \tilde{G}(s)$ in $d$ dimensions[@Mason1995], suggests defining a new quantity: the two-point (2P) mean-squared displacement, MSD$^\textup{2P}$, as [@Crocker2000] $$\textup{MSD}^\textup{2P} = \frac{2r}{a} D_{\parallel}(r, \tau). \label{dist_msd}$$ This is the thermal motion obtained by extrapolating the long-wavelength thermal fluctuations of the medium to the bead radius. If the material is homogeneous, isotropic on length scales significantly smaller than the tracer, incompressible, and connected to the tracers by uniform no-slip boundary conditions over their entire surfaces, the MSD$^\textup{2P}$ will match the conventional MSD$^\textup{1P}$. Any difference between them can provide insights into the local microenvironment experienced by the tracers [@Valentine2001; @Shin2004]. Here we are interested in using correlated motion to measure stress transmission between particles, i.e., the hydrodynamic interaction between them, rather than characterizing the bulk properties of the material they are embedded in. For example, the correlation in motion of two optically trapped beads suspended in water was used to measure the hydrodynamic interaction far from a boundary [@Meiners1999]. The hydrodynamic interaction between colloidal particles near a single rigid boundary was calculated [@Blake1971; @Pozrikidis1992] and measured[@Dufresne2000] for a pair of particles diffusing at a distance $h$ above a wall : $$\begin{aligned} D_{\parallel}(r \gg h) &=& \frac{3{k_{\rm B}T}h^2}{2\pi \eta r^3 },\\ D_{\perp}(r \gg h) &=& \frac{3 {k_{\rm B}T}h^4}{ 4\pi \eta r^5}. \label{Eq:cordif_wall}\end{aligned}$$ These coefficients describe the leading terms of the in-plane correlated diffusion between two colloidal particles. Note that the leading term in the hydrodynamic interaction decays as $\sim r^{-3}$, rather than $\sim r^{-1}$, which is due to the unconserved momentum in the system. Similarly, in other confined geometries hydrodynamic interactions depend differently on distance, yielding different functional forms for the stress transmission [@Diamant09], e.g. $ D(r, \tau) \sim~r^{-2}$ in quasi-two dimensional samples [@Cui2004]. This was demonstrated experimentally for different colloidal suspensions [@Cui2002; @Cui2004; @Sonn-Segev2015]. In order to extract the full information hidden in the correlated motion of tracer particles it is beneficial to compare measurements to a physical model describing the embedding material. For example, the thickness of a soap film[@Prasad2009] or a thin viscoelastic layer[@Ladam2003] could be extracted given a proper model for deformations in a quasi 2D layer with free or rigid boundaries, respectively. Stress propagation in passive in-vitro actomyosin networks {#passive} ========================================================== The measurements: 2P correlations --------------------------------- Let us start by considering stress propagation in an entangled actin networks with a mesh size $\xi_s=300$ nm [@Sonn-Segev2014; @Sonn-Segev2014a]. In Fig. \[fig:Drq\] the correlated motion in the longitudinal and transverse direction, $D_{\parallel}$ and $D_\perp$, respectively, are presented as a function of particle separation. There are several interesting features in these plots: (i) there are two regimes of stress propagation as a function of inter-particle separation, (ii) the crossover distance between the two regimes, $r_c$, is an order of magnitude bigger than the particle diameter (0.49 [$\mu$m]{}) and the mesh size, (iii) the new intermediate regime is characterized by $D_\parallel\sim r^{-3}$ and $D_\perp<0$, as opposed to the well known long distance scaling i.e. $D_{\parallel,\perp}\sim r^{-1}$. These results can be interpreted by reexamining the Stokes problem of a rigid sphere of radius $a$ driven by a constant force $\vec{F}$ through an incompressible fluid of viscosity $\eta$ [@happel91]. The fluid velocity at a distance $r$ from the sphere can be described by a multipole expansion of the force and density fields, in analogy to the multipole expansion commonly done to describe the electrical field arising from a charged sphere. The first term is a force monopole, which is the field that would arise from the perturbed sphere (colloid) if it was infinitely small. The second contribution would have been a force dipole, but for this scenario i.e. a sphere in an isotropic medium, this term vanishes [@Pozrikidis1992]. The third term in the force field is a force quadrupole; its physical meaning is that there is a difference between the force field created by a point particle and one with finite size. The first term in the mass field is a dipole; due to conservation of mass a local increase in density must be combined with a decrease in density nearby. As opposed to the force monopole, which decays as $r^{-1}$, the two subdominant terms in the flow field, the mass dipole and force quadrupole decay as $r^{-3}$ with different signs [@Sonn-Segev2014; @Diamant2015]. Although the functional form of these two contributions is the same, their physical origin is different, as we discuss below. In a simple fluid, such as a Newtonian fluid, the subdominant response becomes significant only at distances comparable to the particle size $a$, the only length scale in the system. Therefore, the subdominant response decays as $a^2/r^3$, and vanishes as $a\rightarrow0$. For a viscoelastic complex fluid the two subdominant contributions become separate, the mass dipole, originating from mass conservation and reflecting the fluid flow in the vicinity of the particle, depends on the local environment of the tracer, while the force quadrapole reflecting momentum transfer through the bulk material depends on the bulk viscosity[@Sonn-Segev2014; @Diamant2015]. In a case where the local environment (solvent) has much lower viscosity than the bulk material viscosity (polymer network), we expect the subdominant contribution to manifest to much larger distances. ![(a) Longitudinal and (b) transverse displacement correlations as a function of particle separation, $r$, at lag time $\tau=0.014$s for $\xi_s=0.3~\mu$m, $a=0.245~\mu$m and ${{\langle \hspace{0.1em} l \hspace{0.1em} \rangle}}=13~\mu$m. The crossover distance $r_c$ (blue dashed line) is defined at the intersection of the fitted dominant ($r^{-1}$) and subdominant ($r^{-3}$) power-law decays of $D_\parallel$. Reproduced from [@Sonn-Segev2014a] with permission from The Royal Society of Chemistry.[]{data-label="fig:Drq"}](passive_drr_dqqa=0_245) For example, even in the case of entangled actin gels (Fig. \[fig:Drq\]) with no addition of cross-linkers or bundling agents we observe the subdominant response up to a distance of approximately $r_c=3.5$ $\mu$m. The prediction in such a case is that: $$\begin{aligned} D_\parallel \sim r^{-3} \textup{ and } D_\perp<0 \textup{ if } r<r_c \nonumber\\ D_\parallel \sim r^{-1} \textup{ and } D_\perp>0 \textup{ if } r>r_c.\end{aligned}$$ This result should hold for any complex fluid with $r_c \sim \eta_b/\eta_\ell$, where $\eta_b$ ($\eta_\ell$) corresponds to the bulk (local) viscosity. Note that for complex fluids, with more than one relevant length scale (e.g. $a,\xi$ for a polymer gel), the response decays as $~\xi^2/r^3$ and therefore does not vanishes as $a\rightarrow0$. In this context we call the effective viscosity experienced by the tracer particle due to complex media (for example, solvent and the polymer network) the local viscosity. The local viscosity experienced by a tracer particle in an actin network decreases with particle size becoming closer to the solvent viscosity (although it will never reach exactly that limit). In Fig. \[fig:Drq\] the theoretically expected power law is seen only for half a decade. Nonetheless, we have observed this exact power law at different lag times in the same experiments, and in numerous other actomyosin networks: with and without cross-linking molecules, with and without myosin, and at different filament lengths. Further support for this power law dependence is that it can be derived base on the condition of mass conservation [@Diamant2015]. Finally, the theoretical analysis discussed above is confirmed below (Fig. \[fig:figure4\]), where a rescaling of the data according results in a single master curve. We note here that the asymptotic behavior of the correlated motion of beads in actin networks has been measured previously, and used to demonstrate the advantage of 2P microrheology in measuring the bulk shear modulus of viscoelastic inhomogeneous materials [@Crocker2000; @Gardel2003]. It was also shown that in such complex fluids intrinsic structural length scales affect the materials’ shear modulus [@Liu2006]. The interpretation: the two fluid model --------------------------------------- In order to understand the mechanical response of actin gels in the intermediate regime we require a theoretical model for such gels. We use the two fluid model of polymer gels [@DeGennes76; @DeGennes76a; @Milner1993; @Levine2000; @Levine2001] for this purpose. In this model the polymer is treated as a dilute viscoelastic network coupled to an incompressible solvent by friction forces. A local mechanical perturbation by a tracer particle will cause the solvent to flow through the polymer network in its proximity. However, at some larger distance, friction forces will cause the polymer network to move together with the fluid, as one continuum medium. There will arise a typical distance separating the flow of fluid against and with the polymer network. This crossover distance, $r_c$, can be calculated within the framework of the two fluid model [@Sonn-Segev2014; @Sonn-Segev2014a; @Diamant2015], and reads: $$r_c = a [2(\eta_b/\eta_\ell)g(\xi_d/a)]^{1/2},\ \ g(x) = x^2+x+1/3, \label{rc}$$ where $\xi_d$ is the dynamic correlation length of the viscoelastic gel. The ratio between the bulk and local viscosity is equivalent to the experimentally measured ratio $H(\tau)=\textup{MSD}^\textup{1P}/\textup{MSD}^\textup{2P}$ which is time (and frequency) dependent in a viscoelastic material [@Sonn-Segev2014; @Sonn-Segev2014a]. This is true, in thermal equilibrium, since $\textup{MSD}^\textup{1P}$ is inversely related to the local viscosity of the network, and $\textup{MSD}^\textup{2P}$ reflects bulk properties[@Sonn-Segev2014a]. Therefore, $r_c \sim H(\tau)^{1/2}$, as was demonstrated in passive actin networks with various particle sizes and mesh sizes (Fig. \[fig:figure4\](a)). In actin $H\sim 100$ [@Sonn-Segev2014a], causing $r_c$ to be ten times larger than the typical length scales in the system ($\xi_s,a$). This means that the mechanical response of actin networks decays faster than originally expected crossing over to a slower decay rate at a distance of a few micrometers (In Fig. \[fig:figure4\] $r_c$ ranges between 2 [$\mu$m]{}to 6.5 [$\mu$m]{}). This decay length is comparable to the size of biological cells. For stiffer cytoskeleton networks containing microtubules $r_c$ can reach values of 15 [$\mu$m]{}[@Pelletier2009]. ![Crossover distance for experiments on passive entangled actin networks at different conditions. (a) For all conditions $r_c$ is linear with $\sqrt{H}$ and increases with either $\xi_s$ or $a$. (b) All experimental results fall on a master curve once $r_c^2$ is normalized by $Ha^2$ and rescaled according to Eq. (\[rc\]). A fit to Eq. (\[rc\]) with $\xi_d=1.25\xi_s$ is presented by the solid line. Open (filled) symbols correspond to $a=0.55$ ($0.245$) $\mu$m. Each symbol corresponds to a different mesh size: $\xi_s=0.21$ (black squares), $0.26$ (magenta triangles), $0.3$ (cyan circles), $0.35$ (blue diamonds), and $0.44$ $\mu$m (red left triangles). The average filament length for all experiments was ${{\langle \hspace{0.1em} l \hspace{0.1em} \rangle}}=13~\mu$m. Adapted from [@Sonn-Segev2014] with permission from the American Physical Society.[]{data-label="fig:figure4"}](master) We can take our analysis one step further and extract structural information from the stress transmission signal, since the intermediate term includes the information about the correlation length of the network. (Eq. \[rc\]). The missing piece of information in Eq. \[rc\] is the ratio between the dynamic correlation length and the structural length scale, i.e., the mesh size. The dynamic correlation length is defined as the length scale over which momentum is absorbed in the system. It is related to the gel’s mesh size, but is not equal to it. This was obtain from the fit to the data in Fig \[fig:figure4\](b) to be, $\xi_d/\xi_s=1.25$. Using this relation we can now extract the mesh size of an actin network from $r_c$. As stated in Sec. \[stress\], the combination of stress propagation measurements and its modeling reveals structural information on the sample. It can be argued that many complex fluids and especially cytoskeleton networks possess more than one typical structural length scale. For example, actin filaments have a persistence length of $5-10$ $\mu$m and a contour length that can vary between 2 $\mu$m to $20$ $\mu$m. The effect of filament length was studied recently [@Sonn-Segev2014a]. Following the analysis described above, the stress propagation signal ($D_\parallel$) and the viscosity ratio $H(\tau)$ were measured for networks made of filaments of well defined length (ranging between 2 to 13 [$\mu$m]{}), all with a mesh size of $\xi_s=0.3$ [$\mu$m]{}. The dynamic correlation length $\xi_d$ was extracted by scaling $r_c$ with $H$ and $a$ for all the networks. Fig \[fig7\] shows that for filaments shorter than $8$ $\mu$m, $\xi_d$ depends also on filament length for a given $\xi_s$, i.e, stress propagation depends on also on $xi_d$. ![Dynamic correlation length, $\xi_d$, as a function of the average filament length, ${{\langle \hspace{0.1em} l \hspace{0.1em} \rangle}}$ (bottom) and actin/CP concentration ratio (top). Actin concentration was held at 1 mg/ml, resulting in a $\xi_s=0.3~\mu$m, and $a=0.245~\mu$m. Reproduced from [@Sonn-Segev2014a] with permission from The Royal Society of Chemistry []{data-label="fig7"}](figure7) Stress propagation in steady state in vitro actomyosin networks {#active} =============================================================== The study of stress propagation by 2P microrheology requires large amounts of data. Such measurements are ideally done in thermodynamic equilibrium and are challenging in rapidly evolving networks. To study of stress propagation in active mater, such as actomyosin networks it is advantageous to work in conditions where the gel arrives at a long-lived active steady state. Formation of active steady state networks ----------------------------------------- Actomyosin networks which arrive at long lived active steady states were created recently by polymerizing a mixture of unlabeled and biotinilated actin monomers in the presence of neutravidin and small myosin minifilaments. An average distance between crosslinkers of $ \approx 3~\mu$m, and a mesh size of $\xi_s=0.3~\mu$m were obtained by stochiometry. Varying degrees of activity were achieved by changing myosin minifilament concentration and size ($N_{myo}=19\pm3$ or $N_{myo}=32\pm5$ two-headed myosin dimers per minifilament) [@Future]. The number of myosin heads per minifilament was estimated from the distribution of minifilament length as measured using CryoEM (see supplementary information in [@Ideses2013]). In order to determine when these gels reach a steady state the motion of fluorescent polystyrene beads ($a= 0.55$ [$\mu$m]{}) was recorded for several hours at intervals of 15 min. The most obvious effect of myosin concentration on these actomyosin networks is to increase their stiffness, as seen from the decrease in the MSD$^{\textup{1P}}$ of the tracer particles with the increase in myosin concentration (Fig.  \[fig:characteristics\]a,b). About 50 min after mixing the various components the MSD$^{\textup{1P}}(\tau=7 ֿֿ \textup{s})$ settled to a steady value for most of the myosin concentrations (Fig.  \[fig:characteristics\]c,d). The ensemble and time average of the MSD$^{\textup{1P}}$ were compared to demonstrate that although the system is not at thermal equilibrium it is ergodic (Fig.  \[fig:characteristics\]e,f) [@Future]. ![Mean squared displacement (MSD) of particles in networks with different \[myosin\]/\[actin\] at two minifilaments sizes. (a) and (b) time and ensemble-averaged MSD of probe particles as a function of lag-time $\tau$ approximately 100 mins after polymerization. Minifilaments are composed of $N_{myo}=19$ (a) or $N_{myo}=32$ (b) myosins heads. (c) and (d) MSD at a lag time of $\tau=7~$s re-measured as a function of experiment time. The experiment time is the time between the onset of gel polymerization and the measurement time. Sizes of mini-filaments are $N_{myo}=19$ (c) and $N_{myo}=32$ (d). Colors and symbols correspond to different \[Myosin\]/\[Actin\] ratios: 0 (blue circles), 0.0017 (red squares), 0.0025 (green triangles), 0.005 (orange diamonds), 0.0083 (violet right triangles), 0.01 (maroon down triangles) 0.012 (magenta stars) and 0.02 (black pluses). (e) and (f) Comparison between time-averaged and ensemble-averaged MSD for networks with \[Myosin\]/\[Actin\]=0.0025 approximately 100 min after polymerization. Sizes of mini-filaments are $N_{myo}=19$ (e) and $N_{myo}=32$ (f), and initial slopes are $\alpha=0.7\pm 0.05$ and $\alpha=0.8\pm 0.05$. []{data-label="fig:characteristics"}](fig_act_2) Effect of motor concentration and minifilament size --------------------------------------------------- The two fluid model description, that was used for the passive networks, should hold for this system as well. Here, active random fluctuations are present in addition to the thermal fluctuations, but the system is still comprised of a polymer network immersed in a solvent. The significant amount of stiffening in these networks due to the motor concentration (Fig. \[fig:characteristics\]a,b) can be attributed to the addition of actin crosslinking sites by myosin minifilaments and to the reduction of slack in the actin filaments [@Future]. However, these changes do not affect the functional form of stress propagation through the active networks (Fig. \[fig:rcross\]). As in passive actin networks (Fig. \[fig:Drq\]), the correlated diffusion in the longitudinal direction decays fast at short distances $D_\parallel\sim r^{-3}$, and slowly at large distances $D_\parallel\sim r^{-1}$ . ![ Correlated motion of beads in actomyosin active networks. (a),(b) Longitudinal displacement correlations as a function of particle separation at lag time $\tau=0.014$ s and \[Myosin\]/\[Actin\]=0.0025. Mini-filaments are constructed by $N_{myo}=19$ (a) or $N_{myo}=32$ (b) myosins heads. The cross-over distance (orange dashed line) is clearly seen as in passive actin networks. (c) and (d): Transverse displacement correlation at the same conditions as in (a) and (b). []{data-label="fig:rcross"}](drr_dqq) The crossover distance, $r_c$ for both myosin minifilament sizes changes slightly with myosin concentration (Fig. \[fig:rcross1\]). It ranges between $4.5-3.5$ $\mu$m and $5.5-5.0$ $\mu$m for $N_{myo}=19$ and $32$ respectively (Fig. \[fig:rcross1\]). $r_c$ for the large minifilaments is bigger than for the smaller minifilaments. Considering that a polymer becomes stiffer with applied stress [@Rubinstein2003] due to a reduction in its configuration entropy, this result is expected, since larger myosin minifilaments can apply stronger forces on the network making it much stiffer and as a consequence increasing $\eta_b/\eta_\ell$. Moreover, at the same myosin to actin concentration ratio, which are kept constant in the experiments, they add less crosslinking sites resulting in a larger mesh size. ![(a) and (b) The cross-over distance, $r_c$, in networks with increasing myosin concentration. (c),(d) The Dynamic correlation length $\xi_d$ as a function of myosin concentration, base on the two fluid model scaling (Eq. \[rc\]). []{data-label="fig:rcross1"}](rc_xi_2) The structural features of these active networks were smaller than the diffraction limit and could not be resolved by optical microscopy, since no bundling proteins were added during preparation. Therefore, a direct observation of the structural evolution of these networks, which is expected for actomyosin networks (see Sec. \[coarsening\]), was not possible here. Nonetheless, insight into the structural evolution of the networks after initiation of polymerization and before they reach steady state can be obtained from $r_c$, based on the two fluid model, Eq. \[rc\], and the assumption that the networks are close enough to thermal equilibrium for the extracted $\xi_d$ to be a good estimate of $\xi_s$. For both systems we see a jump in the dynamic correlation length with addition of myosin minifilaments. We believe this result reflects the expected coarsening of the network due to the presence of motors prior to arriving at a steady state. It was recently shown that myosin stiffens actin networks in two ways, one of which is the by addition of crosslinking sites to the network [@Future]. As a result the mesh size and dynamical correlation length are expected to decrease. This effect can be seen in the $N_{myo}=19$ system for myosin concentrations \[myosin\]/\[Actin\]$>6\cdot 10^{-3}$. Here, the myosin minifilaments concentration becomes comparable to the concentration of biotin/neutravidin crosslinking sites. Stress propagation in evolving in-vitro actomyosin networks {#evolving} =========================================================== So far we have considered stress propagation through the bulk of a complex fluid. It is also interesting to ask how stress propagates directly through the polymer network (and not through the solvent). It is known that the cytoskeleton (i.e. actin filaments) is connected to the extracellular matrix through binding sites[@Cooper2000]. A mechanical signal, passing from the extracellular matrix through these adhesion points into the cell may propagate along the actin network as well as through the bulk. To characterize stress propagation through the actin network we turn to study evolving actomyosin networks comprised of actin, fascin, and large myosin minifilaments ($N_{myo}=150$) [@Ideses2013]. In this actomyosin system (see Fig. \[network\](a)) the fluorescently labeled actin (red) and myosin (green) can be directly observed. These gels follow the evolution stages described in Sec. \[self\] and Fig. \[network\]. After approximately two minutes these gels are fully connected and start coarsening. During the whole coarsening stage the myosin minifilaments remain embedded in the actin network. The correlated motion ($D_{\parallel}, D_{\perp}$) of myosin minifilaments, averaged over the entire coarsening stage is presented in Fig. \[fig:Dii\_ss\_1\](a). As opposed to the measurements of stress propagation between beads which are not attached to the network, there are very strong correlations in the motion of the myosin minifilament. Thus, a relatively low amount of measurements is required to extract reproducible results and good signal to noise ratio, at least for ubiquitous particle separations of the order of ten of micrometers. ![Correlated motion of myosin minifilaments embedded in the actin network (as in [@Ideses2013]) (a) $D_{\parallel}$ and $D_{\perp}$ of an evolving network 280 s after initiation with \[myosin\]/\[actin\]$=0.0067$. (b) $D_{\parallel}$ at various myosin concentrations (M/A denotes myosin to actin concentration ratio).[]{data-label="fig:Dii_ss_1"}](evol_err_noarrow) The functional dependence of $D_{\parallel}$ and $D_{\perp}$ on distance is very different from the bulk response. Connection through and elastic network results in positive motion correlation up to large distances both in the longitudinal and transverse directions (Fig. \[fig:Dii\_ss\_1\](a)). This long range correlation reflects the motion of the network’s center of mass and should therefore persist throughout the sample. Statistics are insufficient to obtain a reliable measure of the functional form of this long range response. At small inter-motor separations statistics is also low, since we observe, most commonly, separations of 5-50 $\mu$m. Interestingly, in cases where the short distance response measurement is reliable, i.e., $D_{\perp}$ in Fig.  \[fig:Dii\_ss\_1\](a) and $D_{\parallel}$ of M/A=0.02 in Fig.  \[fig:Dii\_ss\_1\](b), $D_{\parallel}$ and $D_{\perp}$ increase at short distances to some maximal value before decreasing down to the long range value. At very short distances negative correlations may arise. These can be attributed to a lack of statistics, to crosslinking sites that sustain local torques, or to local contraction effects within the network between nearby motors. We focus on the stronger signal $D_{\parallel}$ to demonstrate the effect of motor concentration (Fig. \[fig:Dii\_ss\_1\](b)). A similar behavior of $D_{\parallel}$ as a function of distance is seen for the various myosin concentrations, and the correlations at large distances are essentially equal. However, at intermediate distances, the lower the myosin concentration the higher the 2P correlations. This result probably reflects the higher probability of the two motors to be affected by the same third motor at smaller motor concentrations, as the inter-motor distance increases with the decrease in motor concentration. We estimate the inter-motor distance $\Delta x_m\approx 10,30,40~\mu$m for \[myosin\]/\[actin\]=0.02,0.0067,0.005 respectively, which supports this interpretation. Conclusions =========== In this review we suggest two point motion correlation of embedded beads as a measure of stress propagation through complex materials. We demonstrate our approach in studying stress propagation in model cytoskeleton networks including actin gels at thermal equilibrium, actomyosin networks which arrive at a mechanical steady state, and fast evolving actomyosin networks. Furthermore, we show that structural information is encoded into the stress propagation signal and can be extract from experiments by comparison to the solution of the stokes problem in a relevant model for the specific complex fluid in question. Stress propagation was characterized here both through the bulk material (Sec. \[passive\] and \[active\]), and directly through the polymer network (Sec. \[evolving\]). The amount of statistics required in order to get a good signal of stress propagation through the polymer network was several orders of magnitude smaller than what was required to recover the propagation signal through the bulk. This is due to the much stronger correlations in motion of two tracers connected directly to an elastic object. In cells, where mechanical signals are commonly used, it is convenient that perturbations applied on the cytoskeleton protein directly propagate well to long distances, while their effect on the surrounding fluid decays fast ($\sim r^{-3}$). Long range perturbations generated by myosin II motors were previously proposed to promote dynamic motor-mediated attraction and fusion of actin bundles. These perturbation were suggested to propagate via a 2D elastic actin network to which the bundles are coupled and via the surrounding fluid[@Gilo2009]. The work reviewed here focused on the linear response of a material to thermal and small active perturbations. Cytoskeleton networks, however, have non-linear elastic properties [@Gardel2006], as do many complex fluids (e.g., [@Waitukaitis2012]). A natural extension of our approach is to actively and strongly perturb one of the tracer particles to characterize the non-linear response of complex fluids, as done in active microrheology. In this method a tracer particle is externally driven, for example by means of a magnetic field [@Amblard1996; @Choi2011] or optical tweezer [@Mizuno2007; @Mizuno2008; @Lee2010]. The response of the media is then separated to an in-phase part reflecting the elastic shear modulus and an out-of-phase part which reflect the loss modulus. The control over the amplitude and strain rate of the mechanical perturbation is essential for probing the non-linear response of the material. There are a few reposts on active microrheology in actin networks. It was shown that active microrheolgy of actin networks at low strain amplitude agrees well with passive microrheology measurements [@Mizuno2008; @Lee2010]. As expected, at larger strain amplitudes, inaccessible to passive microrheology, the known non-linear stiffening of actin networks [@Gardel2006] can be observed by active microrheology [@Lee2010]. Active microrheology of active actomyosin networks was used to characterize the athermal fluctuations of such networks [@Mizuno2007] at the linear mechanical response regime. Generalizing our approach to further study actin networks focusing on their non linear response with and without myosin is expected to yield a better understanding the nature of stress transmission in these systems. References {#references .unnumbered} ========== [10]{} url \#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{} Dominguez R and Holmes K C 2011 [*Annual review of biophysics*]{} [**40**]{} 169–86 Pollard T D and Cooper J A 2009 [*Science (New York, N.Y.)*]{} [**326**]{} 1208–12 Cooper J A 1991 [*Annual review of physiology*]{} [**53**]{} 585–605 ISSN 0066-4278 Blanchoin L, Boujemaa-Paterski R, Sykes C and Plastino J 2014 [ *Physiological reviews*]{} [**94**]{} 235–63 Otterbein L R, Graceffa P and Dominguez R 2001 [*Science (New York, N.Y.)*]{} [**293**]{} 708–11 Wegner A and Isenberg G 1983 [*Proceedings of the National Academy of Sciences of the United States of America*]{} [**80**]{} 4922–5 Lodish H, Berk A, Zipursky S, Matsudaira P, Baltimore D and Darnell J 2000 [The Dynamics of Actin Assembly]{} [*Molecular Cell Biology*]{} (W. H. Freeman) Schafer D A, Jennings P B and Cooper J A 1996 [*The Journal of cell biology*]{} [**135**]{} 169–79 Pollard T D and Cooper J A 1986 [*Ann. Rev. Biochem.*]{} [**55**]{} 987–1035 C 1999 [*[Actin Filaments. Biophysics textbook online]{}*]{} Reymann A C, Boujemaa-Paterski R, Martiel J L, Gu[é]{}rin C, Cao W, Chin H F, [De La Cruz]{} E M, Th[é]{}ry M and Blanchoin L 2012 [*Science (New York, N.Y.)*]{} [**336**]{} 1310–4 Haviv L, Gillo D, Backouche F and Bernheim-Groswasser A 2008 [*J. Mol. Biol.*]{} [**375**]{} 325–330 Ideses Y, Sonn-Segev A, Roichman Y and Bernheim-Groswasser A 2013 [*Soft Matter*]{} [**9**]{} 7127 Backouche F, Haviv L, Groswasser D and Bernheim-Groswasser A 2006 [*Physical biology*]{} [**3**]{} 264–73 Lieleg O, Claessens M M A E and Bausch A R 2010 [*Soft Matter*]{} [**6**]{} 218–225 Strelnikova N, Herren F, Schoenenberger C A and Pfohl T 2016 [*Frontiers in Materials*]{} [**3**]{} 20 Hartman M A, Spudich J A, Abouhamed M, Grobe K, I V and Thelen f X 2012 [ *Journal of cell science*]{} [**125**]{} 1627–32 Syamaladevi D P, Spudich J A and Sowdhamini R 2012 [*Bioinformatics and biology insights*]{} [**6**]{} 11–21 Desnos C, Huet S, Darchen F 2007 [*Biol Cell.*]{} [**99**]{} 411–-23 Kuhne,W. (1864). Untersuchungen uber das Protoplasma und die Contractilitat. Leipzig: W. Engelmann Reisler E, Smith C and Seegan G 1980 [*Journal of Molecular Biology*]{} [ **143**]{} 129–145 Kaminer B and Bell A L 1966 [*Journal of Molecular Biology*]{} [**20**]{} 391–401 Koretz J F 1979 [*Biophysical journal*]{} [**27**]{} 433–46 Davis J S 1988 [*Annual Review of Biophysics and Biophysical Chemistry*]{} [**17**]{} 217–239 Shutova M, Yang C, Vasiliev J M, Svitkina T, Gautel M, Parsons J and Horwitz 2012 [*PLoS ONE*]{} [**7**]{} e40814 Davis J S, Buck J and Greene E P 1982 [*FEBS Letters*]{} [**140**]{} 293–297 Koretz J F 1982 [*Methods in enzymology*]{} [**85 Pt B**]{} 20–55 Pollard T D 1982 [*The Journal of cell biology*]{} [**95**]{} 816–25 Katsura I and Noda H 1973 [*Journal of biochemistry*]{} [**73**]{} 245–56 Sweeney H L and Houdusse A 2010 [*Annual review of biophysics*]{} [**39**]{} 539–57 Finer J T, Simmons R M and Spudich J A 1994 [*Nature*]{} [**368**]{} 113–119 R[ü]{}egg C, Veigel C, Molloy J E, Schmitz S, Sparrow J C and Fink R H A 2002 [*News in physiological sciences : an international journal of physiology produced jointly by the International Union of Physiological Sciences and the American Physiological Society*]{} [**17**]{} 213–8 Kaya M and Higuchi H 2013 [*Cellular and molecular life sciences : CMLS*]{} [**70**]{} 4275–92 Lieleg O, Schmoller K M, Claessens M M A E and Bausch A R 2009 [*Biophys. J.*]{} [**96**]{} 4725–4732 Haviv L, Gov N, Ideses Y and Bernheim-Groswasser A 2008 [*European Biophysics Journal*]{} [**37**]{} 447–454 ISSN 1432-1017 Bendix P M, Koenderink G H, Cuvelier D, Dogic Z, Koeleman B N, Brieher W M, Field C M, Mahadevan L and Weitz D A 2008 [*Biophysical journal*]{} [ **94**]{} 3126–36 Kohler S, Schaller V and Bausch A R 2011 [*Nat Mater*]{} [**10**]{} 462–468 Smith D, Ziebert F, Humphrey D, Duggan C, Steinbeck M, Zimmermann W and K[ä]{}s J 2007 [*Biophysical journal*]{} [**93**]{} 4445–52 Murrell M P and Gardel M L 2012 [*Proceedings of the National Academy of Sciences*]{} [**109**]{} 20820–20825 M, Depken M, Stuhrmann B, Korsten M, MacKintosh F C and Koenderink G H 2011 [*Proc Natl Acad Sci U S A*]{} [**108**]{} 9408–9413 Vogel S K, Petrasek Z, Heinemann F and Schwille P 2013 [*eLife*]{} [**2**]{} e00116 Thoresen T, Lenz M and Gardel M 2013 [*Biophysical Journal*]{} [**104**]{} 655–665 Stuhrmann B, [Soares e Silva]{} M, Depken M, MacKintosh F C and Koenderink G H 2012 [*Physical Review E*]{} [**86**]{} 020901 Koenderink G H, Dogic Z, Nakamura F, Bendix P M, MacKintosh F C, Hartwig J H, Stossel T P and Weitz D A 2009 [*Proceedings of the National Academy of Sciences of the United States of America*]{} [**106**]{} 15192–15197 Liverpool T B, Marchetti M C 2005, [*Europhys. Lett.*]{} [**69**]{}, 846 Dasanayake N L, Michalski P J, and Carlsson A E 2011, [*Phys. Rev. Lett.*]{} [**107**]{}, 118101 Lenz M, Thoresen T, Gardel M L, and Dinner A R 2012, [*Phys. Rev. Lett.*]{} [**108**]{}, 238107 MacKintosh F C and Levine A J 2008, [*Phys. Rev. Lett.*]{} [**100**]{}, 018104 Mizuno D, Tardin C, Schmidt C F, and Mackintosh F C 2007, [*Science*]{} [**315**]{}, 370 Lenz M 2014 [*Physical Review X*]{} [**4**]{} 041002 Alvarado J, Sheinman M, Sharma A, MacKintosh F C and Koenderink G H 2013 [ *Nature Physics*]{} [**9**]{} 591–597 Wang S and Wolynes P G 2012 [*Proceedings of the National Academy of Sciences*]{} [**109**]{} 6446–6451 Style R W, Boltyanskiy R German G K, Hyland C, MacMinn C W, Mertz A F, Wilen L A, Xu Y and Dufresne E R 2014, [*Soft Matter*]{} [**10**]{} 4047–4055 Ladam G, Vonna L and Sackmann E 2003, [*J. Phys. Chem. B*]{} [**107**]{} 8965–8971 Landau L D and Lifshitz E M 1986 [*Theory of Elasticity*]{} 3rd ed (Oxford: Pergamon Press) Crocker J C, Valentine M T, Weeks E R, Gisler T, Kaplan P D, Yodh A G and Weitz D A 2000 [*Phys. Rev. Lett.*]{} [**85**]{} 888–891 Mason T G and Weitz D A 1995 [*Phys. Rev. Lett.*]{} [**74**]{} 1250–1253 Mason T G, Ganesan K, van Zanten J H, Wirtz D and Kuo S C 1997 [*Phys. Rev. Lett.*]{} [**79**]{} 3282–3285 Squires T M and Mason T G 2010 [*Annu. Rev. Fluid Mech.*]{} [**42**]{} 413–438 Sonn-Segev A, Bernheim-Groswasser A, Diamant H and Roichman Y 2014 [ *Physical Review Letters*]{} [**112**]{} 088301 Valentine M, Kaplan P, Thota D, Crocker J, Gisler T, Prud’homme R, Beck M and Weitz D 2001 [*Physical Review E*]{} [**64**]{} 061506 Shin J H, Gardel M L, Mahadevan L, Matsudaira P and Weitz D A 2004 [*Proc. Natl. Acad. Sci. U.S.A*]{} [**101**]{} 9636–9641 Meiners J C and Quake S R 1999 [*Phys. Rev. Lett.*]{} [**82**]{}(10) 2211–2214 Blake J R 1971 [*Proc. Camb. Philos. S-M*]{} [**70**]{} 303 Pozrikidis C 1992 [*Boundary Integral and Singularity Methods for Linearized Viscous Flow*]{} (New York: Cambridge University Press) Dufresne E R, Squires T M, Brenner M P and Grier D G 2000 [*Phys. Rev. Lett.*]{} [**85**]{} 3317–3320 Diamant H 2009 [*Journal of the Physical Society of Japan*]{} [**78**]{} 041002 Cui B, Diamant H, Lin B and Rice S A 2004 [*Phys. Rev. Lett.*]{} [**92**]{} 258301 Cui B, Diamant H and Lin B 2002 [*Physical Review Letters*]{} [**89**]{} 1–4 Sonn-Segev A, B[ł]{}awzdziewicz J, Wajnryb E, Ekiel-Je[ż]{}ewska M L, Diamant H and Roichman Y 2015 [*The Journal of chemical physics*]{} [**143**]{} 074704 Prasad V and Weeks E R 2009, [*Phys. Rev. Lett.*]{} [**102**]{} 178302 Sonn-Segev A, Bernheim-Groswasser A and Roichman Y 2014 [*Soft matter*]{} [**10**]{} 8324–8329 Happel J and Brenner H 1991 [*[Low Reynolds Number Hydrodynamics]{}*]{} (Dordrecht: Kluwer) Diamant H 2015 [*The European Physical Journal E*]{} [**38**]{} 1–11 Gardel M L, Valentine M T, Crocker J C, Bausch A R and Weitz D A 2003 [ *Phys. Rev. Lett.*]{} [**91**]{} 158302 Liu J, Gardel M L, Kroy K, Frey E, Hoffman B D, Crocker J C, Bausch A R and Weitz D A 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 118104 Gennes P G D 1976 [*Macromolecules*]{} [**9**]{} 587–593 Gennes P G D 1976 [*Macromolecules*]{} [**9**]{} 594–598 Milner S T 1993 [*Phys. Rev. E*]{} [**48**]{}(5) 3674–3691 Levine A J and Lubensky T C 2000 [*Phys. Rev. Lett.*]{} [**85**]{} 1774–1777 Levine A J and Lubensky T C 2001 [*Phys. Rev. E*]{} [**63**]{}(4) 041510 Pelletier V, Gal N, Fournier P and Kilfoil M 2009 [*Physical Review Letters*]{} [**102**]{} 188303 Sonn-Segev A Bernheim-Groswasser A R Y 2035 (*Preprint* ) Rubinstein M and Colby R H 2003 [Polymer physics]{} Cooper G M and Hausman R E 2000 [*[The Cell: A Molecular Approach 2nd Edition]{}*]{} ISBN 0878932194 Gillo D, Gilboa B, Gurka R and Bernheim-Groswasser A 2009, [*Phys Biol.*]{} [**6**]{} 036003 Gardel M L, Nakamura F, Hartwig J H, Crocker J C, Stossel T P and Weitz D A 2006 [*Proc. Natl. Acad. Sci. U.S.A*]{} [**103**]{} 1762–1767 Waitukaitis S R and Jaeger H M 2012 [*Nature*]{} [**487**]{} 205–209 Amblard F, Maggs A C, Yurke B, Pargellis A N and Leibler S 1996 [*Physical Review Letters*]{} [**77**]{} 4470–4473 Choi S, Steltenkamp S, Zasadzinski J and Squires T 2011 [*Nature Communications*]{} [**2**]{} 312 Mizuno D, Head D A, MacKintosh F C and Schmidt C F 2008 [*Macromolecules*]{} [**41**]{} 7194–7202 Lee H, Ferrer J M, Nakamura F, Lang M J and Kamm R D 2010 [*Acta Biomaterialia*]{} [**6**]{} 1207–1218

ArXiv

--- abstract: 'We present a scheme for correcting for crosstalk- and noise-induced errors in exchange-coupled singlet-triplet semiconductor double quantum dot qubits. While exchange coupling allows the coupling strength to be controlled independently of the intraqubit exchange couplings, there is also the problem of leakage, which must be addressed. We show that, if a large magnetic field difference is present between the two qubits, leakage is suppressed. We then develop pulse sequences that correct for crosstalk- and noise-induced errors and present parameters describing them for the 24 Clifford gates. We determine the infidelity for both the uncorrected and corrected gates as a function of the error-inducing terms and show that our corrected pulse sequences reduce the error by several orders of magnitude.' author: - Donovan Buterakos - 'Robert E. Throckmorton' - 'S. Das Sarma' title: 'Error correction for gate operations in systems of exchange-coupled singlet-triplet qubits in double quantum dots' --- Introduction ============ Semiconductor-based electron spin qubits are one of several platforms that are currently being actively investigated, both theoretically and experimentally, with the goal of eventually building a quantum computer. These types of qubits are formed from one or more quantum dots with a single electron trapped inside each. A number of different varieties of semiconductor electron spin qubits exist, such as the single-spin exchange qubit[@LossPRA1998; @HuPRA2000; @HuPRA2001; @NowackScience2011; @PlaNature2012; @BraakmanNatNano2013; @PlaNature2013; @VeldhorstNatNano2014; @OtsukaSciRep2016; @ItoSciRep2016], the singlet-triplet double-dot qubit[@LevyPRL2002; @PettaScience2005; @FolettiNatPhys2009; @VanWeperenPRL2011; @MauneNature2012; @ShulmanScience2012; @DialPRL2013; @ShulmanNatCommun2014; @ReedPRL2016; @MartinsPRL2016], the triple-dot exchange-only qubit[@DiVincenzoNature2000; @MedfordNatNano2013; @MedfordPRL2013; @EngSciAdv2015; @ShimPRB2016], and a double-dot “hybrid” qubit with three electrons total[@ShiPRL2012; @KimNature2014; @KimNPJQI2015]. These various spin qubit platforms have been realized in both Si-based and GaAs-based architectures. Other quantum computing platforms include superconducting transmons and trapped ions. While other platforms currently have better coherence times and higher fidelity, allowing a larger number of gates to be performed, semiconductor electron spin qubits allow for faster gates and are compatible with the existing semiconductor industry, thus allowing easier scaling up for eventual real applications beyond just laboratory demonstrations of principles. The greatest challenge for semiconductor-based qubits currently is thus improving coherence time and fidelity. Much progress has been made on this issue, with experiments on singlet-triplet qubits reporting fidelities as high as 99% for single-qubit gates and 90% for two-qubit gates. This still falls short of the goal of at least 99% required to implement surface codes[@FowlerPRA2012], let alone the 99.99% fidelity in all operations required before other error-correcting techniques can begin to be implemented. This unfortunately compares unfavorably with superconducting qubits, which claim fidelities of over 99.9% for single-qubit gates and over 99% for two-qubit gates[@BarendsNature2014], and with ion trap qubits, which claim similar fidelities[@BenhelmNatPhys2008]. However, both types of qubits have gate times on the order of $\mu$s, compared to ns for semiconductor-based electron spin qubits. The fact that so much experimental progress has been made in semiconductor spin qubits over the last decade is therefore promising. The subject is currently highly active with around ten large interdisciplinary groups being involved in spin qubit experiments all over the world. We will focus on the singlet-triplet qubit in this work. These qubits consist of two quantum dots, each with a single electron trapped inside, coupled via exchange coupling and subject to a magnetic field gradient (i.e., magnetic fields that differ at each dot). The exchange coupling can be varied purely electrically, allowing for fast manipulation of the qubit. Typically, the field gradient is created by either depositing a micromagnet nearby or by polarizing the nuclear spins (if possible) and cannot be changed quickly, so it is held constant. However, a recent work[@CollardArXiv] attempts to realize this gradient electrically by tuning the effective $g$ factor within each quantum dot, which would allow for fast manipulation of this gradient. We will assume, however, that this gradient is held constant, as this is the case in most experiments. We will see later that the form of the effective Hamiltonian for this qubit restricts us to rotations about axes in one quadrant of the $xz$ plane, and furthermore does not allow pure $x$ or $z$ rotations, meaning that rotations about other axes must be performed with complex pulse sequences. Not only does this mean that most rotations require more complex pulse sequences to perform, but also that we cannot implement common error correction schemes such as the NMR-inspired Hahn echo technique or its generalization, the Carr-Purcell-Meiboom-Gill (CPMG) technique[@WitzelPRL2007; @WitzelPRB2007; @LeePRL2008]. Other techniques are thus required, and have in fact been developed. We will assume, consistent with the experimental situation, that our qubits are subject to two types of noise—magnetic field noise, present in the field gradient, and charge noise, which manifests as noise in the exchange coupling. We will assume throughout this work that the noise is quasistatic, which is often a good approximation. While the idea of dynamical decoupling in semiconductor spin qubits through specially-designed pulse sequences to improve retention of the qubit state has been around for years[@BluhmPRL2010; @BluhmNatPhys2011; @SergeevichPRA2011; @ShulmanNatCommun2014; @MuhonenNatNanotechnol2014; @MalinowskiNatNanotechnol2017], the problem of error correction in the performance of gate operations was first considered in Ref. , in which the technique of Soft Uniaxial Positive Control for Orthogonal Drift Error (<span style="font-variant:small-caps;">supcode</span>) was introduced, and further developed in later work[@WangPRA2014; @ThrockmortonPRB2017]. This technique makes use of square pulses (i.e., a piecewise constant $J$) both to implement gates and to correct errors to arbitrary order. An even more recent work[@ZengArXiv] considers error correction to arbitrary order using smooth pulses, but considers only magnetic field noise. These works considered singlet-triplet qubits in isolation, without any interqubit coupling. To build a practical quantum computer, one must couple qubits together so that multiqubit gates can be performed as well. The ability to perform arbitrary single-qubit gates and at least one two-qubit gate are necessary for universal quantum computation. The effects of noise on such two-qubit operations has in fact been considered in several works[@WuPRB2017; @WangPRA2014; @WangNPJQI2015; @CalderonVargasPRL2017] (the last of these is actually platform-independent). Such coupling, unfortunately, also introduces a new challenge—crosstalk. The fact that the qubits are coupled means that, while a single-qubit operation is being performed, the changes in the qubit’s state also cause unintended changes in nearby qubits. As a result, we need to correct for crosstalk as well. There are (at least) two ways to couple singlet-triplet qubits. One method is through capacitive coupling, in which two qubits are coupled via interaction of their dipole moments. The singlet state of a singlet-triplet qubit possesses a nonzero dipole moment, but not the triplet state, resulting in a state-dependent electrostatic interaction. This interaction is empirically found[@ShulmanScience2012] to be proportional to the exchange couplings in the two qubits. Correcting for both noise- and crosstalk-induced error in single-qubit gate operations was the subject of a recent work of ours[@ButerakosPRB2018], which uses a technique based on <span style="font-variant:small-caps;">supcode</span> to cancel errors in single-qubit operations to leading order. The other coupling method, which is also used to couple singlet-triplet qubits, is exchange coupling, in which one of the spins in one qubit is coupled to a spin in the other qubit. This has the advantage of allowing independent control of the intraqubit exchange couplings and the interqubit coupling. However, this also allows for leakage of the system out of the computational subspace; this coupling could, for example, put the system into the state, $\ket{\uparrow\uparrow\downarrow\downarrow}$. The subject of correcting errors in single-qubit gates performed on one of a pair of exchange-coupled singlet-triplet qubits is the focus of the present work. The system under consideration is a pair of exchange-coupled singlet-triplet qubits with identical magnetic field gradients, but subject to differing overall magnetic fields. As has been shown, it is not necessary that the magnetic field gradients of the qubits be identical, but we shall assume that they are for simplicity. The differing overall magnetic fields are crucial to allowing us to tune the energy cost of entering the two possible leakage states, $\ket{\uparrow\uparrow\downarrow\downarrow}$ and $\ket{\downarrow\downarrow\uparrow\uparrow}$. We split the Hamiltonian of this system into two parts, one that only connects states within the computational subspace and the magnetic field-induced separation of the leakage states, and those which are responsible for leakage. We then apply perturbation theory to show that a sufficiently large magnetic field can suppress leakage and argue that we can ignore the leakage terms for the remainder of this work. We then outline our approach for developing pulse sequences that correct for noise- and crosstalk-induced errors. This approach is similar to that of Ref. ,where the case of capacitive interqubit coupling was addressed (in contrast to the current work addressing interqubit exchange coupling). It consists of following the naïve single-qubit gate with identity operations on both qubits arranged in such a way as to cancel the error in the gate to first order. We add these pulses in “blocks” consisting of two pulses on each qubit, arranged such that both pairs have the same duration. This differs from Ref.  since, in this previous work, the “blocks” instead consisted of two pulses on one qubit and a single pulse on the other. We find that the arrangement of the “blocks” used in this work allow for shorter pulse sequences, both in time and in number of “blocks” needed. We obtain the parameters needed to correct errors in the naïve pulse sequences introduced in Ref.  and show that these sequences reduce the errors (infidelities) in the gates by [*several orders of magnitude*]{}. However, we find that, for low noise and low coupling, the error still scales linearly, only transitioning to a quadratic dependence for larger values. This is due to some residual leakage, since we neglect the effect of the terms that cause it. We then point out, however, that the range over which we obtain this linear dependence can be reduced by increasing the magnetic field difference between the qubits. The rest of the manuscript is organized as follows. Section II introduces the Hamiltonian that we use in the rest of this work and quantifies the magnitude of the effect of leakage out of the computational subspace. In Sec. III, we present the formulas for the first-order error due to crosstalk and noise. Section IV describes the methods that we use to correct for these errors, and we describe our results in Sec. V. We give our conclusions in Sec. VI, and present the detailed numerical parameters for our error-corrected pulse sequences in the Appendix. Derivation of the Hamiltonian ============================= In the singlet-triplet encoding scheme, each qubit is encoded in the spin degrees of freedom of two electrons bound in a pair of quantum dots. An external magnetic field is applied along the $z$-axis, conserving the $z$-component of the total spin $S_{1z}+S_{2z}$, and thus the computational space is given by the space with $S_z=0$. We define the basis states as $\ket{0}=\frac{\ket{\uparrow\downarrow}+\ket{\downarrow\uparrow}}{\sqrt{2}}$, $\ket{1}=\frac{\ket{\uparrow\downarrow}-\ket{\downarrow\uparrow}}{\sqrt{2}}$, though other conventions exist. The effective Hamiltonian for a single qubit in this basis is given by $$H=hX+J(t)Z,\label{eqn:singleh}$$ where $X$ and $Z$ are the Pauli matrices, $h$ is the magnetic field gradient between the two quantum dots, and $J$ is the exchange coupling. The strength of the magnetic field gradient can only be varied slowly compared to the qubit coherence time, and so we treat $h$ as constant, controlling the qubit only by varying the strength of $J$. Critically, the sign of $J$ remains the same, meaning that only forward rotations about axes lying in one quadrant of the $xz$-plane can be performed. We now consider a system of two singlet-triplet qubits consisting of a linear array of four quantum dots with exchange coupling between nearest neighbors. Other geometries can also be considered, but all will lead to a Hamiltonian of the same basic form. The exchange coupling between the middle two quantum dots serves as a controllable $X_1X_2$ coupling between the two qubits, allowing for two-qubit gates to be performed[@LiPRB2012]. Unfortunately, the presence of this coupling interferes with single qubit operations by introducing crosstalk between the two qubits. While this can be mitigated by simply reducing the strength of the interaction, it can not be entirely eliminated, since the interaction cannot be completely turned off. In addition, for future quantum computing applications with many qubits working together in a circuit, it is essentially impossible to turn off all inter-qubit couplings while performing single qubit operations on specific qubits. This exchange coupling can also potentially introduce leakage outside of the computational subspace. Although the $z$-component of total spin is still conserved, the mixing of spin states between qubits allows the system to enter, e.g., the state $\ket{\uparrow\uparrow\downarrow\downarrow}$, and conversely. However, leakage can be guarded against by applying a large magnetic field difference between the two ST qubits[@LiPRB2012]. We quantify exactly to what degree this protects against leakage by using perturbation theory. Let us number the spins in one of our qubits 1 and 2, and those in the other qubit 3 and 4. We now add the exchange coupling, $J_{23}\vec{S}_2\cdot\vec{S}_3$, to the Hamiltonian for the full system, thus coupling the two qubits. We now expand this term out in terms of the $x$, $y$, and $z$ components of the spins. The term $J_{23}S_{2z}S_{3z}$ simply allows us to perform two-qubit operations; in terms of our computational basis, this term becomes $\tfrac{1}{4}J_{23}X_1X_2$. Unfortunately, this is also the term that results in crosstalk between the two qubits. The terms involving the $x$ and $y$ components, on the other hand, cause leakage out of the computational subspace; it is the effect of these terms that we now quantify. In particular, we will show that a large magnetic field difference between the two qubits helps to suppress leakage. Let $B$ be the energy splitting of each leakage state due to the magnetic field difference between the two qubits. Denote the basis states of the computational subspace by $\ket{n}$, where $n$ runs from 1 to 4, and denote the two leakage states by $\ket{\xi}$, where $\xi$ runs over the two states $\ket{\uparrow\uparrow\downarrow\downarrow}$ and $\ket{\downarrow\downarrow\uparrow\uparrow}$. Let $H_4$ denote the terms of the Hamiltonian that connect states within the 4-dimensional computational subspace, and, for simplicity, assume the basis $\ket{n}$ diagonalizes $H_4$, so $H_4\ket{n}=E_n\ket{n}$. We let the unperturbed Hamiltonian be given by $H_4$ plus the terms corresponding to the energies of the leakage states, and we let the perturbation be given by the terms that couple the computational states to the leakage states which we label $k_{n\xi}$ as follows: $$H_0=H_4+B\ket{\uparrow\uparrow\downarrow\downarrow}\bra{\uparrow\uparrow\downarrow\downarrow}-B\ket{\downarrow\downarrow\uparrow\uparrow}\bra{\downarrow\downarrow\uparrow\uparrow}$$ $$H'=\sum_{n,\xi}k_{n\xi}\ket{\xi}\bra{n}+k_{n\xi}^*\ket{n}\bra{\xi}$$ The relative energy scales are such that $B\gg E_n\gg k_{n\xi}$. Then using perturbation theory, the leading order corrections to the eigenstates and energies are as follows: $$\ket{n^1}=\sum_\xi\frac{\braket{\xi^0|H'|n^0}}{E_n^0-E_\xi^0}\ket{\xi^0}=\sum_\xi\frac{k_{n\xi}}{E_n\pm B}\ket{\xi^0}=O\bigg(\frac{k_{n\xi}}{B}\bigg)$$ $$E_n^2=\sum_\xi\frac{|\braket{\xi^0|H'|n^0}|^2}{E_n^0-E_\xi^0}=\sum_\xi\frac{|k_{n\xi}|^2}{E_n\pm B}=O\bigg(\frac{k_{n\xi}^2}{B}\bigg) \label{eqn:pertenergy}$$ Thus, if a strong magnetic field difference is applied between the two singlet-triplet qubits, the effect of leakage due to exchange coupling is of order $k/B$. This is sufficiently small to ignore for applications of single-qubit gates, since $B$ can be made large and the value of $k$ is already minimized to reduce crosstalk. Since a sufficiently large magnetic field difference makes any terms which lead to leakage become negligible, we need only consider the projection of the exchange term $J_{23}\vec{S}_2\cdot\vec{S}_3$ into the computational space. For simplicity, we define $k=-J_{23}/4$, so that the interaction takes the form $kX_1X_2$. Thus the effective Hamiltonian is given by $$H_4=h_1X_1+h_2X_2+J_1Z_1+J_2Z_2+kX_1X_2.$$ First order expansions of crosstalk and noise ============================================= We closely follow the technique developed in Ref. , to perform an expansion of the evolution operator $e^{-itH}$ to first order in $k$. This expansion uses commutation relations of Pauli matrices in order to separate the single-qubit parts of the evolution operator, $e^{-it(h_iX_i+J_iZ_i)}$, from the two-qubit cross terms. To simplify the resulting expressions, we use the shorthand $a_i=\sqrt{h_i^2+J_i^2}$, and introduce the following rotated basis, chosen so that the single qubit evolution for qubit $i$ reduces to an $X_i'$ rotation. $$\begin{aligned} \begin{split} &X_i'=(h_iX_i+J_iZ_i)/a_i\\ &Y_i'=Y_i\\ &Z_i'=(-J_iX_i+h_iZ_i)/a_i \label{eqn:paulirotation} \end{split}\end{aligned}$$ This expansion allows the evolution operator to be written in the form $$e^{-itH}=e^{-it(a_1X_1'+a_2X_2')}\left (1-i\sum_{i'j'}\Delta_{i'j'}^{\text{ex}}\sigma_{i'}\otimes\sigma_{j'}\right ), \label{eqn:concisekerror}$$ where $\sigma_{i'}$ runs over the rotated Pauli matrices $X_1'$, $Y_1'$, $Z_1'$, and similarly $\sigma_{j'}$ runs over $X_2'$, $Y_2'$, $Z_2'$. Performing the expansion yields the coefficients $\Delta_{i'j'}^{\text{ex}}$ as follows: $$\begin{aligned} \Delta_{X_1'X_2'}^{\text{ex}}=k&\frac{h_1h_2}{a_1a_2}t,\nonumber\\ \begin{split} \Delta_{Y_1'Y_2'}^{\text{ex}}=k&\frac{J_1J_2}{2a_1a_2}\Big\{-t\operatorname{sinc}{[2(a_1+a_2)t]}\\&+t\operatorname{sinc}{[2(a_1-a_2)t]}\Big\}, \end{split}\nonumber\\ \begin{split} \Delta_{Z_1'Z_2'}^{\text{ex}}=k&\frac{J_1J_2}{2a_1a_2}\Big\{t\operatorname{sinc}{[2(a_1+a_2)t]}\\&+t\operatorname{sinc}{[2(a_1-a_2)t]}\Big\}, \end{split}\nonumber\\ \Delta_{X_1'Y_2'}^{\text{ex}}=-&k\frac{h_1J_2}{a_1}t^2\operatorname{sinc}^2{(a_2t)},\nonumber\\ \Delta_{X_1'Z_2'}^{\text{ex}}=-&k\frac{h_1J_2}{a_1a_2}t\operatorname{sinc}{(2a_2t)},\nonumber\\ \begin{split} \Delta_{Y_1'Z_2'}^{\text{ex}}=k&\frac{J_1J_2}{2a_1a_2}\Big\{(a_1+a_2)t^2\operatorname{sinc}^2{[(a_1+a_2)t]}\\&+(a_1-a_2)t^2\operatorname{sinc}^2{[(a_1-a_2)t]}\Big\}. \end{split} \label{eqn:kerror}\end{aligned}$$ The other three terms ($\Delta_{Y_1'X_2'}^{\text{ex}}$, $\Delta_{Z_1'X_2'}^{\text{ex}}$, and $\Delta_{Z_1'Y_2'}^{\text{ex}}$) are omitted for the sake of brevity, but can be obtained by interchanging the subscripts 1 and 2 in the expressions above. Since $J_i$ is time dependent, the rotated matrices $X_i'$, $Y_i'$, $Z_i'$ are also time dependent, and thus are transformed back into the standard basis before proceeding further into the calculation. In addition to crosstalk, the singlet-triplet qubit system encounters noise from two separate sources, namely, field and charge noise. We work in the quasistatic approximation, where we assume that the noise changes slowly compared to the gate implementation time. These noise sources enter into the Hamiltonian via corrections to the values $h_i$ and $J_i$. Since the magnetic field itself does not vary over the implementation of the gate, the field noise simply causes a small constant term $\textit{dh}_i$ to be added to $h_i$. Charge noise, however, is more subtle, since the value of $J_i$ is changed over the course of the pulse in order to implement each specific gate. If the exchange interaction is adjusted via detuning control, that is, by changing the energy difference $\epsilon_i$ between the two dots comprising a qubit, then the change in $J_i$ is given by $\textit{dJ}_i=\frac{\partial J_i}{\partial\epsilon_i}\,d\epsilon_i$, where $d\epsilon_i$ is caused by the charge noise and stays constant in the quasistatic limit. Empirically, $J_i$ is found to have an exponential dependence on $\epsilon_i$, at least in the regime of operation, which means that $\textit{dJ}_i$ is proportional to $J_i$. It is possible that some systems could have a different dependence of $J_i$ on $\epsilon_i$, but it has been shown that <span style="font-variant:small-caps;">supcode</span> can easily accommodate these other cases[@WangPRA2014]. For specificity, we consider the case of $dJ_i$ being proportional to $J_i$, i.e., $J_i$ being exponential in $\epsilon_i$. The expansion of the evolution operator for a single qubit to first order in noise terms has been performed[@KestnerPRL2013], and since the noise sources as well as the coupling between qubits are small, any effect of one on the other can be ignored. Thus we simply combine the crosstalk expansion terms we obtained with the previously derived noise terms $\Delta_i^{\text{q}n}$, so that, to first order, the full noisy evolution operator is given by $$\begin{aligned} \begin{split} &e^{-itH}=e^{-it(h_1X_1+J_1Z_1)}e^{-it(h_2X_2+J_2Z_2)}\times\\ &\bigg[1-i\sum_i(\Delta_i^{\text{q1}}\sigma_i\otimes 1 +\Delta_i^{\text{q2}}1\otimes\sigma_i) -i\sum_{ij}\Delta_{ij}^{\text{ex}}\sigma_i\otimes\sigma_j\bigg] \label{eqn:evolutionoperator} \end{split}\end{aligned}$$ with $\Delta_i^{\text{q}n}$ given by $$\begin{aligned} \begin{split} \Delta_x^{\text{q}n}=&\frac{2h_n^2a_nt+J_n^2\sin{2a_nt}}{2a_n^3} \mathit{dh}_n\\&+\frac{h_nJ_n(2a_nt-\sin{2a_nt})}{2a_n^3}\mathit{dJ}_n, \end{split}\nonumber\\ \begin{split} \Delta_y^{\text{q}n}=&\frac{J_n(\cos{2a_nt}-1)}{2a_n^2} \mathit{dh}_n\\&+\frac{h_n(1-\cos{2a_nt})}{2a_n^2}\mathit{dJ}_n, \end{split}\nonumber\\ \begin{split} \Delta_z^{\text{q}n}=&\frac{h_nJ_n(2a_nt-\sin{2a_nt})}{2a_n^3} \mathit{dh}_n\\&+\frac{2J_n^2a_nt+h_n^2\sin{2a_nt}}{2a_n^3}\mathit{dJ}_n. \end{split} \label{eqn:singleerror}\end{aligned}$$ The evolution operator given in Eq. , is of the form $U=R(1+\Delta)$, where $R$ is an ideal rotation of qubits 1 and 2 along particular axes of the Bloch sphere, and $\Delta$ is some small error produced by noise and crosstalk. Uncorrected rotations along axes in the first quadrant of the $xz$-plane can be performed simply by allowing the system to evolve at a fixed value of $J_1$ and $J_2$ for a given amount of time $t$. In order to perform rotations about other axes or to correct for crosstalk, it is necessary to perform several uncorrected rotations in a row (i.e., to allow the system to evolve for some time $t_1$, then change the values of $J_1$ and $J_2$ and allow the system to continue evolving for some time $t_2$, etc.). In this case, it is necessary to calculate the errors for each segment of the gate separately and combine the results, since the values of $J_1$ and $J_2$ are different for each part. To first order, the result of performing an uncorrected gate $U_2$ followed by $U_1$ is $$U_1U_2=R_1(1+\Delta_1)R_2(1+\Delta_2)=R_1R_2(1+R_2^\dagger\Delta_1R_2+\Delta_2). \label{eqn:adderrors}$$ For the pulses we generate, we will set $h_1=h_2=h$; however, the same method can be applied to systems where $h_1$ and $h_2$ are different. In either case, the noise terms $\mathit{dh}_1$ and $\mathit{dh}_2$ are independent. Error Cancellation ================== As with previous <span style="font-variant:small-caps;">supcode</span> pulses, our strategy involves applying an initial uncorrected gate followed by an uncorrected identity operation in such a way that the errors cancel out to leading order. Single qubit rotations have been studied extensively in Ref. , so for the initial uncorrected gate, we use the equations presented in that work to generate the desired rotation on qubit 1, and perform a $2\pi$ rotation on qubit 2 at the same time. The uncorrected identity operation which follows can be somewhat complex, and in fact, a sufficient degree of complexity is required in order for it to have the freedom to cancel all sources of error for an arbitrary gate. The general form of this identity operation is the key to generating fast, efficient pulse sequences. For a single-qubit system[@WangPRA2014], a sequence of interrupted $2\pi$ rotations was used, where the axes of rotation were tuned to precisely cancel out the error. Specifically, the total error of the pulse was calculated in terms of parameters $j_n$ corresponding to the values of $J$ at each part of the gate, and then the total error was set equal to zero, forming a complicated set of equations which were solved numerically for the parameters $j_n$. Correcting a two-qubit system is significantly more complicated because the amount of time needed to perform $2\pi$ rotations differs depending on the axis of rotation, and if qubits 1 and 2 are rotated about different axes concurrently, one operation will finish before the other. Freedom to vary the axes of rotation is needed to allow for numerical solutions which eliminate error, but varying these values can affect the timing between qubits 1 and 2, specifically which segment of the pulse on qubit 1 coincides with a given segment on qubit 2. This makes expressing the total error as a function of a set of parameters, a key step in the error correction process, very difficult. In order to avoid this problem, Ref.  uses a base identity operation consisting of two $2\pi$ rotations on one qubit, and a $4\pi$ rotation on the other, chosen such that the time taken to perform the $4\pi$ rotation equals the total time needed to perform both $2\pi$ rotations. While this form of an identity operation solves the problem of timing, it can be somewhat constricting, leading to longer, slower pulses. As stated previously, the major difficulty to performing dynamical decoupling on singlet-triplet qubit systems is the restriction of rotations to the first quadrant of the $xz$-plane. However, in order to generate efficient pulses, it is necessary to have a large amount of freedom in rotating qubits during the pulse. For singlet-triplet systems, this is best accomplished by allowing sequences with values of $J_i$ alternating between large and small. The scheme that uses a $4\pi$ rotation does not allow for strictly alternating values of $J_i$, since the basic pattern was a nested sequence of $2\pi$, $2\pi$, $4\pi$ rotations, corresponding to a cycle between small, large, and medium $J_i$. Additionally, the value of $J_i$ during the $4\pi$ rotation is completely constrained by the parameters chosen for the opposite qubit, meaning that the qubits cannot be individually controlled. ![The base uncorrected identity which consists of two $2\pi$ rotations on each qubit (top and bottom). $t_{1n}$ and $t_{1n}'$ define one $2\pi$ rotation on qubit 1, and this rotation is interrupted by the $2\pi$ rotation given by $t_{2n}$ and $t_{2n}'$. Similarly, $t_{3n}$, $t_{3n}'$, $t_{4n}$, and $t_{4n}'$ define two $2\pi$ rotations on qubit 2. The ellipsis in the middle indicates where copies of this base identity with larger values of $n$ interrupt this identity.[]{data-label="fig:improvedpulse"}](figimprovedpulse.pdf){width="\columnwidth"} We are able to improve upon the previous method by replacing the $4\pi$ rotation with two $2\pi$ rotations, while still enforcing the constraint that the total time of these two rotations must equal the total time of the two $2\pi$ rotations on the other qubit, as shown in the figure (\[fig:improvedpulse\]). This allows an alternating pattern between large and small values of $J_i$, and reduces the dependence of one qubit’s control on the other. To simplify some of the resulting equations, we use the times of each segment of the pulse as the free parameters, rather than the pulse height $j_i$ used in previous works. Ultimately these are equivalent, since for a fixed value of $h$, the strength of the exchange interaction $J_i$ forms a one-to-one correspondence with the time needed to perform a $2\pi$ rotation, as discussed below. We define times $t_{1n}$ through $t_{4n}$ and $t_{1n}'$ through $t_{4n}'$ as in Fig. \[fig:improvedpulse\], where the segments of the pulse corresponding to $t_{in}'$ and $t_{in}'$ form a single interrupted $2\pi$ rotation. The subscript $n$ distinguishes between the different nested copies of the base identity operation. We impose the constraint that $t_{1n}+t_{2n}=t_{3n}+t_{4n}$, and similarly for $t'$, meaning that each copy of the base identity has 6 independent parameters. The values of $j_{in}$ are determined in terms of the total time $t_{in}+t_{in}'$ by calculating what value of $j_{in}$ gives a $2\pi$ rotation which takes the given amount of time, as follows: $$j_{in}=\sqrt{\Big(\frac{\pi}{t_{in}+t_{in}'}\Big)^2-h^2}. \label{eqn:jprime}$$ To write the full sequence of operations comprising the nested identity operation, let $U(J_1,J_2,t)$ be an uncorrected rotation at the given values $J_1$, and $J_2$, for a time $t$. The sequence of operations before and after the interruption, shown on the left and right of Fig. \[fig:improvedpulse\], we denote as $A_n$ and $A_n'$ respectively. These each consist of three rotations: the first and third corresponding to the first and second $2\pi$ rotations on each qubit, and the middle corresponding to the overlap region where part of the first $2\pi$ rotation on one qubit is being performed concurrently with the second $2\pi$ rotation on the other. This overlap requires a selection statement dependent on which $2\pi$ rotations are overlapping, which is determined by the times $t_{1n}$ and $t_{3n}$. The sequences $A_n$ and $A_n'$ are given by $$A_n= \begin{cases} U(j_{2n},j_{4n},t_{2n})U(j_{1n},j_{4n},t_{1n}-t_{3n})\\ \qquad\qquad\qquad\times U(j_{1n},j_{3n},t_{3n})\text{\qquad\quad if $t_{1n}>t_{3n}$,}\\ U(j_{2n},j_{4n},t_{4n})U(j_{2n},j_{3n},t_{3n}-t_{1n})\\ \qquad\qquad\qquad\times U(j_{1n},j_{3n},t_{1n})\text{\qquad\quad if $t_{1n}<t_{3n},$} \end{cases} \label{eqn:idtleft}$$ $$A_n'= \begin{cases} U(j_{1n},j_{3n},t_{3n}')U(j_{1n},j_{4n},t_{1n}'-t_{3n}')\\ \qquad\qquad\qquad\times U(j_{2n},j_{4n},t_{2n}')\text{\qquad\quad if $t_{1n}'>t_{3n}'$,}\\ U(j_{1n},j_{3n},t_{1n}')U(j_{2n},j_{3n},t_{3n}'-t_{1n}')\\ \qquad\qquad\qquad\times U(j_{2n},j_{4n},t_{4n}')\text{\qquad\quad if $t_{1n}'<t_{3n}'$.} \end{cases} \label{eqn:idtright}$$ We nest multiple copies of the base identity operation in order to create enough degrees of freedom to find a solution that cancels the error of the initial rotation. We find that 5 is the minimum number of copies needed, and so the sequence of rotations which comprise the full uncorrected identity, denoted $I^{(5)}$, is as follows: $$I^{(5)}=\prod_{n=1}^5A_n'\prod_{n=5}^1A_n. \label{eqn:identityfull}$$ The error for each uncorrected rotation is calculated in terms of the parameters $t_{in}$ and $t_{in}'$ by Eq. , and added together using Eq. , and the result is added to the error of the initial pulse. The norm of this total first order error is numerically minimized over the parameters $t_{in}$ and $t_{in}'$, and a minimum sufficiently close to zero indicates that the crosstalk and static noise error has been canceled using the pulse sequence given by the values of $t_{in}$ and $t_{in}'$. Experimental constraints on the value of $J_i$ can be accounted for by defining bounds on the total time of a $2\pi$ rotation $t_\text{min/max}$ in terms of the constraining values of $J_\text{min/max}$: $$t_\text{min/max}=\frac{\pi}{\sqrt{h^2+J_\text{max/min}^2}}.$$ Then, during the numerical minimization, the constraint on the time for each $2\pi$ rotation $t_\text{min}<t_{in}+t_{in}'<t_\text{max}$ is applied, along with the physical constraint that $t_{in},t_{in}'>0$. This ensures that the derived pulse sequence respects experimental limitations. For the pulse sequences we derived, we used values of $J_\text{min}$ and $J_\text{max}$ equal to $\frac{1}{30}$ and $30$ respectively, but other constraints can be used in much the same manner. Numerical Results ================= ![[**Top:**]{} A corrected pulse sequence for the gate $e^{-\frac{2\pi}{3}\frac{X+Y+Z}{\sqrt{3}}}$ using the identity operation presented in this work. [**Bottom:**]{} For comparison, a corrected pulse sequence for the same gate using the identity with $4\pi$ rotations.[]{data-label="fig:newmethod"}](ca02log.pdf "fig:"){width="\columnwidth"} ![[**Top:**]{} A corrected pulse sequence for the gate $e^{-\frac{2\pi}{3}\frac{X+Y+Z}{\sqrt{3}}}$ using the identity operation presented in this work. [**Bottom:**]{} For comparison, a corrected pulse sequence for the same gate using the identity with $4\pi$ rotations.[]{data-label="fig:newmethod"}](ca02old.pdf "fig:"){width="\columnwidth"} We generated pulses which correct against first order error for the 24 Clifford gates. The parameters precisely defining these pulses are given in the Appendix. For comparison, we also generated pulses using the previous method which uses $4\pi$ rotations. Using the identity operation with two $2\pi$ rotations, we are able to reduce the total length of the error correcting identity from $36\pi$ to $20\pi$, corresponding to roughly a 40% decrease in the length of the pulse, as shown in figure \[fig:newmethod\]. ![Error scaling of a naive uncorrected rotation compared with the corrected pulse sequence for the gate $e^{-\frac{2\pi}{3}\frac{X+Y+Z}{\sqrt{3}}}$ at a constant value of $B=100h$. [**Top:**]{} Error with random values of $(\mathit{dh}_i/h, \mathit{dJ}_i/J_i, k/h)$. [**Bottom:**]{} Error plotted against only $k/h$, with $\mathit{dh}_i=\mathit{dJ}_i=0$.[]{data-label="fig:errorcheck"}](lerrorplotca02.pdf "fig:"){width="\columnwidth"} ![Error scaling of a naive uncorrected rotation compared with the corrected pulse sequence for the gate $e^{-\frac{2\pi}{3}\frac{X+Y+Z}{\sqrt{3}}}$ at a constant value of $B=100h$. [**Top:**]{} Error with random values of $(\mathit{dh}_i/h, \mathit{dJ}_i/J_i, k/h)$. [**Bottom:**]{} Error plotted against only $k/h$, with $\mathit{dh}_i=\mathit{dJ}_i=0$.[]{data-label="fig:errorcheck"}](konlyplot.pdf "fig:"){width="\columnwidth"} A full dynamical decoupling randomized benchmarking analysis is needed to test the pulses against experimental noise. Such an analysis, while feasible, is extremely computationally demanding, and therefore should only be carried out in the context of an actual experimental realization of the pulse sequence developed in the current work. However, we can demonstrate that under a quasistatic noise approximation, the sequences we derived completely cancel first order error. This is done by choosing random values for the errors $\mathit{dh}_i/h$, $\mathit{dJ}_i/J_i$ and the relative strength of the exchange coupling $k/h$. From these, the uncorrected rotations performed for each segment of the pulse are evaluated by numerically exponentiating the full 6-dimensional Hamiltonian which includes the noise terms, leakage states, and magnetic field splitting $B$. By multiplying together the matrices corresponding to each individual segment, the matrix resulting from applying the full pulse with the given noise terms can be found. We denote this matrix $P(\mathit{dh}_i, \mathit{dJ}_i, k)$ and compare it against the ideal Clifford gate which the pulse implements, obtaining a difference $\Delta_P$. The norm of $\Delta_P$, defined as $\sqrt{\text{Tr}(\Delta_P^\dagger\Delta_P)}$, is related to the infidelity of the pulse $P$ averaged over all initial states, and thus is a good measure of total error of the pulse. In order to determine how the error of the pulse scales with initial noise terms, the norm of $\Delta_P$ is plotted against the norm of $(\mathit{dh}_i/h,\mathit{dJ}_i/J_i,k/h)$. In the top portion of Fig. \[fig:errorcheck\], we show such a plot for a value of $B$ equal to $100h$. As shown, the total error is second order for initial error values greater than roughly $10^{-2}$, but becomes linear below that point as leakage errors begin to dominate. Increasing or decreasing the value of $B$ will cause this crossover point to shift left or right respectively. One can isolate the behavior of the error in terms of the interaction strength $k$ alone by performing a similar calculation setting $\mathit{dh}_i$ and $\mathit{dJ}_i$ to 0. This shows the error scaling of the corrected and uncorrected pulses more clearly. Specifically, from the bottom of Fig. \[fig:errorcheck\], we again see that the error scales quadratically above $10^{-2}$ and linearly below that point, yielding an improvement of nearly two orders of magnitude. The reason for the spread in data points in the top graph compared to the bottom, is that for randomly selected noise values, $(\mathit{dh}_i/h, \mathit{dJ}_i/J_i)$ it is possible for different second order error terms to add constructively or destructively, thus adding some variation to the total error of the pulse. Restricting to only one variable produces a much finer line, as this is no longer the case. In both cases, we see the elimination of first order error for noise values above $1/B$. Conclusion ========== We have demonstrated a method for correcting crosstalk- and noise-induced error in single-qubit gates in exchange-coupled singlet-triplet qubits. Unlike the capacitively-coupled qubits studied in Ref. , we can tune the coupling between the qubits independently of the intraqubit exchange coupling used to perform gates. However, leakage out of the computational subspace is a problem here, which does not arise for capacitive coupling. We first showed that a sufficiently large magnetic field difference between the two qubits helps to suppress leakage, and then proceeded to develop pulse sequences that cancel crosstalk- and noise-induced error to first order. Our methods are similar to those used in Ref.  for capacitively coupled singlet-triplet qubits—we perform an uncorrected gate, and then follow it with a sequence of uncorrected identity operations designed in such a way as to cancel crosstalk- and noise-induced error in the gate to first order. The basic building blocks of our sequences, however, are different—rather than perform a single $4\pi$ rotation on the “idle” qubit, we perform two $2\pi$ rotations, each with a different value of the intraqubit exchange coupling. This allows for shorter and faster sequences; we now only require at least 5 of these “blocks”, fewer than what was needed with the older method. We find that our sequences do, in fact, reduce the error in our qubits by several orders of magnitude. We notice, however, that, for low noise and crosstalk, the error is still first order in our measure of the total error described above, but then crosses over to second-order behavior for larger values. This indicates that, compared to the analogous sequences for capacitively-coupled qubits, our ability to use the sequences presented in this work to combat error is more limited. While, as usual, our results show that reduction of noise and crosstalk during a single-qubit gate are important for improving fidelity, we also show that creating a large magnetic field difference between the two qubits will help to achieve this goal in the case of exchange-coupled qubits. While a number of techniques exist for reducing magnetic field noise, such as polarization of nuclear spins in GaAs or working with isotopically-purified Si, which has very few magnetic impurities, charge noise is much more difficult to handle, in no small part due to the fact that the origin of this noise is poorly understood. We should also point out that, throughout this work, we made a number of approximations. First of all, we assumed that the magnetic field gradients on the two qubits were the same, which we believe is a reasonable assumption to make. In principle, they could differ, either intentionally or due to natural variation in the field gradients produced at each qubit, whatever the method used to do so might be. Our method already corrects for small, unintentional, variations, since these can essentially be included as part of the noise term. Larger variations, however, would require us to modify our pulse sequences. This also introduces complications similar to those in the capacitively-coupled case[@ButerakosPRB2018]. Second of all, we assumed that the pulses were perfect square pulses. Such pulses are impossible in reality, as there will be a finite rise or fall time. The effect of this finite ramping up or down has been studied in the case of a single isolated qubit[@WangPRA2014], and it was found that the sequences derived therein for correcting noise-induced error still performed well. We thus expect that this will continue to hold true even for the sequences derived in this work. Finally, we assumed that the noise in the system was quasistatic, i.e., it was possible to neglect the time variation in the noise. While this tends to be a good approximation, it is found that, in reality, both types of noise in singlet-triplet qubits exhibit power-law spectra. To be exact, both types of noise follow a $1/f^\alpha$ spectrum, with $\alpha=2.6$ for the magnetic field noise and $0.7$ for the charge noise[@DialPRL2013; @MedfordPRL2012]. Filtering out the high-frequency noise components, especially in the charge noise, will bring a system closer to the quasistatic limit that we worked in. It has been shown, however, that, even in this case, sequences similar to ours for a single isolated qubit[@WangPRA2014; @ThrockmortonPRB2017] still correct errors due to time-dependent noise, and thus we expect the same to hold true for the sequences developed here. Developing a means of combatting noise that takes into full account the time dependence of any specific noise spectrum will likely yield even better results, but doing so is beyond the scope of this work. Even though we considered the case of two coupled qubits here, an eventual practical quantum computer will have billions of qubits, all with some degree of coupling. As a result, extensions of the techniques developed here to the case of a larger number of qubits will be necessary. Since the coupling between two qubits falls off with distance, we expect that one really only needs to correct for error due to qubits up to a distance of, say, the fifth-nearest neighbors, if that. Correction of errors due to qubits further away should be possible through other error-correction techniques. This work is supported by the Laboratory for Physical Sciences. Parameters for Dynamical Decoupling Pulse Sequences for the 24 Clifford Gates ============================================================================= In Tables \[Tab:RotParamsI\]–\[Tab:RotParamsIII\] we list parameters defining pulse sequences for the 24 Clifford gates. The top portion of the tables consists of values $j_i^{\text{rot}}$, $t_i^{\text{rot}}$, $j'^{\text{rot}}$ which encode the initial rotations as $$R=\prod_{i=1}^N U(j_i^{\text{rot}},j'^{\text{rot}},t_i^{\text{rot}})$$ for $N$ equal to 1, 3, or 5, depending on the number of values given in the table. The rest of the tables are the values of $t_{in}$ and $t'_{in}$ which define the uncorrected identity by Eqs. –. Only values of $i$ from 1 to 3 are shown with $t_{4n}$ being defined as $t_{1n}+t_{2n}-t_{3n}$, and similarly for $t_{4n}'$. Axis $I$ $x$ $x$ $x$ $y$ $y$ $y$ $z$ -------------------- --------- --------- --------- ---------- --------- --------- ---------- --------- Angle 0 $\pi/2$ $\pi$ $3\pi/2$ $\pi/2$ $\pi$ $3\pi/2$ $\pi/2$ $j'^{\text{rot}}$ 30.0000 3.37447 1.57530 0.80106 0.82357 0.80164 0.77982 6.50963 $j_1^{\text{rot}}$ 30.0000 0.19724 0.19336 0.20358 30.0000 30.0000 30.0000 30.0000 $j_2^{\text{rot}}$ — 30.0000 30.0000 30.0000 0.93548 0.93548 0.93548 6.54409 $j_3^{\text{rot}}$ — 0.19724 0.19336 0.20358 30.0000 30.0000 30.0000 30.0000 $j_4^{\text{rot}}$ — — — — 0.93548 0.93548 0.93548 — $j_5^{\text{rot}}$ — — — — 30.0000 30.0000 30.0000 — $t_1^{\text{rot}}$ 0.10466 0.39633 0.79278 1.17605 0.02617 0.02617 0.02617 0.01640 $t_2^{\text{rot}}$ — 0.09997 0.09813 0.09981 1.14711 1.14711 1.14711 0.44421 $t_3^{\text{rot}}$ — 0.39633 0.79278 1.17605 0.02617 0.05233 0.07850 0.01640 $t_4^{\text{rot}}$ — — — — 1.14711 1.14711 1.14711 — $t_5^{\text{rot}}$ — — — — 0.07850 0.07850 0.07850 — $t_{1,1}$ 0.23932 0.52518 0.17782 0.35794 0.87280 1.40951 1.47408 2.08110 $t_{2,1}$ 1.09125 1.69460 2.08781 1.29553 0.18612 0.20762 0.35782 0.22124 $t_{3,1}$ 1.12263 1.71424 1.94402 1.20225 0.24035 0.28494 0.23108 0.56868 $t'_{1,1}$ 0.59780 0.35620 0.40442 0.25259 2.16299 1.40150 1.30849 1.04347 $t'_{2,1}$ 1.61793 1.40675 1.05123 1.31383 0.21544 0.48157 0.45315 0.18968 $t'_{3,1}$ 1.98543 1.25467 0.69678 1.26393 0.41602 0.17172 0.28890 0.27459 $t_{1,2}$ 0.44412 0.30496 0.21167 0.33209 1.80820 1.28984 1.17247 1.28622 $t_{2,2}$ 1.43763 1.30979 1.62248 1.40209 0.23931 0.44525 0.33085 0.24893 $t_{3,2}$ 1.58580 1.20774 1.51527 1.14195 0.61849 0.19688 0.27989 0.39266 $t'_{1,2}$ 0.43189 0.32979 0.22604 0.44226 1.32490 1.43301 1.61739 1.81344 $t'_{2,2}$ 1.33166 1.33003 1.32999 1.68187 0.25344 0.24189 0.17401 0.24925 $t'_{3,2}$ 1.49696 1.20594 1.39791 1.56451 0.26449 0.19837 0.22329 0.46107 $t_{1,3}$ 0.59437 0.19836 0.32108 0.23737 1.63651 1.90912 0.78061 1.32820 $t_{2,3}$ 1.27812 1.69950 1.08601 1.62730 0.31245 0.20040 0.99287 0.21810 $t_{3,3}$ 1.71539 1.47222 0.47707 1.62737 0.53524 0.39193 0.31686 0.81853 $t'_{1,3}$ 0.83936 0.20294 0.25174 0.24861 1.46388 0.77040 0.87515 1.69784 $t'_{2,3}$ 0.74344 1.43150 2.03496 1.40707 0.36946 0.94415 0.93132 0.16985 $t'_{3,3}$ 1.35372 0.77597 1.92016 0.99473 0.81303 0.36259 0.31753 0.35261 $t_{1,4}$ 0.25739 0.22260 0.34850 0.31382 1.76187 1.65753 1.52861 1.27658 $t_{2,4}$ 1.59299 1.72292 1.31460 1.94008 0.30317 0.31011 0.21738 0.12688 $t_{3,4}$ 1.71721 1.56652 1.04504 1.46824 0.30507 0.22075 0.29048 0.42355 $t'_{1,4}$ 0.34269 0.25157 0.26769 0.35810 1.18533 1.19014 1.53603 1.71460 $t'_{2,4}$ 1.05712 1.40351 1.82480 1.16347 0.35652 0.43290 0.40722 0.11908 $t'_{3,4}$ 1.26890 1.30144 1.54555 1.04779 0.51251 0.25596 0.27451 0.30832 $t_{1,5}$ 0.40810 0.12889 0.30221 0.13525 1.93710 1.49564 1.37219 1.31394 $t_{2,5}$ 1.78749 1.16500 1.46658 1.36368 0.36629 0.25970 0.25657 0.47911 $t_{3,5}$ 1.77536 0.85654 1.49514 1.17617 0.73136 0.29417 0.22679 0.63768 $t'_{1,5}$ 0.66870 0.13444 0.31150 0.15842 1.18187 1.43701 1.46942 1.79038 $t'_{2,5}$ 1.13501 1.94727 1.26679 1.70282 0.37723 0.19528 0.42305 0.42390 $t'_{3,5}$ 1.35550 1.71163 1.36833 1.59387 0.50954 0.32472 0.23331 0.42474 Axis $z$ $z$ $x+y$ $x-y$ $x+z$ $x-z$ $y+z$ $y-z$ -------------------- --------- ---------- --------- --------- --------- --------- --------- --------- Angle $\pi$ $3\pi/2$ $\pi$ $\pi$ $\pi$ $\pi$ $\pi$ $\pi$ $j'^{\text{rot}}$ 0.36056 0.17551 0.63069 0.63069 2.64575 2.22703 2.30059 2.30059 $j_1^{\text{rot}}$ 30.0000 30.0000 0.67417 0.67417 1.00000 30.0000 30.0000 30.0000 $j_2^{\text{rot}}$ 0.38039 0.26157 30.0000 30.0000 — 0.87487 0.93444 0.93444 $j_3^{\text{rot}}$ 30.0000 30.0000 0.67417 0.67417 — 30.0000 30.0000 30.0000 $j_4^{\text{rot}}$ — — — — — — — — $j_5^{\text{rot}}$ — — — — — — — — $t_1^{\text{rot}}$ 0.02640 0.03937 1.74236 0.86255 1.11072 0.05233 0.02672 0.07794 $t_2^{\text{rot}}$ 2.90256 3.01556 0.05233 0.05233 — 1.18222 1.14770 1.14770 $t_3^{\text{rot}}$ 0.02640 0.03937 0.86255 1.74236 — 0.05233 0.07794 0.02672 $t_4^{\text{rot}}$ — — — — — — — — $t_5^{\text{rot}}$ — — — — — — — — $t_{1,1}$ 0.99823 1.92843 0.08833 1.28673 0.34596 0.79030 1.60569 0.23302 $t_{2,1}$ 0.60853 0.34890 2.00643 0.68412 1.70052 1.81814 0.11982 1.50661 $t_{3,1}$ 0.23483 0.72084 1.67281 0.22703 1.56475 2.12906 0.48991 1.26268 $t'_{1,1}$ 1.73635 1.20417 0.12953 1.29749 0.37416 0.12950 1.42417 0.59732 $t'_{2,1}$ 0.26384 0.39221 1.07374 0.35098 1.41921 1.30677 0.12906 1.52730 $t'_{3,1}$ 0.24694 0.24100 0.89230 0.28196 1.37933 0.89247 0.26335 1.65259 $t_{1,2}$ 1.61084 1.59488 0.27424 1.42591 0.33335 0.45526 1.58591 0.18225 $t_{2,2}$ 0.25952 0.45350 1.48830 0.42358 1.34677 1.44068 0.36992 1.45365 $t_{3,2}$ 0.23031 0.13759 1.38719 0.40661 1.21093 1.24880 0.51445 1.25939 $t'_{1,2}$ 1.48086 1.52179 0.21616 1.39520 0.34354 0.41969 1.54260 0.18602 $t'_{2,2}$ 0.18545 0.34452 1.57898 0.74742 1.61747 1.45752 0.42459 1.46460 $t'_{3,2}$ 0.19828 0.77318 1.28815 0.47538 1.28189 1.35509 0.47459 1.25762 $t_{1,3}$ 1.31731 1.82596 0.35433 1.29316 0.28838 0.21660 1.39782 0.21928 $t_{2,3}$ 0.24558 0.86655 1.35594 0.59180 1.68930 1.46350 0.27381 1.56973 $t_{3,3}$ 0.02527 0.91830 0.77629 0.27573 1.34067 0.84143 0.92875 0.98139 $t'_{1,3}$ 1.81151 1.28752 0.27864 1.40956 0.25486 0.24862 1.71559 0.21539 $t'_{2,3}$ 0.57198 0.69158 1.78281 0.34085 1.37178 1.66983 0.19586 1.56537 $t'_{3,3}$ 0.98100 0.66898 1.83730 0.23121 1.02614 1.51736 0.61196 1.08867 $t_{1,4}$ 1.39452 0.83871 0.42834 1.16557 0.17440 0.33921 1.39500 0.23414 $t_{2,4}$ 0.34113 0.23818 1.37340 0.79534 1.63848 1.34803 0.24753 1.46441 $t_{3,4}$ 0.18321 0.20384 1.39414 0.36738 1.45126 1.37034 0.43543 1.39353 $t'_{1,4}$ 1.41114 2.28866 0.42670 1.48273 0.23327 0.28588 1.56864 0.24909 $t'_{2,4}$ 0.45458 0.12850 1.67357 0.44575 1.47534 1.73744 0.29270 1.45691 $t'_{3,4}$ 0.28846 0.33058 1.46928 0.44121 1.26173 1.56477 0.39172 1.42224 $t_{1,5}$ 1.63184 1.44467 0.14450 1.34021 0.18444 0.17733 1.29598 0.14930 $t_{2,5}$ 0.57060 0.50947 1.57740 0.39524 1.42391 1.37012 0.70988 1.64764 $t_{3,5}$ 0.73779 0.26562 1.33720 0.23217 1.18487 0.95442 0.71168 1.33522 $t'_{1,5}$ 1.42744 1.51242 0.13516 1.20153 0.17807 0.17531 1.81005 0.16545 $t'_{2,5}$ 0.49945 0.44739 1.39521 0.33676 1.69454 1.76074 0.49729 1.43169 $t'_{3,5}$ 0.35619 0.65637 1.10140 0.18018 1.50998 1.56945 0.53217 1.17211 Axis $x+y+z$ $x+y+z$ $-x+y+z$ $-x+y+z$ $x-y+z$ $x-y+z$ $x+y-z$ $x+y-z$ -------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- Angle $2\pi/3$ $4\pi/3$ $2\pi/3$ $4\pi/3$ $2\pi/3$ $4\pi/3$ $2\pi/3$ $4\pi/3$ $j'^{\text{rot}}$ 1.25195 1.21590 0.26126 0.30530 1.25195 1.21590 0.30530 0.26126 $j_1^{\text{rot}}$ 1.28889 1.28889 0.34503 0.34503 1.28889 1.28889 0.34503 0.34503 $j_2^{\text{rot}}$ 30.0000 30.0000 30.0000 30.0000 30.0000 30.0000 30.0000 30.0000 $j_3^{\text{rot}}$ 1.28889 1.28889 0.34503 0.34503 1.28889 1.28889 0.34503 0.34503 $j_4^{\text{rot}}$ — — — — — — — — $j_5^{\text{rot}}$ — — — — — — — — $t_1^{\text{rot}}$ 1.39062 1.39062 1.16990 1.16990 0.53517 0.53517 1.79990 1.79990 $t_2^{\text{rot}}$ 0.03489 0.06977 0.06977 0.03489 0.03489 0.06977 0.03489 0.06977 $t_3^{\text{rot}}$ 0.53517 0.53517 1.79990 1.79990 1.39062 1.39062 1.16990 1.16990 $t_4^{\text{rot}}$ — — — — — — — — $t_5^{\text{rot}}$ — — — — — — — — $t_{1,1}$ 1.50682 0.23907 2.09919 2.13904 0.25877 1.32694 0.96335 0.44496 $t_{2,1}$ 0.43378 1.79424 0.13944 0.14355 1.37466 0.09653 0.14066 1.19443 $t_{3,1}$ 0.36195 1.89588 0.27167 0.24009 1.39970 0.43533 0.15304 1.30631 $t'_{1,1}$ 1.42114 0.33431 0.81692 0.88218 0.77381 1.78280 2.05010 0.26043 $t'_{2,1}$ 0.60324 1.00912 0.12582 0.11347 1.17007 0.10471 0.14852 1.26224 $t'_{3,1}$ 0.48662 1.20899 0.18390 0.22136 1.71261 0.39275 0.70179 1.27329 $t_{1,2}$ 1.39992 0.62971 1.49855 1.12516 0.28032 1.86516 1.48547 0.18537 $t_{2,2}$ 0.42414 0.77535 0.51557 0.47319 1.74845 0.39374 0.14925 1.52920 $t_{3,2}$ 0.27850 1.08828 0.50164 0.16185 1.71345 0.67478 0.30620 1.08443 $t'_{1,2}$ 1.21485 0.47335 1.63508 1.99977 0.67062 1.23605 1.64317 0.43359 $t'_{2,2}$ 0.42281 1.70063 0.55703 0.44445 1.12901 0.35980 0.23526 1.60726 $t'_{3,2}$ 0.22864 1.98788 0.87039 0.90586 1.40689 0.32649 0.34039 1.58324 $t_{1,3}$ 0.62378 0.64882 1.66255 1.55661 0.52880 1.27017 1.99104 0.18084 $t_{2,3}$ 1.12751 0.92558 0.24453 0.23475 1.32471 0.17439 0.22327 1.80390 $t_{3,3}$ 0.27546 1.36298 0.21519 0.61544 1.74611 0.18039 0.44019 1.81096 $t'_{1,3}$ 1.86465 0.39490 1.46305 1.56039 0.40115 1.78846 1.06507 0.21559 $t'_{2,3}$ 0.19255 1.60297 0.23591 0.22036 1.10171 0.15650 0.22715 1.32750 $t'_{3,3}$ 0.41572 1.75840 0.54332 0.17008 1.38539 0.46378 0.25086 0.79889 $t_{1,4}$ 1.12639 0.87640 1.58559 1.62250 0.41300 1.68174 1.35793 0.48565 $t_{2,4}$ 0.40157 1.20571 0.45452 0.39473 1.29101 0.16459 0.27905 2.07706 $t_{3,4}$ 0.29799 1.65093 0.92537 0.31961 1.34976 0.80909 0.05876 1.69366 $t'_{1,4}$ 1.72699 0.39587 1.51482 1.50224 0.75062 1.42172 1.63523 0.50811 $t'_{2,4}$ 0.40525 1.51421 0.54323 0.32635 1.11254 0.19048 0.20301 1.04977 $t'_{3,4}$ 0.23737 1.47462 0.36858 0.74476 1.60877 0.18202 0.61873 1.40129 $t_{1,5}$ 1.25033 0.28166 2.10165 2.07986 0.48603 1.22964 1.34488 0.20260 $t_{2,5}$ 0.36179 1.76795 0.24481 0.27539 1.22698 0.51213 0.45962 0.96152 $t_{3,5}$ 0.24346 1.66999 0.26932 0.44037 1.29670 0.30894 0.57191 0.49989 $t'_{1,5}$ 1.57572 0.37654 1.00130 0.99860 0.43093 1.57113 1.65222 0.23722 $t'_{2,5}$ 0.32856 1.26919 0.32996 0.31567 1.51202 0.46363 0.33684 2.15458 $t'_{3,5}$ 0.25086 1.39094 0.50461 0.40620 1.71000 0.43901 0.22441 2.01752 [99]{} D. Loss and D. P. DiVincenzo, , 120 (1998). X. Hu and S. Das Sarma, , 062301 (2000). X. Hu and S. Das Sarma, , 042312 (2001). K. Nowack, M. Shaffei, M. Laforest, G. E. D. K. Prawiroatmodjo, L. R. Schreiber, C. Reichl, W. Wegscheider, and L. M. K. Vandersypen, Science [**333**]{}, 1269 (2011). J. J. Pla, K. Y. Tan, J. P. Dehollain, W. H. Lim, J. J. L. Morton, D. N. Jamieson, A. S. Dzurak, and A. Morello, Nature (London) [**489**]{}, 541 (2012). F. R. Braakman, P. Barthelemy, C. Reichl, W. Wegscheider, and L. M. K. Vandersypen, Nat. Nanotechnol. 8, [**432**]{} (2013). J. J. Pla, K. Y. Tan, J. P. Dehollain, W. H. Lim, J. J. L. Morton, F. A. Zwanenburg, D. N. Jamieson, A. S. Dzurak, and A. Morello, Nature (London) [**496**]{}, 334 (2013). M. Veldhorst, J. C. C. Hwang, C. H. Yang, A. W. Leenstra, B. de Ronde, J. P. Dehollain, J. T. Muhonen, F. E. Hudson, K. M. Itoh, A. Morello, and A. S. Dzurak, Nat. Nanotechnol. 9, 981 (2014). T. Otsuka, T. Nakajima, M. R. Delbecq, S. Amaha, J. Yoneda, K. Takeda, G. Allison, T. Ito, R. Sugawara, A. Noiri, A. Ludwig, A. D. Wieck, and S. Tarucha, Sci. Rep. [**6**]{}, 31820 (2016). T. Ito, T. Otsuka, S. Amaha, M. R. Delbecq, T. Nakajima, J. Yoneda, K. Takeda, G. Allison, A. Noiri, K. Kawasaki, and S. Tarucha, Sci. Rep. 6, 39113 (2016). J. Levy, , 147902 (2002). J. Petta, A. Johnson, J. Taylor, E. Laird, A. Yacoby, M. Lukin, C. Marcus, M. Hanson, and A. Gossard, Science [**309**]{}, 2180 (2005). S. Foletti, H. Bluhm, D. Mahalu, V. Umansky, and A. Yacoby, Nat. Phys. [**5**]{}, 903 (2009). I. van Weperen, B. D. Armstrong, E. A. Laird, J. Medford, C. M. Marcus, M. P. Hanson, and A. C. Gossard, , 030506 (2011). B. M. Maune, M. G. Borselli, B. Huang, T. D. Ladd, P. W. Deelman, K. S. Holabird, A. A. Kiselev, I. Alvarado-Rodriguez, R. S. Ross, A. E. Schmitz, M. Sokolich, C. A. Watson, M. F. Gyure, and A. T. Hunter, Nature (London) [**481**]{}, 344 (2012). M. D. Shulman, O. E. Dial, S. P. Harvey, H. Bluhm, V. Umansky, and A. Yacoby, Science [**336**]{}, 202 (2012). O. E. Dial, M. D. Shulman, S. P. Harvey, H. Bluhm, V. Umansky, and A. Yacoby, , 146804 (2013). M. D. Shulman, S. P. Harvey, J. M. Nichol, S. D. Bartlett, A. C. Doherty, V. Umansky, and A. Yacoby, Nat. Commun. [**5**]{}, 5156 (2014). M. D. Reed, B. M. Maune, R. W. Andrews, M. G. Borselli, K. Eng, M. P. Jura, A. A. Kiselev, T. D. Ladd, S. T. Merkel, I. Milosavljevic, E. J. Pritchett, M. T. Rakher, R. S. Ross, A. E. Schmitz, A. Smith, J. A. Wright, M. F. Gyure, and A. T. Hunter, , 110402 (2016). F. Martins, F. K. Malinowski, P. D. Nissen, E. Barnes, S. Fallahi, G. C. Gardner, M. J. Manfra, C. M. Marcus, and F. Kuemmeth, , 116801 (2016). D. P. DiVincenzo, D. Bacon, J. Kempe, G. Burkard, and K. B. Whaley, Nature (London) [**408**]{}, 339 (2000). J. Medford, J. Beil, J. M. Taylor, S. D. Bartlett, A. C. Doherty, E. I. Rashba, D. P. DiVincenzo, H. Lu, A. C. Gossard, and C. M. Marcus, Nat. Nanotechnol. [**8**]{}, 654 (2013). J. Medford, J. Beil, J. M. Taylor, E. I. Rashba, H. Lu, A. C. Gossard, and C. M. Marcus, , 050501 (2013). K. Eng, T. D. Ladd, A. Smith, M. G. Borselli, A. A. Kiselev, B. H. Fong, K. S. Holabird, T. M. Hazard, B. Huang, P. W. Deelman, I. Milosavljevic, A. E. Schmitz, R. S. Ross, M. F. Gyure, and A. T. Hunter, Sci. Adv. [**1**]{}, e1500214 (2015). Y.-P. Shim and C. Tahan, , 121410(R) (2016). Z. Shi, C. B. Simmons, J. R. Prance, J. K. Gamble, T. S. Koh, Y.-P. Shim, X. Hu, D. E. Savage, M. G. Lagally, M. A. Eriksson, M. Friesen, and S. N. Coppersmith, , 140503 (2012). D. Kim, Z. Shi, C. B. Simmons, D. R. Ward, J. R. Prance, T. S. Koh, J. K. Gamble, D. E. Savage, M. G. Lagally, M. Friesen, S. N. Coppersmith, and M. A. Eriksson, Nature (London) [**511**]{}, 70 (2014). D. Kim, D. R. Ward, C. B. Simmons, D. E. Savage, M. G. Lagally, M. Friesen, S. N. Coppersmith, and M. A. Eriksson, npj Quant. Inf. [**1**]{}, 15004 (2015). A. G. Fowler, M. Mariantoni, J. M. Martinis, A. N. Cleland, , 032324 (2012). R. Barends, J. Kelly, A. Megrant, A. Veitia, D. Sank, E. Jeffrey, T. C. White, J. Mutus, A. G. Fowler, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth, C. Neill, P. O’Malley, P. Roushan, A. Vainsencher, J. Wenner, A. N. Korotkov, A. N. Cleland, and J. M. Martinis, Nature [**508**]{}, 500 (2014). J. Benhelm, G. Kirchmair, C. F. Roos, and R. Blatt, Nat. Phys. [**4**]{}, 463 (2008). P. Harvey-Collard, R. M. Jock, N. T. Jacobson, A. D. Baczewski, A. M. Mounce, M. J. Curry, D. R. Ward, J. M. Anderson, R. P. Manginell, J. R. Wendt, M. Rudolph, T. Pluym, M. P. Lilly, M. Pioro-Ladrière and M. S. Carroll, IEEE International Electron Devices Meeting (IEDM), 36.5.1-36.5.4 (2017). W. M. Witzel and S. Das Sarma, , 077601 (2007). W. M. Witzel and S. Das Sarma, , 241303 (2007). B. Lee, W. M. Witzel, and S. Das Sarma, , 160505 (2008). H. Bluhm, S. Foletti, D. Mahalu, V. Umansky, A. Yacoby, , 216803 (2010). H. Bluhm, S. Foletti, I. Neder, M. Rudner, D. Mahalu, V. Umansky, and A. Yacoby, Nat. Phys. [**7**]{}, 109 (2011). A. Sergeevich, A. Chandran, J. Combes, S. D. Bartlett, and H. M. Wiseman, , 052315 (2011). J. T. Muhonen, J. P. Dehollain, A. Laucht, F. E. Hudson, T. Sekiguchi, K. M. Itoh, D. N. Jamieson, J. C. McCallum, A. S. Dzurak, and A. Morello, Nat. Nanotechnol. [**9**]{}, 986 (2014). F. K. Malinowski, F. Martins, P. D. Nissen, E. Barnes, Ł. Cywiński, M. S. Rudner, S. Fallahi, G. C. Gardner, M. J. Manfra, C. M. Marcus, and F. Kuemmeth, Nat. Nanotechnol. [**12**]{}, 16 (2017). X. Wang, L. S. Bishop, J. P. Kestner, E. Barnes, K. Sun, and S. Das Sarma, Nat. Commun. [**3**]{}, 997 (2012). X. Wang, L. S. Bishop, E. Barnes, J. P. Kestner, and S. Das Sarma, , 022310 (2014). R. E. Throckmorton, C. Zhang, X.-C. Yang, X. Wang, E. Barnes, and S. Das Sarma, , 195424 (2017). J. Zeng, X.-H. Deng, A. Russo, and E. Barnes, New J. Phys. 20, 033011 (2018). Y.-L. Wu and S. Das Sarma, , 165301 (2017). X. Wang, E. Barnes, and S. Das Sarma, npj Quant. Inf. [**1**]{}, 15003 (2015). F. A. Calderon-Vargas and J. P. Kestner, , 150502 (2017). D. Buterakos, R. E. Throckmorton, and S. Das Sarma, , 045431 (2018). R. Li, X. Hu, and J. Q. You, , 205306 (2012). J. P. Kestner, X. Wang, L. S. Bishop, E. Barnes, and S. Das Sarma, , 140502 (2013). J. Medford, Ł. Cywiński, C. Barthel, C. M. Marcus, M. P. Hanson, and A. C. Gossard, , 086802 (2012).

ArXiv

--- abstract: 'We construct and analyze a microscopic model for insulating rock salt ordered double perovskites, with the chemical formula A$_2$BB’O$_6$, where the B’ atom has a 4d$^1$ or 5d$^1$ electronic configuration and forms a face centered cubic (fcc) lattice. The combination of the triply-degenerate $t_{2g}$ orbital and strong spin-orbit coupling forms local quadruplets with an effective spin moment $j=3/2$. Moreover, due to strongly orbital-dependent exchange, the effective spins have substantial biquadratic and bicubic interactions (fourth and sixth order in the spins, respectively). This leads, at the mean field level, to three main phases: an unusual antiferromagnet with dominant octupolar order, a ferromagnetic phase with magnetization along the $[110]$ direction, and a non-magnetic but quadrupolar ordered phase, which is stabilized by thermal fluctuations and intermediate temperatures. All these phases have a two sublattice structure described by the ordering wavevector ${{\boldsymbol}Q} =2\pi (001)$. We consider quantum fluctuations and argue that in the regime of dominant antiferromagnetic exchange, a non-magnetic valence bond solid or quantum spin liquid state may be favored instead. Candidate quantum spin liquid states and their basic properties are described. We also address the effect of single-site anisotropy driven by lattice distortions. Existing and possible future experiments are discussed in light of these results.' author: - Gang Chen - Rodrigo Pereira - Leon Balents bibliography: - 'ref.bib' date: - - title: 'Exotic phases induced by strong spin-orbit coupling in ordered double perovskites' --- Introduction {#sec:sec1} ============ In magnetic Mott insulators with quenched orbital degrees of freedom, weak spin-orbit coupling (SOC) only leads to a small correction to the usual spin exchange Hamiltonian in the form of single-site anisotropy and Dzyaloshinskii-Moriya interactions.[@PhysRev12091; @Dzyaloshinsky1958241] In the presence of strong SOC, however, a completely different physical picture emerges, in which spin itself is not a good quantum number, and magnetic anisotropy is usually large. Generally, strong SOC is common in the Lanthanides, in which the relevant 4f-electrons are very tightly bound to the nucleus. The tight binding shields the electrons from crystal fields, which tend to split the orbital degeneracies involved in SOC, and moreover reduces exchange, which also competes with SOC. While more rare, strong SOC is becoming an increasing focus in d-electron systems, in which electrons are more delocalized than in the Lanthanides, and more diverse phenomena can be expected. For instance, strong SOC can be expected in 5d transition metal compounds, which have large intrinsic atomic SOC due to their high atomic weight. In this category, many Ir-based magnets have been studied recently[@PhysRevB.78.094403; @PhysRevLett.102.017205]. Lighter transition metals may also exhibit strong SOC if competing effects such as crystal fields and exchange are suppressed, e.g. by choosing crystal structures with high-symmetry and well-separated magnetic ions, respectively. An example of this type is the “spin-orbital liquid” state observed in the Fe-based spinel FeSc$_2$S$_4$,[@loidl:nsnmr; @fritsch:prl04; @loidl:ns] which is believed to be driven by SOC. [@PhysRevLett.102.096406; @PhysRevB.80.224409; @PhysRevB.82.041105] In this paper, we consider the case of insulating magnetic ordered double perovskites. Structurally, ordered double perovskites (with the chemical formula A$_2$BB’O$_6$) are derived from the usual perovskites ABO$_3$ by selectively replacing half the B ions with another species, denoted B’. We focus on the case in which the B ions are non-magnetic and the B’ ones are magnetic. Because of the difference in the valence charges and ionic radius between B and B’ ions, the magnetic B’ ions form an fcc lattice structure with a lattice constant double of the original cubic one. Many ordered double perovskites incorporate strong intrinsic SOC, as B’ ions are commonly 4d and 5d transition metals. Moreover, the large B’-B’ separation weakens exchange, similarly to FeSc$_2$S$_4$. Here, we construct an appropriate microscopic model for the most quantum of these materials (a list may be found in Table \[tab:Tab1\]), in which the magnetic ion contains a single unpaired $S=1/2$ spin. The physics is strongly influenced by the combination of the orbital degeneracy of the $t_{2g}$ multiplet, which acts as an effective $\ell=1$ orbital angular moment. Due to strong SOC, this combines with the $S=1/2$ spin to induce an effective total angular momentum $j=3/2$ description of the system. Moreover, due to the orbitally-dependent exchange, the interaction of these $j=3/2$ contains large biquadratic (fourth order in spin operators) and triquadratic (sixth order in spin operators) interactions. These support exotic phases not easily found in systems with dominant bilinear spin exchange. Analysis of the microscopic model shows that the strong SOC enhances quantum fluctuations and leads to several interesting phases: (1) an unconventional antiferromagnet (denoted AFM) in which the magnetic [*octupole*]{} and [*quadrupole*]{} moments rather than the dipole moment are dominant, (2) an unusual non-collinear ferromagnet (denoted FM110) with a doubled unit cell and magnetization along the $[110]$ axis, (3) a (biaxial) “spin nematic” phase with quadrupolar order but unbroken time reversal symmetry and, more speculatively, (4) a possible quantum spin liquid (QSL) phase. Phases (1), (2), and (4) are low temperature phases and persist as ground states, while the spin nematic, phase (3), occurs in a broad intermediate temperature range below the paramagnetic state but above any magnetic ordering temperature. ![(Color online) Mean field $T=0$ phase diagram for the model Hamiltonian in Eq. . AFM denotes the “antiferromagnetic” ground state given by Eq.  and Eq. , FM110 denotes the ferromagnetic ground state with an easy axis oriented along $[ 110]$, given by Eq. . The FM100 state, which is ferromagnetic with easy axis along $[ 100]$ appears in the narrow band between the two phase phase boundaries. In the figure, $J=1$.[]{data-label="fig:phase_diagram"}](fig1.pdf){width="8cm"} States with magnetic multipole order are more often observed in f-electron systems where crystal field effects are less important than SOC.[@RevModPhys.81.807] As a consequence, the atomic wavefunctions are total angular momentum eigenstates, in which the spin and orbital degrees of freedom are highly entangled. This leads to highly non-Heisenberg exchange between the local moments, which is described by interaction of higher magnetic multipole operators. Such interactions may drive multipolar order, as suggested for instance in URu$_2$Si$_2$.[@Kotliar] Recently this has been suggested to also occur in d electron systems with unquenched orbital degeneracy and sufficient SOC.[@PhysRevLett.103.067205]  We find a similar mechanism at work in the AFM phase. A ferromagnetic state is not in itself unusual, though such is relatively uncommon in insulators. However, cubic ferromagnets with an easy axis oriented along the $[110]$ direction is quite uncommon. This can be understood from the Landau theory for a ferromagnet: the usual fourth order cubic anisotropy term favors either $[100]$ or $[111]$ orientation, depending upon its sign, but never $[110]$. To obtain a $[110]$ easy axis, one requires sixth order or higher terms to be substantial, making this rare indeed. Remarkably, such $[110]$ anisotropy has been observed in experiments on Ba$_2$NaOsO$_6$.[@Erickson] Both the above states, when heated above their magnetic ordering temperatures, allow on symmetry grounds for an intermediate phase which is time-reversal symmetric but with quadrupolar order – the spin nematic. Applying the mean field theory at $T>0$, we indeed find such a phase in a broad range of parameter space. While spin nematic states have been suggested previously in NiGa$_2$S$_4$[@stoudenmire:214436; @läuchli2006quadrupolar; @tsunetsugu-06; @senthil-06; @nakatsuji-05], it has not been established in that material. The mechanism for quadrupolar order here is much more transparent and robust than in that case. The above three phases, while somewhat unconventional, may be obtained within a mean-field analysis. A QSL state, however, cannot be described by any mean field theory, and is considerably more exotic. The search for a QSL, which is a state in which quantum fluctuations prevent spins from ordering even at zero temperature, is a long-standing problem in fundamental physics.[@balents:nat] Since the possibility of a QSL was suggested by Anderson in the early 1970s,[@Anderson1973153] this has been an active area for theory and experiment. Despite the current maturity of the theory for QSL,[@PhysRevB.65.165113] the experimental confirmation of the existence of such an exotic phase is still elusive. Very commonly geometrical frustration is thought to be a driving mechanism for QSL formation, and consequently most research (both theoretically and experimentally) has been devoted to systems of this type, such as triangular,[@PhysRevB.73.155115] kagome,[@PhysRevB.75.184406] hyperkagome[@PhysRevB.78.094403; @PhysRevLett.99.137207; @PhysRevLett.101.197202] and pyrochlore lattices.[@PhysRevB.69.064404] Here we suggest a different route, in which quantum fluctuations are enhanced primarily by strong SOC, rather than geometrical frustration. In fact, the magnetic ions in ordered double perovskites reside on a face centered cubic (fcc) sublattice, which can be viewed as edge-sharing tetrahedra, and is somewhat geometrically frustrated. Without strong SOC, however, this frustration is weak, and indeed the classical Heisenberg antiferromagnet on the fcc lattice is known to magnetically order into a state with the ordering wavevector $2\pi(001)$.[@henley] The tendency of the simple fcc antiferromagnet to order may be partially attributed to its large coordination number ($z=12$), which leads to mean-field like behavior. By contrast, strong SOC induces effective exchange interactions very different from Heisenberg type, with strong directional dependence that may make a QSL more favorable. To make this suggestion more concrete, we propose a natural wavefunction for a QSL in our model, and discuss the physical properties of such a state. We now outline the main results of the paper, and how they are presented in the following sections. In Sec. \[sec:sec2\], we show that strong SOC leads to an effective $j=3/2$ local moment on each B’ site. We write down a model Hamiltonian which includes three interactions: nearest neighbor (NN) antiferromagnetic (AFM) exchange, $J$, NN ferromagnetic (FM) exchange, $J'$, and electric quadrupolar interaction, $V$. These interactions are all [*projected*]{} down to the effective $j=3/2$ manifold, which induces many terms beyond the usual quadratic exchange. Indeed, because of the four-dimensional basis of spin-$3/2$ states, the resulting Hamiltonian can be thought of as an anisotropic $\Gamma$ matrix model.[@PhysRevB.69.235206] We then discuss the symmetry properties of the projected Hamiltonian. Surprisingly, we find that, in the limit of vanishing FM exchange, the Hamiltonian has a “hidden” global SU$(2)$ symmetry despite its complicated appearance. In Sec. \[sec:sec3\], we consider the mean field ground states of the model, characterized by local (single-site) order parameters. In Sec. \[sec:sec31\] we begin by considering the more accessible limit in which strong uniaxial single-site anisotropy (due e.g. to a tetragonal distortion of the crystal) lifts the “orbital” four-fold degeneracy of $j=3/2$ quadruplets down to easy-axis or easy-plane Kramer’s doublets. In these limits, the effective Hamiltonian in the reduced phase space is mapped onto that of an XXZ antiferromagnet which can be understood even without mean field theory. Next, in Sec. \[sec:sec32\], we carry out $T=0$ mean field theory for the case of cubic symmetry. Here we find the AFM state and two ferromagnetic states (the FM110 state and another state with a $[ 100]$ easy axis). The $T=0$ mean field phase diagram is shown in Fig. \[fig:phase\_diagram\]. Finally, having described the situations with strong and vanishing single-site anisotropy, we determine in Sec. \[sec:interm-anis\] the mean-field phase diagram for intermediate anisotropy. In Sec. \[sec:sec4\] we identify the multipolar order parameters of the three ordered phases and analyze the $T>0$ behavior by mean-field theory. Here we find the quadrupolar phase, and discuss several phase transitions which occur. We also discuss the behavior of the magnetic susceptibility in different parameter regimes. In Sec. \[sec:sec5\], we consider quantum effects beyond the mean field theory. First, we carry out a spin wave calculation, which determines the collective mode structure, and also shows that in the regime where nearest-neighbor antiferromagnetic exchange is dominant (small $J'$ and $V$) quantum fluctuations are large and may destabilize the ordered AFM phase. Therefore, we consider possible non-magnetic ground states, both of valence bond solid and quantum spin liquid (QSL) type. We formulate a slave-fermion theory with four-component spin $s=3/2$ fermions, in such a way that at mean field level the hidden SU$(2)$ symmetry is preserved and the correct ground state, the analog of a singlet in the usual Heisenberg model, is obtained for a single pair of nearest-neighbor sites. The corresponding mean field theory naturally includes the intrinsic spatial anisotropy of the strong SOC limit. We analyze two different mean field ansätze, with zero and $\pi$-flux. In both cases the mean field Hamiltonian respects all the symmetries of the original spin Hamiltonian. The $\pi$-flux state is found to have lower mean field energy. For both states, the spinons are at quarter-filling, leading to a spinon Fermi sea. There is no Fermi surface nesting and we expect that this spinon Fermi surface should be stable against weak perturbations. Predictions based on the picture of spinon Fermi surface are made. Finally in Sec. \[sec:sec6\], we compare our theoretical prediction with current experimental findings and suggest further directions for theory and experiment. Model and Symmetry {#sec:sec2} ================== Spin-orbit interaction and hybridization of atomic orbitals ----------------------------------------------------------- The magnetic ions B’ (Os$^{7+}$, Re$^{6+}$, Mo$^{5+}$) found in the ordered double perovskites in Table \[tab:Tab1\] all have one electron in the triply degenerate $t_{2g}$ multiplet. The atomic spin-orbit interaction projected down to the $t_{2g}$ triplet is written as $${{\mathcal}H}_{\text{so}} = - \lambda \; {\bm l}\cdot {{\boldsymbol}S} \;,$$ in which the total angular momentum quantum numbers of these operators are $l=1,S=1/2$. The effective orbital angular momentum ${{\boldsymbol}l}$ comes from the projection of orbital angular momentum ${{\boldsymbol}L}$ onto the $t_{2g}$ triplets, $$\mathcal{P}_{t_{2g}} {{\boldsymbol}L}\mathcal{P}_{t_{2g}} = - {{\boldsymbol}l} \;.$$ Here $\mathcal{P}_{t_{2g}} = \sum_{a=yz,xz,xy} |a\rangle \langle a|$ is the projection operator to the $t_{2g}$ manifold. The eigenstates of $l^z$ with eigenvalues $m=0,\pm 1$ and $S^z$ with eigenvalues $\sigma=\pm 1/2 \equiv \uparrow,\downarrow$, written in terms of the usual $t_{2g}$ states are $$|0,\sigma\rangle=|d^\sigma_{xy}\rangle;\quad |\pm1,\sigma\rangle=\frac{\mp|d^\sigma_{yz}\rangle- i|d^\sigma_{xz}\rangle}{\sqrt2}.$$ This interaction favors $j=3/2$ (${\boldsymbol}{j} = {\boldsymbol}{l}+ {\boldsymbol}{S}$) quadruplets over $j=1/2$ doublets by an energy separation $3\lambda/2$. In the strong spin-orbit interaction limit, the local Hilbert space is restricted to four low-lying states $$|d_\alpha\rangle=\sum_{m,\sigma}C_{m\sigma}^\alpha |m,\sigma\rangle,$$ where $\alpha=\pm3/2,\pm1/2$ is the $j^z$ eigenvalue and $$C_{m\sigma}^\alpha=\left\langle l=1,S=\frac12;m,\sigma \left| l=1,S=\frac12;j=\frac32,\alpha\right.\right\rangle$$ is a Clebsch-Gordan coefficient. In the materials under consideration, $\lambda$ is indeed a very large energy scale (fraction of an eV), justifying the strong SOC limit. Every operator expressed in terms of spin and orbitals must therefore be projected into this subspace and its projection can be written in terms of $j=3/2$ angular momentum operator. For example, $$\begin{aligned} {{\mathcal}P}_{\frac{3}{2}}\;{\boldsymbol}{S}\;{{\mathcal}P}_{\frac{3}{2}} &=& \frac{1}{3} \; {\boldsymbol}{j}, \\ {{\mathcal}P}_{\frac{3}{2}}\;{\boldsymbol}{l}\;{{\mathcal}P}_{\frac{3}{2}} &=& \frac{2}{3} \; {\boldsymbol}{j} \;. \end{aligned}$$ Here ${{\mathcal}P}_{\frac{3}{2}}$ is the projection operator into the $j=3/2$ quadruplets. Furthermore, for the magnetic moment ${{\boldsymbol}M}$ for electrons in atomic $d$ orbitals, we have $${{\boldsymbol}M} \equiv {{\mathcal}P}_{\frac{3}{2}} [2 {\boldsymbol}{ S} + (-{\boldsymbol}{ l})]{{\mathcal}P}_{\frac{3}{2}} = 0 \;.\label{zeromom}$$ The vanishing magnetic moment is quite remarkable and partially explains why the compounds have small magnetic moments in comparison with spin-$\frac{1}{2}$ systems without orbital degeneracy. In reality, the measured magnetic moments are nonzero because the atomic 4d or 5d orbitals strongly hybridize with p orbitals at the oxygen sites that form an octahedron surrounding each B’ site. For instance, for Ba$_2$NaOsO$_6$, the hybridization energy is estimated to be of the order of electron volts [@lee:epl2007; @Erickson] and comparable to the energy gap between Os d and O p states. For this reason, it is more appropriate to think in terms of molecular orbitals with mixed d and p character. For example, molecular $xy$ orbitals are written as $$|D^\sigma_{i,xy}\rangle=\frac{|d^\sigma_{i,xy}\rangle+r|p^\sigma_{i,xy}\rangle}{\sqrt{1+r^2}},\label{eq:1}$$ where $|d^\sigma_{i,xy}\rangle$ is the state corresponding to one electron in the $xy$ orbital and spin $\sigma$ on site $i$, and $|p^\sigma_{i,xy}\rangle$ is a linear combination (with $xy$ symmetry) of states that have a singlet on the $ d_{xy}$ orbital and one hole on an oxygen site $$|p^\sigma_{i,xy}\rangle=\frac12\left(|p^\sigma_{i+\hat{e}_x,y}\rangle+|p^\sigma_{i+\hat{e}_y,x}\rangle+|p^\sigma_{i-\hat{e}_x,y}\rangle+|p^\sigma_{i-\hat{e}_y,x}\rangle\right),$$ where $\hat{e}_{x,y}$ are real space vectors from the B’ site to neighboring oxygens along $x$ or $y$ directions. The mixing parameter $r$ is of order $t_{dp}/\Delta$, where $t_{dp}$ is the hopping matrix element between d and p orbitals and $\Delta$ is the gap to oxygen p states. In the limit of strong spin-orbit interaction, we must project into four low-lying molecular orbitals which are a superposition of the four atomic states with $j=3/2$ and p states $$|D_{i,\alpha}\rangle=\sum_{m,\sigma}C_{m\sigma}^\alpha \left| D_{i,m}^\sigma\right\rangle.$$ While the atomic magnetic moment in Eq. (\[zeromom\]) vanishes, there is a nonzero contribution to the molecular ${\boldsymbol}{M}$ from holes in p orbitals. After taking the projection into $j=3/2$ states, the coupling of the molecular orbital to a magnetic field reads $${\mathcal}{H}_{Z}=-g \mu_B \, {\boldsymbol}{h}\cdot {\boldsymbol}{j},$$ where $g=r^2/[3(1+r^2)]$ is the Landé factor, and $\mu_B$ is the Bohr magneton. Exchange interactions and electric quadrupolar interaction ---------------------------------------------------------- In the last subsection, we discussed the effect of strong spin-orbit interaction in determining the local degrees of freedom and pointed out that every operator must be projected into the $j=3/2$ quadruplets. In this subsection, we introduce the interactions between the local moments, and discuss the mechanics of the projection. The first interaction to consider is nearest-neighbor antiferromagnetic exchange, through the virtual transfer of electrons through intermediate oxygen p orbitals. These processes are strongly restricted by symmetry. For example, in XY planes, only electrons residing on $d_{xy}$ orbitals can virtually hop to neighboring sites via $p_x$ and $p_y$ orbitals of the intermediate oxygen sites. The exchange path and relevant orbitals are depicted in Fig. \[fig:fig1\]. Alternatively, one can interpret this process as kinetic exchange between molecular $D_{xy}$ orbitals, which are mixtures of the transition metal d state and p states on the neighboring four oxygens (see Eq. ). As a consequence, the antiferromagnetic exchange interaction can be written $\mathcal{H}_{\text{ex-1}} =\mathcal{H}_{\text{ex-1}}^{\text{XY}} +\mathcal{H}_{\text{ex-1}}^{\text{YZ}} +\mathcal{H}_{\text{ex-1}}^{\text{XZ}} $, where $${\mathcal H}_{\text{ex-1}}^{\text{XY}} = J \sum_{\langle ij \rangle \in \text{XY}}\left( {\boldsymbol}{S}_{i,xy} \cdot {\boldsymbol}{S}_{j,xy} - \frac{1}{4}\; n_{i,xy} n_{j,xy}\right) \;,$$ where the sum is over nearest neighbor sites in the XY planes, and the corresponding terms for YZ and XZ planes are obtained by the obvious cubic permutation. Here the operators ${\boldsymbol}{S}_{i,xy}$ and $n_{i,xy}$ denote the spin residing on $xy$ orbital and orbital occupation number at site $i$, respectively. In terms of spin and orbital angular momentum operators acting on site $i$, $$\begin{aligned} {\boldsymbol}{S}_{i,xy}&=&{\boldsymbol}{S}_i[1-(l_i^z)^2],\\ n_{i,xy}&=&1-(l_i^z)^2, \label{eq:spinnumber}\end{aligned}$$ Throughout this paper, we use the subindices ($i,xy$) to denote the site and orbitals, superindex ($\mu=x,y,z$) to denote the spin component, and capital letters (XY, XZ, YZ) to denote the planes. With these definitions, we note that the single occupancy condition at each site, which defines the Mott insulating state, becomes $$n_{i,xy}+n_{i,xz}+n_{i,yz} = 1. \label{singleoc}$$ Moreover, from Eq. , orbitally-resolved spins satisfy $${\boldsymbol}{S}_{i,xy} + {\boldsymbol}{S}_{i,yz}+ {\boldsymbol}{S}_{i,xz} = {\boldsymbol}{S}_i \;. \label{eq:spinconstraint}$$ The second interaction is the nearest-neighbor ferromagnetic exchange interaction. This interaction is due to the spin transfer through orthogonal orbitals at the intermediate oxygen sites in the exchange path, as shown in Fig. \[fig:fig1\]. For two sites $i,j$ in the XY plane, this ferromagnetic exchange is written as $$\begin{aligned} {\mathcal H}_{\text{ex-2},ij}^{\text{XY}} &=& - J' \left[ {\boldsymbol}{S}_{i,xy} \cdot ({\boldsymbol}{S}_{j,yz}+{\boldsymbol}{S}_{j,xz}) +\langle i \leftrightarrow j\rangle \right] \nonumber \\ & & - \frac{3J'}{4} \left[n_{i,xy} (n_{j,xz} +n_{j,yz}) +\langle i \leftrightarrow j\rangle \right] \;,\end{aligned}$$ where the $xy$ orbital only interacts with $yz$ and $xz$ orbitals at neighboring sites. Applying the single-occupancy constraint, the nearest-neighbor ferromagnetic exchange interaction can be simplified, up to a constant, to $$\begin{aligned} {\mathcal H}_{\text{ex-2}}^{\text{XY}} &=& - J' \sum_{\langle ij \rangle \in \text{XY}} \left[ {\boldsymbol}{S}_{i,xy} \cdot ({\boldsymbol}{S}_{j,yz}+{\boldsymbol}{S}_{j,xz}) +\langle i \leftrightarrow j\rangle \right] \nonumber \\ & & + \frac{3J'}{2} \sum_{\langle ij \rangle} n_{i,xy} n_{j,xy} \;. \label{exchangeFM2}\end{aligned}$$ Microscopically, $J'/J \sim {{\mathcal}O}(J_H/U_p) $ where $J_H$ and $U_p$ are the Hund’s coupling and Hubbard Coulomb interaction at the oxygen site, respectively. The third interaction is the electric quadrupole-quadrupole interaction. The 4d or 5d electron carries an electric quadrupole moment, and the interaction between these moments may not be negligible because of the long spatial extent of the molecular orbitals. Calculating the direct electrostatic energy between all possible orbital configurations for two electrons residing in neighboring sites in an XY plane, we obtain the quadrupole-quadrupole interaction $$\begin{aligned} {{\mathcal}H}_{\text{quad},ij}^{\text{XY}}& =& V n_{i,xy}n_{j,xy}\nonumber\\ &&-\frac{V}2[n_{i,xy}(n_{j,yz}+n_{i,xz})+(i\leftrightarrow j)]\nonumber\\ &&-\frac{13V}{12}(n_{i,yz}n_{j,yz}+n_{i,xz}n_{j,xz})\nonumber\\ &&+\frac{19V}{12}(n_{i,yz}n_{j,xz}+n_{i,xz}n_{j,yz}).\end{aligned}$$ Here $V>0$ is defined as the Coulomb repulsion between two nearest-neighbor $xy$ orbitals on XY planes. If $Q$ is the magnitude of the electric quadrupole and $a$ is the lattice constant of the fcc lattice, we have $V=9\sqrt2Q^2/a^5$. In general, the main contribution to $Q$ comes from the charge at the oxygen sites, hence the larger the hybridization, the larger the value of $V$. Using the single-occupancy constraint in Eq. (\[singleoc\]) and summing over sites, the quadrupole-quadrupole interaction simplifies to $$\begin{aligned} {{\mathcal}H}_{\text{quad}}^{\text{XY}} &=& \sum_{\langle ij \rangle\in \text{XY}}\left[- \frac{4V}{3} (n_{i,xz} - n_{i,yz}) (n_{j,xz} - n_{j,yz}) \right.\nonumber\\ &&\left.+\frac{9V}{4} n_{i,xy} n_{j,xy} \right], \label{quadrupole2}\end{aligned}$$ in which we have ignored an unimportant constant. The minimal Hamiltonian for the cubic system contains all three of these exchange interactions in addition to the on-site SOC, $${{\mathcal}H} = {{\mathcal}H}_{\text{ex-1}} + {{\mathcal}H}_{\text{ex-2}} + {{\mathcal}H}_{\text{quad}} + {{\mathcal}H}_{\text{so}} \;.$$ Since we are interested in the limit of strong spin-orbit interaction, we need to project ${{\mathcal}H}$ onto the $j=3/2$ quadruplets at every site. As an example, we write down the projection for ${\boldsymbol}{S}_{i,xy}$ and $n_{i,xy}$, $$\begin{aligned} \label{eq:projxy} \tilde{S}_{i,xy}^x &=&\frac{1}{4} j_i^{x} -\frac{1}{3} {j_i^z j_i^{x} j_i^z} \\ \tilde{S}_{i,xy}^y &=& \frac{1}{4}{j_i^{y}}- \frac{1}{3}{j_i^z j_i^{y} j_i^z}\\ \tilde{S}_{i,xy}^z &=& \frac{3}{4} j_i^z -\frac{1}{3} j_i^z j_i^{z} j_i^z \\ \tilde{n}_{i,xy} &=& \frac{3}{4} - \frac{1}{3}(j_i^z)^2,\end{aligned}$$ in which, $ \tilde{{\mathcal}O} \equiv {{\mathcal}P}_{\frac{3}{2}}\; {{\mathcal}O} \; {{\mathcal}P}_{\frac{3}{2}} $. Spin and occupation number operators for other orbitals can be readily generated by a cubic permutation. After the projection, the minimal Hamiltonian reduces, up to a constant, to $$\tilde{{\mathcal}H} = \tilde{{\mathcal}H}_{\text{ex-1}} + \tilde{{\mathcal}H}_{\text{ex-2}} + \tilde{{\mathcal}H}_{\text{quad}} \;.\label{model}$$ As one may notice, the projected Hamiltonian contains 4-spin and 6-spin interactions in addition to the usual quadratic 2-spin interactions if it is expressed in terms of the effective spin moment ${{\boldsymbol}j}_i$. One can view these multiple spin terms as the interaction between magnetic multipoles (quadrupole and octupole) at different sites. Such multipolar Hamiltonians are much less familiar than the usual quadratic exchange forms, and some caution should be used. In particular, experience with similar models shows that such interactions can magnify quantum effects, for instance leading to the appearance of a quadrupolar phase in the biquadratic case[@läuchli2006quadrupolar]. Hence, the naïve classical approximation – replacing ${\boldsymbol}{j}$’s by classical vectors – is inadvisable, and we will proceed differently below. Symmetry properties of the Hamiltonian -------------------------------------- Before we move on to discuss the ground state of the Hamiltonian $\tilde{{\mathcal}H}$ in Eq. (\[model\]), we need to have some understanding about its symmetry properties. We start from the NN AFM exchange interaction $\tilde{{\mathcal}H}_{\text{ex-1}}$. The latter has an apparent cubic space group symmetry. The total angular momentum ${\boldsymbol}{J}=\sum_i{\boldsymbol}{j}_i$ is *not* conserved, $[\tilde{{\mathcal}H}_{\text{ex-1}},{\boldsymbol}{J}]\neq0$. Nevertheless, $\tilde{{\mathcal}H}_{\text{ex-1}}$ surprisingly has a “hidden” SU$(2)$ symmetry. The three generators of this global *continuous* symmetry are defined as follows, $$G^{\mu} = \sum_iG^\mu_i= \sum_i \left[\frac{7}{6} j_i^{\mu} - \frac{2}{3} (j_i^{\mu})^3 \right],$$ with $\mu = x,y,z$. One can readily check that these generators commute with $\tilde{{\mathcal}H}_{\text{ex-1}}$, $$[G^{\mu}, \tilde{{\mathcal}H}_{\text{ex-1}}] =0 ,$$ and satisfy the SU$(2)$ algebra, $$[G^{\mu}, G^{\nu}] = i \epsilon_{\mu\nu\lambda} G^{\lambda} \;.$$ In addition, the Casimir operator $\mathbf{G}^2$ also commutes with $\tilde{{\mathcal}H}_{\text{ex-1}}$. The physical meaning of these generators is easy to see if one expresses $G^{x,y,z}$ in matrix form. For a single site, $$\begin{aligned} \label{eq:hiddensu2} G^x_i & = & -\frac{1}{2} \left[\begin{array}{llll} & & & 1 \\ & & 1 & \\ & 1 & & \\ 1 & & & \end{array} \right] = \frac{1}{2} (-\sigma^x)_{14} \oplus (-\sigma^x)_{23} \\ G^y_i & = & \frac{1}{2} \left[\begin{array}{llll} & & & -i \\ & & i & \\ & -i & & \\ i & & & \end{array} \right] = \frac{1}{2} (\sigma^y)_{14} \oplus (-\sigma^y)_{23} \\ G^z_i & = & \frac{1}{2} \left[\begin{array}{llll} -1 & & \\ & 1 & & \\ & & -1 & \\ & & & 1 \end{array} \right] = \frac{1}{2} (-\sigma^z)_{14} \oplus (\sigma^z)_{23},\end{aligned}$$ in which the empty matrix entries are zero and we have expressed these generators as the direct sum of two Pauli matrices, one ($\boldsymbol{\sigma}_{14}$) for the subspace of $j^z_i = \pm 3/2$ states and the other ($\boldsymbol{\sigma}_{23}$) for the subspace of $j^z_i = \pm 1/2$ states. One intuitive way to think about these SU$(2)$ generators is that they transform the spin components in the $j^z_i = \pm 3/2$ subspace together with $j^z_i = \pm 1/2$ subspace. This is a global symmetry of $\tilde{{\mathcal}H}_{\text{ex-1}}$. Now we consider the other two interactions, $\tilde{\mathcal H}_{\text{ex-2}}$ and $\tilde{\mathcal H}_{\text{quad}}$. We find that the electric quadrupole-quadrupole interaction $\tilde{\mathcal H}_{\text{quad}}$ also commutes with $\mathbf{G}$. On the other hand, the ferromagnetic exchange interaction $\tilde{{\mathcal}H}_{\text{ex-2}}$ breaks this SU$(2)$ symmetry; thus $$[\tilde{\mathcal H},\mathbf{G}]\propto J'.$$ For $J'\ll J,V$, we have an approximate continuous symmetry. Mean-field ground states {#sec:sec3} ======================== In this section, we study the ground state of the model Hamiltonian $\tilde{{\mathcal}H}$ in Eq. . We begin in Sec. \[sec:sec31\] by getting some intuition from considering a perturbed model with strong easy-axis or easy-plane anisotropy. This starting point also has experimental motivation as several ordered double perovskites in Tab. \[tab:Tab1\] develop such anisotropies that are driven by lattice distortions. Armed with the understanding of the anisotropic cases, we proceed to analyze the case of cubic symmetry using mean-field theory in Sec. \[sec:sec32\]. We consider briefly intermediate strength anisotropy in Sec. \[sec:interm-anis\]. In general, Curie-Weiss mean-field theory consists of decoupling all inter-site interactions to obtain self-consistent single-site Hamiltonians. At zero temperature, this is equivalent to assuming a product form for the wavefunction, i.e. $$\label{eq:20} |\Psi\rangle = \otimes_i |\psi_i\rangle,$$ where the product is over sites, and $|\psi_i\rangle$ is an arbitrary $j=3/2$ ket. One calculates the mean-field ground state energy as the expectation value of the Hamiltonian in this state, and minimizes it. Thus the mean-field approximation can also be considered as a simple variational one. The case with strong anisotropy {#sec:sec31} ------------------------------- In this subsection, we add to the Hamiltonian $\tilde{{\mathcal}H}$ in Eq. (\[model\]) a strong anisotropic term, $${{\mathcal}H}_{\text{ani}} = - D \sum_i (j_i^z)^2 \;,$$ in which, $D$ can be positive or negative, representing easy-axis or easy-plane anisotropy, respectively. Although this interaction is anisotropic in spin space, it still respects the “hidden” global SU(2) symmetry. ### Easy-axis anisotropy {#sec:easy-axis-anisotropy} Let us start with easy-axis anisotropy $D>0$. Assuming the anisotropy is very strong $D \gg J,J',V$ and $D \ll \lambda$, which favors $j_i^z = \pm 3/2$ states, we can safely project the Hamiltonian $\tilde{{\mathcal}H}$ into the latter two-dimensional subspace. We then obtain the effective Hamiltonian $$\begin{aligned} {{\mathcal}H}_{\text{eff-1}} & = & \sum_{\langle ij \rangle \in \text{XZ,YZ}} \left[\left(\frac{J}{4}+\frac{J'}{2}\right) \boldsymbol{T}_i \cdot \boldsymbol{T}_j - J' {T}_i^z {T}_j^z\right ] \nonumber \\ & & + {{\mathcal}N}\left(-\frac{J}{4}+\frac{3J'}{2} + \frac{11V}{12}\right), \label{eq:projz}\end{aligned}$$ where we have introduced pseudospin-$1/2$ operators $\boldsymbol{T}_i$ acting on the basis $j_i^z = \pm 3/2$ with $T_i^z = \pm 1/2$ corresponding to $j_i^z =\pm 3/2$, respectively. And ${{\mathcal}N}$ is the total number of sites. In the reduced space, the original spin vector reduces to the pseudospin in the following way: $$(j^x,j^y,j^z) \Rightarrow 3(0,0,T^z) \;.$$ Notice that after this projection the interaction on the horizontal bonds in XY planes disappear in the effective Hamiltonian ${{\mathcal}H}_{\text{eff-1}}$. This can be understood in terms of the original orbital picture as the easy-axis anisotropy lifts the degeneracy of $t_{2g}$ triplets, favoring $xz$ and $yz$ orbitals to be occupied. As a result, [*the above effective Hamiltonian is operating on a bond-depleted fcc lattice, which is in fact a unfrustrated bipartite bcc lattice*]{}. In ${{\mathcal}H}_{\text{eff-1}}$ because of the in-plane anisotropy introduced by the FM exchange, the ground state of ${{\mathcal}H}_{\text{eff-1}}$ is “antiferromagnetically” ordered in the $(T^x,T^y)$ plane with an ordering wavevector ${\boldsymbol}{Q} = 2\pi (001)$. We denote this as the AFM state. The corresponding mean-field ground state is just the direct product, $$|\Psi(\phi)\rangle=\prod_i|\psi_i(\phi)\rangle, \label{eq:state_aniz1}$$ where $$| \psi_i (\phi)\rangle = \frac1{\sqrt2}[| j_i^z=\tfrac{3}{2} \rangle + (-)^{2z_i} e^{i \phi}|j_i^z= \tfrac{3}{2} \rangle]. \label{eq:state_aniz2}$$ with an arbitrary phase $\phi$. The arbitrariness of the phase comes from the U$(1)$ symmetry of the projected effective Hamiltonian Eq. . However, as discussed in the previous section, the continuous symmetry in the original Hamiltonian is broken completely when $J'\neq 0$. Therefore, the U(1) symmetry of Hamiltonian in Eq.  is a by-product of the projection. Because we are in the subspace of $j^z = \pm 3/2$, the orbital occupation is $$(\langle \tilde{n}_{i,yz} \rangle, \langle \tilde{n}_{i,xz} \rangle, \langle\tilde{n}_{i,xy} \rangle) = (1/2,1/2,0) .$$ It is also important to note that the ground state in Eq.  is not a conventional Néel state as it has a vanishing static magnetic dipole moment, $$\langle \Psi | {{\boldsymbol}j}_i| \Psi \rangle = 0 \;!$$ The $\phi$ dependence only shows up in the spin operators of a specific orbital, $$\begin{aligned} \langle\Psi | \tilde {{\boldsymbol}S}_{i,yz} | \Psi \rangle &=& -\frac{1}{4}(-)^{2z_i}(\cos{\phi},\sin{\phi},0 ),\\ \langle\Psi | \tilde{{\boldsymbol}S}_{i,xz} | \Psi \rangle &=& \frac{1}{4}(-)^{2z_i}(\cos{\phi},\sin{\phi},0 ), \\ \langle\Psi | \tilde{{\boldsymbol}S}_{i,xy} | \Psi \rangle &=& (0,0,0).\end{aligned}$$ ### Easy-plane anisotropy {#sec:easyplane} Now we consider easy-plane anisotropy $D<0$. We also assume the anisotropy is very strong $|D| \gg J,J',V$ and $|D| \ll \lambda$, which favors $j_i^z = \pm 1/2$ states and obtain the effective Hamiltonian after projection into the $j_i^z =\pm 1/2$ subspace, $$\begin{aligned} {{\mathcal}H}_{\text{eff-2}} & = & \sum_{\langle ij \rangle \in \text{XY} } \frac{4}{9} ( J \boldsymbol{T}_i \cdot \boldsymbol{T}_j + J' T_i^z T_j^z )+ \sum_{\langle ij \rangle \in \text{XZ} } \left[ \frac{J}{36} \boldsymbol{T}_i \cdot \boldsymbol{T}_j \right. \nonumber \\ &+&\left. J'(-\frac{1}{6} T_i^x T_j^x + \frac{5}{18} T_i^y T_j^y +\frac{1}{6} T_i^z T_j^z )\right] \nonumber \\ &+& \sum_{\langle ij \rangle \in \text{YZ} } \left[ \frac{J}{36} \boldsymbol{T}_i \cdot \boldsymbol{T}_j + J' ( \frac{5}{18} T_i^x T_j^x -\frac{1}{6} T_i^y T_j^y \right. \nonumber \\ &+& \left. \frac{1}{6} T_i^zT_j^z) \right] + {{\mathcal}N} (- \frac{J}{4} + \frac{3J'}{2} + \frac{11V}{12} ) \;.\end{aligned}$$ Here the pseudospin-$1/2$ operator $\boldsymbol{T} $ is acting on the subspace of $j^z = \pm 1/2$ with $T^z = \pm 1/2 $ representing $j^z = \pm 1/2$, respectively. In the reduced spin space, the original spin vector is reduced to the pseudospin in the following way: $$(j^x,j^y,j^z) \Rightarrow (2 T^x, 2 T^y,T^z)\label{eq:3} \;.$$ We can now find the mean-field ground state of this Hamiltonian. For an effective $S=1/2$ model of this type, this is equivalent to the classical approximation. Classically, we can find the minimum energy states by the Luttinger-Tisza method. This amounts to looking for the eigenvalues of the spin Hamiltonian regarded as a quadratic form, and seeking a classical spin solution which is built of a superposition only of those eigenvectors which have minimum energy eigenvalues. The result in this case is that, for $0<J'<J$, there are two classes of solution, all collinear spin states. These are: (i) states with ${{\boldsymbol}Q}=2\pi(100)$ and the pseudo-spin axis in the $yz$ plane, and (ii) states with ${{\boldsymbol}Q}=2\pi(010)$ and the pseudo-spin axis in the $xz$ plane. As for the easy-axis case, there is an accidental degeneracy of spin orientations within the plane normal to ${{\boldsymbol}Q}$. Note that while the pseudospin orients freely along a circle in this plane, the magnetization orients along an ellipse due to the factor of 2 in Eq. . One readily expresses the ground state in the unprojected Hilbert space. For example, taking ${{\boldsymbol}Q} = 2\pi (010)$ and pseudospin pointing along $x$ direction, then $$\label{eq:easyplanstate} |\Psi\rangle = \prod_i | \psi_i \rangle$$ with $$\label{eq:easyplanestate} | \psi_i \rangle = \frac{1}{\sqrt{2}} [ | j_i^z = \tfrac{1}{2} \rangle + (-)^{2y_i} | j_i^z = -\tfrac{1}{2} \rangle ] \;.$$ This is once again an antiferromagnetic state, and to distinguish it from the one which obtains for Ising anisotropy, we denote it AFM’. The defining difference of the AFM’ and the AFM state discussed previously is that, the former has a non-zero dipole moment, while, at least within mean field theory, the latter does not. The cubic case {#sec:sec32} -------------- Having understood the cases with strong easy-axis and easy-plane anisotropies, let us now turn to the Hamiltonian $\tilde{{\mathcal}H}$ with cubic lattice symmetry in Eq. . Both ground states of the Hamiltonian with strong easy-axis or easy-plane anisotropy comprise two-sublattice structure with an ordering wavevector equivalent to ${{\boldsymbol}Q} = 2\pi (001)$. It is therefore natural to guess that the same two sublattice structure is also obtained in the cubic case. While we have not proven this, we have investigated more general mean-field ground states, allowing for much larger unit cells, but found in every case that the minimum energy is found for the two-sublattice configuration. Therefore, in what follows, we assume the two sublattice structure with ordering ${{\boldsymbol}Q} = 2\pi (001)$ (which is equivalent to $2\pi (100)$ and $2\pi(010)$ in the cubic case). We make no further assumptions, and minimize the energy with respect to an arbitrary wavefunction on each of the two sublattices. The resulting variational phase diagram is depicted in Fig. \[fig:phase\_diagram\]. ### Antiferromagnetic (AFM) state {#sec:antif-afm-state} In Fig. \[fig:phase\_diagram\], for small $J'/J$ and $V/J$, we find a phase, denoted AFM, which is the natural continuation of the AFM phases encountered in the anisotropic limits. Here, as in those cases, the states on the two sublattices are simply related by a time reversal transformation, and indeed the ground state has the same form as that found in the easy-axis case, Eqs. ,. The appearance of time-reversed pairs of sites is natural, since the largest interaction, $\tilde{{\mathcal}H}_{\text{ex-1}}$, is dominated by the spin-spin exchange term. Interestingly, one finds that the ground state has a [*continuous*]{} degeneracy: the phase $\phi$ in Eq.  can be arbitrary. Since the Hamiltonian with non-vanishing $J'$ has no continuous symmetry, this degeneracy appears to be accidental. Since it has the same form as we found in Sec. \[sec:easy-axis-anisotropy\], we continue to use the label AFM here for this state (and in Fig. \[fig:phase\_diagram\]). ### Ferromagnetic 110 (FM110) state {#sec:ferr-110-fm110} With large $J'/J$ and $V/J$, the orbital-orbital interaction has more weight in the Hamiltonian $\tilde{{\mathcal}H}$, and the nature of the ground state changes. One should note that even the pure orbital-orbital interaction is not trivial and classical, since the orbital occupation numbers no longer commute after projection down to the $j=3/2$ quadruplets. However, in mean-field theory one may still treat the expectation values classically. Note that the largest terms in the orbital-orbital interaction are those which are diagonal in the orbital basis, namely, the second term in Eq.  and the second term in Eq. . To minimize the diagonal orbital interaction like $\tilde{n}_{i,xz} \tilde{n}_{j,xz}$, a schematic recipe is to maximize $\tilde{n}_{i,xz} $ while minimizing $\tilde{n}_{j,xz} $. This is necessary because one cannot minimize both $\tilde{n}_{i,xz} $ and $\tilde{n}_{j,xz} $ simultaneously, since, due to the single-occupancy constraint, the other diagonal terms such as $\tilde{n}_{i,yz} \tilde{n}_{j,yz}$ would then be increased. Since the occupation numbers of the same orbital must be taken different on different sites, and these occupation numbers are time-reversal invariant, the states on the two sublattices cannot be time-reversed counterparts. Consequently, there is a competition between the orbital-orbital interactions ($\tilde{{\mathcal}H}_{\text{ex-2}}$ and $\tilde{{\mathcal}H}_{\text{quad}}$) and the nearest neighbor antiferromagnetic exchange interaction ($\tilde{{\mathcal}H}_{\text{ex-1}}$). In the large $J'/J$ and $V/J$ region, when the orbital interactions dominate, we find however that time-reversal symmetry is still broken, and since these states are not composed of time-reversed pairs, the result is an uncompensated net ferromagnetic moment. In the majority of phase space, we find the ground state is characterized by three parameters, $r$, $\phi_1$ and $\phi_2$, $$\begin{aligned} |\psi_A \rangle_{\text{FM110}} & = & \frac{r}{ \sqrt{2}} ( e^{i\phi_1} |j^z=\tfrac{1}{2} \rangle + e^{i(\phi_2-\phi_1)}|j^z= -\tfrac{1}{2} \rangle ) + \sqrt{\frac{1-r^2}{2}} \left( e^{i \phi_2} |j^z=\tfrac{3}{2} \rangle +|j^z=-\tfrac{3}{2} \rangle \right) \\ |\psi_B \rangle_{\text{FM110}} & = & \frac{r}{ \sqrt{2}} ( - e^{-i\phi_1} |j^z=\tfrac{1}{2} \rangle + i e^{i(\phi_1 -\phi_2)}|j^z=-\tfrac{1}{2} \rangle ) + \sqrt{\frac{1-r^2}{2}} (-i e^{-i \phi_2 } |j^z=\tfrac{3}{2} \rangle +|j^z= -\tfrac{3}{2} \rangle ) \;, \label{eq:stateFM1}\end{aligned}$$ in which, “A” and “B” represent the two sublattices, and $r, \phi_1$ and $\phi_2$ are determined by minimizing the mean field energy. Note that in Eq.  the three parameters $r, \phi_1$ and $\phi_2$ are uniquely determined by $J'/J$ and $V/J$. So the orbital occupations can be readily generated, $$\begin{aligned} (\langle \tilde{n}_{A,yz} \rangle, \langle \tilde{n}_{A,xz} \rangle, \langle\tilde{n}_{A,xy} \rangle )_{\text{FM110}} &=& (\frac{1}{2} -\frac{r^2}{3}-\frac{r\sqrt{1-r^2} }{\sqrt{3}} \cos{\phi_1},\frac{1}{2} -\frac{r^2}{3}+\frac{r \sqrt{1-r^2}}{\sqrt{3}} \cos{\phi_1}, \frac{2r^2}{3}), \\ (\langle \tilde{n}_{B,yz} \rangle, \langle \tilde{n}_{B,xz} \rangle, \langle\tilde{n}_{B,xy} \rangle )_{\text{FM110}} &=& (\frac{1}{2} -\frac{r^2}{3}+\frac{r\sqrt{1-r^2} }{\sqrt{3}} \cos{\phi_1},\frac{1}{2} -\frac{r^2}{3}-\frac{r\sqrt{1-r^2} }{\sqrt{3}} \cos{\phi_1}, \frac{2r^2}{3}) .\end{aligned}$$ It is interesting to see the spin vectors of two sublattices are symmetric about $[1\bar{1}0]$ direction, $$\begin{aligned} \langle {{\boldsymbol}j}_A \rangle_{\text{FM110}} &=& r \left( \sqrt{3-3r^2} \cos{(\phi_1-\phi_2)} + r \cos{(2\phi_1-2\phi_2)}, \sqrt{3-3r^2} \sin{(\phi_1-\phi_2)} -r \sin{(2\phi_1-2\phi_2)}, 0 \right) \\ \langle {{\boldsymbol}j}_B \rangle_{\text{FM110}} &=& r \left(- \sqrt{3-3r^2} \sin{(\phi_1-\phi_2)} + r \sin{(2\phi_1-2\phi_2)}, -\sqrt{3-3r^2} \cos{(\phi_1-\phi_2)} - r \cos{(2\phi_1-2\phi_2)},0 \right)\end{aligned}$$ so the system has a non-vanishing net spin polarization, that is $$\tfrac{1}{2} \langle {{\boldsymbol}j}_A + {{\boldsymbol}j}_B\rangle_{\text{FM110}} = \frac{r}{2} \left[ \sqrt{3-3r^2} (\cos{(\phi_1-\phi_2)} - \sin{(\phi_1-\phi_2)} ) + r (\cos{(2\phi_1-\phi_2)} + \sin{(2\phi_1-\phi_2)} )\right] (1,-1,0)\;.$$ This direction of polarization is equivalent to $[110]$ by a 90 degree rotation, so we denote this a FM110 state. It occupies the corresponding region in Fig. \[fig:phase\_diagram\]. ### Ferromagnetic 100 (FM100) state {#sec:ferr-100-state} Between the AFM and FM110 states, a narrow region of intermediate phase intervenes (see Fig. \[fig:phase\_diagram\]). Numerically we find this phase is characterized by two parameters $r_1$ and $r_2$, $$\begin{aligned} |\psi_A \rangle_{\text{FM100}} & = & \frac{r_1}{ \sqrt{2}} ( |j^z=\tfrac{1}{2} \rangle + |j^z= -\tfrac{1}{2} \rangle ) + \sqrt{\frac{1-r_1^2}{2}} \left( |j^z=\tfrac{3}{2} \rangle +|j^z=-\tfrac{3}{2} \rangle \right) \\ |\psi_B \rangle_{\text{FM100}} & = & \frac{r_2}{ \sqrt{2}} (- |j^z=\tfrac{1}{2} \rangle + |j^z=-\tfrac{1}{2} \rangle ) + \sqrt{\frac{1-r_2^2}{2}} (-|j^z=\tfrac{3}{2} \rangle +|j^z= -\tfrac{3}{2} \rangle ) \;. \label{eq:FM100}\end{aligned}$$ The parameters $r_1$ and $r_2$ are determined by $J'/J$ and $V/J$ and in this intermediate phase $r_1 \neq r_2$. A second ground state is obtained then by interchanging $r_1$ and $r_2$. The orbital occupation numbers and spin vectors are give by $$\begin{aligned} (\langle \tilde{n}_{A,yz} \rangle, \langle \tilde{n}_{A,xz} \rangle, \langle\tilde{n}_{A,xy} \rangle )_{\text{FM100}} &=& (\frac{1}{2} -\frac{r_1^2}{3}-\frac{r_1}{\sqrt{3}} \sqrt{1-r_1^2},\frac{1}{2} -\frac{r_1^2}{3}+\frac{r_1}{\sqrt{3}} \sqrt{1-r_1^2}, \frac{2r_1^2}{3}), \\ (\langle \tilde{n}_{B,yz} \rangle, \langle \tilde{n}_{B,xz} \rangle, \langle\tilde{n}_{B,xy} \rangle )_{\text{FM100}} &=& (\frac{1}{2} -\frac{r_2^2}{3}+\frac{r_2}{\sqrt{3}} \sqrt{1-r_2^2} ,\frac{1}{2} -\frac{r_2^2}{3}-\frac{r_2}{\sqrt{3}} \sqrt{1-r_2^2} , \frac{2r_2^2}{3}) \end{aligned}$$ and $$\begin{aligned} \langle {{\boldsymbol}j}_A \rangle_{\text{FM100}} &=& \left(r_1^2 + r_1 \sqrt{3-3r_1^2} , 0, 0 \right) \\ \langle {{\boldsymbol}j}_B \rangle_{\text{FM100}} &=& \left(r_2^2 - r_2 \sqrt{3-3r_2^2}, 0, 0 \right)\;.\end{aligned}$$ We see that the net spin polarization is along the $[100]$ direction. Due to cubic symmetry, all possible $[ 100]$ directions are possible. By analogy with the previous phase, we denote this phase FM100. It occupies the narrow region shown in Fig. \[fig:phase\_diagram\]. ### Transitions {#sec:transitions} The intermediate FM100 state in Eq.  is smoothly connected to the AFM state, which is obtained by setting $r_1=r_2=0$. By contrast, it cannot be connected to the FM110 state. This indicates that the transition between FM100 to AFM is continuous while the transition from FM100 to FM110 is first-order. Indeed, this can also be clearly seen from the behavior of the spin and orbital order parameters across these transitions (see Fig. \[fig:orbitalspin\]). Both the spin and orbital order parameters jump when the system goes from FM100 phase to FM110 phase. The approach to the line $J'=0$, while not a transition [*per se*]{}, does represent a change of behavior. In particular, on this line, the hidden SU(2) symmetry is restored, and new ground states may be obtained from the above three phases by SU(2) rotations. This allows, for instance, for AFM states to develop with non-vanishing magnetic dipole order in the ground state. However, for arbitrarily small $J'$ the SU(2)-induced degeneracy is broken and the results quoted above hold. Intermediate anisotropy {#sec:interm-anis} ----------------------- We briefly address here the evolution of the ground states with increasing $|D|$ between the cubic and strongly anisotropic limits. For easy-axis anisotropy, $D>0$, this process is relatively simple. The AFM phase (with ${{\boldsymbol}Q}=2\pi(001)$) is favored by this sign of anisotropy, and therefore, with increasing $D$, it expands at the expense of the FM110 and FM100 states. Indeed, for very large $D$, ferromagnetic states occur only for unphysically large $J'$ and $V$. In the case of easy-plane anisotropy, $D<0$, the phase diagram is more subtle. For weak $|D|$, the main effect is to break the symmetry between the formerly equivalent $[ 100]$ wavevectors. In this case, states with minimal $\langle j_z^2 \rangle$ are favored, which prefers ${{\boldsymbol}Q}=2\pi(100), 2\pi(010)$ rather than ${{\boldsymbol}Q}=2\pi(001)$. If we begin in the AFM state for $D=0$, this aligns the pseudospin in the plane normal to this wavevector. The phase degeneracy which obtains for the cubic case is broken by the anisotropy, and a definite alignment is obtained. Moreover, as states with $j_z=\pm 1/2$ are increasingly mixed into the ground state, a non-vanishing dipole moment, proportional to the pseudospin, is induced. The magnitude of this staggered magnetization grows continuously with $|D|$, eventually as $D\rightarrow -\infty$, approaching the value obtained in Sec. \[sec:easyplane\]. This local moment is oriented in the plane normal to ${{\boldsymbol}Q}$, and can take values distributed over an [*ellipse*]{} in this plane. In the large $|D|$ limit, this ratio of the major (perpendicular to ${{\boldsymbol}Q}$ and to $z$) and minor ($z$) axes of the ellipse approaches 2, corresponding to the accidental degeneracy discussed in Sec. \[sec:easyplane\] . Because the state for non-zero $D$ evolves smoothly into this limit, and has a non-zero local moment, we denote it an AFM’ state, following the earlier notation. Beginning in the FM110 state at $D=0$, one observes two subsequent transitions. First, small $|D|$ orients the magnetization normal to ${{\boldsymbol}Q}$. For concreteness consider ${{\boldsymbol}Q}=2\pi(100)$, in which case one obtains a ferromagnetic magnetization of the form ${{\boldsymbol}m}=(0,m_1,m_2)$. For $D=0^-$, $m_1=m_2$, but subsequently $m_2$ decreases such that $m_2<m_1$. We denote this state FM110\*. An example for the orientation of net polarization in the FM110\* by varying $V$ is given in Fig. \[fig:orientation\]. Eventually once some critical anisotropy is reached, $m_2$ vanishes continuously. At this point the magnetization is aligned along the $(010)$ axis. For yet larger anisotropy, eventually the ferromagnetic magnetization vanishes entirely, and the ground state switches to the AFM’ state. An example of the transitions from FM110\* to FM100 then to AFM’ by varying the easy-plane anisotropy $D$ is given in Fig. \[fig:orientation1\]. Finally, starting in the FM100 state, the magnetization immediately switches to the $(010)$ direction. This is the same phase as the intermediate phase observed starting from the FM110 phase. Thus with further increase in anisotropy, the ground state switches to the AFM’ state. One may also visualize the evolution of the ground states with anisotropy by considering planar phase diagrams at fixed $D$. With increasing positive $D$ (Ising anisotropy), the AFM state is stabilized, and simply expands in the $J'-V$ plane, pushing the FM100 and FM110 states outward. For increasing negative $D$, apart from the fact that the AFM and FM110 states evolve into the AFM’ and FM110\* states, the behavior is similar: the AFM’ state expands at the expense of the ferromagnetic states. ![ (Color online) The orientation of the net spin polarization for the FM110$^*$ phase, at $T=0$. $\theta$ is angle between the net spin polarization and nearest $[110]$ direction. In the figure, $J'=0.2, D=-0.05, J=1$. $\cos {\theta}$ increases from $1/\sqrt{2}$ for the FM100 phase to $1$ for the FM110$^*$ phase as $V$ goes through the phase transition point.[]{data-label="fig:orientation"}](fig7.pdf){width="6.5cm"} ![ (Color online) The net polarization versus the easy plane anisotropy, at $T=0$. The upper curve (in red) is for the $y$ component of the net polarization. And the lower curve (in blue) is for the $z$ component of the net polarization. The ordering wavevector is ${{\boldsymbol}Q} = 2\pi (100)$. In the graph, $J'=V=0.4$ and $J=1$. When $0<|D|\lesssim 0.36$, the system is in FM110$^*$ phase; when $0.36<|D|<0.41$, the system is in the FM100 phase; when $|D|\gtrsim 0.41$, the system is in AFM’ phase.[]{data-label="fig:orientation1"}](fig8.pdf){width="6cm"} Multipolar orders and $T>0$ behavior {#sec:sec4} ==================================== Order parameters {#sec:order-parameters} ---------------- In this section, we extend the analysis of the previous section to non-zero temperature. To do so, we employ the usual extension of mean field theory to include thermal fluctuations. To characterize the phases encountered in this treatment, it is natural to introduce several types of order parameter. First, on a single site $i$, we may measure the dipole moment, which is proportional to ${{\boldsymbol}j}_i$. However, we may also measure the next two multipoles: the quadrupole moment, proportional to $$\label{eq:4} Q_{i}^{\mu\nu} = \left\langle j_i^\mu j_i^\nu\right\rangle - \frac{j(j+1)}{3} \delta^{\mu\nu},$$ and the octupole moment $$\label{eq:5} O_i^{\mu\nu\lambda} = \left\langle j_i^\mu j_i^\nu j_i^\lambda \right\rangle.$$ A typical magnetic state has a non-vanishing local dipole moment, which inevitably induces some higher multipole order parameters (see below). However, one sometimes encounters purely multipole states, in which $\langle j_i^\mu\rangle=0$ but $Q_i^{\mu\nu}$ and/or $O_i^{\mu\nu\lambda}$ are/is non-vanishing. Moment Symmetry Operator ------------ ------------ ---------------------------------------------------------------------------------- Dipole $\Gamma_4$ $M^x = j^x$ $M^y = j^y$ $M^z = j^z$ Quadrupole $\Gamma_3$ $Q^{3z^2} = [3(j^z)^2- {\bm j}^2]/\sqrt{3} $ $Q^{x^2-y^2} = ( j^x)^2-(j^y)^2 $ $\Gamma_5$ $Q^{xy} = \overline{j^x j^y}/2 $ $Q^{yz} = \overline{j^y j^z}/2 $ $Q^{xz} = \overline{j^z j^x}/2 $ Octupole $\Gamma_2$ $ T_{xyz} = \sqrt{15}/6 \overline{j^x j^y j^z} $ $\Gamma_4$ $ T^x_{\alpha} = (j^x)^3 - [\overline{j^x (j^y)^2} + \overline{(j^z)^2 j^x}]/2 $ $ T^y_{\alpha} = (j^y)^3 - [\overline{j^y (j^z)^2} + \overline{(j^x)^2 j^y}]/2 $ $ T^z_{\alpha} = (j^z)^3 - [\overline{j^z( j^x)^2} + \overline{(j^y)^2 j^z}]/2 $ $\Gamma_5$ $ T^x_{\beta} = \sqrt{15}[\overline{j^x (j^y)^2} - \overline{(j^z)^2 j^x}]/6 $ $ T^y_{\beta} = \sqrt{15}[\overline{j^y (j^z)^2} - \overline{(j^x)^2 j^y}]/6 $ $ T^z_{\beta} = \sqrt{15}[\overline{j^z (j^x)^2} - \overline{(j^y)^2 j^z}]/6 $ : Multipole moments within a cubic $\Gamma_8$ quartet. Bars over symbols indicate the sum with respect to all the possible permutations of the indices, e.g. $\overline{j^x (j^y)^2} = j^x (j^y)^2 + j^y j^x j^y + (j^y)^2 j^x$. Adapted from Ref.  and Ref. .[]{data-label="tab:Tab2"} The components of these tensor can be decomposed into irreducible representations of the cubic group (characterizing the symmetry of the ideal double perovskite structure). This decomposition is described fully in Table \[tab:Tab2\]. Here we note in particular the two-dimensional $\Gamma_3$ representation $$\begin{aligned} \label{eq:6} Q_{i}^{3z^2} & = & \frac{1}{\sqrt{3}}\left\langle 3(j_i^z)^2 - j(j+1)\right\rangle , \nonumber \\ Q_i^{x^2-y^2} & = & \left\langle (j_i^x)^2 - (j_i^y)^2\right\rangle,\end{aligned}$$ which are analogous to the $e_g$ orbitals in atomic physics. The remaining three independent components of $Q_i^{\mu\nu}$ ($j_i^x j_i^y + j_i^y j_i^x$ etc.) form a three-dimensional representation analogous to the $t_{2g}$ orbitals, but do not appear in our analysis. Another important way to break up the tensor order parameters is into combinations which appear in the spin Hamiltonian. Specifically, these are the orbital occupation operators, $\tilde{n}_{i,yz},\tilde{n}_{i,xz},\tilde{n}_{i,xy}$, and the orbitally-resolved spin operators, $\tilde{S}^\mu_{i,yz},\tilde{S}^\mu_{i,xz},\tilde{S}^\mu_{i,xy}$. These can be expressed in terms of the multipoles describe above. For the occupation numbers, $$\begin{aligned} \label{eq:7} \tilde{n}_{i,yz} & = & \frac{1}{3} + \frac{1}{6\sqrt{3}}Q_i^{3z^2} - \frac{1}{6} Q_i^{x^2-y^2}, \nonumber \\ \tilde{n}_{i,xz} & = & \frac{1}{3} + \frac{1}{6\sqrt{3}}Q_i^{3z^2} + \frac{1}{6} Q_i^{x^2-y^2}, \nonumber \\ \tilde{n}_{i,xy} & = & \frac{1}{3} - \frac{1}{3\sqrt{3}}Q_i^{3z^2} .\end{aligned}$$ The orbitally-resolved spins decompose as $$\begin{aligned} \label{eq:8} \tilde{S}_{i,yz}^x & = & \frac{1}{15} j^x_i - \frac{2}{15} T_{i,x}^\alpha \nonumber \\ \tilde{S}_{i,yz}^y & = & \frac{2}{15} j^y_i + \frac{1}{15} T_{i,\alpha}^y + \frac{1}{3\sqrt{15}} T_{i,\beta}^y, \nonumber \\ \tilde{S}_{i,yz}^z & = & \frac{2}{15} j^z_i + \frac{1}{15} T_{i,\alpha}^z - \frac{1}{3\sqrt{15}} T_{i,\beta}^z, \nonumber \\ \tilde{S}_{i,xz}^x & = & \frac{2}{15} j^x_i + \frac{1}{15} T_{i,\alpha}^x - \frac{1}{3\sqrt{15}} T_{i,\beta}^x, \nonumber \\ \tilde{S}_{i,xz}^y & = & \frac{1}{15} j^y_i - \frac{2}{15} T_{i,\alpha}^y \nonumber \\ \tilde{S}_{i,xz}^z & = & \frac{2}{15} j^z_i + \frac{1}{15} T_{i,\alpha}^z + \frac{1}{3\sqrt{15}} T_{i,\beta}^z, \nonumber \\ \tilde{S}_{i,xy}^x & = & \frac{2}{15} j^x_i + \frac{1}{15} T_{i,\alpha}^x + \frac{1}{3\sqrt{15}} T_{i,\beta}^x, \nonumber \\ \tilde{S}_{i,xy}^y & = & \frac{2}{15} j^y_i + \frac{1}{15} T_{i,\alpha}^y - \frac{1}{3\sqrt{15}} T_{i,\beta}^y, \nonumber \\ \tilde{S}_{i,xy}^z & = & \frac{1}{15} j^z_i - \frac{2}{15} T_{i,\alpha}^z \end{aligned}$$ Cubic system: phases {#sec:cubic-system:-phases} -------------------- We first discuss the phases occurring in the cubic system at $T>0$. The ground states discussed earlier are all stable to small thermal fluctuations, and hence persist at low temperature. Thus we expect, broadly speaking, an antiferromagnetic (AFM) and ferromagnetic (FM110/FM100) region at low temperature. Of course, at temperatures much larger than $J$, one has a disordered paramagnetic phase. Interestingly, an additional phase appears at intermediate temperature. This is a non-magnetic [*quadrupolar ordered*]{} phase. To see how this arises, we describe the mean-field procedure and its results. Mean field theory is formulated in the usual way. We self-consistently decouple interactions between different sites $i$ and $j$ as follows: $$\begin{aligned} \hat{{\mathcal}O}_i \cdot \hat{{\mathcal}O}_j & \Rightarrow & \hat{{\mathcal}O}_i \cdot \langle \hat{{\mathcal}O}_j \rangle + \langle \hat{{\mathcal}O}_i \rangle\cdot \hat{{\mathcal}O}_j \\ & & - \langle\hat{{\mathcal}O}_i \rangle \cdot \langle \hat{{\mathcal}O}_j \rangle.\nonumber \;,\end{aligned}$$ where $\hat{{\mathcal}O}_i$ and $\hat{{\mathcal}O}_j$ are two operators at site $i$ and $j$, respectively. These operators are nothing but the orbital occupation numbers and orbitally resolved spins, which are related to the multipolar operators by Eqs. - . Decoupling [*all*]{} pairwise interactions between sites in this way, we then obtain a set of single-site problems for each $j=3/2$. Note that these single-site problems involve not just the usual Weiss exchange field, but also “multipolar fields”, which act as effective second and third order spin anisotropies. The mean-field equations determine self-consistent values of the orbital occupation numbers and orbitally resolved fields. As it is straightforward to formulate the mean-field equations, and solve them numerically, we do not give the details of these calculations here. A distinct class of solutions describes each phase. For the antiferromagnetic phase, we find the following operators are non-zero: $$\begin{aligned} \label{eq:9} \langle {\bm j}_i \rangle & = & \pm n (u_1,u_2,0), \\ \langle Q_i^{3z^2}\rangle & = & q, \\ \langle {\bm T}_{i,\alpha} \rangle &= & \pm t_\alpha (u_1,u_2,0), \\ \langle {\bm T}_{i,\beta} \rangle &= & \pm t_\beta (-u_1,u_2,0),\end{aligned}$$ where we have taken ${{\boldsymbol}Q}=2\pi(0,0,1)$ for concreteness, and the upper and lower signs refer to the A and B sublattices, respectively. The parameters $n,q,t_\alpha,$ and $t_\beta$ are positive at all $T>0$ in the AFM phase. However, note that $n$ vanishes in the limit $T\rightarrow 0$, in agreement with the vanishing dipole moment discussed earlier for the AFM ground state. In the FM110 state, the non-zero expectation values are: $$\begin{aligned} \label{eq:10} \langle {\bm j}_i \rangle & = & m(1,1,0) \pm n (1,-1,0), \\ \langle Q_i^{3z^2} \rangle & = & q, \\ \langle Q_i^{x^2-y^2}\rangle & = & \mp q', \\ \langle {\bm T}_{i,\alpha}\rangle & = & t_\alpha(1,1,0) \pm \tilde{t}_\alpha(1,-1,0), \\ \langle {\bm T}_{i,\beta}\rangle & = & t_\beta(1,-1,0) \pm \tilde{t}_\beta(1,1,0), \end{aligned}$$ where again we took ${{\boldsymbol}Q}=2\pi(0,0,1)$ and the upper/lower signs refer to the A/B sublattices. In this case the parameters $m,n,q,q',t_\alpha,\tilde{t}_\alpha,t_\beta,\tilde{t}_\beta$ are all non-zero at temperatures within the FM110 phase including $T=0$. The third ordered phase dominating the phase diagram is the quadrupolar one (For the purposes of this section, we ignore the FM100 phase, which extends into a narrow region of ferromagnetic state with variable polarization direction at $T>0$, as it occupies a very small volume of the phase diagram). In the quadrupolar state, there is only a single non-vanishing order parameter: $$\label{eq:11} \langle Q_i^{x^2-y^2}\rangle = \mp q' .$$ Let us discuss the symmetries of these three states. In the AFM and FM110 phases, time reversal symmetry is broken. However, the net magnetization vanishes in the AFM state. In the AFM state, this is guaranteed by invariance under the combined operations of translation (such as by $(0,1/2,1/2)$, which interchanges the A and B sublattices) and time-reversal. No such symmetry can be combined with time-reversal in the FM110 case. Various point group symmetries are also present in the AFM and FM110 phases, but we do not describe this in detail. In the quadrupolar case, time-reversal symmetry is unbroken, which is sufficient to require the dipolar and octupolar order parameters to vanish. Only point group symmetries are broken by the quadrupolar order. Four-fold ($C_4$) rotations about the $x$ or $y$ axes, and three-fold ($C_3$) rotations about $[ 111]$ axes are broken in this state. While the $C_4$ rotation about the $z$ axis is also broken, the combination of this $C_4$ rotation and a translation exchanging the A and B sublattices remains a symmetry of the quadrupolar state. A standard classification scheme for quadrupolar states is to examine the examine the eigenvalues of the $Q_i^{\mu\nu}$ matrix. These must sum to zero because the matrix is traceless. States in which there are only two distinct eigenvalues, i.e. ${\rm eigs}(Q) = \{ q,q,-2q\}$ are called [*nematics*]{}, and correspond to the situation in which one principal axis is distinguished from the other two, which remain identical. In the most general case, there are three distinct eigenvalues, i.e. ${\rm eigs}(Q) = \{ q_1,q_2,-q_1-q_2\}$, with $q_1\neq q_2$. This is called a [ *biaxial nematic*]{}, and is a state in which all three principal axes are distinct. The quadrupolar state obtained here is such a biaxial nematic. Physically, the local susceptibility in this state takes distinct values $\chi_{\rm local}^{xx}, \chi_{\rm local}^{yy},\chi_{\rm local}^{zz}$ for fields along each of the axes. However, note from Eq.  that the quadrupolar order parameter changes sign between the two sublattices. Thus we should properly call this state an [*antiferro-biaxial nematic*]{}. Due to the staggered ordering, the bulk susceptibility does not distinguish all three axes. Instead, there are only two distinct components, $\chi^{xx}=\chi^{yy} \neq \chi^{zz}$. The difference between the two components of the susceptibility serves as a simple macroscopic means to observe quadrupolar ordering. Cubic system: phase diagram and transitions {#sec:cubic-system:-phase} ------------------------------------------- By solving the mean-field equations numerically, we have determined the phase diagram for the cubic case. Parts of it can be understood analytically. Suppose that the transitions from the high temperature normal phase to the quadrupolar and AFM phases are second order. This appears to be always true for the quadrupolar phase, while it true for the AFM for most parameters, but weakly violated in some regions. With this assumption, we can determine the critical temperatures for these transitions by the usual condition of marginal stability (vanishing of the quadratic term in the Landau theory) of the free energy. We find that the critical temperature for the quadrupolar state is $$\label{eq:12} T_c^{\rm quad} = \frac{43V+18J'- 3J}{18},$$ and that for the AFM state is $$\label{eq:13} T_c^{\rm AFM} = \frac{J + 10J' + \sqrt{73J^2 + 164 J J' + 100 (J')^2}}{36}.$$ Without the assumption that the transitions are continuous, the critical temperature could be higher. Thus Eqs. (\[eq:12\],\[eq:13\]) give lower bounds for the transition temperatures, strictly speaking. Extending the two-dimensional $T=0$ phase diagram in Fig. \[fig:phase\_diagram\] into a third dimension of temperature, the quadrupolar phase occurs “above” the portion for which $T_c^{\rm quad}> T_c^{\rm AFM}$, which implies $V>V_c$, with $$\label{eq:14} V_c = \frac{7J - 26J' + \sqrt{73J^2 + 164 J J' + 100 (J')^2}}{86}.$$ The curve $V_c(J')$ defines an almost straight line in the 2d phase diagram, as shown in Fig. \[fig:thermal\_phase\]. In fact, Eq.  slightly underestimates $V_c$, as it assumes the normal to AFM transition is continuous, when it is in fact weakly first order in this vicinity. However, the true $V_c$ found numerically is only a few percent larger. For $V<V_c$, no quadrupolar phase occurs. Instead, the first ordering transition from high temperature is into the AFM state. This is true even when the ground state is ferromagnetic, so that in this case (when $V<V_c$) the system first orders into the AFM state, and then at lower temperature switches to the FM110 phase. ![(Color online) Zero temperature two-dimensional phase diagram in the cubic case (same as Fig. \[fig:phase\_diagram\]), overlaid with the regions of different $T>0$ behavior. Dashed lines are obtained from mean field numerics. The solid line (in orange) is defined by Eq. . In region I, there is a single transition to the AFM state. In region II, the system supports an intermediate temperature quadrupolar ordered phase. The transition from the normal state to the quadrupolar phase is second order. In this region, there is a first order transition to the AFM phase on further cooling, for parameters such that the latter is the ground state. Otherwise, the lower temperature transition is to a ferromagnetic (predominantly FM110) state. In region III, the system first turns from normal phase to the AFM state then to a ferromagnetic (predominantly FM110) state on further cooling.[]{data-label="fig:thermal_phase"}](fig9.pdf){width="8cm"} We now discuss the nature of the transitions. The quadrupolar ordering transition is, as already mentioned, continuous (see Figs. \[fig:ops1030\], \[fig:ops3030\]). It is described by a single scalar order parameter, (equal to $q'$ in Eq. ), for each of the three $[100]$ wavevectors, describing the associated staggered quadrupole moment. Formally, $$\label{eq:15} \phi_a = (-1)^{2x_{i}^a} \langle Q^{x^2-y^2}_i \rangle,$$ where $x_i^1=x_i, x_i^2=y_i,x_i^3=z_i$. According to symmetry, the Landau free energy for $\phi_a$ has the same form as that for an O(3) magnetic transition with cubic anisotropy. Beyond mean field theory, this is believed to support a three dimensional transition in the O(3) universality class. ![(Color online) Temperature dependence of order parameters for $J'/J=0.2$, $V/J=0.1$. For these parameters, there is a direct, continuous, normal to AFM transition, at $T/J \approx 0.38$. The three curves show: squares (red online) $|\langle {\bm T}_{A,\alpha}-{\bm T}_{B,\alpha}\rangle|/2$, circles (blue online) $\langle Q_A^{3z^2}+Q_B^{3z^2}\rangle /2$, and diamonds (yellow online) $|\langle {\bm j}_A - {\bm j}_B\rangle |/2$. Note: in Figs.7-10, the symbols are [*not*]{} the data points (which are much more dense) – they simply label the different curves.[]{data-label="fig:ops2010"}](fig10.pdf){width="8cm"} ![(Color online) Temperature dependence of order parameters for $J'/J=0.1$, $V/J=0.3$. For these parameters, there is a continuous normal to quadrupolar transition, at $T/J \approx 0.65$, followed by a first order transition to the AFM state at $T/J\approx 0.37$. The four order parameters plotted are: squares (red online) $|\langle {\bm T}_{A,\alpha}-{\bm T}_{B,\alpha}\rangle|/2$, circles (blue online) $\langle Q_A^{3z^2}+Q_B^{3z^2}\rangle /2$, diamonds (yellow online) $|\langle {\bm j}_A - {\bm j}_B\rangle |/2$, and triangles (green online) $\langle Q_A^{x^2-y^2}-Q_B^{x^2-y^2}\rangle/2$.[]{data-label="fig:ops1030"}](fig11.pdf){width="8cm"} ![(Color online) Temperature dependence of order parameters for $J'/J=0.3$, $V/J=0.3$. For these parameters, there is a continuous normal to quadrupolar transition, at $T/J \approx 0.85$, followed by a continuous quadrupolar to FM110 transition at $T/J\approx 0.33$. The four order parameters plotted are: squares (red online) $|\langle {\bm T}_{A,\alpha}-{\bm T}_{B,\alpha}\rangle|/2$, circles (blue online) $\langle Q_A^{3z^2}+Q_B^{3z^2}\rangle /2$, diamonds (yellow online) $|\langle {\bm j}_A + {\bm j}_B\rangle |/2$, and triangles (green online) $\langle Q_A^{x^2-y^2}-Q_B^{x^2-y^2}\rangle/2$.[]{data-label="fig:ops3030"}](fig12.pdf){width="8cm"} ![(Color online) Temperature dependence of order parameters for $J'/J=0.40$, $V/J=0.05$. For these parameters, there is a continuous normal to AFM transition, at $T/J \approx 0.49$, followed by a continuous AFM to FM110 transition at $T/J\approx 0.34$. The four order parameters plotted are: squares (red online) $|\langle {\bm T}_{A,\alpha}-{\bm T}_{B,\alpha}\rangle|/2$, circles (blue online) $\langle Q_A^{3z^2}+Q_B^{3z^2}\rangle /2$, diamonds (yellow online) $|\langle {\bm j}_A + {\bm j}_B\rangle |/2$, and triangles (green online) $\langle Q_A^{x^2-y^2}-Q_B^{x^2-y^2}\rangle/2$.[]{data-label="fig:ops4005"}](fig13.pdf){width="8cm"} The transition from the normal to the AFM state is continuous in mean field theory for small $V$ (see Fig. \[fig:ops2010\]), becoming weakly first order for larger $V$, close to $V_c$ where the intermediate quadrupolar phase emerges. The normal-AFM transition is characterized, for each of the three wavevectors, by a two-component primary order parameter, which could be taken to be the two components of ${\bm T}_{A,\alpha}$ normal to ${\bm Q}$. In principle, the degeneracy of the ordering pattern within this “XY” plane normal to ${\bm Q}$ is, as we have remarked, accidental, and should be removed by additional effects. We do not, however, observe this degeneracy lifting within mean field theory for the present model. With the degeneracy, the transition should be therefore described by the free energy for some six component order parameter. As we do not understand the degeneracy lifting mechanism at present, we do not attempt here to establish the true critical properties for this transition (when it is continuous) with fluctuations taken into account. The quadrupolar to FM110 transition is continuous in mean field theory (see Fig. \[fig:ops3030\]). This could be anticipated by examining the form of the order parameters in the FM110 phase. We note that the antiferro-biaxial nematic order parameter of the quadrupolar state is already non-vanishing in the FM110 phase. Hence, we might naturally expect, upon heating, that thermal fluctuations first restore time-reversal symmetry, yielding the quadrupolar phase, before fully restoring all symmetry in the normal state. To determine the nature of the order parameter for this transition, note that the wavevector and local anisotropy axes are already established in the quadrupolar state. Hence the direction of the uniform and staggered magnetizations are already determined, up to a sign and interchange, above the transition. For instance, for the quadrupolar state in Eq. , with ${\bm Q}=2\pi(001)$, the uniform magnetization can lie along $\pm (110)$ and the staggered magnetization along $\pm(1\overline{1}0)$, or vice-versa. Thus the symmetry breaking from the quadrupolar to the FM110 state is described by two Ising order parameters. We therefore expect this transition, beyond mean field theory, to be similar to that of an Ashkin-Teller or similar models. The quadrupolar to AFM transition appears strongly first order (see Fig. \[fig:ops1030\]). This is in agreement with the expectations of Landau theory, as the symmetry of the AFM phase is not a subgroup of the symmetry of the quadrupolar one. In terms of order parameters, this is evident since $\langle Q^{x^2-y^2}_i \rangle$ is non-zero in the quadrupolar phase but zero in the AFM one, while the magnetic order parameters are zero in the quadrupolar phase but non-zero in the AFM. Fine tuning of the free energy would be required to arrange both these types of order to change at the same temperature in a continuous fashion. In region III one encounters a transition from the AFM to FM110 state. This appears to be continuous in mean field theory (see Fig. \[fig:ops4005\]). One can understand this by noting that the AFM solution can be regarded as a subset of the FM110 one, if the unit vector $(u_1,u_2,0)$ is taken to be along $(1,-1,0)$. Then the transition to the FM110 is described by the emergence of a non-zero $m$. Like the normal to AFM transition, because we have not understood the degeneracy-breaking mechanism in the AFM state, we do not attempt to analyze this transition beyond MFT. Effects of anisotropy {#sec:effects-anisotropy} --------------------- We now consider the effects of anisotropy on the $T>0$ phase diagram, focusing on the case of weak $|D| \ll J,J',V$. We have already considered the effects of $D$ on the AFM and FM110 states in Sec. \[sec:transitions\]. We saw that easy-axis anisotropy favors states with the wavevector ${{\boldsymbol}Q}$ parallel to the $z$ (easy) axis. This is because the anisotropy couples directly to the $Q_i^{3z^2}$ field: $$\label{eq:16} H_D = - D\sum_i (j_i^z)^2 = {\rm const.} - \frac{D}{\sqrt{3}} \sum_i Q_i^{3z^2}.$$ Both the AFM and FM110 states have a non-zero and constant expectation value of $Q_i^{3z^2}$, which is maximized in this orientation. Conversely, easy-plane anisotropy favors states with the wavevector ${{\boldsymbol}Q}$ perpendicular to the $z$ (hard) axis, for the same reason. ![Temperature dependence of the free energy difference between states with wavevector parallel and perpendicular to the Ising axis, in the presence of a weak anisotropy $|D|=0.05$. Here $J'/J=V/J=0.3$. Solid line: $\Delta F = F({\bm Q}\parallel {\bm\hat z};D=.05) - F({\bm Q}\perp {\bm\hat z};D=.05)$. Dotted line: $\Delta F = F({\bm Q}\parallel {\bm\hat z};D=-.05) - F({\bm Q}\perp {\bm\hat z};D=-.05)$. One sees that in the quadrupolar phase, both signs of anisotropy favor the wavevector aligned with the z axis. In the FM110 phase, however, this is favored only for $D>0$. For $D<0$ (easy plane anisotropy), the state with wavevector normal to z is preferred. Note also that the energy difference is much larger in the FM110, consistent with the expected linear and quadratic dependence on $D$ in FM110 and quadrupolar phases, respectively. []{data-label="fig:DeltaF"}](fig14.pdf){width="8cm"} We now repeat this analysis for the quadrupolar state. Here the situation is more subtle because $\langle Q_i^{3z^2}\rangle$ vanishes in the quadrupolar state. Moreover, the cubic rotations (e.g. $\langle Q_i^{3x^2}\rangle$), while not vanishing, give zero net contribution due to the opposite signs on the A and B sublattices. This means that the splitting of the different wavevector states vanishes at linear order in $D$. There is instead a quadratic contribution, which, numerically, we find favors the states with ${\bm Q}$ parallel to $z$ (see Fig. \[fig:DeltaF\]). Being quadratic in $D$, this same configuration is favored [*for both the easy-axis and easy-plane case*]{}. Thus we have the interesting situation that for easy-plane anisotropy, the wavevector orients parallel to $z$ in the quadrupolar phase, but perpendicular to $z$ in the low temperature phase. Note that the quadrupolar phase remains distinct from the normal phase even with non-zero $D$, as it continues to break symmetries, notably translational invariance. Magnetic susceptibility {#sec:field} ----------------------- In this subsection, we discuss the magnetic response at $T>0$, which is an important indicator, especially of the quadrupolar ordering transition. At high temperature, of course, one observes Curie-Weiss behavior. For the general Hamiltonian with anisotropy $D$, there are two different Curie-Weiss temperatures, for fields parallel and perpendicular to $z$: $$\begin{aligned} \label{eq:17} \Theta_{CW}^{zz} & = & -\frac{J}{5} + \frac{32 J'}{45} + \frac{4D}{5}, \nonumber \\ \Theta_{CW}^{xx} & = & -\frac{J}{5} + \frac{32 J'}{45} - \frac{2D}{5}.\end{aligned}$$ These are obtained from the high temperature expansion of the susceptibility up to $O(1/T^2)$. These expressions may be useful in extracting exchange constants from experiment. Interestingly, if one calculates the powder average average susceptibility, the contributions of the anisotropy cancel at this order and the Curie-Weiss temperature measured in this way is independent of $D$. It is also interesting to note that, in the region of larger $V/J$ and small $J'/J$, one obtains a ferromagnetic ground state with an antiferromagnetic (negative) Curie-Weiss temperature. On lowering temperature, the susceptibility shows distinct behaviors in the different parts of the phase diagram. We focus here for simplicity on the cubic system, starting with region I. Here the susceptibility displays the usual cusp associated with antiferromagnetic order, at the normal to AFM transition. The inverse susceptibility is plotted in Fig. \[fig:chiinv0101\] for $J'=V=0.1J$, in the midst of region I. It shows a minimum at the transition, and pronounced curvature below the transition temperature, saturating to a large constant value in the $T\rightarrow 0$ limit. We note that the large zero temperature susceptibility is not related to gapless excitations, but is a general consequence of strong SOC, and should be expected in all parameter regimes of this model. ![(Color online) Inverse susceptibility at the normal to AFM transition for $J'=V=0.1J$. Blue (upper) curve: $1/\chi_{xx}$, red (lower) curve: $1/\chi_{zz}$. []{data-label="fig:chiinv0101"}](fig15.pdf){width="7.5cm"} ![(Color online) Inverse susceptibility for for $J'=0.2J$, $V=0.3J$. Blue (lower) curve: $1/\chi_{xx}$, red (upper) curve: $1/\chi_{zz}$, yellow (middle) curve: $1/\chi_{\rm powder}$. For these parameters the quadrupolar transition is at $T/J \approx 0.75$, and the ferromagnetic transition is at $T/J \approx 0.23$. []{data-label="fig:chiinv0203"}](fig16.pdf){width="7.5cm"} Next consider region II. Here, one observes a cusp at the normal to quadrupolar transition. This cusp is, however, rather different from the one just mentioned. Specifically, it is [*not*]{} a minimum of $1/\chi$, and instead separates two distinct “Curie-Weiss” regimes in which $1/\chi$ is linear but with different, positive, slopes (i.e. different effective magnetic moments). The presence of a lower temperature Curie-Weiss regime is a signature of quadrupolar order. This is because the quadrupolar mean field splits only the point group degeneracy of the spins, but preserves a local Kramer’s doublet. This doublet gives rise to a Curie law. An example is plotted in Fig. \[fig:chiinv0203\]. As the quadrupolar order lowers the symmetry of the system to tetragonal, we see actually two different effective moments in susceptibility parallel to the wavevector ${\bm Q}$ ($\chi_{zz}$) and perpendicular to it ($\chi_{xx}=\chi_{yy}$). We observe that the effective magnetic moment seen in $\chi_{xx}$ is typically [ *enhanced*]{} in the quadrupolar phase, while it is suppressed in $\chi_{zz}$, both relative to the isotropic effective magnetic moment in the normal phase. At still lower temperature, one encounters the ferromagnetic phases. Here of course the susceptibility for the easy directions diverges. Focusing on the dominant FM110 phase, one sees that since the easy direction is in the (001) plane selected by the quadrupolar order, $\chi_{zz}$ does not diverge, but $\chi_{xx}$ and $\chi_{yy}$ do. Beyond mean-field: spin waves and non-magnetic ground states {#sec:sec5} ============================================================ In Sec. \[sec:sec33\], we obtained the mean field phase diagram. Here we consider quantum effects beyond mean field. We first consider spin wave fluctuations, and obtain the collective mode spectrum in linear spin wave theory. From this, we obtain the quantum correction to the order parameter, and, in the ideal case of $J'=V=0$, we will see that this is very large and invalidates the mean field theory in the vicinity of this parameter regime. This suggests the possibility of very different states dominated by quantum fluctuations. We then explore this possibility, considering some candidate non-magnetic ground states of our model. First, we consider the quantum ground states of pairs of sites, unveiling a pseudo-singlet structure, analogous to the $S=0$ singlet ground states for pairs of antiferromagnetically interacting spins with SU(2) symmetry. This leads naturally to the possibility of “valence bond” states built from these pseudo-singlets. We consider both a static, Valence Bond Solid (VBS) state, and states in which the valence bonds are fluctuating, in which case we obtain a Quantum Spin Liquid (QSL) state. Very little theoretical work has been done on QSL states in systems with strong spin orbit coupling, i.e. with strongly broken SU(2) symmetry. As such, the structure of possible QSL states in the present model requires particular investigation. Guided by the pseudo-singlet structure, and the hidden SU(2) symmetry of the model, we construct candidate QSL states for the full Hamiltonian, $\tilde{{\mathcal}H}_{\text{ex-1}}$, by a slave-particle technique. Spin waves {#sec:sec33} ---------- In the previous two sections, we have discussed the state phase diagram based on mean-field theory. Here, we perform a linear spin wave analysis, which perturbatively describes the effect of quantum fluctuations on the various phases obtained so far, and also predicts the structure of collective modes, which might, e.g., be observed in inelastic neutron scattering. Finally, because we have not explored the full space of mean-field states, the calculation also provides an important check that the phases we have found are at least metastable. The conventional Holstein-Primakoff (HP) transformation for spin-S operators cannot be directly applied for the three variational ground states because none of the three states is a fully polarized state for any projection of the spin angular momentum operator ${{\boldsymbol}j} $. This is especially severe for the AFM state, for which the spin expectation value simply vanishes. Instead, we formulate an “SU(4) spin wave theory”, by rewriting the Hamiltonian, , in a bilinear form in terms of the 15 generators of the SU(4) group. To do so, we introduce, for any local basis for the single-site Hilbert space $\{ |n\rangle\},n=1,2,3,4$, the complete set of operators[@PhysRevLett.81.3527] $$\label{eq:2} {\mathcal}S^n_m = |m\rangle \langle n|.$$ These SU(4) generators obey the algebra $[{\mathcal}S^n_m,{\mathcal}S^l_k]=\delta_{nk}{\mathcal}S^l_m-\delta_{ml}{\mathcal}S^n_k$. We can then use the HP transformation for the generators of SU(4). In this transformation, one selects a particular state in the four-dimensional basis to be the vacuum and introduces three bosons associated with excitations to the three other states. For the AFM phase, the classical ground state is given by Eq. . Hence we take, on the A sublattice ($z_i$ integer), the basis $$\begin{aligned} |1\rangle_A &=&\frac{1}{\sqrt2}(|3/2\rangle+e^{i\phi}|-3/2\rangle),\quad|2\rangle_A = |1/2\rangle,\nonumber\\ |3\rangle_A &=&|-1/2\rangle,\quad|4\rangle_A=\frac{1}{\sqrt2}(|3/2\rangle-e^{i\phi}|-3/2\rangle),\end{aligned}$$ while for sublattice B ($z_i$ half integer), $$\begin{aligned} |1\rangle_B &=&\frac{1}{\sqrt2}(|3/2\rangle-e^{i\phi}|-3/2\rangle),\quad|2\rangle_B = |-1/2\rangle,\nonumber\\ |3\rangle_B &=&|1/2\rangle, \quad|4\rangle_B = \frac{1}{\sqrt2}(|3/2\rangle+e^{i\phi}|-3/2\rangle).\end{aligned}$$ The Hamiltonian $\tilde{{\mathcal}{H}}$ in Eq.  in this basis has a quadratic form, $$\tilde{{\mathcal}{H}}=\sum_{\langle ij\rangle} C_{klmn}(i,j)\,{{\mathcal}S}^l_k(i)\,{{\mathcal}S}^n_m(j),\label{HSS}$$ where the coefficients $C_{klmn}(i,j)$ (which are straightforward to obtain, so we do not give them explicitly) depend linearly on $J$, $J'$ and $V$. To introduce the HP transformation on sublattice A (B), we choose $|1\rangle$ as the vacuum which is annihilated by three “magnon” annihilation operators $a_n$ ($b_n$), $n=2,3,4$. One can think these three bosons as descending from mixed spin and orbitals fluctuations of the Hamiltonian before the ${{\mathcal}P}_{\frac{3}{2}}$ projection. The HP transformation is defined as[@PhysRevB.60.6584], for $i$ in the A sublattice, $$\begin{aligned} {{\mathcal}S}^1_1(i)&=&M-\sum_{n\neq1} a^\dagger_n(i)a^{\phantom\dagger}_n(i), \\ {{\mathcal}S}^1_n(i)&=&a^\dagger_n(i)\sqrt{ M-\sum_{l\neq1} a^\dagger_l(i) a^{\phantom\dagger}_l(i)},\qquad (n\neq1)\\ {{\mathcal}S}^l_n(i)&=&a^\dagger_n(i)a^{\phantom\dagger}_l(i). \qquad\qquad\qquad\qquad \,\,\,(l,n\neq 1) , \label{HPSU4}\end{aligned}$$ while for $i$ in the B sublattice, the same formula holds with $a_n(i)$ replaced by $b_n(i)$. In the above equations $M$ is defined as the number of columns in the Young tableaux for the representation of SU(4). In our case (fundamental representation), we must set $M=1$. In the generalization to arbitrary $M$, the classical limit where the classical ground state becomes exact is $M\to\infty$. However, we apply this HP transformation directly for $M=1$. Inserting this into , we expand it to obtain a quadratic form in the bosonic operators. The constant term in the expansion gives the classical ground state energy, $$\frac{E_{\text{AFM}}}{{\mathcal}{N}}=M^2\left(-\frac{J}{2}+J'+\frac{11V}{12}\right),$$ independent of the phase $\phi$. The quadratic terms lead to quantum corrections. Defining the Fourier transform of the bosonic operators, the spin-wave Hamiltonian can be organized in the form $\sum_\mathbf{k}H_\mathbf{k}$ with $$H_\mathbf{k}=(\;{\mathcal}{A}^\dagger_\mathbf{k}\quad{\mathcal}{A}^{\phantom\dagger}_\mathbf{-k}\;)\left(\begin{array}{cc}F_\mathbf{k}&G^\dagger_\mathbf{k}\\G_\mathbf{k}&F_\mathbf{k}\end{array}\right)\left(\begin{array}{c}{\mathcal}{A}^{\phantom\dagger}_\mathbf{k}\\ {\mathcal}{A}^{\dagger}_\mathbf{-k}\end{array}\right)$$ where ${\mathcal}{A}_{\mathbf{k}}=(a_{2\mathbf{k}},a_{3\mathbf{k}},a_{4\mathbf{k}},b_{2\mathbf{k}},b_{3\mathbf{k}},b_{4\mathbf{k}})$ is the vector of magnon annihilation operators and $F_\mathbf{k}$ and $G_\mathbf{k}$ are $6\times6$ matrices. This spin-wave Hamiltonian can be diagonalized by standard methods.[@PhysRevB.76.064418] For the AFM ground state, Eq. , we obtain a gapless “magnon” mode, as depicted in Fig. \[fig:spinwave\_AFM\]. This gapless mode is associated with the continuous accidental degeneracy, and indeed occurs for arbitrary $\phi$. By contrast, in the FM110 and FM100 phases, one observes a gap for all the spin wave modes (see Fig. \[fig:spinwave\_FM110\] and Fig. \[fig:spinwave\_FM100\]). The gap in the FM110 phase increases with $J'$, as expected since this corresponds to increasingly violated SU(2) symmetry. In all cases, the modes are all well-defined with positive real frequencies, indicating the stability of the phases in the classical sense: i.e. that we have properly found local energy minima of the mean field theory. Finally, having obtained the spin wave modes, we can evaluate the quantum corrections. It is most interesting to consider the reduction of the order parameter by quantum fluctuations. We can define this by considering the probability to find a given spin in its mean-field ground state. This is nothing but the vacuum state of the HP bosons. Hence this probability is given, for a site on the A sublattice, by $$\label{eq:21} P_{gs}(i) = \langle 1 - \sum_{n\neq 1} a_n^\dagger(i) a_n^{\vphantom\dagger}(i) \rangle.$$ This quantity is directly analogous to the staggered magnetization in the usual HP treatment of a quantum antiferromagnet. We therefore denote $\Delta M = 1-P_{gs}(i) = \sum_{n\neq 1} \langle a_n^\dagger(i) a_n^{\vphantom\dagger}(i)\rangle$. This is obtained, at $T=0$, by integrating the zero point contribution to the boson number from each spin wave mode. The quadratic spin-wave Hamiltonian is diagonalized by a Bogoliubov transformation $ {\boldsymbol}{{\mathcal}Q}_{\bf k}$, $$({{\mathcal}C}_{\bf k}, {{\mathcal}C}^{\dagger}_{- \bf k} )^{T}= {\boldsymbol}{{\mathcal}Q}_{\bf k} ({{\mathcal}A}_{\bf k}, {{\mathcal}A}^{\dagger}_{- \bf k} )^{T} \;,$$ in which, ${{\mathcal}C}_{\bf k} = (c_{1{\bf k}},c_{2{\bf k}},c_{3{\bf k}},c_{4{\bf k}},c_{5{\bf k}},c_{6{\bf k}})$, and $ {\boldsymbol}{{\mathcal}Q}_{\bf k}$ is a $12 \times 12$ matrix. From this we obtain the quantum correction $$\begin{aligned} \Delta M &=& \frac{1}{{\mathcal}N} \sum_{n \neq 1}[\sum_{i \in A} a_n^{\dagger} (i) a_n(i) + \sum_{i \in B} b_n^{\dagger} (i) b_n(i)] \nonumber \\ &=& \frac{1}{2}\{ \frac{1}{ {\mathcal}N }\sum_{\bf k} \sum_{i=1}^{6} [ {\boldsymbol}{{\mathcal}Q}^{\dagger} {\boldsymbol}{{\mathcal}Q} ]_{ii}-3 \}.\end{aligned}$$ Numerically, we find that this quantum correction is maximal for $J'=V=0$, and is given by $\Delta M \approx 1.7$ at this point. This is much larger than $1$, implying that the fluctuations at this point are large and that the mean field theory is at least quantitatively invalid. For increasing $J'$ and $V$ the correction becomes significantly smaller, and mean field theory may be reliable. In the vicinity of the $J'=V=0$, one may expect a very different ground state, incorporating strong quantum fluctuations. We explore some possible [ *non-magnetic*]{} ground states in the remainder of this section. ![Spin wave spectrum for the AFM phase at $J'=0.1$ and $V=0.2$ along \[001\] momentum direction. And $J=1$. There is one low-lying gapless mode. The fcc lattice constant is set to be $a=1$. And the phase $\phi =0$ for the ground state in Eq. . []{data-label="fig:spinwave_AFM"}](fig17.pdf){width="6.0cm"} ![Spin wave spectrum for the FM110 phase at $J'=0.3$ and $V=0.2$ along \[001\] momentum direction. The lowest excitation mode has an energy gap $\Delta = 0.241$ at $k=0$. In the graph, $J=1$. []{data-label="fig:spinwave_FM110"}](fig18.pdf){width="6.0cm"} ![Spin wave spectrum for the FM100 phase at $J'=0.1$ and $V=0.4$ along \[001\] momentum direction. The lowest excitation mode has an energy gap $\Delta = 0.0224$ at $k=0$. In the graph, $J=1$. []{data-label="fig:spinwave_FM100"}](fig19.pdf){width="6.0cm"} Pseudo-singlets in different planes ----------------------------------- We start our analysis of non-magnetic states by considering two sites in the XY plane, which interact with the Hamiltonian $\tilde{{\mathcal}H}_{\text{ex-1}}^{\text{XY}}(i,j)$. Remarkably, the ground state has a form identical to an SU(2) spin singlet, if written in terms of pseudospin-$1/2$ states $j^z = \pm 1/2$: $$\label{eq:singletxy} |\text{XY}\rangle_{ij} = \frac{1}{\sqrt{2}} \left(|\tfrac{1}{2} \rangle_i |- \tfrac{1}{2} \rangle_j - |-\tfrac{1}{2} \rangle_i |\tfrac{1}{2} \rangle_j\right) \;.$$ One may understand this result by writing down the projected spin and occupation number operators in $xy$ orbitals, in the basis of $j^z$ eigenstates (see Eq. \[eq:projxy\]): $$\begin{aligned} \tilde{{\boldsymbol}S}_{xy}& = & \frac{1}{3} \left( \begin{array}{c|c|c} 0 & & \\ \hline & \bm{\sigma} & \\ \hline & & 0 \end{array} \right),\label{Sxymatrix} \\ \tilde{n}_{xy} &=& \frac{2}{3} \left( \begin{array}{c|c|c} 0 & & \\ \hline & I_2 & \\ \hline & & 0 \end{array} \right)\label{Ixymatrix} \;.\end{aligned}$$ Here $\bm{\sigma}$ is the vector of Pauli matrices and $I_2$ is the $2 \times 2$ identity matrix. One may consider $ \tilde{{\boldsymbol}S}_{xy}$ as an effective spin-$1/2$ operator in the subspace of $j^z = \pm 1/2$ states, which naturally explains the SU$(2)$ singlet in Eq. . For the XZ and YZ planes, one simply needs to apply a cubic permutation to the results obtained for XY planes, or more formally, apply a unitary transformation that rotates about the $[111 ]$ axis by $\pm 2\pi/3$, $$\begin{aligned} \tilde{S}^{\mu'}_{xz} & = & U^\dagger \tilde{S}^\mu_{xy} U, \label{rotatexz}\\ \tilde{S}^{\mu''}_{yz} & = & U \tilde{ S}^\mu_{xy} U^\dagger \label{rotateyz} \;,\end{aligned}$$ with $$U=\exp\left(-i\frac{2\pi}{3}\frac{j^x+j^y+j^z}{\sqrt{3}}\right).\label{C3rotation}$$ The upper indices $\mu'=p(\mu)$ and $\mu''=p^{-1}(\mu)$ in Eqs.  and denote cyclic and anti-cyclic permutations of $x,y,z$, respectively \[i.e. $p:(x,y,z)\to (y,z,x)$, with inverse $p^{-1}:(x,y,z)\to (z,x,y)$\]. The two-site ground states in the XZ and YZ planes are the pseudo-singlets in the subspace of $j^y = \pm 1/2$ states and $j^x = \pm 1/2$ states, respectively. Valence bond solid state {#sec:valence-bond-solid} ------------------------ It is natural to consider a product state of such pseudo-singlet “valence bonds” (also called “dimers”) as a candidate (prototypical variational) non-magnetic ground state. To do so, we must divide the spins into two neighboring sublattices, which will be paired. This by necessity breaks lattice symmetries. Such a state is called a Valence Bond Solid, or VBS, state. At the level of valence bond product states, many possible arrangements of the dimers are degenerate. This degeneracy is artificial and will be broken if the wavefunctions are improved. We will not investigate this in any detail, and just consider the simplest VBS state in which the dimers form a “columnar” arrangement within a single (001) plane. See Fig. \[fig:columnar\]. ![(color online) Columnar Valence Bond Solid (VBS) state within an XY plane. The dashed square indicates the face of a conventional cubic unit cell, while the solid lines connect the FCC nearest neighbors within the plane, which form a 45$^\circ$ rotated square lattice.[]{data-label="fig:columnar"}](fig20.pdf){width="6.0cm"} The variational energy of such a state (actually any state with a planar arrangement of dimers has the same energy) is readily evaluated. We obtain the energy per site $E_{VBS} /{{\mathcal}N}= \langle VBS| \tilde{H}_{ex-1}|VBS\rangle/{{\mathcal}N} = - 5/12 J =-0.42J$. This is slightly higher than the mean-field ground state energy of the AFM state, $E_{AFM}^{MF}/{{\mathcal}N} = - J/2$. However, the large quantum fluctuations are expected to destabilize the latter state, and perhaps might stabilize the VBS one. So such a VBS state seems competitive, and may be considered as a possibility for future exploration. QSLs and Fermionic mean field theory ------------------------------------ The most general approach that has been applied to describe QSL states is the slave particle method, in which auxiliary fermions are introduced, and the ground state for the spin system is described by some projection of a nominally simple fermionic state into the physical spin Hilbert space. This results, in the usual SU(2)-invariant case, in wavefunctions which are composed of superpositions of products of SU(2) singlets. Here, the appearance of two-site pseudo-singlet ground states points to the possibility of applying a similar fermionic mean field theory. In this section, we implement this technique for the full antiferromagnetic exchange interaction. We first introduce the auxiliary fermionic creation operators, whose quanta we call “spinons”[@PhysRevLett.81.3527] $$|\alpha\rangle_i=f^\dagger_{i\alpha}|\textrm{vacuum}\rangle, \quad\alpha=1,...4,$$ where for convenience we have relabeled the states $j^z=\frac{3}{2},\frac{1}{2},-\frac{1}{2},-\frac{3}{2}$ by $\alpha=1,2,3,4$, respectively. The physical Hilbert space is constructed from states with one fermion at each site, which imposes the constraint $$\sum_{\alpha=1}^4f^\dagger_{i\alpha}f^{\phantom\dagger}_{i\alpha}=1. \label{constr}$$ In this notation, the spin and number operators become $$\begin{aligned} \tilde{\bm S}_{i,xy} & \Rightarrow & F^\dagger_i \tilde{\bm S}_{i,xy} F^{\phantom\dagger}_i, \\ \tilde{ n}_{i,xy} & \Rightarrow & F^\dagger_i \tilde{n}_{i,xy} F^{\phantom\dagger}_i, \end{aligned}$$ where on the right-hand side it is to be understood that the matrices in Eqs.  and act on the vector of spinon operators $$F_i=(f_{i1},f_{i2},f_{i3},f_{i4})^{\textrm{T}}.$$ Similar expressions can readily be written for operators in XZ and YZ planes. Thus the Hamiltonian in terms of spinons reads $$\begin{aligned} \tilde{{\mathcal}H}_{\text{ex-1}}^{\text{XY}} (ij) &=& J [(F_i^{\dagger} \tilde{\bm S}_{i,xy} F_i) \cdot (F_j^{\dagger} \tilde{\bm S}_{j,xy} F_j) \nonumber\\ && - \frac{1}{4} (F_i^{\dagger} \tilde{n}_{i,xy} F_i) \cdot (F_j^{\dagger} \tilde{n}_{j,xy} F_j) ] \nonumber \\ &=& \frac{2J}{9} \sum_{\alpha,\beta=2,3} \left[ -f^{\dagger}_{i,\alpha} f_{j,\alpha} f^{\dagger}_{j,\beta} f_{i,\beta} \right. \nonumber \\ && \left. + f^{\dagger}_{i,\alpha} f_{j,\beta} f^{\dagger}_{j,\beta} f_{i,\alpha} - f^{\dagger}_{i,\alpha} f_{i,\beta} \delta_{\alpha \beta} \right] \;,\end{aligned}$$ in which, $\bm{\sigma}_{\alpha\beta} \cdot \bm{\sigma}_{\alpha'\beta'} = 2 \delta_{\alpha \beta'} \delta_{\alpha'\beta} - \delta_{\alpha \beta} \delta_{\alpha'\beta'}$ has been used. Similar spinon Hamiltonians can also be written down for XZ and YZ planes. When we write down the full antiferromagnetic exchange Hamiltonian and sum over XY, YZ and XZ planes, we find that the single-site terms, which are quadratic in spinon operators, sum up to a constant once we impose the single occupancy constraint. We are then left with the terms that are quartic in spinon operators. We now follow the standard procedure of slave particle mean field theory to decouple the quartic terms in the spinon Hamiltonian and write down a mean field ansatz. We start with the exchange Hamiltonian in the XY plane, $\tilde{{\mathcal}H}_{\text{ex-1}}^{\text{XY}}$. We require a mean field ansatz for the fermionic bond expectation values, $\langle f^\dagger_{i\alpha}f^{\phantom\dagger}_{j\beta}\rangle$. Noting the structure of the two site pseudo-singlet in this plane, we choose an ansatz which reproduces a quantum ground state of this type. Specifically, $$\chi_{ij;\alpha\beta}\equiv \langle f^\dagger_{i\alpha}f^{\phantom\dagger}_{j\beta}\rangle=\chi_{ij}(\mathcal{I}_{xy})_{\beta\alpha},\qquad \langle ij\rangle \in \text{XY} \label{chiijansatz}$$ with $$\mathcal{I}_{xy}= \left( \begin{array}{c|c|c} 0 & & \\ \hline & I_2 & \\ \hline & & 0 \end{array} \right) \;.$$ Note that, by construction, this expectation value is invariant under the hidden SU(2) symmetry. The $\chi_{ij}$ on the XZ and YZ planes are determined by symmetry $$\begin{aligned} \label{eq:ansatzxzyz} \begin{array}{ll} \vspace{2mm} \chi_{ij;\alpha\beta} \equiv \langle f^\dagger_{i\alpha}f_{j\beta}\rangle = \chi_{ij}(\mathcal{I}_{xz})_{\beta\alpha}, & \langle ij\rangle \in \text{XZ} \;, \\ \chi_{ij;\alpha\beta} \equiv \langle f^\dagger_{i\alpha}f_{j\beta}\rangle = \chi_{ij}(\mathcal{I}_{yz})_{\beta\alpha}, & \langle ij\rangle \in \text{YZ} \;, \end{array}\end{aligned}$$ and $$\begin{aligned} \label{eq:idxzyz} \mathcal{I}_{xz} & = & U^{\dagger} \mathcal{I}_{xy} U \\ \mathcal{I}_{yz} & = & U \mathcal{I}_{xy} U^{\dagger}\end{aligned}$$ with the unitary transformation introduced in Eq. . We then arrive at the mean field Hamiltonian $$\begin{aligned} {{\mathcal}H}_{\text{MF}}^{\text{XY}} &=& - \tilde{J} \sum_{\langle ij \rangle \in \text{XY}} \left[ (\chi_{ij} F^\dagger_{j} \mathcal{I}_{xy} F^{\phantom\dagger}_{i}+h.c.)-2|\chi_{ij}|^2\right] \nonumber\\ & & +\sum_i \Lambda_i(F^\dagger_iF^{\phantom\dagger}_i-1) \label{meanfield}\end{aligned}$$ with $ \tilde{J} \equiv 2J/9 $. Here $\Lambda_i$ are the Lagrange multipliers related to the single-occupancy constraint in Eq. . $\mathcal{H}_{\text{MF}}^{\text{XZ}}$ and $\mathcal{H}_{\text{MF}}^{\text{YZ}} $ can be readily written down using Eq. \[eq:ansatzxzyz\] and Eq. . ### uniform spin liquid {#sec:uniform-spin-liquid} As discussed in Sec. \[sec:sec2\], the antiferromagnetic exchange Hamiltonian has a “hidden” global SU$(2)$ symmetry, $[G^{\mu}, \tilde{{\mathcal}H}_{\text{ex-1}}]=0$. It is easy to find that the full mean field Hamiltonian we have here respects this “hidden” global SU$(2)$ symmetry. We seek a quantum spin liquid ground state which does not break any symmetries of the original Hamiltonian. Translational invariance imposes $\Lambda_i=\Lambda=$ const. First we consider the ansatz for a uniform spin liquid $$\chi_{ij}=\chi_{ji}=\chi ,$$ for $i,j$ nearest neighbors on the fcc lattice. This naturally respects point group and time reversal symmetries. The Hamiltonian in Eq.  is then diagonalized by Fourier transform $$f_{\alpha}(\mathbf{k})=\sum_je^{-i\mathbf{k}\cdot \mathbf{R}_j}f_{j\alpha}.$$ We find $$\begin{aligned} \mathcal{H}_{MF}&=&\sum_{\lambda=1,2}\sum_{\mathbf{k}}\epsilon_\lambda(\mathbf{k})[\tilde f^\dagger_{\lambda+}(\mathbf{k})\tilde f^{\phantom\dagger}_{\lambda+}(\mathbf{k})+\tilde f^\dagger_{\lambda-}(\mathbf{k})\tilde f^{\phantom\dagger}_{\lambda-}(\mathbf{k})]\nonumber\\ &&+12{{\mathcal}N}\tilde{J}\chi^2,\end{aligned}$$ where $\lambda=1,2$ label doubly degenerate bands with dispersion $$\begin{aligned} && \epsilon_{1,2}(\mathbf{k})\equiv \tilde{J}\chi\,\tilde{\epsilon}_{1,2}(\mathbf{k})= -2\tilde{J}\chi\Big[C_xC_y+C_yC_z+C_zC_x \nonumber \\ && \pm\sqrt{C_x^2C_y^2+C_y^2C_z^2+C_z^2C_x^2-C_xC_yC_z(C_x\!+\!C_y\!+\!C_z)}\Big]. \nonumber \\ &&\end{aligned}$$ Here $C_x=\cos (k_x/2)$, $C_y=\cos (k_y/2)$ and $C_z=\cos (k_z/2)$. The double degeneracy of the two bands is due to Kramer’s degeneracy, since ${\bm j}_i$ is a spin-3/2 operator and the Hamiltonian has time reversal symmetry. The ground state wave function at the mean field level is described by a Fermi sea of spinons $$|\Psi_{MF}\rangle=\prod_{\lambda=1,2}\prod_{\mathbf{k}}\tilde f^\dagger_{\lambda+}(\mathbf{k})\tilde f^\dagger_{\lambda-}(\mathbf{k})|\textrm{vacuum}\rangle,\label{fermisurface}$$ for all $\mathbf{k}$ below the Fermi surface. The mean field ground state energy per site is $$\frac{E_{MF}}{{{\mathcal}N} \tilde{J}}=2\chi\sum_{\lambda=1,2}\int\frac{d^3 k}{(2\pi)^3}\,\theta[\mu-\tilde\epsilon_\lambda(\mathbf{k})] \tilde\epsilon_\lambda(\mathbf{k})+12 \chi^2,\label{EMF}$$ where the integral is over the first Brillouin zone of the fcc lattice and the dimensionless chemical potential $\mu$ is fixed by the quarter filling condition $$2\sum_{\lambda=1,2 }\int\frac{d^3 k}{(2\pi)^3}\,\theta[\mu-\tilde\epsilon_\lambda(\mathbf{k})] =1.$$ The wave function in Eq. (\[fermisurface\]) must be Gutzwiller-projected into the physical Hilbert space with one spinon per site $$|\Psi\rangle=\mathcal{P}_{n_i=1}|\Psi_{MF}\rangle.$$ Here we simply evaluate the ground state energy at the mean field level. Minimizing Eq. (\[EMF\]) with respect to the parameter $\chi$, we find $$\chi^* =-\frac{1}{12}\sum_{\lambda=1,2 }\int\frac{d^3 k}{(2\pi)^3}\,\theta[\mu-\tilde\epsilon_\lambda(\mathbf{k})] \tilde\epsilon_\lambda(\mathbf{k}),$$ and $$\frac{E_{MF}}{{\mathcal}N}=-\frac{8}{3}J(\chi^*)^2.$$ The spinon density is at quarter filling for $\mu\approx -1.58J$. The mean field energy for the uniform spin liquid state is then $E^{(0)}_{MF}/{{\mathcal}N}\approx -0.041 J $. ### $\pi$-flux spin liquid {#sec:pi-flux-spin} We now consider the ansatz for the $\pi$ flux spin liquid state illustrated in Fig. \[fig:piflux\]. In order to preserve time reversal symmetry, the phase of the $\chi_{ij}$ at each bond can only assume the values $0$ or $\pi$. We divide the fcc lattice into four cubic lattices $$\begin{aligned} \mathbf{r}^A_j&=&(0,0,0)+\mathbf{R}_j,\nonumber\\ \mathbf{r}^B_j&=&(1/2,1/2,0)+\mathbf{R}_j,\nonumber\\ \mathbf{r}^C_j&=&(0,1/2,1/2)+\mathbf{R}_j,\nonumber\\ \mathbf{r}^D_j&=&(1/2,0,1/2)+\mathbf{R}_j,\nonumber\end{aligned}$$ where $\mathbf{R}_j$ is a unit vector in the cubic lattice with lattice parameter $a=1$. We denote by $F_{jA}$ the vector of spinon annihilation operators at site $j$ of sublattice $A$, and similarly for the other sublattices. We assign $\chi_{ij}=+\chi$ to the bonds connecting sites in sublattice $A$ to all its nearest neighbors in sublattices $B,C,D$, and $\chi_{ij}=-\chi$ to the bonds connecting two sites that belong to sublattices $B,C$ or $D$. In other words, this ansatz corresponds to assigning $-\chi$ to the three bonds in the $BCD$ base of each tetrahedron in the fcc lattice and $+\chi$ to the three bonds connecting the $BCD$ base to the $A$ vertex. As a result, there is $\pi$ flux through every triangle and zero flux through every square in the fcc lattice. While this ansatz is clearly invariant under point group transformations about $A$ sites, it is also invariant under lattice translations, despite the fact that this permutes the 4 sublattices. This is because the corresponding changes in $\chi_{ij}$ can be removed by a gauge transformation. For instance, the gauge transformation $$\begin{aligned} &&F_{jA}\to -F_{jA},\quad F_{jB}\to -F_{jB},\nonumber\\&&F_{jC}\to F_{jC},\quad F_{jD}\to F_{jD},\end{aligned}$$ exchanges the signs of $\chi_{ij}$ between sublattices $A$ and $B$. It follows that the $\pi$ flux ansatz is invariant under point group symmetries about any site of the fcc lattice and therefore respects all symmetries of the original Hamiltonian. ![(Color online) Conventional unit cell of the fcc lattice divided into four sublattices. The ansatz for the $\pi$ flux state corresponds to assigning hopping amplitude $+\chi$ to the bonds represented by solid lines and $-\chi$ to the bonds represented by the dashed lines. []{data-label="fig:piflux"}](fig21.pdf){width="6.0cm"} Minimizing the energy for the mean field Hamiltonian with four sublattices, we find 8 doubly degenerate bands for the $\pi$ flux state. Quarter filling is reached for dimensionless chemical potential $\mu\approx -1.68J$. The mean field energy is $$\frac{E_{MF}}{{{\mathcal}N}\tilde{J}}=2\chi\sum_{\lambda=1}^8\int\frac{d^3 k}{(2\pi)^3}\,\theta[\mu-\tilde\epsilon_\lambda(\mathbf{k})] \tilde\epsilon_\lambda(\mathbf{k})+12 \chi^2,$$ where the integral is over the reduced Brillouin zone of the cubic lattice. We find $E^{(\pi)}_{MF}/{{\mathcal}N}=-0.053J$. This is lower than the energy for the uniform state. We may also compare this to the [ *mean field*]{} energy of a VBS state (it is not so meaningful to compare these slave particle mean field energies directly to the variational energies quoted earlier for the Weiss mean field and VBS states). For a mean-field VBS state, we take $\chi_{ij}$ non-zero only on a set of non-overlapping dimers. In this case, we obtain $E_{VBS}^{MF}/{{\mathcal}N}=-1/18J\approx-0.055J$. This is slightly lower than the QSL states, but we expect that the energy of the spin liquid states will be lowered by the Gutzwiller projection, since the latter is known to enhance spin-spin correlations.[@citeulike:5750703] We note that both spin liquid states have Fermi surfaces which are not nested, and have no obvious instabilities. The states are also stable against bond anisotropy which enhances the hopping in a given plane. Perturbations to NN AFM exchange, such as next-nearest-neighbor interactions, will in general require more general ansätze for the bond matrix $\chi_{ij}$. Nonetheless, as long as the perturbations are in some sense small, the $\chi_{ij}$ assumed in Eq. (\[chiijansatz\]), in which hopping in a given plane occurs predominantly for two out of four spinon species, should be a good starting point for approximations. Discussion {#sec:sec6} ========== In this paper, we have introduced and analyzed a model to describe localized electrons in a 4d$^1$ or 5d$^1$ configuration on an fcc lattice, in which strong spin-orbit coupling and the $t_{2g}$ orbital degeneracy combine to produce an effective $j=3/2$ description. The model contains three interactions – nearest neighbor antiferromagnetic and ferromagnetic exchange, and electric quadrupolar interactions – and in addition may include the effect of structural anisotropy. We obtain the (Weiss) mean-field phase diagram, which includes 3 main phases, which all have a two-sublattice ${{\boldsymbol}Q}= 2\pi (001)$ structure. In all the phases, large multipolar order parameters in addition to the usual magnetic dipolar order are present. Most remarkably, we find a broad regime of time-reversal invariant but quadrupolar ordered phase at intermediate temperatures. A spin-wave analysis indicates that quantum fluctuations are strong when nearest-neighbor antiferromagnetic exchange is dominant, and in this case, we suggest possible quantum spin liquid and valence bond solid phases. Experimental ramifications {#sec:exper-ramif} -------------------------- The theory developed here can be applied and tested in a multitude of ways. Here we discuss a few of the main experimental properties which might be measured. First, there is the spatial symmetry breaking of the ordered phases. All the ordered states, at least in the cubic case, break lattice symmetries, and in particular double the unit cell. This can be tested in experiments such as neutron and x-ray scattering. We discuss in some further detail the most intriguing case of the quadrupolar phase, which is non-magnetic. It constitutes a type of real (time-reversal invariant) orbital ordering. It leads to a spontaneous splitting of the local quadruplet, breaking it down to an elemental Kramer’s doublet. As such, this is not entirely distinct from a Jahn-Teller effect, in which ionic motions would lead to such splitting. In particular, even though in our model atomic displacements are not involved in an essential way, they would be expected at least to respond to the orbital order. In principle, this could be measured by scattering (x-rays or neutrons) which accurately measure the crystal structure and symmetry. From the order parameter description of the quadrupolar phase, we can obtain the corresponding space group and crystal structure parameters to be sought in such a measurement. In particular, we find that the quadrupolar ordered phase corresponds to the tetragonal space group P4$_2$/mnm (number 136). In this space group, apart from the doubling of the unit cell, one finds that all the A sites, B sites, and B’ sites remain equivalent. However, the oxygens are no longer equivalent, but split into three classes, occupying the 4e, 4f, and 4g Wyckoff positions. Each of these positions has one degree of freedom which is not fixed by symmetry. Physically, the oxygens remain constrained to the simple cubic axes of the perovskite reference unit cell, but may move by different amounts along each of the three axes. This is two more degrees of freedom than is found in the cubic Fm$\overline{3}$m (number 225) space group, in which the oxygens maintain an ideal octahedron equidistant from each B (or B’) site. While symmetry requires these structural changes, we do not have at present an estimate for their magnitude, which could be weak if coupling to the lattice is not strong. Another quantity we have already discussed in Sec. \[sec:field\] is the magnetic susceptibility, which shows signatures of the quadrupolar and ordering transitions. One complication is that the susceptibility is in many cases highly anisotropic, and one may not be sure what component(s) are being measured in practice. Specifically, one may expect, if the system is ideal and fully in equilibrium, that the broken symmetry order parameters can be reoriented by the magnetic field, in such a way that they minimize the free energy. This will typically favor orientations which maximize the magnetic susceptibility. For instance, in the quadrupolar phase for cubic symmetry, this is an orientation with ${{\boldsymbol}Q}$ perpendicular to the field. However, such reorientation involves motion of domain walls and very large numbers of spins, and can easily be prevented by pinning or at least be incomplete in practice. Thus some diversity of behavior may be expected in experiment, as well as possibly hysteretic behavior even in the non-magnetic state. When the crystal is non-cubic, one may explore the influence of single-ion anisotropy on the magnetic susceptibility. A naïve application of Eq.  would immediately imply that single-ion terms do not contribute to the Curie-Weiss temperature as measured in the powder susceptibility. However, we caution that these equations hold only in the true high-temperature regime, in which $T \gg |D|$. If $T$ is smaller than or comparable to $|D|$, higher order terms in the high temperature expansion are non-trivial, and a non-vanishing fitted Curie-Weiss temperature may result from $D$ alone. Let us consider the powder susceptibility $\overline{\chi}$ for independent ions (i.e. neglecting exchange). One has $$\label{eq:18} \overline{\chi} = \frac{\chi_{zz}}{3} + \frac{2\chi_{xx}}{3} = \frac{3}{4T} + \frac{\tanh{D/T}}{2D},$$ in units of $g^2\mu_B^2$. We then suppose a linear fit to $1/\overline{\chi}$ versus $T$ is made over a narrow region in the neighborhood of the temperature $T_{\rm fit}$, and extrapolated to find the Curie-Weiss temperature as the intercept of the horizontal axis. The result is $$\label{eq:19} \Theta_{\text{CW}}(T_{\rm fit}) = - \frac{2T_{\rm fit} \left( T_{\rm fit} \sinh (\frac{2D}{T_{\rm fit}})-2 D\right) }{D\left(3 \cosh(\frac{2D}{T_{\rm fit}})+7\right)}.$$ Note that the fitted Curie-Weiss temperature is always negative, and is independent of the sign of $D$. It reaches a maximum in magnitude (at fixed $D$) of $\Theta_{\text{CW}} \approx -0.18 |D|$ when $T_{\rm fit} \approx 0.88|D|$, and only approaches zero very slowly when $T_{\rm fit} \gg |D|$: $\Theta_{\text{CW}} \sim - \frac{4D^2}{15T_{\rm fit}}$. Conversely, at a fixed fitting temperature, the maximum achievable Curie-Weiss temperature is $\Theta_{\text{CW}} \approx -0.26 T_{\rm fit}$, when $|D|=1.84T_{\rm fit}$. Materials survey {#sec:materials-survey} ---------------- We now turn to a discussion of specific materials which have been studied experimentally. ### Ba$_2$YMoO$_6$ {#sec:ba_2ymoo_6} We begin with the material Ba$_2$YMoO$_6$, which has been suggested experimentally to be an exotic “valence bond glass” or to have a “collective spin singlet” ground state. The expected separation between the $j=3/2$ and $j=1/2$ states in this material is over 2000K, so that the effective $j=3/2$ description used here should be excellent. Two recent experimental papers[@arxiv0542; @arxiv1665] observed an unusual behavior of the magnetic susceptibility, with [*two*]{} Curie regimes, such that $1/\chi$ is linear both above 100K and below 50K. Moreover, the magnetic specific heat shows a peak around 50K, with Ref.  estimating the total magnetic entropy approximately equal to R$\ln$4, as expected for $j=3/2$. Both these results suggest the existence of some single-ion anisotropy, which would explain the existence of two Curie regimes because it splits the 4-fold degeneracy of the $j=3/2$ states but leaves a 2-fold Kramer’s doublet at temperatures below $|D|$, which still gives a Curie signal. However, the cubic symmetry observed experimentally seems to rule out such an explanation. Moreover, the form of the powder susceptibility in Refs.  is qualitatively different from that expected for either fixed easy-plane or easy-axis anisotropy. These difficulties are resolved if one considers the possibility of [*spontaneous*]{} anisotropy, which indeed is the primary characteristic of the quadrupolar ordered state. For example, the mean-field susceptibility for the cubic model with $J'=0.2J, V=0.3J$ is plotted in Fig. \[fig:chiinv0203\]. At temperatures above the FM110 phase, one indeed observes two Curie regimes in the susceptibility, with a larger Curie constant at low temperature, as seen in the experiments. The kink in $\chi$ coincides with the quadrupolar ordering transition, and there is a peak in the specific heat at this temperature, also as observed in experiment. The theoretical specific heat has a second peak at lower temperatures, associated with magnetic ordering and exhaustion of the unsplit Kramer’s doublet. We suggest that this peak is below the lowest temperatures measured, or perhaps is avoided due to disorder, and the spins falling out of equilibrium at low temperature. The fact that the Curie-Weiss temperature extracted [*below*]{} 50K is only -2.3K corroborates the notion that any magnetic ordering may be too low to observe or be obscured by the effects of disorder. Compound $B'$ config. crystal structure $\Theta_{\text{CW}}$ $\mu_{\text{eff}}(\mu_B)$ magnetic transition frustration parameter $f$ Ref ----------------- -------------------- --------------------------------------- ---------------------- --------------------------- ---------------------------- --------------------------- ------ Ba$_2$YMoO$_6$ Mo$^{5+}(4d^1)$ cubic $-91$K $1.34$ PM down to $2$K $f \gtrsim 45$ \[\] Ba$_2$YMoO$_6$ Mo$^{5+}(4d^1)$ cubic $-160$K $1.40$ PM down to $2$K $f \gtrsim 80$ \[\] Ba$_2$YMoO$_6$ Mo$^{5+}(4d^1)$ cubic $-219$K $1.72$ PM down to $2$K $f \gtrsim 100$ \[\] La$_2$LiMoO$_6$ Mo$^{5+}$ (4d$^1$) monoclinic -45K 1.42 PM to 2K $ f\gtrsim 20$ \[\] Sr$_2$MgReO$_6$ Re$^{6+}(5d^1)$ tetragonal $-426$K $1.72$ spin glass, $T_G \sim 50$K $f \gtrsim 8$ \[\] Sr$_2$CaReO$_6$ Re$^{6+}(5d^1)$ monoclinic $-443$K $1.659$ spin glass, $T_G\sim 14$K $f \gtrsim 30$ \[\] Ba$_2$CaReO$_6$ Re$^{6+}(5d^1)$ cubic to tetragonal (at $T\sim 120$K) $-38.8$K $0.744$ AFM $T_N =15.4$K $f\sim 2$ \[\] Ba$_2$LiOsO$_6$ Os$^{7+}(5d^1)$ cubic $-40.48$K $0.733$ AFM $T_N\sim8$K $f \gtrsim 5$ \[\] Ba$_2$NaOsO$_6$ Os$^{7+}(5d^1)$ cubic $-32.45$K $0.677$ FM $T_N\sim8$K $f \gtrsim 4 $ \[\] Ba$_2$NaOsO$_6$ Os$^{7+}(5d^1)$ cubic $\sim-10$K $\sim 0.6$ FM $T_N=6.8$K $f \gtrsim 4 $ \[\] : A list of ordered double perovskites. Note the discrepancy in Curie-Weiss temperature and $\mu_{ \text{eff} } $ may originate from the experimental fitting of data at different temperature range. []{data-label="tab:Tab1"} ### La$_2$LiMoO$_6$ {#sec:la_2limoo_6} La$_2$LiMoO$_6$ is monoclinic, the deviation from cubic symmetry arising primarily from rotations of the octahedra. The local coordination of the Mo sites is nearly perfectly octahedral with a weak tetragonal compression. The nature of crystal field effects, if significant, is unclear at present. Magnetically, the susceptibility shows, like Ba$_2$YMoO$_6$, two apparent Curie regimes, separated by a kink at approximately 150K. However, opposite to that material, La$_2$LiMoO$_6$ shows a smaller effective moment at low temperature compared to high temperature. In addition, the high temperature Curie-Weiss temperature is $\Theta_{\text{CW}} \approx -45K$, significantly smaller than the kink temperature. Irreversibility distinguishing the behavior of the ZFC/FC susceptibility appears below 25K. The appearance of two Curie regimes again suggests either fixed or spontaneous magnetic anisotropy setting in around 150K. However, the [*reduction*]{} of the effective moment below the kink in $\chi^{-1}$ is puzzling. We did not find this behavior in the powder susceptibility within our model, with or without anisotropy modeled by $D$. As remarked above, however, the actual nature of the crystal field anisotropy in La$_2$LiMoO$_6$ is unclear. If it is significant and different in form from the $D$ term, this might explain the behavior. Single crystal studies would be helpful in elucidating the situation. ### Sr$_2$CaReO$_6$ and Sr$_2$MgReO$_6$ {#sec:sr_2c-sr_2mgr} Sr$_2$CaReO$_6$ and Sr$_2$MgReO$_6$ have distorted perovskite structures, with monoclinic and tetragonal symmetry, respectively.[@wiebe:prb2002; @wiebe:prb2003] Experimentally, the materials are notable for their very high antiferromagnetic Curie-Weiss temperature, $-\Theta_{\text{CW}}\gtrsim$ 400K. Susceptibility and specific heat measurements show anomalies suggestive of freezing and/or short-range ordering at 14K and 50K, for Sr$_2$CaReO$_6$ and Sr$_2$MgReO$_6$, respectively. Two possible interpretations of this behavior are: (1) the Curie-Weiss temperature is dominated by strong exchange, but fluctuations largely suppress ordering, or (2) the Curie-Weiss temperature is due largely to single-ion effects, and the true exchange scale is comparable to the observed anomalies in $\chi$ and $c_v$. In the former scenario, the key question is why these two materials show so much larger exchange than do the other compounds in this family. From the point of view of this work, attributing the Curie-Weiss temperature to exchange alone would imply $J$ is actually comparable to the SOC, so that the projection to $j=3/2$ may even be suspect. The Curie-Weiss temperatures are sufficiently large that one may suspect that the $5d$ electrons are not so well localized, and the system is close to a Mott transition. It would be interesting to measure their optical properties to address this possibility. The latter explanation seems possible, as both materials show significant deviations from the cubic structure: Sr$_2$CaReO$_6$ is monoclinic, while Sr$_2$MgReO$_6$ is tetragonal. The actual distortions of the octahedra are rather small in both cases, the Re-O distance varying by only about 0.02$\AA$ at room temperature. However, there are significant rotations and tilts of the octahedra, and crystal field splittings of the $j=3/2$ quadruplet are certainly allowed. Examination of the Re-O bond lengths suggests easy-axis anisotropy. From Eq. , we see that in principle a negative Curie-Weiss temperature could be attributed to $D$. However, from the present model we cannot obtain such a large value, which in these two materials is comparable or larger than the fitting temperature. Nevertheless, we may imagine that some combination of exchange and single-ion anisotropy may conspire to produce the observed behavior. If we assume a large easy-axis anisotropy, we would then expect, based on the the analysis in Sec. \[sec:easy-axis-anisotropy\], to have an AFM ground state. The anomalies might be related to this ordering. Experimentally, spin freezing and irreversibility is observed, but without clear signs of long-range ordering. The experimentalists caution that, due to the small magnetic moment of the Re$^{6+}$ ions, a small ordered component could not be ruled out in either material.[@wiebe:prb2003; @wiebe:prb2002] Indeed, in the AFM state, a very small moment is expected, due to the primacy of octupolar order. While this is promising, we note that it is likely that several effects not in our model play a role. First, the structure of the materials is not a simple compression of the cubic structure, and so the crystal fields might have a significantly different form from the simple $D$ term. This is especially true in Sr$_2$CaReO$_6$, which has the more distorted monoclinic structure. Second, in Sr$_2$MgReO$_6$, the Re-O-Mg bond angles are very different in the XY plane (160$^\circ$) and normal to it (180$^\circ$),[@wiebe:prb2003] so substantial spatial anisotropy in the exchange couplings may be present. This is not included in our model. Also in Sr$_2$MgReO$_6$, the ZFC and FC susceptibility actually diverge already around 300K, which suggests a high degree of disorder in this material, which might be responsible for converting the AFM to a glassy state. One indication supporting an antiferromagnetic ground state is the observation, in Sr$_2$CaReO$_6$, of a $T^3$ magnetic contribution to the specific heat, in contrast to the usual linear one characteristic of a spin glass. The $T^3$ behavior would naturally be expected from spin waves in the AFM state, which as we have noted displays gapless spin waves, at least in the semi-classical approximation. Such $T^3$ behavior might even persist if the AFM order had a finite correlation length, due to Halperin-Saslow modes[@halperin1977hts], as recently postulated in NiGa$_2$S$_4$[@podolsky2008hsm]. A $T$-linear specific heat was observed in Sr$_2$MgReO$_6$, albeit with a small coefficient.[@wiebe:prb2003] As we have already remarked, however, this material is likely to be more disordered, consistent with the more conventional spin glass-like specific heat. While these considerations seem reasonable, they are hardly definitive. Further studies, particularly on single crystals, would be most helpful in clarifying the physics of these materials. ### Ba$_2$CaReO$_6$ {#sec:ba_2careo_6} In Ba$_2$CaReO$_6$, there is a structural transition from a high temperature cubic phase to a low temperature tetragonal one, with a doubled unit cell, at $T=120\text{K}$[@yamamura:jssc2006]. The experimentalists have fitted the low temperature structure to the I4/m space group. From this fit, they found an elongation along the $c$ (or $z$) axis, but a slight compression of the ReO$_6$ octahedra. One may consider two possibilities. Either this is indeed the correct symmetry, in which case it must have structural origin not related to the 5d electrons, [*or*]{} this transition in fact coincides with the quadrupolar ordering described here, which also gives a tetragonal state with the same unit cell. The P4$_2$/mnm space group was not considered in Ref.. A small, negative Curie-Weiss temperature $\Theta_{\text{CW}} = -$39K was measured by fitting the susceptibility in the high temperature cubic phase , indicating that $J'$ should be not too large. A predominantly antiferromagnetic ordering transition was observed at $T=$15K, which is consistent with our expectations in the small $J'$ and $V$ regime (recall that an antiferromagnetic state is expected both with and without single ion anisotropy in this parameter range). It would be interesting to compare the predicted magnetic structure in the AFM or AFM’ phase with experiment, by carrying out neutron scattering and NMR measurements. ### Ba$_2$NaOsO$_6$ {#sec:ba_2naoso_6} Ba$_2$NaOsO$_6$,[@Erickson] is one of only two examples in this class in which single crystal experiments have been performed, to our knowledge. A transition at $T=$6.8K was found to a ferromagnetic state with easy axis along a $[ 110 ] $ direction. Within the experimental resolution, the material was found to remain cubic down to the lowest measured temperature. A fit to Curie-Weiss behavior found a negative Curie-Weiss temperature, $-13$K$<\Theta_{\text{CW}}<-$10K, depending upon field orientation. We note that this is compatible with Eq. , and suggests that the material is in the regime of larger $V$ and smaller $J'$. In this region, we expect a high temperature quadrupolar transition, well above the ferromagnetic state. While no such transition is observed in the experiment, the magnetic specific heat of such a transition may be masked by the lattice contribution at higher temperature, and the signature in susceptibility may be subtle. Several other indications are in favor of this scenario. First, the susceptibility continues to display pronounced anisotropy, favoring the $[ 110 ] $ direction, up to at least 200K, well above the ferromagnetic transition. This would indeed be expected in the quadrupolar ordered state. Second, the magnetic specific heat divided by temperature, integrated over the peak up to about 15K gives a magnetic entropy of approximately $R\ln 2$, which is only half the expected entropy for the $j=3/2$ quadruplet. This entropy may be released over a significantly higher temperature range, up to the quadrupolar ordering transition. Third, the observation itself of ferromagnetism with a $[110]$ easy axis is a marked success of our theory. This type of anisotropy is not natural from the standpoint of Landau theory, within which cubic anisotropy manifests itself at leading order as a term in the free energy of the form $v[(M^x)^4+(M^y)^4+(M^z)^4]$ (${\bm M}$ is the magnetization), which, depending upon the sign of $v$, generates a $[ 111]$ or $[ 100]$ easy axis. A continuous Landau transition to the FM110 state is instead made possible by the fact that the quadrupolar order already breaks the cubic symmetry in the paramagnetic state. It would be interesting to further probe the system to establish in more detail the correspondence (or lack thereof) with our theoretical predictions. The predicted tetragonal distortion of the cubic structure would be a natural quantity to seek in experiment. This also manifests locally in the magnetism, since although the net magnetization is aligned with the $[ 110 ]$ axis, the local spin expectation values are not. This might be measurable for instance by a local probe such as NMR. ### Ba$_2$LiOsO$_6$ {#sec:ba_2lioso_6} Ba$_2$LiOsO$_6$ has also been grown in single crystal form. The structure was determined to be cubic by x-ray diffraction at room temperature.[@stitzer:sss2002] Aside from this diffraction data, only bulk magnetic susceptibility results are available. One observes a negative Curie-Weiss temperature $\Theta_{\text{CW}} \approx -40K$, and an apparent antiferromagnetic transition at $T_N \approx 8K$. This appears largely consistent with the expected behavior in region I of Fig. \[fig:thermal\_phase\]. Comparison and future work {#sec:comp-with-other} -------------------------- We are aware of only one other theoretical work studying this class of materials. Lee and Pickett[@lee:epl2007] performed electronic structure calculations for Ba$_2$NaOsO$_6$ and Ba$_2$LiOsO$_6$, emphasizing the role of SOC. We completely agree with the conclusion that SOC plays a crucial role in the magnetism. However, the magnetic structure and phase transitions were not addressed. Our study is much more comprehensive, and gives a great deal of guidance both for future theory and experiment. Many experimental suggestions have already been made. In particular verification of the quadrupolar ordering transition would be especially exciting. On the theoretical side, the problem of the effects of quantum fluctuations in the small $J'$ and $V$ limit remains rather open. It would be remarkable if a spin liquid or valence bond solid state could be established for this highly non-SU(2) symmetric and nominally “large” spin $j=3/2$ model. To do so would require some hard theoretical work applying more quantitative numerical methods to our model Hamiltonian. A natural extension of this work would be to consider the “higher spin” analogs of these materials, with 4d$^2$ or 5d$^2$ electronic states. As there is still partial occupation of the $t_{2g}$ orbitals in this case, we expect SOC again to play a dominant role, and interesting multipolar physics is likely present. We thank John Greedan, Ian Fisher and Ram Seshadri for useful discussions. This work was supported by the DOE through Basic Energy Sciences grant DE-FG02-08ER46524. LB’s research facilities at the KITP were supported by the National Science Foundation grant NSF PHY-0551164.

ArXiv

--- abstract: 'The general perturbative expression for the lateral Casimir force between two plates covered by longitudinal corrugations of arbitrary shape is obtained. This expression is applicable for corrugation periods larger than the separation distance. The cases of asymmetric corrugations are considered, which allow to increase the maximum to minimum force ratio and affect the character of equilibrium points. This opens new opportunities to control the lateral Casimir forces for use in microelectromechanical devices based entirely on the vacuum fluctuation properties.' author: - 'E. V. Blagov,${}^{1}$ G. L. Klimchitskaya,${}^{2}$ U. Mohideen,${}^{3}$ and V. M. Mostepanenko${}^{1,2}$' title: On the control of the lateral Casimir force between corrugated surfaces --- It is well known that the Casimir force [@1] arises due to electromagnetic quantum fluctuations and acts between closely spaced surfaces. For short distances of order a nanometer, the limiting form of the Casimir force results in the van der Waals force. In fact the Casimir force is a collective quantum phenomenon which results from the alteration of the zero-point photon spectrum by surfaces of material bodies. H. B. G. Casimir [@1] was the first to calculate an extraordinary property that two parallel uncharged metallic plates placed in vacuum at some separation $a$ would feel an attractive force per unit area $F_0(a)=-\pi^2\hbar c/(240a^4)$, where $\hbar$ is the Planck constant and $c$ is the velocity of light. The Casimir force exhibits important yet exotic dependencies on the shape of boundary surfaces, which are not simple extensions from flat boundaries (see monographs [@2; @3; @4] and reviews [@5; @6]). Recently a number of experiments were performed on measuring the normal Casimir force, acting perpendicular to the interacting surfaces [@7; @8; @9; @10; @11; @12; @13; @14; @15]. The highest precision of order of 1% of the measured force was achieved in the measurements by means of atomic force microscope [@8; @9; @10; @11] and microelectromechanical torsional oscillator [@14; @15]. The extent of agreement between experiment and theory was used to set stringent limits on predictions of fundamental physical theories [@15; @16; @17; @18; @19]. The Casimir force can be the dominant force acting between surfaces for separation distances on the order or less than a few hundred nanometers. The Casimir effect was used to produce mechanical motion of microfabricated silicon plate [@20]. This is the first case of microelectromechanical device which shows actuation by the Casimir force. Note also that the Casimir forces acting between closely spaced surfaces play the crucial role in the fabrication and yield of microdevices due to the phenomena of adhesion and stiction [@21]. Similar to the normal Casimir force, the lateral Casimir force originates from the modifications of zero-point oscillations by material boundaries. It acts tangential to the two surfaces with nanoscale periodic corrugations. To first order the origin of the lateral Casimir force can be simply understood as follows. As the Casimir force is a strong function of distance ($F_0\sim a^{-4}$), the natural tendency is for corrugated plates to align themselves such that peaks of the top plate are directly over the peaks of the bottom plate. This force to align the peaks is then the origin of the lateral force acting in the horizontal direction between the plates. The lateral Casimir force for anisotropic boundaries was investigated theoretically and a harmonic dependence on corresponding angle was predicted [@3; @22; @23]. For two aligned sinusoidally corrugated plates made of ideal metal the lateral Casimir force was discussed in Refs. [@24; @25] and a harmonic dependence on a phase shift between corrugations was found. In Ref. [@26] the first experimental observation of the lateral Casimir force between the sinusoidally corrugated plate and large corrugated sphere was reported. The predicted sinusoidal dependence on the phase shift between corrugations was confirmed. The analysis of the theoretical dependence for the lateral Casimir force was performed in Ref. [@26] and the optimum values of the parameters providing the maximum values of the lateral force in the case of sinusoidal corrugations were found. In this report we investigate the lateral Casimir force between two parallel plates covered by longitudinal grooves of arbitrary shape. This opens new opportunities to change the magnitude of the lateral Casimir force and obtain asymmetric lateral forces with a more complicated character of equilibrium points leading to a rich variety of surface interactions. As a consequence, it may become possible to control friction through nanoscale modification of the contact surfaces to increase the yield of microdevices [@21]. The normal Casimir force acting between plates covered with small distortions has been the subject of much attention in literature (see Ref. [@6] for review). In Ref. [@27] a general method was provided which enables one to find corrections to the normal Casimir force in the configuration of two parallel plates with small deviations from plane parallel geometry. This method is based on the pairwise summation of the interatomic Casimir potentials with a subsequent normalization of the obtained interaction constant [@3; @6]. In doing so, the amplitudes of deviations from plane parallel geometry $A_{1,2}$ are assumed to be small as compared to the separation $a$ between plates. To fix the definition of the separation between the nonplanar plates, we describe their surfaces by the equations $$z_1^{(s)}=A_1f_1(x,y), \qquad z_2^{(s)}=a+A_2f_2(x,y), \label{eq1}$$ where $a$ is the mean value of the distance between the square plates with a side $2L$ and the coordinate system is connected with the lower plate. The values of the amplitudes are chosen in such a way that $\max|f_i(x,y)|=1$. The zero point in the $z$-axis is so chosen that $$\begin{aligned} && \langle z_1^{(s)}\rangle\equiv A_1\langle f_1(x,y)\rangle =\frac{A_1}{4L^2}\int_{-L}^{L}dx\int_{-L}^{L}dyf_1(x,y)=0, \nonumber \\ && \langle z_2^{(s)}\rangle\equiv a+A_2\langle f_2(x,y)\rangle=a. \label{eq2}\end{aligned}$$ The normal Casimir force between the plates covered with distortions was represented by a perturbative expansion up to the fourth order [@27] $$F(a)=F_0(a)\sum\limits_{k=0}^{4}\sum\limits_{l=0}^{4-k} c_{kl}\left(\frac{A_1}{a}\right)^k\left(\frac{A_2}{a}\right)^l, \label{eq3}$$ where $F_0(a)$ is the Casimir force per unit area of flat plates defined earlier, $c_{00}=1$ and $c_{01}=c_{10}=0$ \[this follows from our choice (\[eq2\])\]. Explicit expressions for the other coefficients $c_{kl}$ are found in Ref. [@27] (see also Refs. [@3; @6]). They are rather combersome. If, however, both functions $f_1$ and $f_2$ are periodic in two variables with periods much larger than $a$, but much less than $2L$, simple approximate expressions for $c_{kl}$ follow leading to $$\begin{aligned} &&F(a)=F_0(a)\left[ \vphantom{\left(\langle f_1^2\rangle A_1^2\right)}\right. 1+ \frac{10}{a^2}\left( \langle f_1^2\rangle A_1^2-2\langle f_1f_2\rangle A_1A_2 +\langle f_2^2\rangle A_2^2\right) \nonumber \\ &&\phantom{aaa} +\frac{20}{a^3}\left( \langle f_1^3\rangle A_1^3-3\langle f_1^2f_2\rangle A_1^2A_2 +3\langle f_1f_2^2\rangle A_1A_2^2 -\langle f_2^3\rangle A_2^3\right) \label{eq4} \\ &&\phantom{aaa} +\frac{35}{a^4}\left.\left( \langle f_1^4\rangle A_1^4-4\langle f_1^3f_2\rangle A_1^3A_2 +6\langle f_1^2f_2^2\rangle A_1^2A_2^2- 4\langle f_1f_2^3\rangle A_1A_2^3 +\langle f_2^4\rangle A_2^4\right)\right]. \nonumber\end{aligned}$$ Here the averaging as in Eq. (\[eq2\]) is done over the periods. From Eq. (\[eq4\]) the Casimir energy per unit area can be found $$\begin{aligned} &&E(a)=\int_{a}^{\infty}F(a^{\prime})da^{\prime}= E_0(a)\left[ \vphantom{\left(\langle f_1^2\rangle A_1^2\right)}\right. 1+ \frac{6}{a^2}\left( \langle f_1^2\rangle A_1^2-2\langle f_1f_2\rangle A_1A_2 +\langle f_2^2\rangle A_2^2\right) \nonumber \\ &&\phantom{aaa} +\frac{10}{a^3}\left( \langle f_1^3\rangle A_1^3-3\langle f_1^2f_2\rangle A_1^2A_2 +3\langle f_1f_2^2\rangle A_1A_2^2 -\langle f_2^3\rangle A_2^3\right) \label{eq5} \\ &&\phantom{aaa} +\frac{15}{a^4}\left.\left( \langle f_1^4\rangle A_1^4-4\langle f_1^3f_2\rangle A_1^3A_2 +6\langle f_1^2f_2^2\rangle A_1^2A_2^2- 4\langle f_1f_2^3\rangle A_1A_2^3 +\langle f_2^4\rangle A_2^4\right)\right], \nonumber\end{aligned}$$ where $E_0(a)=-\pi^2\hbar c/(720a^3)$ is the Casimir energy between two flat plates. Now let functions $f_{1,2}$ depend on only one variable, say $x$, i.e. they describe longitudinal corrugations extending in the $y$ direction. Let their periods be equal, $\Lambda_1=\Lambda_2=\Lambda$, but there may exist some phase shift $x_0$. In this case $E(a)=E(a,x_0)$ and the lateral Casimir force arises directed along $x$-axis. It can be found from Eq. (\[eq5\]) as $$\begin{aligned} &&F^{lat}(a,x_0)=-\frac{\partial E(a,x_0)}{\partial x_0}= F_0(a)\frac{2A_1A_2}{a^2}\left[ 2\frac{\partial}{\partial x_0}\langle f_1f_2\rangle \right. \nonumber \\ &&\phantom{aaa} +5\left( \frac{A_1}{a}\frac{\partial}{\partial x_0}\langle f_1^2f_2\rangle -\frac{A_2}{a}\frac{\partial}{\partial x_0}\langle f_1f_2^2\rangle \right) \label{eq6} \\ &&\phantom{aaa} +10\left.\left( \frac{A_1^2}{a^2}\frac{\partial}{\partial x_0}\langle f_1^3f_2\rangle -\frac{3}{2}\frac{A_1A_2}{a^2}\frac{\partial}{\partial x_0} \langle f_1^2f_2^2\rangle +\frac{A_2^2}{a^2}\frac{\partial}{\partial x_0} \langle f_1f_2^3\rangle\right)\right]. \nonumber\end{aligned}$$ Eq. (\[eq6\]) permits to find the lateral Casimir force between plates covered by the longitudinal corrugations of any shape. Under the condition used $\Lambda\gg a$ the precision of the pairwise summation method is very high and the results practically coincide with the exact ones (this was demonstrated in Ref. [@25] for the case of sinusoidal corrugations). The same conclusion can be obtained by the method of Ref. [@28]. The advantage of Eq. (\[eq6\]) is the possibility to quickly obtain the results for corrugations of any shape and to generalize this formalism for the case of real metals of finite conductivity (see Ref. [@26]). As was shown in Ref. [@26] for the sinusoidal corrugations, the dependence of the lateral force on the phase shift is almost sinusoidal. Specifically, the points of the stable and unstable equilibrium (where $F^{lat}=0$) alternate with half the corrugation period. The magnitudes of the maximum and minimum values of $F^{lat}$ are equal to each other so that their ratio is equal to unity. These characteristic features of the lateral Casimir force are preserved for any longitudinal corrugation whose shape is symmetric relatively to some vertical axis. In applications to microelectromechanical systems, a more complicated nature of the lateral Casimir force may be desirable, i.e. different magnitudes of maximum and minimum values and a more complicated character of the points of equilibrium. This can be achieved by use of asymmetric longitudinal grooves on the plate surfaces. We consider saw toothed corrugations with equal amplitudes shown in Fig. 1,a. Within the periods (from 0 to $\Lambda$ for $f_1$ and from $x_0$ to $x_0+\Lambda$ for $f_2$) the analytical representations of the corrugation functions of Fig. 1,a are $$f_1(x)=\frac{2x}{\Lambda}-1, \qquad f_2(x)=1-\frac{2(x-x_0)}{\Lambda}. \label{eq7}$$ Calculating all matrix elements from Eq. (\[eq6\]) over the period from 0 to $\Lambda$, one obtains $$\begin{aligned} && \langle f_1f_2\rangle=-\frac{1}{3}+2\frac{x_0}{\Lambda}- 2\frac{x_0^2}{\Lambda^2}, \nonumber \\ && \langle f_1^2f_2\rangle= \langle f_1f_2^2\rangle=-\frac{4}{3}\frac{x_0}{\Lambda}+ 4\frac{x_0^2}{\Lambda^2}- \frac{8}{3}\frac{x_0^3}{\Lambda^3}, \label{eq8} \\ && \langle f_1^3f_2\rangle= \langle f_1f_2^3\rangle=-\frac{1}{5}+2\frac{x_0}{\Lambda}- 6\frac{x_0^2}{\Lambda^2}+ 8\frac{x_0^3}{\Lambda^3}-4\frac{x_0^4}{\Lambda^4}, \nonumber \\ && \langle f_1^2f_2^2\rangle=\frac{1}{5} -\frac{8}{3}\frac{x_0^2}{\Lambda^2}+ \frac{16}{3}\frac{x_0^3}{\Lambda^3}- \frac{8}{3}\frac{x_0^4}{\Lambda^4}. \nonumber\end{aligned}$$ Substitution of Eq. (\[eq8\]) into Eq. (\[eq6\]) with $A_1=A_2=A$ leads to the result $$F^{lat}(a,x_0)=8|F_0(a)|\frac{A^2}{a\Lambda} \left(2\frac{x_0}{\Lambda}-1\right) \left[1+10\frac{A^2}{a^2}\left(1-2\frac{x_0}{\Lambda} +2\frac{x_0^2}{\Lambda^2}\right)\right]. \label{eq9}$$ As an illustration, Fig. 2,a shows the dependence of $F^{lat}/|F_0|$ on $x_0/\Lambda$ computed by Eq. (\[eq9\]) with the typical values of parameters $A/a=0.3$ and $a/\Lambda=0.2$. From this figure it is seen that the lateral force is still symmetric and the points of unstable equilibrium ($x_0/\Lambda=0.5,\,1.5,\,\ldots$) are exactly in the middle between the points of stable equilibrium ($x_0/\Lambda=0,\,1,\,2,\,\ldots$). Like in the case of the sinusoidal corrugations, the lateral force is negative over one half of the period and positive over another half. The magnitudes of maximum and minimum values of the force are equal. At the same time, the case of saw toothed structures is different from sinusoidal corrugations because here the extremum values of the lateral force are achieved near the points of stable equilibrium (where the force is discontinuous). The points of stable equilibrium in this configuration are especially stable. Even small deviation from the stable equilibrium (where the value of the lateral force is taken equal to zero, i.e., half a sum of the limiting values from the left and from the right) leads to a large lateral force, restoring the state of equilibrium. Thus, the plates with saw toothed corrugations could be used in microdevices where the lateral displacements of the elements should be avoided (e.g. in devices such as micromirrors, microgears, micropumps, microsensors and microvalves). Now we consider even more asymmetric longitudinal corrugations on the lower plate allowing to obtain different magnitudes for the maximum and minimum values of the lateral Casimir force. On the upper plate the same corrugations as in the previous example are preserved. The profiles are shown in Fig. 1,b. The new parameter $\Delta=l_x/\Lambda$ characterizes the extent of asymmetry. If $l_x=0$ ($\Delta=0$), the profiles in Fig. 1,b coincide with those in Fig. 1,a. The function $f_1$, describing the lower plate, can be presented as (within one period) $$f_1(x)=\left\{ {\begin{array}{rr} -\frac{1-\Delta}{1+\Delta}, & 0<x\leq l_x, \\ \frac{2}{1-\Delta^2}\frac{x}{\Lambda}- \frac{1+\Delta^2}{1-\Delta^2},& l_x<x\leq\Lambda. \end{array}} \right. \label{eq10}$$ The function $f_2$ for the upper plate is given by Eq. (\[eq7\]). Note that according to Eq. (\[eq2\]) both functions have zero mean values over the period. The expression for the lateral Casimir force is obtained by calculating the matrix elements from Eq. (\[eq6\]) over the period (0,$\Lambda$). Here the result takes a different form depending on whether $x_0\leq l_x$ or $x_0\geq l_x$. As before, we consider $A_1=A_2$. Then for $x_0\leq l_x$ one obtains $$\begin{aligned} && F^{lat}(a,x_0)=-8|f_0(a)|\,\frac{A^2}{a\lambda}\, \frac{1-\Delta}{1+\Delta} \left\{ \vphantom{\left[\frac{1+5\Delta^2+4\Delta^3+ \Delta^4}{(1+\Delta)^2}\right]} 1+\frac{10}{3}\frac{A}{a}\frac{1}{1+\Delta} \left[\Delta(3+\Delta)-3\frac{x_0}{\Lambda}(1+\Delta)\right]\right. \nonumber \\ &&\phantom{aaa} \left. +10\frac{A^2}{a^2}\left[\frac{1+5\Delta^2+4\Delta^3+ \Delta^4}{(1+\Delta)^2}-4\frac{x_0}{\Lambda} \frac{\Delta(3+\Delta)}{1+\Delta}+6\frac{x_0^2}{\Lambda^2} \right]\right\}. \label{eq11}\end{aligned}$$ For large phase shifts $x_0\geq l_x$ the expression for the lateral Casimir force is a bit more complicated $$F^{lat}(a,x_0)=8|f_0(a)|\frac{A^2}{a\lambda}\frac{1}{1-\Delta^2} \left(2\frac{x_0}{\Lambda}-1-\Delta^2\right) \left(1+\frac{10}{3}\frac{A}{a}X_1+ 10\frac{A^2}{a^2}X_2\right), \label{eq12}$$ where the coefficients $X_{1,2}$ are given by $$\begin{aligned} && X_1=-\frac{\Delta^2\left[2-3\Delta+3\Delta^2+\Delta^3- 3\left(1+\Delta^2\right)x_0/\Lambda+ 3x_0^2/\Lambda^2\right]}{\left(1-\Delta^2\right)\left(1+ \Delta^2-2x_0/\Lambda\right)}, \label{eq13} \\ && X_2=\frac{1}{\left(1-\Delta^2\right)^2\left(1+ \Delta^2-2x_0/\Lambda\right)}\left[ \vphantom{\frac{x_0^2}{\Lambda^2}\left(1+\Delta^6\right)} 1-\Delta^2+10\Delta^4-12\Delta^5+\Delta^6\right. \nonumber \\ && \phantom{aaa} +4\Delta^7+\Delta^8-4\frac{x_0}{\Lambda}\left( 1-\Delta^2+3\Delta^3-4\Delta^5+3\Delta^6+\Delta^7\right) \nonumber \\ && \phantom{aaa} \left. +6\frac{x_0^2}{\Lambda^2}\left(1+\Delta^6\right) -4\frac{x_0^3}{\Lambda^3}\left(1-\Delta^2+\Delta^4\right) \right]. \nonumber\end{aligned}$$ Direct calculation shows that for $x_0=l_x$ Eqs. (\[eq11\]) and (\[eq12\]) lead to one and the same result. Furthermore, for $\Delta=0$ one obtains $X_1=0$, $X_2=1-2(x_0/\Lambda)+2(x_0/\Lambda)^2$, so that Eq. (\[eq12\]) coincides with Eq. (\[eq9\]) as required (because the profiles in Figs. 1,a,b coincide when $\Delta=0$). The coordinate of the point of unstable equilibrium, where $F^{lat}(a,{\tilde{x}}_0)=0$, is found from Eq. (\[eq12\]) $${\tilde{x}}_0=\frac{\Lambda\left(1+\Delta^2\right)}{2}. \label{eq14}$$ For example, in Fig. 2,b one period of the relative lateral Casimir force $F^{lat}/|F_0|$ is plotted as a function of $x_0/\Lambda$ for $\Delta=1/2$ ($l_x=0.5\Lambda$), $A/a=0.3$ and $a/\Lambda=0.2$. According to Eq. (\[eq14\]), the point of unstable equilibrium is ${\tilde{x}}_0=5\Lambda/8$, i.e. it is asymmetric being shifted from the middle of a period. It is seen also that the magnitudes of maximum and minimum values of the lateral Casimir force in this case differ by 1.9. Needless to say, the work done by the lateral force over one period is equal to zero as it should be in any adiabatic process. If $\Delta$ is further increased, the position of the points of unstable equilibrium will be even more shifted to the right boundary of the period with an increase of the ratio of magnitudes of the maximum to minimum lateral forces. In the above the method of pairwise summation and perturbation theory in relative distortion amplitudes were applied to obtain the general expression for the lateral Casimir force between metallic plates with longitudinal corrugations of arbitrary profile. The pairwise summation works well when the corrugation period is several times larger than the separation between plates. The obtained expressions were used in the case of asymmetric saw tooth like structures which lack the right-left symmetry of sinusoidal corrugations. It was shown that with the proper choice of the corrugation shape and parameters, it is possible to change not only the magnitude of the lateral Casimir force, but make it asymmetric and affect the character of points of equilibrium. In particular, the maximum to minimum force ratio can be increased by several times. This opens new opportunities to control the lateral Casimir force for the diversified applications in microelectromechanical devices based entirely on the vacuum fluctuation properties of quantum electrodynamics. This work was supported by the National Institute of Standards and Technology, through a Precision Measurement Grant and the University of California and Los Alamos National Laboratory through the LANL-CARE program. G.L.K. and V.M.M. were also partially supported by CNPq and Finep (Brazil). [99]{} H. B. G. Casimir, [ Proc. K. Ned. Akad. Wet.]{} [**51**]{}, 793 (1948). P. W. Milonni, [*The Quantum Vacuum*]{} (Academic Press, San Diego, 1994). V. M. Mostepanenko and N. N. Trunov, [*The Casimir Effect and its Applications*]{} (Clarendon Press, Oxford, 1997). K. A. Milton, [*The Casimir Effect*]{} (World Scientific, Singapore, 2001). M. Kardar and R. Golestanian, Rev. Mod. Phys. [**71**]{}, 1233 (1999). M. Bordag, U. Mohideen, and V. M. Mostepanenko, [ Phys. Rep.]{} [**353**]{}, 1 (2001). S. K. Lamoreaux, [ Phys. Rev. Lett.]{} [**78**]{}, 5 (1997). U. Mohideen and A. Roy, [ Phys. Rev. Lett.]{} [**81**]{}, 4549 (1998); G. L. Klimchitskaya, A. Roy, U. Mohideen, and V. M. Mostepanenko, [ Phys. Rev. A]{} [**60**]{}, 3487 (1999). A. Roy and U. Mohideen, [ Phys. Rev. Lett.]{} [**82**]{}, 4380 (1999). A. Roy, C.-Y. Lin, and U. Mohideen, [ Phys. Rev. D]{} [**60**]{}, 111101(R) (1999). B. W. Harris, F. Chen, and U. Mohideen, Phys. Rev. A [**62**]{}, 052109 (2000). T. Ederth, Phys. Rev. A [**62**]{}, 062104 (2000). G. Bressi, G. Carugno, R. Onofrio, and G. Ruoso, Phys. Rev. Lett. [**88**]{}, 041804 (2002). R. S. Decca, D. López, E. Fischbach, and D. E. Krause, [Phys. Rev. Lett.]{} [**91**]{}, 050402 (2003). R. S. Decca, E. Fischbach, G. L. Klimchitskaya, D. E. Krause, D. López, and V. M. Mostepanenko, [Phys. Rev. D]{} [**68**]{}, 116003 (2003). M. Bordag, B. Geyer, G. L. Klimchitskaya, and V. M. Mostepanenko, [Phys. Rev. D]{} [**58**]{}, 075003 (1998); [**60**]{}, 055004 (1999); [**62**]{}, 011701(R) (2000). V. M. Mostepanenko and M. Novello, [Phys. Rev. D]{} [**63**]{}, 115003 (2001). E. Fischbach, D. E. Krause, V. M. Mostepanenko, and M. Novello, [Phys. Rev. D]{} [**64**]{}, 075010 (2001). G. L. Klimchitskaya and U. Mohideen, Int. J. Mod. Phys. A [**17**]{}, 4143 (2002). H. B. Chan, V. A. Aksyuk, R. N. Kleiman, D. J. Bishop, and F. Capasso, Science [**291**]{}, 1941 (2001); Phys. Rev. Lett. [**87**]{}, 211801 (2001). E. Buks and M. L. Roukes, Phys. Rev. B [**63**]{}, 033402 (2001). Yu. S. Barash, Izv. Vuzov. Ser. Radiofiz. [**16**]{}, 1086 (1973) \[Sov. Radiophys. [**16**]{}, 945 (1973)\]. S. J. van Enk, [Phys. Rev. A]{} [**52**]{}, 2569 (1995). R. Golestanian and M. Kardar, [Phys. Rev. Lett.]{} [**78**]{}, 3421 (1997); [Phys. Rev. A]{} [**58**]{}, 1713 (1998). T. Emig, A. Hanke, R. Golestanian, and M. Kardar, Phys. Rev. Lett. [**87**]{}, 260402 (2001); Phys. Rev. A [**67**]{}, 022114 (2003). F. Chen, U. Mohideen, G. L. Klimchitskaya, and V. M. Mostepanenko, Phys. Rev. Lett. [**88**]{}, 101801 (2002); Phys. Rev. A [**66**]{}, 032113 (2002). M. Bordag, G. L. Klimchitskaya, and V. M. Mostepanenko, [Int. J. Mod. Phys. A]{} [**10**]{}, 2661 (1995). C. Genet, A. Lambrecht, P. Maia Neto, and S. Reynaud, Europhys. Lett. [**62**]{}, 484 (2003).  ![ The relative lateral Casimir force as a function of a relative phase shift for the corrugations of Fig. 1,a and Fig. 1,b \[(a), (b), respectively\]. ](figL-2)

ArXiv

[**** ]{}\ Vasiliki Liakoni^1\*¶^, Alireza Modirshanechi^1\*¶^, Wulfram Gerstner^1^, Johanni Brea^1^\ **1** EPFL, School of Computer and Communication Sciences and School of Life Sciences, Lausanne, Switzerland\ ¶These authors contributed equally to this work. \* [email protected] (VL) and [email protected] (AM) Abstract {#abstract .unnumbered} ======== Surprise-based learning allows agents to rapidly adapt to non-stationary stochastic environments characterized by stationary periods separated by sudden changes. We show that exact Bayesian inference in a hierarchical model gives rise to a surprise-modulated trade-off between forgetting old observations and integrating them with the new ones. The modulation depends on a probability ratio, called the “Bayes Factor Surprise” that tests the prior belief against the current belief. We demonstrate that in several existing approximate algorithms the Bayes Factor Surprise modulates the rate of adaptation to new observations. We derive three novel surprised-based algorithms, one in the family of particle filters, one in the family of variational learning, and the other in the family of message passing, that are biologically plausible, have constant scaling in observation sequence length and particularly simple update dynamics for any distribution in the exponential family. Empirical results show that these surprise-based algorithms estimate parameters better than alternative approximate approaches and reach levels of performance comparable to computationally more expensive algorithms. The Bayes Factor Surprise is related to but different from Shannon Surprise. In two hypothetical experiments, we make testable predictions for physiological or behavioral indicators that dissociate the Bayes Factor Surprise from Shannon Surprise. The theoretical insight of casting various approaches as surprise-based learning, as well as the proposed online algorithms, may be applied to the analysis of animal and human behavior, and to reinforcement learning in non-stationary environments. Introduction {#intro .unnumbered} ============ Animals, humans, and similarly reinforcement learning agents may safely assume that the world is stochastic and stationary during some intervals of time interrupted by change points. The position of leafs on a tree, a stock market index, or the time it takes to travel from A to B in a crowded city is often well captured by stationary stochastic processes for extended periods of time. Then sudden changes may happen, such that the distribution of leaf positions becomes different due to a storm, the stock market index is affected by the enforcement of a new law, or a blocked road causes additional traffic jams. The violation of an agent’s expectation caused by such sudden changes is perceived by the agent as surprise, which can be seen as a measure of how much the agent’s current belief differs from reality. Surprise, with its physiological manifestations in pupil dilation [@preuschoff2011pupil; @nassar2012rational] and EEG signals [@modirshanechi2019trial; @ostwald2012evidence; @mars2008trial], is believed to modulate learning, potentially through the release of specific neurotransmitters [@angela2005uncertainty; @gerstner2018eligibility], so as to allow animals and humans to adapt quickly to sudden changes. The quick adaptation to novel situations has been demonstrated in a variety of learning experiments [@nassar2012rational; @nassar2010approximately; @behrens2007learning; @angela2005uncertainty; @glaze2015normative; @heilbron2019confidence]. The bulk of computational work on surprise-based learning can be separated into two groups. Studies in the field of computational neuroscience have focused on biological plausibility with little emphasis on the accuracy of learning [@nassar2012rational; @angela2005uncertainty; @nassar2010approximately; @faraji2018balancing; @friston2017active; @schwartenbeck2013exploration; @friston2010free; @behrens2007learning; @bogacz2017tutorial], whereas exact and approximate Bayesian online methods [@adams2007bayesian; @fearnhead2007line] for change point detection and parameter estimation have been developed without any focus on biological plausibility [@aminikhanghahi2017survey; @wilson2010bayesian; @cummings2018differentially; @lin2017sharp; @masegosa2017bayesian]. In this work, we take a top-down approach to surprise-based learning. We start with a generative model of change points similar to the one that has been the starting point of multiple experiments [@nassar2012rational; @nassar2010approximately; @behrens2007learning; @angela2005uncertainty; @glaze2015normative; @heilbron2019confidence; @findling2019imprecise]. We demonstrate that Bayesian inference on such a generative model can be interpreted as modulation of learning by surprise; we show that this modulation leads to a natural definition of surprise which is different, but closely related to Shannon Surprise [@shannon1948mathematical]. Moreover, we derive three novel approximate online algorithms with update rules that inherit the surprise-modulated adaptation rate of exact Bayesian inference. The overall goal of the present study is to give a Bayesian interpretation for surprise-based learning in the brain, and to find approximate methods that are computationally efficient and biologically plausible while maintaining the learning accuracy at a high level. As a by-product, our approach provides theoretical insights on commonalities and differences among existing surprise-based and approximate Bayesian approaches. Importantly, our approach makes specific experimental predictions. In the Results section, we first introduce the generative model, and qualitatively summarize our contributions. We then present our surprise-based interpretation of Bayesian inference and our three approximate algorithms. Next, we use simulations to compare our algorithms with existing ones on two different tasks inspired by and closely related to real experiments [@nassar2010approximately; @nassar2012rational; @behrens2007learning; @mars2008trial; @ostwald2012evidence]. At the end of the results section, we formalize a few experimentally testable predictions of our theory and illustrate them with simulations. A brief review of related studies as well as a few directions for further work are supplied in the Discussion section. Results {#Results .unnumbered} ======= In order to study learning in an environment that exhibits occasional and abrupt changes, we consider a hierarchical generative model (A) in discrete time, similar to existing model environments [@nassar2012rational; @nassar2010approximately; @behrens2007learning; @angela2005uncertainty]. At each time point $t$, the observation $Y_t = y_t$ comes from a probability distribution $P_Y(y_t|\theta_t)$ with parameter $\Theta_t = \theta_t$, where both $y_t$ and $\theta_t$ can be multi-dimensional (we indicate random variables by capital letters, and values by small letters). Abrupt changes of the environment correspond to sudden changes of the parameter $\Theta_t$. At every time $t$, there is a change probability $p_c \in (0,1)$ for the parameter $\Theta_t$ to be drawn from its prior distribution $\pi^{(0)}$ independently of its previous value, and a probability $1 - p_c$ to stay the same as $\Theta_{t-1}$. A change at time $t$ is specified by the event $\Delta H_t = 1$; otherwise $\Delta H_t = 0$ (see A and Methods for details). Given a sequence of observations $Y_{1:t} = y_{1:t} \equiv (y_1, \ldots, y_t)$, the *agent’s belief* $\pi^{(t)}(\theta)$ about the parameter $\theta$ at time $t$ is defined as the posterior probability distribution $\textbf{P}(\Theta_t=\theta|Y_{1:t} = y_{1:t})$ of the parameter $\Theta_t$. In the online learning setting studied here, the agent’s goal is to update the belief $\pi^{(t)}(\theta)$ to the new belief $\pi^{(t+1)}(\theta)$, or an approximation thereof, upon observing $Y_{t+1} = y_{t+1}$. A simplified real-world example of such an environment is illustrated in B. Imagine that every day a friend of yours meets you at the coffee shop (B left). The time of arrival of your friend (i.e. $Y_t$) exhibits some variability, due to various sources of stochasticity (e.g. traffic and your friend’s daily workload), but it has a stable average over time (i.e. $\Theta_t$). However, if during two or three weeks the road is blocked due to repairs (B right), your friend arrives later since he has to take a detour. The start of the road block is translated to $\Delta H_{t+1} = 1$ in our framework, and the sudden change in the arrival time of your friend to a sudden change from $\Theta_t = \theta$ to $\Theta_{t+1} = \theta'$. Even without any explicit discussion with your friend about this situation and only by observing his or her arrival time, you can notice the abrupt change and hence adapt your schedule to the new situation. In this section, we go from theoretical results over algorithms to simulations and experimental predictions. Qualitative summary of two main results {#contrib .unnumbered} --------------------------------------- Before we present our results in detail, let us summarize our two main findings. First, exact Bayesian inference on the generative model discussed above (see and Methods) leads to an explicit trade-off between (i) integrating a new observation $y^{new}$ with the old belief $\pi^{\text{old}}$ into a distribution $\pi^{\text{integration}}$ and (ii) forgetting the past observations, so as to restart with the belief $\pi^{\text{reset}}$ which relies only on the new observation and the prior $\pi^{(0)}$ $$\label{Eq:tradeoff} \pi^{\text{new}}(\theta) = (1 - \gamma) \, \pi^{\text{integration}}(\theta| y^{\text{new}}, \pi^{\text{old}}) + \gamma \, \pi^{\text{reset}}(\theta | y^{\text{new}}, \pi^{(0)}) .$$ This trade-off is governed by a *surprise-modulated adaptation rate* $$\label{Eq:Gamma_def} \gamma ( \text{S}, m ) = \frac{m \text{S}}{1 + m \text{S}} \in [0,1]\, ,$$ where $\text{S} \geq 0$ can be interpreted as the surprise of the most recent observation, and $m \geq 0 $ is a parameter controlling the effect of surprise on learning. The exact definitions of $\pi^{\text{integration}}$, $\pi^{\text{reset}}$, and $\text{S}$ will be given in the next section; the first result is that a split as in with a weighting factor (“adaptation rate” $\gamma$) as in is exact and always possible for the class of environments defined by our hierarchical generative model. Second, three novel approximate algorithms inherit the surprise-modulated adaptation rate from the exact Bayesian approach, i.e. and . The first algorithm uses the ideas of particle filtering [@gordon1993novel] for an efficient approximation for our hierarchical generative model. We refer to our approximate algorithm as Particle Filtering with $N$ particles and abbreviate it by pf$N$ (see Algo. 1). The second algorithm adapts an earlier algorithm of surprise minimization learning (SMiLe, [@faraji2018balancing]) to variational learning. We refer to our novel algorithm as Variational SMiLe and abbreviate it by VarSMiLe (see Algo. 2). The third algorithm is based on message passing with a finite number of messages $N$ [@adams2007bayesian]. We refer to this algorithm as MP$N$ (see Algo. 3). All algorithms are computationally efficient and biologically plausible, since Particle Filtering has a neuronal implementation [@kutschireiter2017nonlinear; @shi2009neural; @huang2014neurons; @legenstein2014ensembles], MP$N$ can be seen as a greedy version of pf$N$ without sampling, and Variational SMiLe can be implemented by simple neo-Hebbian [@lisman2011neohebbian] update rules. Simulation results show that the performance of the three approximate algorithms is comparable to and more robust across environments than other state-of-the-art approximations. Online Bayesian inference modulated by surprise {#recbayesinf .unnumbered} ----------------------------------------------- According to the definition of the hierarchical generative model (see and Methods), the value $y_{t+1}$ of the observation at time $t+1$ depends only on the parameters $\theta_{t+1}$, and is (given $\theta_{t+1}$) independent of earlier observations and earlier parameter values. We exploit this Markovian property and update, using Bayes’ rule, the belief $\pi^{(t)}(\theta)\equiv \textbf{P}(\Theta_{t}=\theta|Y_{1:t} = y_{1:t})$ at time $t$ to the new belief at time $t+1$ $$\label{Eq:Belief_Def_2} \pi^{(t+1)}(\theta) = \frac{ P_Y(y_{t+1}|\theta) \textbf{P}(\Theta_{t+1}=\theta|Y_{1:t} = y_{1:t})}{\textbf{P}(Y_{t+1} = y_{t+1}|Y_{1:t} = y_{1:t})} \, .$$ So far, remains rather abstract. The aim of this section is to rewrite it in the form of . The first term in the numerator of is the likelihood of the current observation given the parameter $\Theta_{t+1}$ (specified by the time-invariant function $P_Y(y|\theta)$ in the definition of the generative model - see Methods), and the second term is the agent’s estimated probability distribution of $\Theta_{t+1}$ before observing $Y_{t+1} = y_{t+1}$. Since there is always the possibility of an abrupt change, the second term is not the agent’s previous belief $\pi^{(t)}$, but $\textbf{P}(\Theta_{t+1}=\theta|Y_{1:t} = y_{1:t}) = (1-p_c) \pi^{(t)}(\theta) + p_c \pi^{(0)}(\theta)$. As a result, it is possible to find a recursive formula for updating the belief. For the derivation of this recursive rule, we define the following terms. The probability of observing $Y_t = y_t$ with a belief $\pi^{(t')}$ is denoted as $$\label{Eq:p_y_t} P(y_t; \pi^{(t')}) = \int P_Y(y_t|\theta) \pi^{(t')}(\theta) d\theta \, ,$$ where $P_{Y}$ is the time-invariant likelihood function. Note that two particularly interesting cases of are $P(y_{t+1}; \pi^{(t)})$, i.e. the probability of a new observation $y_{t+1}$ with the current belief $\pi^{(t)}$, and $P(y_{t+1}; \pi^{(0)})$, i.e. the probability of a new observation $y_{t+1}$ with the prior belief $\pi^{(0)}$. The “Bayes Factor Surprise” $\textbf{S}_{\mathrm{BF}}$ of the observation $Y_{t+1} = y_{t+1}$ is defined as the ratio of the probability of observing $Y_{t+1} = y_{t+1}$ given $\Delta H_{t+1} = 1$ (i.e. when there is a change), to the probability of observing $Y_{t+1} = y_{t+1}$ given $\Delta H_{t+1} = 0$ (i.e. when there is no change), i.e. $$\label{Eq:S_GM} \textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)}) = \frac{P(y_{t+1}; \pi^{(0)})}{P(y_{t+1}; \pi^{(t)})}\, .$$ This definition of surprise measures how much more probable the current observation is under the naive prior $\pi^{(0)}$ relative to the current belief $\pi^{(t)}$ (see the Discussion section for further interpretation). This probability ratio is the Bayes factor [@efron2016computer] that tests the prior belief $\pi^{(0)}$ against the current belief $\pi^{(t)}$. We emphasize that our definition of surprise is not arbitrary, but essential in order to write the exact inference on the generative model in the compact recursive form indicated in and . Under the assumption of no change $\Delta H_{t+1} = 0$, and using the most recent belief $\pi^{(t)}$ as prior, the exact Bayesian update for $\pi^{(t+1)}$ is denoted as $$\label{Eq:pi_B} \pi_B^{(t+1)}(\theta) = \frac{P_Y(y_{t+1}|\theta) \pi^{(t)}(\theta) } {P(y_{t+1}; \pi^{(t)})}\, .$$ Note that $\pi_B^{(t+1)}(\theta)$ corresponds to the term $\pi^{\text{integration}}$ of ; it describes the incorporation of the new information into the current belief via Bayesian updating. Using the above definitions and , we have for the generative model of .A the following proposition. Exact Bayesian inference on the generative model is equivalent to the recursive update rule $$\label{Eq:Bayesian_Rec_Formula} \pi^{(t+1)}(\theta) = \Big(1 - \gamma \Big( \textbf{S}_{\mathrm{BF}}^{(t+1)}, \frac{p_c}{1-p_c} \Big) \Big) \pi^{(t+1)}_B(\theta) + \gamma \Big( \textbf{S}_{\mathrm{BF}}^{(t+1)}, \frac{p_c}{1-p_c} \Big) P(\theta|y_{t+1})\, ,$$ where $\textbf{S}_{\mathrm{BF}}^{(t+1)} = \textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)})$ is the Bayes Factor Surprise and $$\label{Eq:cond_prob_ThetaY} P(\theta|y_{t+1}) = \frac{P_Y(y_{t+1}|\theta) \pi^{(0)}(\theta)} {P(y_{t+1}; \pi^{(0)})}\,$$ is the posterior if we take $y_{t+1}$ as the only observation. Note that $P(\theta|y_{t+1})$ corresponds to $\pi^{reset}$ in . The adaptation rate $\gamma$ in is $\gamma_{t+1} = \gamma(\text{S}, m)$ as in , with $\text{S} = \textbf{S}_{\mathrm{BF}}^{(t+1)}$ as in and $m = \frac{p_c}{1-p_c}$. The recursive formula of shows an explicit trade-off between integrating the new sample with the old information and forgetting the previous observations. The weight $\gamma_{t+1}$ of this convex sum is modulated by the surprise $\textbf{S}_{\mathrm{BF}}^{(t+1)}$ caused by the new observation. Since the parameter of modulation $m$ is equal to $\frac{p_c}{1-p_c}$, the effect of surprise on learning increases when the environment is more volatile, i.e. when the change probability $p_c$ increases. Despite the simplicity of the recursive formula in , the updated belief $\pi^{(t+1)}$ is generally not in the same family of distributions as the previous belief $\pi^{(t)}$, e.g. the result of averaging two normal distributions is not a normal distribution. Hence it is in general impossible to find a simple and exact update rule for e.g. some sufficient statistic. In the following sections, we investigate three approximations (Algo. 1-3) that have simple update rules and finite memory demands, so that the updated belief remains tractable over a long sequence of observations. Particle Filtering (Algo. 1) {#pf .unnumbered} ---------------------------- The history of change points up to time $t$ is a binary sequence, e.g. $\Delta h_{1:t} = \{ 0,0,0,1,0,1,1 \}$, where the value $1$ indicates a change in the corresponding time step. An exact Bayesian update can be performed by marginalization over the history. As a result of this marginalization, the agent’s belief is $\pi^{(t+1)}(\theta) = \sum_{\Delta h_{1:t+1}} \textbf{P}(\theta | \Delta h_{1:t+1}, y_{1:t+1}) \textbf{P}(\Delta h_{1:t+1}| y_{1:t+1})$, where we dropped the explicit mentioning of the random variables, e.g. $Y_{1:t+1}$, and display only their values, e.g. $y_{1:t+1}$, to shorten notation. The first factor is simple to compute, because when $\Delta h_{1:t+1}$ is known, inference depends only on the observations after the last change point. However, since the computation of the term $\textbf{P}(\Delta h_{1:t+1}| y_{1:t+1})$ is difficult and the summation over all hidden states is computationally costly, we approximate in this section this term via particle filtering [@gordon1993novel], i.e. $$\label{Eq:pDh_pf_approx} \textbf{P}(\Delta h_{1:t+1} | y_{1:t+1}) \approx \sum_{i=1}^{N} w_{t+1}^{(i)} \delta (\Delta h_{1:t+1} - \Delta h_{1:t+1}^{(i)} ) \, ,$$ where $\{ \Delta h_{1:t+1}^{(i)} \}_{i=1}^N$ is a set of $N$ realizations (particles) drawn from a proposal distribution $Q(\Delta h_{1:t+1} | y_{1:t+1})$, and $\{ w_{t+1}^{(i)} \}_{i=1}^N$ are their corresponding weights at time $t+1$. Hence the approximated belief is $$\label{Eq:pi_pf_approx} {\hat{\pi}}^{(t+1)}(\theta) = \sum_{i=1}^{N} w_{t+1}^{(i)} {\hat{\pi}}_i^{(t+1)}(\theta) = \sum_{i=1}^{N} w_{t+1}^{(i)} \textbf{P}(\theta | \Delta h_{1:t+1}^{(i)}, y_{1:t+1})\, ,$$ where ${\hat{\pi}}_i^{(t+1)}(\theta)$ is the approximated belief corresponding to particle $i$. The update procedure includes two steps: (i) updating the weights, and (ii) sampling the new hidden state $\Delta h_{t+1}$ for each particle. The first step amounts to $$\begin{aligned} \label{Eq:wB} & w_{t+1}^{(i)} = (1-\gamma_{t+1})w_{B, t+1}^{(i)} + \gamma_{t+1}w_{t}^{(i)}\, ,\\ & w_{B, t+1}^{(i)} = \frac{P (y_{t+1}; {\hat{\pi}}_i^{(t)} ) }{P (y_{t+1}; {\hat{\pi}}^{(t)} ) } w_{t}^{(i)}\, , \end{aligned}$$ where $\gamma_{t+1} = \gamma \big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)}) , m \big)$ with $ m = \frac{p_c}{1-p_c}$ (cf. ), and $\{ w_{B, t+1}^{(i)} \}_{i=1}^N$ are the weights corresponding to the Bayesian update ${\hat{\pi}}_{B}^{(t+1)}$ of (see Methods). In the second step we sample each particle’s hidden state $\Delta h^{(i)}_{t+1}$ from a proposal distribution $Q(\Delta h_{t+1}^{(i)}| \Delta h_{1:t}^{(i)}, y_{1:t+1})$. To go from the sequence $\{ \Delta h_{1:t}^{(i)} \}_{i=1}^N$ to $\{ \Delta h_{1:t+1}^{(i)} \}_{i=1}^N$ we always keep the old sequence up to time $t$, and for each particle $i$, we add a new element $\Delta h_{t+1}^{(i)} \in \{ 0, 1 \}$ representing no change $\Delta h_{1:t+1}^{(i)} = [ \Delta h_{1:t}^{(i)}, 0 ]$ or change $\Delta h_{1:t+1}^{(i)} = [ \Delta h_{1:t}^{(i)}, 1 ]$. Note, however, that it is not needed to keep the whole sequences $\Delta h_{1:t+1}^{(i)}$ in memory, but instead one can use $\Delta h_{t+1}^{(i)}$ to update ${\hat{\pi}}_{i}^{(t)}$ to ${\hat{\pi}}_{i}^{(t+1)}$. The change probability depends on the new observation $y_{t+1}$ and is given by (see Methods) $$\begin{aligned} \label{Eq:proposal_final} Q(\Delta h_{t+1}^{(i)} = 1| \Delta h_{1:t}^{(i)}, y_{1:t+1}) = \gamma \Big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}_i^{(t)}), \frac{p_c}{1-p_c} \Big)\, . \end{aligned}$$ Interestingly, the above formulas are in the same spirit as . For the weight update there is a trade-off between an exact Bayesian update and keeping the value of the previous time step, controlled by a adaptation rate modulated exactly in the same way as in . Note that in contrast to , the trade-off for the particles’ weights is not between forgetting and integrating, but between maintaining the previous knowledge and integrating. However, the change probability () for sampling is an increasing function of surprise. As a result, although the weights are updated less for surprising events, a higher surprise causes a higher probability for change, indicated by $\Delta h_{t+1}^{(i)} = 1$. This is eventually identical to forgetting, since for a particle $i$ with $\Delta h_{t+1}^{(i)} = 1$, the associated belief is ${\hat{\pi}}_i^{(t+1)} = P(\theta|y_{t+1})$, which is equivalent to a reset of the belief as in . In order to avoid degeneracy of the weights, we employed the Sequential Importance Resampling algorithm [@gordon1993novel; @sarkka2013bayesian] in our implementation of particle filtering (see Methods). Equations \[Eq:pi\_pf\_approx\] and \[Eq:wB\] can be applied to the case where the likelihood function $P_Y(y|\theta)$ is in the exponential family and $\pi^{(0)}$ is its conjugate prior. The resulting algorithm (Algorithm \[Alg:PF\]) has a particularly simple update rule for the belief parameters (see Methods for details). Specify $P_Y(y|\theta)$, $\textbf{P}_{\pi}\big(\Theta = \theta; \chi, \nu \big)$, and $\phi(y)$ where $P_Y \in \{\text{exponential family} \}$, $\textbf{P}_{\pi} \in \{\text{conjugate priors of } P_Y\}$, and $\phi(y)$ is the sufficient statistic. Specify $m=p_c/(1-p_c)$, $N$, and $N_{\text{thrs}}$ Initialize $\chi^{(0)}$, $\nu^{(0)}$, $w^{(i)}_{0}$ $\forall i \in \{1 ... N \}$, and $t \gets 0$. Observe $y_{t+1}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}_i^{(t)})$ using with $\chi_i^{(t)}$, $\nu_i^{(t)}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)}) = \big[ \sum_{i=1}^{N}w^{(i)}_{t} [\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}_i^{(t)})]^{-1}\big]^{-1}$ [ ]{} Compute $\gamma_{t+1} = \gamma \big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)}) , m \big)$ [ ]{} Compute the Bayesian weight $w_{B, t+1}^{(i)}$ using [ ]{} Sample [ ]{} $N_{\text{eff}} \gets (\sum_{i=1}^{N} w_{t+1}^{(i)^2})^{-1}$ If $N_{\text{eff}} \leq N_{\text{thrs}}$: resample [ ]{} $\chi_i^{(t+1)} \gets \chi_i^{(t)} + \phi(y_{t+1})$ and $\nu_i^{(t+1)} \gets \nu_i^{(t)} + 1$ $\chi_i^{(t+1)} \gets \chi^{(0)} + \phi(y_{t+1})$ and $\nu_i^{(t+1)} \gets \nu^{(0)}+ 1$ [ ]{} ${\hat{\pi}}^{(t+1)}(\theta) = \sum_{i=1}^{N} w_{t+1}^{(i)} \textbf{P}_{\pi}\big(\Theta = \theta; \chi_i^{(t+1)}, \nu_i^{(t+1)}\big)$ [ ]{} $t \gets t+1$ Variational SMiLe Rule (Algo. 2) {#var_smiles .unnumbered} -------------------------------- In order to keep the updated belief in the same family as the previous beliefs, one possible, but heuristic, approximation consists in applying the weighted averaging of the exact Bayesian update rule () to the logarithm of the beliefs rather than the beliefs themselves, i.e. $$\label{Eq:Var_SMiLe_Log_Rec_2nd_Ver} \text{log} \big( {\hat{\pi}}^{(t+1)}(\theta) \big) = (1-\gamma_{t+1}) \text{ log} \big( {\hat{\pi}}_B^{(t+1)}(\theta) \big) + \gamma_{t+1} \text{ log} \big( P(\theta|y_{t+1}) \big) + \text{Const.}\, ,$$ where $\gamma_{t+1} = \gamma \big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)}) , m \big)$ is given by with a free parameter $m>0$ which can be tuned to each environment. By doing so, we still have the explicit trade-off between two terms as in ; yet an advantageous consequence of averaging over logarithms is that, if the likelihood function $P_Y(y|\theta)$ is in the exponential family of distributions, and if the initial belief $\pi^{(0)}$ is its conjugate prior, then ${\hat{\pi}}^{(t+1)}$ and $\pi^{(0)}$ are members of the same family. In this particular case, we arrive at a simple update rule for the parameters of ${\hat{\pi}}^{(t+1)}$ (see Algorithm \[Alg:Var\_SMiLe\] for pseudocode and Methods for details). One way to interpret the update rule of is to rewrite it as the solution of a constraint optimization problem. The new belief ${\hat{\pi}}^{(t+1)}$ is a variational approximation of the Bayesian update ${\hat{\pi}}_B^{(t+1)}$ (see Methods) $$\begin{aligned} \label{Eq:Var_SMiLe_Min_1_2nd_Ver} {\hat{\pi}}^{(t+1)}(\theta) &= \arg \min_q \textbf{D}_{KL} \big[ q(\theta) || {\hat{\pi}}_B^{(t+1)}(\theta) \big], \end{aligned}$$ with a family of functions $q(\theta)$ constrained by the Kullback-Leibler divergence $$\begin{aligned} \label{Eq:Var_SMiLe_Min_1_2nd_Ver_Const} \textbf{D}_{KL} \big[ q(\theta) || P(\theta|y_{t+1}) \big] \leq B_{t+1}\, , \end{aligned}$$ where the bound $B_{t+1} \in \big[ 0, \textbf{D}_{KL}[ {\hat{\pi}}_B^{(t+1)}(\theta) || P(\theta|y_{t+1}) ] \big]$ is a decreasing function of the surprise $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)})$ (see Methods for proof), and $P(\theta|y_{t+1})$ is given by . The optimization form of the update rule in equations and gives a geometrical intuition about the abstract form of : the belief is updated by balancing its KL divergence from the integrating belief (i.e. ${\hat{\pi}}_B^{(t+1)}(\theta)$) against its KL divergence from the belief after a reset and one-step integration of the last data point only (i.e. $P(\theta|y_{t+1})$). Because of the similarity of the constraint optimization problem in and to the Surprise Minimization Learning rule “SMiLe” [@faraji2018balancing], we call this algorithm “Variational Surprise Minimization Learning" rule, or in short “Variational SMiLe" rule. Similar methods have been developed in machine learning [@ozkan2013marginalized; @masegosa2017bayesian]. Our variational method, and particularly its surprise-modulated adaptation rate, is complementary to these works, since the earlier studies made different assumptions about the generative model and used different approaches for deriving the learning rule, e.g. a mean-field approximation combined with coordinate descent for minimization of the variational loss function (i.e. free energy) in [@masegosa2017bayesian]. Specify $P_Y(y|\theta)$, $\textbf{P}_{\pi}\big(\Theta = \theta; \chi, \nu \big)$, and $\phi(y)$ where $P_Y \in \{\text{exponential family} \}$, $\textbf{P}_{\pi} \in \{\text{conjugate priors of } P_Y\}$, and $\phi(y)$ is the sufficient statistic. Specify $m$. Initialize $\chi^{(0)}$, $\nu^{(0)}$, and $t \gets 0$. Observe $y_{t+1}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)})$ using [ ]{} Compute $\gamma_{t+1} = \gamma \big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)}), m \big)$ [ ]{} ${\hat{\pi}}^{(t+1)}(\theta) = \textbf{P}_{\pi}\big(\Theta = \theta; \chi^{(t+1)}, \nu^{(t+1)}\big)$ [ ]{} $t \gets t+1$ Message-Passing $N$ (Algo. 3) {#sec:adam_mac_variants .unnumbered} ----------------------------- For a hierarchical generative model similar to ours, a message passing algorithm has been used to perform exact Bayesian inference [@adams2007bayesian]. Let us count the number of change points up to time $t$ by the random variable $H_t = \sum_{k=1}^{t} \Delta H_k$. Following the idea of [@adams2007bayesian], we define the random variable $R_t = \text{max} \{ n \in \mathbb{N} : H_{t-n+1} = H_{t} \}$ in order to describe the time since the last change point. The exact Bayesian expression for $\pi^{(t)}(\theta)$ can be written as $$\label{Eq:adam_mac_MPA} \begin{aligned} \pi^{(t)}(\theta) &= \sum_{r_t=1}^{t} \textbf{P}(R_t = r_t |Y_{1:t}=y_{1:t}) \textbf{P}(\Theta_{t+1} = \theta | R_t = r_t, Y_{1:t}=y_{1:t}). \end{aligned}$$ This equation is similar to Particle Filtering () with a number of particles equal to $t$. Upon each observation of a new sample $Y_{t+1}$, a new particle is generated and added to the set of particles, modeling the possibility of a change point occurring at $t+1$. To have a formulation similar to the one of Particle Filtering we rewrite the belief as $$\begin{aligned} \pi^{(t)}(\theta) &= \sum_{k=0}^{t-1} w^{(k)}_t \textbf{P}(\Theta_{t+1} = \theta | R_t = t-k, Y_{1:t}=y_{1:t}), \end{aligned}$$ where $w^{(k)}_t = \textbf{P}(R_t = t-k |Y_{1:t}=y_{1:t})$ is the weight of the particle $k$ at time $t$. To update the belief after observing $Y_{t+1} = y_{t+1}$, we use the exact Bayesian recursive formula (). The resulting update rule (see Methods for derivation) for the weights for $ 0 \leq k \leq t-1$ is $$\begin{aligned} \label{Eq:MP_w_oldparticles} & w^{(k)}_{t+1} = (1-\gamma_{t+1}) w^{(k)}_{B,t+1} = (1-\gamma_{t+1}) \frac{ P(y_{t+1}; \pi_k^{(t)}) } { P(y_{t+1}; \pi^{(t)}) } w^{(k)}_t, \end{aligned}$$ and for the newly added particle $t$ $$\begin{aligned} \label{Eq:MP_w_newparticle} & w^{(t)}_{t+1} = \gamma_{t+1}, \end{aligned}$$ where $\gamma_{t+1} = \gamma \big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)}), \frac{p_c}{1-p_c} \big)$ (cf. ). The updates of the equations and are essentially the same as the ones of Particle Filtering, and entail the same surprise modulation and the same trade-off. The only difference is that, while in Particle Filtering the trade-off between integration and reset is accomplished via sampling, in the message passing algorithm, it is accomplished by adding at each time step a new particle with weight $\gamma_{t+1}$. The computational complexity and memory requirements of the complete message passing algorithm increase linearly with time $t$ [@adams2007bayesian]. To deal with this issue and have the same computation and memory demands as Particle Filtering, we implemented a message passing algorithm of the form of , but with a fixed number $N$ of particles, chosen as those with the highest weights $w^{(k)}_{t}$. We call our modification of the message passing algorithm of [@adams2007bayesian] “Message Passing $N$” and abbreviate it by “MP$N$”. To deal with the computational complexity and memory requirements, one may alternatively keep only the particles with weights greater than a cut-off threshold [@adams2007bayesian]. However, such a constant cut-off leads to a varying number (smaller or equal to $t$) of particles in time. Our approximation MP$N$ can therefore be seen as a variation of the algorithm in [@adams2007bayesian] with fixed number of particles $N$, and hence a variable cut-off threshold. The work of [@fearnhead2007line] follows the same principle as [@adams2007bayesian], but employs stratified resampling to eliminate particles with negligible weights, in order to reduce the total number of particles. Their resampling algorithm involves solving a complicated non-linear equation at each time, which makes it unsuitable for a biological implementation. In addition, we experienced that in some cases the small errors introduced in the resampling step of the algorithm of [@fearnhead2007line] accumulated and led to a worse performance than our MP$N$ algorithm which simply keeps the $N$ particles with the highest weight at each time step. For the case where the likelihood function $P_Y(y|\theta)$ is in the exponential family and $\pi^{(0)}$ is its conjugate prior, the resulting algorithm of MP$N$ has a simple update rule for the belief parameters (see Algorithm \[Alg:MPN\] and Methods for details). In our simulations, we also implemented the full message passing algorithm of [@adams2007bayesian] with an almost zero cut-off (machine precision), which we consider as our benchmark “Exact Bayes”, as well as the stratified optimal resampling algorithm of [@fearnhead2007line], called “SOR$N$”. Note that the Exact Bayes algorithm is equivalent to our MP$N$ algorithm when $N$ is equal to $t$ and not fixed in time, and our MP$N$ algorithm is a greedy version of SOR$N$. Specify $P_Y(y|\theta)$, $\textbf{P}_{\pi}\big(\Theta = \theta; \chi, \nu \big)$, and $\phi(y)$ where $P_Y \in \{\text{exponential family} \}$, $\textbf{P}_{\pi} \in \{\text{conjugate priors of } P_Y\}$, and $\phi(y)$ is the sufficient statistic. Specify $m=p_c/(1-p_c)$, and $N$. Initialize $\chi_1^{(0)}$, $\nu_1^{(0)}$, $w^{(1)}_{0} = 1$ and $t \gets 0$. Until $N=t$, do the exact message passing algorithm of and Observe $y_{t+1}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}, {\hat{\pi}}_i^{(t)})$ using with $\chi_i^{(t)}$, $\nu_i^{(t)}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}, {\hat{\pi}}^{(t)})$ as the weighted ($w^{(i)}_{t}$) harmonic mean of $\textbf{S}_{\mathrm{BF}}(y_{t+1}, {\hat{\pi}}_i^{(t)})$ [ ]{} Compute $\gamma_{t+1} = \gamma \big( \textbf{S}_{\mathrm{BF}}(y_{t+1}, {\hat{\pi}}^{(t)}) , m \big)$ [ ]{} Compute the Bayesian weight $w_{B, t+1}^{(i)}$ using [ ]{} [ ]{} $\chi_i^{(t+1)} \gets \chi_i^{(t)} + \phi(y_{t+1})$ and $\nu_i^{(t+1)} \gets \nu_i^{(t)} + 1$ $\chi_{N+1}^{(t+1)} \gets \chi^{(0)} + \phi(y_{t+1})$ and $\nu_{N+1}^{(t+1)} \gets \nu^{(0)}+ 1$ [ ]{} Keep the $N$ particles with highest weights among $w_{t+1}^{(1:N+1)}$, rename and normalize their weights [ ]{} ${\hat{\pi}}^{(t+1)}(\theta) = \sum_{i=1}^{N} w_{t+1}^{(i)} \textbf{P}_{\pi}\big(\Theta = \theta; \chi_i^{(t+1)}, \nu_i^{(t+1)}\big)$ [ ]{} $t \gets t+1$ Surprise-modulation as a framework for other algorithms {#surprise-modulation-as-a-framework-for-other-algorithms .unnumbered} ------------------------------------------------------- Other existing algorithms [@adams2007bayesian; @fearnhead2007line; @nassar2010approximately; @nassar2012rational; @faraji2018balancing] can also be formulated in the surprise-modulation framework of and (see Methods). Moreover, in order to allow for a transparent discussion and for fair comparisons in simulations we extended the algorithms of [@nassar2010approximately; @nassar2012rational] to a more general setting. Here we give a brief summary of the algorithms we considered. A detailed analysis is provided in subsection “Surprise-modulation as a framework for other algorithms” in the Methods. The algorithms of [@nassar2010approximately; @nassar2012rational] were originally designed for a Gaussian estimation task (see Simulations for details of the task) with a broad uniform prior. We extended them to the more general case of Gaussian tasks with Gaussian priors, and we call our extended versions Nas10$^{*}$ and Nas12$^{*}$ for [@nassar2010approximately] and [@nassar2012rational] respectively. Both algorithms have the same surprise-modulation as in and , and are closely related to Particle Filtering with a single particle (pf$1$) (see Methods). To summarize, the algorithms Exact Bayes and SOR$N$ come from the field of change point detection, and while the former has high memory demands, the latter has the same memory demands as our algorithms pf$N$ and MP$N$. The algorithms Nas10$^{*}$, Nas12$^{*}$, and SMiLe, on the other hand, come from the human learning literature and are more biologically oriented. Simulations {#simulations .unnumbered} ----------- With the goal of gaining a better understanding of different approximate algorithms, we evaluated the departure of their performance from the exact Bayesian algorithm in terms of mean squared error (MSE) of $\Theta_t$ (see Methods), on two tasks inspired by and closely related to real experiments [@nassar2010approximately; @nassar2012rational; @behrens2007learning; @mars2008trial; @ostwald2012evidence; @maheu2019brain]: a Gaussian and a Categorical estimation task. We compared our three novel algorithms VarSMiLe, Particle Filtering (pf$N$, where $N$ is the number of particles), and Message Passing with finite number of particles $N$ (MP$N$) to the online Bayesian Message Passing algorithm [@adams2007bayesian] (Exact Bayes), which yields the optimal solution with ${\hat{\Theta}}_t = {\hat{\Theta}}_t^{\text{Opt}}$. Furthermore, we included in the comparison the stratified optimal resampling algorithm [@fearnhead2007line] (SOR$N$, where $N$ is the number of particles), our variant of [@nassar2010approximately] (Nas10$^{*}$) and of [@nassar2012rational] (Nas12$^{*}$), the Surprise-Minimization Learning algorithm of [@faraji2018balancing] (SMiLe), as well as a simple Leaky Integrator (Leaky - see Methods). ### Gaussian estimation task {#gaussian-estimation-task .unnumbered} The task is a generalized version of the experiment of [@nassar2010approximately; @nassar2012rational]. The goal of the agent is to estimate the mean $\theta_{t} = \mu_{t}$ of observed samples, which are drawn from a Gaussian distribution with known variance $\sigma^2$, i.e. $y_{t+1}|\mu_{t+1} \sim \mathcal{N}(\mu_{t+1}, \sigma^2)$. The mean $\mu_{t+1}$ is itself drawn from a Gaussian distribution $\mu_{t+1} \sim \mathcal{N}(0, 1)$ whenever the environment changes. An example of the task can be seen in A. We simulated the task for all combinations of $\sigma \in \{ 0.1, 0.5, 1, 2, 5 \}$ and $p_c \in \{ 0.1, 0.05, 0.01, 0.005, 0.001, 0.0001 \}$. For each combination of $\sigma$ and $p_c$, we first tuned the free parameter of each algorithm, i.e. $m$ for SMiLe and Variational SMiLe, the leak parameter for the Leaky Integrator, and the $p_c$ of Nas10$^{*}$ and Nas12$^{*}$, by minimizing the MSE on three random initializations of the task. For the Particle Filter (pf$N$), the Exact Bayes, the MP$N$, and the SOR$N$ we empirically checked that the true $p_c$ of the environment was indeed the value that gave the best performance, and we used this value for the simulations. We evaluated the performance of the algorithms on ten different random task instances of $10^5$ steps each for $p_c \in \{ 0.1, 0.05, 0.01, 0.005 \}$ and $10^6$ steps each for $p_c \in \{ 0.001, 0.0001 \}$ (in order to sample more change points). Note that the parameter $\sigma$ is not estimated and its actual value is used by all algorithms except the Leaky Integrator. In B we show the $\textbf{MSE} [ {\hat{\Theta}}_t | R_t = n]$ in estimating the parameter after $n$ steps since the last change point, for each algorithm, computed over multiple changes, for two exemplar task settings. The Particle Filter with 20 particles (pf20), the VarSmiLe and the Nas12$^{*}$ have an overall performance very close to that of the Exact Bayes algorithm (i.e. $\textbf{MSE} [ {\hat{\Theta}}_t^{\text{Opt}} | R_t = n ]$), with much lower memory requirements. VarSMiLe sometimes slightly outperforms the other two early after an environmental change (B, right), but shows slightly higher error values at later phases. The MP$N$ algorithm is the closest one to the optimal solution (i.e. Exact Bayes) for low $\sigma$ (B, left), but its performance is much worse for the case of high $\sigma$ and low $p_c$ (B, right). For the Stratisfied Optimal Resampling (SOR) we observe a counter-intuitive behavior in the regime of low $\sigma$; the inclusion of more particles leads to worse performance (B, left). At higher $\sigma$ levels the performance of SOR20 is close to optimal and better than the MP20 in later time-steps. This may be due to the fact that the MP$N$ discards particles in a deterministic and greedy way (i.e. the one with the lowest weight), whereas for the SOR there is a component of randomness in the process of particle elimination, which may be important for environments with higher stochasticity. For the Leaky Integrator we observe a trade-off between good performance in the transient phase and the stationary phase; a fixed leak value cannot fulfill both requirements. The SMiLe rule, by construction, never narrows its belief ${\hat{\pi}}(\theta)$ below some minimal value, which allows it to have a low error immediately after a change, but leads later to high errors. Its performance deteriorates for higher $\sigma$ (B, right). The Nas10$^{*}$ performs well for low, but not for higher values of $\sigma$. Despite the fact that a Particle Filter with 1 particle (pf1) is in expectation similar to Nas10$^{*}$ and Nas12$^{*}$ (see Methods), it performs worse than these two algorithms on trial-by-trial measures. Still, it performs better than the MP1 and identically to the SOR1. In A, we have plotted the average of $\textbf{MSE} [ {\hat{\Theta}}_t^{\text{Opt}} ]$ of the Exact Bayes algorithm over the whole simulation time for each of the considered $\sigma$ and $p_c$ levels. The difference between the other algorithms and this benchmark is called $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ]$ (see Methods) and is plotted in C–F. All algorithms except for the SOR20 have lower average error values for low $\sigma$ and low $p_c$, than high $\sigma$ and high $p_c$. The Particle Filter pf20 and the Message Passing MP20 have the smallest difference from the optimal solution. The average error of MP20 is higher than that of pf20 for high $\sigma$ and low $p_c$, whereas pf20 is more robust across levels of environmental parameters. The worst case performance for pf20 is $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ] = 0.033$ for $\sigma = 5$ and $p_c = 0.0001$, and for SOR20 it is $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ] = 0.061$ for $\sigma = 0.1$ and $p_c = 0.1$. The difference between these two worst case scenarios is significant with ($p\text{-value}= 2.79 \times 10^{-6}$, two-sample t-test, 10 random seeds for each algorithm). Next in performance is the algorithm Nas12$^{*}$ and VarSmiLe. VarSMiLe exhibits its largest deviation from the optimal solution for high $\sigma$ and low $p_c$, but is still more resilient compared to the MP$N$ algorithms for this type of environments. Among the algorithms with only one unit of memory demands, ie. pf1, MP1, SOR1, VarSMiLe, SMiLe, Leaky, Nas10$^{*}$ and Nas12$^{*}$, the winners are VarSmiLe and Nas12$^{*}$. The SOR20 has low error overall, but unexpectedly high error for environmental settings that are presumably more relevant for biological agents (intervals of low stochasticity marked by abrupt changes). The simple Leaky Integrator performs well at low $\sigma$ and $p_c$ but deviates more from the optimal solution as these parameters increase (F). The SMiLe rule performs best at lower $\sigma$, i.e. in more deterministic environments. A summary graph, where we collect the $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ]$ across all levels of $\sigma$ and $p_c$, is shown in . We can see that pf20, Nas12$^{*}$, and VarSMiLe give the lowest worst case (lowest maximum value) $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ]$ and are statistically better than the other 8 algorithms (the errorbars indicate the standard error of the mean across the ten random task instances). ### Categorical estimation task {#sub:cat .unnumbered} The task is inspired by the experiments of [@behrens2007learning; @mars2008trial; @ostwald2012evidence; @maheu2019brain]. The goal of the agent is to estimate the occurrence probability of five possible states. Each observation $y_{t+1} \in \{ 1, ..., 5 \}$ is drawn from a categorical distribution with parameters $\theta_{t+1} = \boldsymbol{p}_{t+1}$, i.e. $y_{t+1}|\boldsymbol{p}_{t+1} \sim \text{Cat}(y_{t+1} ; \boldsymbol{p}_{t+1})$. When there is a change $\Delta H_{t+1} = 1$ in the environment, the parameters $\boldsymbol{p}_{t+1}$ are drawn from a Dirichlet distribution $\text{Dir}(s \cdot \boldsymbol{1})$, where $s \in (0, \infty)$ is the stochasticity parameter. An illustration of this task is depicted in A. We considered the combinations of stochasticity levels $s \in \{ 0.01, 0.1, 0.14, 0.25, 1, 2, 5 \}$ and change probability levels $p_c \in \{ 0.1, 0.05, 0.01, 0.005, 0.001, 0.0001 \}$. The algorithms of [@nassar2012rational; @nassar2010approximately] were specifically developed for a Gaussian estimation task and cannot be applied here. All other algorithms were first optimized for each combination of environmental parameters before an experiment starts, and then evaluated on ten different random task instances, for $10^5$ steps each for $p_c \in \{ 0.1, 0.05, 0.01, 0.005 \}$ and $10^6$ steps each for $p_c \in \{ 0.001, 0.0001 \}$. The parameter $s$ is not estimated and its actual value is used by all algorithms except the Leaky Integrator. The Particle Filter pf20, the MP20 and the SOR20 have a performance closest to that of Exact Bayes, i.e. the optimal solution (B). VarSMiLe is the next in the ranking, with a behavior after a change similar to the Gaussian task. pf20 performs better for $s>2$ and MP20 performs better for $s \leq 2$ (). For this task the biologically less plausible SOR20 is the winner in performance and it behaves most consistently across environmental parameters. Its worst case performance is $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ] = 8.16 \times 10^{-5}$ for $s = 2$ and $p_c = 0.01$, and the worst case performance for pf20 is $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ] = 0.0048$ for $s = 0.25$ and $p_c = 0.005$ ($p\text{-value}= 1.148 \times 10^{-12}$, two-sample t-test, 10 random seeds for each algorithm). For all the other algorithms, except for MP20, the highest deviations from the optimal solution are observed for medium stochasticity levels (B–F). When the environment is nearly deterministic (e.g. $s = 0.001$ so that the parameter vectors $\boldsymbol{p}_{t}$ have almost all mass concentrated in one component), or highly stochastic (e.g. $s > 1$ so that nearly uniform categorical distributions are likely to be sampled), these algorithms achieve higher performance, while the Particle Filter is the algorithm that is most resilient to extreme choices of the stochasticity parameter $s$. For VarSMiLe in particular, the lowest mean error is achieved for high $s$ and high $p_c$ or low $s$ and low $p_c$. A summary graph, with the $\Delta \textbf{MSE} [ {\hat{\Theta}}_t ]$ across all levels of $s$ and $p_c$, can be seen in . The algorithms with the lowest “worst case” are SOR20 and pf20. The top-4 algorithms SOR20, pf20, MP20 and VarSmiLe are significantly better than the others (the errorbars indicate the standard error of the mean across the ten random task instances), whereas MP$1$ and SMiLe have the largest error with a maximum at 0.53. ### Summary of simulation results {#summary-of-simulation-results .unnumbered} In summary, our simulation results of the two tasks collectively suggest that our Particle Filtering (pf$N$) and Message Passing (MP$N$) algorithms achieve a high level of performance, very close to the one of biologically less plausible algorithms with higher (Exact Bayes) and same (SOR$N$) memory demands. Moreover, their behavior is more consistent across tasks. Finally, among the algorithms with memory demands of one unit, VarSMiLe performs best. ### Robustness against suboptimal parameter choice {#robustness-against-suboptimal-parameter-choice .unnumbered} In all algorithms we considered, the environment’s hyper-parameters are assumed to be known. We can distinguish between two types of hyper-parameters in our generative model: 1. the parameters of the likelihood function (e.g. $\sigma$ in the Gaussian task), and 2. the $p_c$ and the parameters of the conjugate prior (e.g. $s$ in the Categorical task). Hyper-parameters of the first type can be added to the parameter vector $\theta$ and be inferred with the same algorithm. However, learning the second type of hyper-parameters is not straightforward. By assuming that these hyper-parameters are learned more slowly than $\theta$, one can fine-tune them after each $n$ (e.g. 10) change points, while change points can be detected by looking at the particles for the Particle Filter and at the peaks of surprise values for VarSMiLe. Other approaches to hyper-parameter estimation can be found in [@george2017principled; @liu2001combined; @doucet2003parameter; @wilson2010bayesian]. When the hyper-parameters are fixed, a mismatch between the assumed values and the true values is a possible source of errors. In this section, we investigate the robustness of the algorithms to a mismatch between the assumed and the actual probability of change points. To do so, we first tuned each algorithm’s parameter for an environment with a change probability $p_c$, and then tested the algorithms in environments with different change probabilities, while keeping the parameter fixed. For each new environment with a different change probability, we calculated the difference between the MSE of these fixed parameters and the optimal MSE, i.e. the resulting MSE for the case that the Exact Bayes’ parameter is tuned for the actual $p_c$. More precisely, if we denote as $\textbf{MSE} [ {\hat{\Theta}}_t ; p_c' , p_c ]$ the MSE of an algorithm with parameters tuned for an environment with $p_c'$, applied in an environment with $p_c$, we calculated the mean regret, defined as $\textbf{MSE} [ {\hat{\Theta}}_t ; p_c' , p_c ] - \textbf{MSE} [ {\hat{\Theta}}_t^{\text{Opt}} , p_c ]$ over time; note that the second term is equal to $\textbf{MSE} [ {\hat{\Theta}}_t ; p_c , p_c ]$ when the algorithm Exact Bayes is used for estimation. The lower the values and the flatter the curve of the mean regret, the better the performance and the robustness of the algorithm in the face of lacking knowledge of the environment. The slope of the curve indicates the degree of deviations of the performance as we move away from the optimally tuned parameter. We ran three random (and same for all algorithms) tasks initializations for each $p_c$ level. In we plot the mean regret for each algorithm for the Gaussian task for four pairs of $s$ and $p_c'$ levels. For $\sigma=0.1$ and $p_c' = 0.04$ (A) the Exact Bayes and the MP20 show the highest robustness (smallest regret) and are closely followed by the pf20, VarSMiLe, and Nas12$^{*}$ (note the regret’s small range of values). The lower the actual $p_c$, the higher the regret, but still the changes are very small. The curves for the SMiLe and the Leaky Integrator are also relatively flat, but the mean regret is much higher. The SOR$20$ is the least robust algorithm. Similar observations can be made for $\sigma=0.1$ and $p_c' = 0.004$ (B). In this case, the performance of all algorithms deteriorates strongly when the actual $p_c$ is higher than the assumed one. However, for $\sigma=5$ (C and D), the ranking of algorithms changes. The SOR$20$ is very robust for this level of stochasticity. The pf$20$ and MP$20$ perform similarly for $p_c = 0.04$, but for lower $p_c'$ the pf$20$ is more robust and the MP$20$ exhibits high fluctuations in its performance. The Nas12$^{*}$ is quite robust at this $\sigma$ level. Overall for Exact Bayes, SOR$20$, pf$20$, VarSMiLe and Nas12$^{*}$, a mismatch of the assumed $p_c$ from the actual one does not deteriorate the performance dramatically for $\sigma=5$, $p_c' = 0.004$ (D). The SMiLe and the Leaky Integrator outperform the other algorithms for higher $p_c'$ if $p_c < p_c'$ (C). A potential reason is that the optimal behavior for the Leaky Integrator (according to the tuned parameters) is to constantly integrate new observations into its belief (i.e. to act like a Perfect Integrator) regardless of the $p_c'$ level. This feature makes it blind to the $p_c$ and therefore very robust against the lack of knowledge of it (C). In summary, most of the time, the mean regret for Exact Bayes and MP20 is less than the mean regret for pf20 and VarSMiLe. However, the variability in the mean regret for pf20 and VarSMiLe is smaller, and their curves are flatter across $p_c$ levels, which makes their performance more predictable. The results for the Categorical estimation task are similar to those of the Gaussian task, with the difference that the SOR$20$ is very robust for this case (). Experimental prediction {#experimental-prediction .unnumbered} ----------------------- It has been experimentally shown that some important behavioral and physiological indicators statistically correlate with a measure of surprise or a prediction error. Examples of such indicators are the pupil diameter [@preuschoff2011pupil; @nassar2012rational; @joshi2019pupil], the amplitude of the P300, N400, and MMN components of EEG [@mars2008trial; @ostwald2012evidence; @lieder2013modelling; @kopp2013electrophysiological; @meyniel2016human; @modirshanechi2019trial; @musiolekmodeling], the amplitude of MEG in specific time windows [@maheu2019brain], BOLD responses in fMRI [@konovalov2018neurocomputational; @loued2020anterior], and reaction time [@huettel2002perceiving; @meyniel2016human]. The surprise measure is usually the negative log-probability of the observation, known as Shannon Surprise [@shannon1948mathematical], and denoted here as $\textbf{S}_{\mathrm{Sh}}$. However, as we show in this section, as long as there is an uninformative prior over observations, Shannon Surprise $\textbf{S}_{\mathrm{Sh}}$ is just an invertible function of our modulated adaptation rate $\gamma$ and hence an invertible function of the Bayes Factor Surprise $\textbf{S}_{\mathrm{BF}}$. Thus, based on the results of previous works [@nassar2010approximately; @nassar2012rational; @meyniel2016human; @ostwald2012evidence; @modirshanechi2019trial], that always used uninformative priors, one cannot determine whether the aforementioned physiological and behavioral indicators correlate with $\textbf{S}_{\mathrm{Sh}}$ or $\textbf{S}_{\mathrm{BF}}$. In this section, we first investigate the theoretical differences between the Bayes Factor Surprise $\textbf{S}_{\mathrm{BF}}$ and Shannon Surprise $\textbf{S}_{\mathrm{Sh}}$. Then, based on their observed differences, we formulate two experimentally testable predictions, with a detailed experimental protocol. Our predictions make it possible to discriminate between the two measures of surprise, and to determine whether physiological or behavioral measurements are signatures of $\textbf{S}_{\mathrm{BF}}$ or of $\textbf{S}_{\mathrm{Sh}}$. ### Theoretical difference between $\textbf{S}_{\mathrm{BF}}$ and $\textbf{S}_{\mathrm{Sh}}$ {#theoretical-difference-between-textbfs_mathrmbf-and-textbfs_mathrmsh .unnumbered} Shannon Surprise [@shannon1948mathematical] is defined as $$\begin{aligned} \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(t)}) &= - \text{log} \Big( \textbf{P}(Y_{t+1} = y_{t+1}|Y_{1:t} = y_{1:t}) \Big) \end{aligned}$$ where for computing $\textbf{P}(Y_{t+1} = y_{t+1}|Y_{1:t} = y_{1:t})$, one should know the structure of the generative model. For the generative model of .A, we find $\textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(t)}) = - \text{log} \Big( (1-p_c) P(y_{t+1}; \pi^{(t)}) + p_c P(y_{t+1}; \pi^{(0)}) \Big)$. While the Bayes Factor Surprise $\textbf{S}_{\mathrm{BF}}$ depends on a ratio between the probability of the new observation under the prior and the current beliefs, Shannon Surprise depends on a weighted sum of these probabilities. Interestingly, it is possible to express (see Methods for derivation) the adaptation rate $\gamma_{t+1}$ as a function of the “difference in Shannon Surprise” $$\begin{aligned} \label{gamma_Shannon} &\gamma_{t+1} = p_c \text{exp} \Big( \Delta \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(t)}, \pi^{(0)}) \Big),\\ & \text{ where } \Delta \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(t)}, \pi^{(0)}) = \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(t)}) - \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(0)}), \end{aligned}$$ where $\gamma_{t+1} = \gamma \big(\textbf{S}_{\mathrm{BF}}^{(t+1)}, m \big)$ depends on the Bayes Factor Surprise and the saturation parameter $m$ (cf. ). Equation \[gamma\_Shannon\] shows that the modulated adaptation rate is not just a function of Shannon Surprise upon observing $Y_{t+1} = y_{t+1}$, but a function of the *difference* between the Shannon Surprise of this observation under the current and under the prior beliefs. In the next subsections, we exploit differences between $\textbf{S}_{\mathrm{BF}}$ and $\textbf{S}_{\mathrm{Sh}}$ to formulate our experimentally testable predictions. ### Experimental protocol {#experimental-protocol .unnumbered} Consider the variant of the Gaussian task of [@nassar2010approximately; @nassar2012rational] which we used in our simulation, i.e. $P_Y(y|\theta) = \mathcal{N}(y; \theta, \sigma^2)$ and $\pi^{(0)}(\theta) = \mathcal{N}(\theta; 0, 1)$. Human subjects are asked to predict the next observation $y_{t+1}$ given what they have observed so far, i.e. $y_{1:t}$. The experimental procedure is as follows: 1. Fix the hyper parameters $\sigma^2$ and $p_c$. 2. At each time $t$, show the observation $y_t$ (produced in the aforementioned way) to the subject, and measure a physiological or behavioral indicator $M_t$, e.g. pupil diameter [@nassar2010approximately; @nassar2012rational]. 3. At each time $t$, after observing $y_t$, ask the subject to predict the next observation ${\hat{y}}_{t+1}$ and their confidence $C_t$ about their prediction. Note that the only difference between our task and the task of [@nassar2010approximately; @nassar2012rational] is the choice of prior for $\theta$ (i.e. Gaussian instead of uniform). The assumption is that, according to the previous studies, there is a *positive* correlation between $M_t$ and a measure of surprise. ### Prediction 1 {#prediction-1 .unnumbered} Based on the results of [@nassar2010approximately; @nassar2012rational], in such a Gaussian task, the best fit for subjects’ prediction ${\hat{y}}_{t+1}$ is ${\hat{\theta}}_t$, and the confidence $C_t$ is a monotonic function of ${\hat{\sigma}}_t$. In order to formalize our experimental prediction, we define, at time $t$, the prediction error as $\delta_t = y_{t} - {\hat{y}}_{t}$ and the “sign bias” as $s_t = \text{sign}(\delta_t {\hat{y}}_{t})$. The variable $s_t$ is a crucial variable for our analysis. It shows whether the prediction ${\hat{y}}_{t}$ is an overestimation in absolute value ($s_t=+1$) or an underestimation in absolute value ($s_t=-1$). .A shows a schematic for the case that both the current and prior beliefs are Gaussian distributions. The two observations indicated by dashed lines have same absolute error $|\delta_t|$, but differ in the sign bias $s$. Given an absolute prediction value ${\hat{y}}>0$, an absolute prediction error $\delta>0$, a confidence value $C>0$, and a sign bias $s \in \{ -1, 1 \}$, we can compute the average of $M_t$ over time for the time points with $|{\hat{y}}_t| \approx {\hat{y}}$, $|\delta_t| \approx \delta$, $C_t \approx C$, and $s_t = s$, which we denote as $\Bar{M}_1({\hat{y}},\delta,s,C)$ – the index 1 stands for the experimental prediction 1. The approximation notation $\approx$ is used for continuous variables instead of equality, due to practical limitations, i.e. for obtaining adequate number of samples for averaging. Note that for our theoretical proofs we use equality, but in our simulation we include the practical limitations of a real experiment, and hence, use an approximation. The formal definitions can be found in Methods. It is worth noting that the quantity $\Bar{M}_1({\hat{y}},\delta,s,C)$ is model independent; it’s calculation does not require any assumption on the learning algorithm the subject may employ. Depending on whether the measurement $\Bar{M}_1({\hat{y}},\delta,s,C)$ reflects $\textbf{S}_{\mathrm{Sh}}$ or $\textbf{S}_{\mathrm{BF}}$, its relationship to the defined four variables (i.e. ${\hat{y}}$, $\delta$, $s$, $C$) is qualitatively and quantitatively different. In order to prove and illustrate our prediction, let us consider each subject as an agent enabled with one of the learning algorithms that we discussed. Similar to above, given an absolute prediction ${\hat{\theta}}>0$ (corresponding to the subjects’ absolute prediction ${\hat{y}}$), an absolute prediction error $\delta>0$, a standard deviation $\sigma_C$ (corresponding to the subjects’ confidence value $C$), and a sign bias $s \in \{ -1, 1 \}$, we can compute the average Shannon Surprise $\textbf{S}_{\mathrm{Sh}}(y_{t}; {\hat{\pi}}^{(t-1)})$ and the average Bayes Factor Surprise $\textbf{S}_{\mathrm{BF}}(y_{t}; {\hat{\pi}}^{(t-1)})$ over time, for the time points with $|{\hat{\theta}}_{t-1}| \approx {\hat{\theta}}$, $|\delta_t| \approx \delta $, ${\hat{\sigma}}_t \approx \sigma_C$, and $s_t = s$, which we denote as $\Bar{\textbf{S}}_{\mathrm{Sh}}({\hat{\theta}},\delta,s,\sigma_C)$ and $\Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}}, \delta,s,\sigma_C)$ respectively. We can show theoretically (see Methods) and in simulations (see .B and Methods) that for any value of ${\hat{\theta}}$, $\delta$, and $\sigma_C$, we have $\Bar{\textbf{S}}_{\mathrm{Sh}}({\hat{\theta}},\delta,s=+1,\sigma_C) > \Bar{\textbf{S}}_{\mathrm{Sh}}({\hat{\theta}},\delta,s=-1,\sigma_C)$ for the Shannon Surprise, and exactly the opposite relation, i.e. $\Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=+1,\sigma_C) < \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=-1,\sigma_C)$ for the Bayes Factor Surprise. Moreover, this effect increases with increasing $\delta$. It should be noted that such an effect is due to the essential difference of $\textbf{S}_{\mathrm{Sh}}$ and $\textbf{S}_{\mathrm{BF}}$ in using the prior belief $\pi^{(0)}(\theta)$. Our experimental prediction is theoretically provable for the cases that each subject’s belief ${\hat{\pi}}^{(t)}$ is a Gaussian distribution, which is the case if they employ VarSMiLe, Nas10$^{*}$, Nas12$^{*}$, pf1, MP1, or Leaky Integrator as their learning rule (see Methods). For the cases that different learning rules (e.g. pf20) are used, where the posterior belief is a weighted sum of Gaussians, the theoretical analysis is more complicated, but our simulations show the same results (see .B and Methods). Therefore, independent of the learning rule, we have the same experimental prediction on the manifestation of different surprise measures on physiological signals, such as pupil dilation. Our first experimental prediction can be summarized as a set of hypotheses shown in Table \[table:Hyp\_3\]. **Hypothesis** **Prediction** --------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------- The indicator reflects $\textbf{S}_{\mathrm{BF}}$ $ \Delta \Bar{M}_1({\hat{\theta}},\delta,C) < 0$ and $\frac{\partial \Delta \Bar{M}_1({\hat{\theta}},\delta,C)}{\partial \delta} < 0$ The indicator reflects $\textbf{S}_{\mathrm{Sh}}$ $ \Delta \Bar{M}_1({\hat{\theta}},\delta,C) > 0$ and $\frac{\partial \Delta \Bar{M}_1({\hat{\theta}},\delta,C)}{\partial \delta} > 0$ The prior is not used for inference $ \Delta \Bar{M}_1({\hat{\theta}},\delta,C) = 0$ : **Experimental Hypotheses and Predictions 1.** $ \Delta \Bar{M}_1({\hat{\theta}},\delta,C)$ stands for $ \Bar{M}_1({\hat{\theta}},\delta,s=+1,C) - \Bar{M}_1({\hat{\theta}},\delta,s=-1,C)$[]{data-label="table:Hyp_3"} ### Prediction 2 {#prediction-2 .unnumbered} Our second prediction follows the same experimental procedure as the one for the first prediction. The main difference is that for the second prediction we need to fit a model to the experimental data. Given one of the learning algorithms, the fitting procedure can be done by tuning the free parameters of the algorithm with the goal of minimizing the mean squared error between the model’s prediction ${\hat{\theta}}_t$ and a subject’s prediction ${\hat{y}}_{t+1}$ (similar to [@nassar2010approximately; @nassar2012rational]) or with the goal of maximizing the likelihood of subject’s prediction ${\hat{\pi}}^{(t)}({\hat{y}}_{t+1})$. Our prediction is independent of the learning algorithm, but in an actual experiment, we recommend to use model selection to find the model that fits the human data best. Having a fitted model, we can compute the probabilities $P(y_{t+1};{\hat{\pi}}^{(t)})$ and $P(y_{t+1};{\hat{\pi}}^{(0)})$. For the case that these probabilities are equal, i.e. $P(y_{t+1};{\hat{\pi}}^{(t)}) = P(y_{t+1};{\hat{\pi}}^{(0)}) = p$, the Bayes Factor Surprise $\textbf{S}_{\mathrm{BF}}$ is equal to 1, independent of the value of $p$ (cf. ). However, the Shannon Surprise $\textbf{S}_{\mathrm{Sh}}$ is equal to $- \log p$, and varies with $p$. .A shows a schematic for the case that both current and prior beliefs are Gaussian distributions. Two cases for which we have $P(y_{t+1};{\hat{\pi}}^{(t)}) = P(y_{t+1};{\hat{\pi}}^{(0)}) = p$, for two different $p$ values, are marked by black dots at the intersections of the curves. Given a probability $p>0$, we can compute the average of $M_t$ over time for the time points with $P(y_{t+1};{\hat{\pi}}^{(t)}) \approx p$ and $P(y_{t+1};{\hat{\pi}}^{(0)}) \approx p$, which we denote as $\Bar{M}_2(p)$ – the index 2 stands for the experimental prediction 2. Analogous to the first prediction, the approximation notation $\approx$ is used due to practical limitations. Then, if $\Bar{M}_2(p)$ is independent of $p$, its behavior is consistent with $\textbf{S}_{\mathrm{BF}}$, while if it decreases by increasing $p$, it can be a signature of $\textbf{S}_{\mathrm{Sh}}$. Our second experimental prediction can be summarized as two hypotheses shown in Table \[table:Hyp\_4\]. Note that in contrast to our first prediction, with the assumption that the standard deviation of the prior belief is fitted using the behavioral data, we do not consider the hypothesis that the prior is not used for inference, since this is indistinguishable from a very large variance of the prior belief. In order to illustrate the possible results and the feasibility of the experiment, we ran a simulation and computed $\Bar{\textbf{S}}_{\mathrm{BF}}(p)$ and $\Bar{\textbf{S}}_{\mathrm{Sh}}(p)$ for the time points with $P(y_{t+1};{\hat{\pi}}^{(t)}) \approx p$ and $P(y_{t+1};{\hat{\pi}}^{(0)}) \approx p$ (see Methods for details). The results of the simulation are shown in .B. **Hypothesis** **Prediction** --------------------------------------------------- ------------------------------------------------ The indicator reflects $\textbf{S}_{\mathrm{BF}}$ $\frac{\partial \Bar{M}_2(p)}{\partial p} = 0$ The indicator reflects $\textbf{S}_{\mathrm{Sh}}$ $\frac{\partial \Bar{M}_2(p)}{\partial p} < 0$ : **Experimental Hypotheses and Predictions 2.**[]{data-label="table:Hyp_4"} Discussion {#discussion .unnumbered} ========== We have shown that performing exact Bayesian inference on a generative world model naturally leads to a definition of surprise and a surprise-modulated adaptation rate. We have proposed three approximate algorithms (pf$N$, VarSMiLe, and MP$N$) for learning in non-stationary environments, which all exhibit the surprise-modulated adaptation rate of the exact Bayesian approach and are biologically plausible. Empirically we observed that our algorithms achieve levels of performance comparable to approximate Bayesian methods with higher memory demands [@adams2007bayesian], and are more resilient across different environments compared to methods with similar memory demands [@fearnhead2007line; @nassar2010approximately; @nassar2012rational; @faraji2018balancing]. Learning in a volatile environment has been studied for a long time in the fields of Bayesian learning, neuroscience, and signal processing. In the following, we discuss the biological plausibility of our work, and we briefly review some of the previously developed algorithms, with particular focus on the ones that have studied environments which can be modeled with a generative model similar to the one in . We then discuss further our results, and propose directions for further work on surprise-based learning. Biological interpretation {#biological-interpretation .unnumbered} ------------------------- Humans are able to quickly adapt to changes [@nassar2012rational; @nassar2010approximately; @behrens2007learning], but human behaviour is also often observed to be suboptimal, compared to the normative approach of the exact Bayesian inference [@mathys2011bayesian; @wilson2013mixture; @nassar2010approximately; @glaze2015normative; @prat2020human]. In general, biological agents have limited resources and possibly inaccurate hyper-parameter assumptions, giving rise to sub-optimal behaviour. We have shown that our algorithms’ accuracy degrades with a sub-optimal choice of hyper-parameters and with a decreasing number of particles in the sampling-based ones, which might be a possible explanation of suboptimal human behavior. In particular, Particle Filtering has been employed in explaining the behavior of human subjects in changing environments; [@daw2008pigeon] uses a single particle, [@findling2019imprecise] combines Particle Filtering with a noisy inference style, and [@prat2020human] uses it for a task with temporal structure. At the level of neuronal implementation, Particle Filtering has been shown to be biologically plausible [@kutschireiter2017nonlinear; @shi2009neural; @huang2014neurons; @legenstein2014ensembles], thus pf$N$ and – its greedy version – MP$N$ could be implemented neurally. Variational SMiLe includes a simple updating scheme (for distributions in the exponential family) that can be implemented with neo-Hebbian update rules [@lisman2011neohebbian]. Our theoretical framework builds on the body of literature on neo-Hebbian three-factor learning rules [@lisman2011neohebbian; @fremaux2016neuromodulated; @gerstner2018eligibility], where a third factor indicating reward or surprise enables or modulates a synaptic change or a belief update [@angela2005uncertainty; @angela2012change]. We have shown how Bayesian or approximate Bayesian inference naturally lead to such a third factor that modulates learning via the surprise modulated adaptation rate $\gamma (S, m)$. This may offer novel interpretations of behavioural and neurophysiological data, and help in understanding how three-factor learning computations may be implemented in the brain. Related work {#related_work .unnumbered} ------------ #### Exact Bayesian inference. As already described in the “Message-Passing $N$” section of the Results, for the generative model in , it is possible to find an exact online Bayesian update of the belief using a message passing algorithm [@adams2007bayesian]. The space and time complexity of the algorithm increases linearly with $t$, which makes it unsuitable for an online learning setting. However, approximations like dropping messages below a certain threshold [@adams2007bayesian] or stratified resampling [@fearnhead2007line] allow to reduce the computational complexity. The former has a variable number of particles in time, and the latter needs solving a complicated non-linear equation at each time step in order to reduce the number of particles to $N$ (called SOR$N$ in the results section). Our message passing algorithm with finite number of particles (messages) $N$ (MP$N$, Algo. 3) is closely related to these algorithms and can be seen as a biologically more plausible variant of the other two. All three algorithms have the same update rules given by and . Hence the algorithms of both [@adams2007bayesian] and [@fearnhead2007line] have the same surprise modulation as our MP$N$. Their difference lies in their approaches to eliminate less “important” particles. #### Leaky integration and variations of delta-rules. In order to estimate some statistics, leaky integration of new observations is a particularly simple form of a trade-off between integrating and forgetting. After a transient phase, the update of a leaky integrator takes the form of a delta-rule that can be seen as an approximation of exact Bayesian updates [@meyniel2016human; @heilbron2019confidence; @angela2009sequential]. This update rule was found to be biologically plausible and consistent with human behavioral data [@meyniel2016human; @angela2009sequential]. However, [@behrens2007learning] and [@heilbron2019confidence] demonstrated that in some situations, the exact Bayesian model is significantly better than leaky integration in explaining human behavior. The inflexibility of leaky integration with a single, constant leak parameter can be overcome by a weighted combination of multiple leaky integrators [@wilson2013mixture], where the weights are updated in a similar fashion as in the exact online methods [@adams2007bayesian; @fearnhead2007line], or by considering an adaptive leak parameter [@nassar2012rational; @nassar2010approximately]. We have shown that the two algorithms of [@nassar2010approximately; @nassar2012rational] can be generalized to Gaussian prior beliefs (Nas10$^{*}$ and Nas12$^{*}$). Our results show that these algorithms also inherit the surprise-modulation of the exact Bayesian inference. Our surprise dependent adaptation rate $\gamma$ can be interpreted as a surprise-modulated leak parameter. #### Other approaches. Learning in the presence of abrupt changes has also been considered without explicit assumptions about the underlying generative model. One approach uses a surprise-modulated adaptation rate [@faraji2018balancing] similar to . The Surprise-Minimization Learning algorithm of [@faraji2018balancing] (SMiLe) has an optimization form of updating similar to the one of VarSMiLe ( and ). The adaptation rate modulation, however, is based on the Confidence Corrected Surprise [@faraji2018balancing] rather than the Bayes Factor Surprise, and the trade-off in its update rule is between resetting and staying with the latest belief rather than between resetting and integrating (see Methods). Other approaches use different generative models, e.g. conditional sampling of the parameters also when there is a change [@angela2005uncertainty; @glaze2015normative], deeper hierarchy without fixed change probability $p_c$ [@wilson2010bayesian], or models with drift in the parameters [@mathys2011bayesian; @gershman2014statistical]. A recent work shows that inference on the generative model of can explain human behavior well even when the true generative model of the environment is different and more complicated [@findling2019imprecise]. They develop a heuristic approach to add noise in the inference process of a Particle Filter. Their algorithm can be interpreted as a surprise-modulated Particle Filter, where the added noise scales with a measure of surprise (conceptually equivalent to Bayesian surprise [@storck1995reinforcement; @schmidhuber2010formal; @itti2006bayesian]). Moreover, another recent work [@prat2020human] shows that approximate sampling algorithms (like Particle Filtering) can explain human behavior better than their alternatives in tasks closely related to the generative model of . The signal processing literature provides further methods to address the problem of learning in nonstationary environments with abrupt changes; see [@aminikhanghahi2017survey] for a review, and [@lin2017sharp; @cummings2018differentially; @ozkan2013marginalized; @masegosa2017bayesian] for a few recent examples. Surprise-modulation as a generic phenomenon {#surprise-modulation-as-a-generic-phenomenon .unnumbered} ------------------------------------------- Learning rate modulation similar to the one in has been previously proposed in the neuroscience literature with either heuristic arguments [@faraji2018balancing] or with Bayesian arguments for a particular experimental task, e.g. when samples are drawn from a Gaussian distribution [@nassar2010approximately; @nassar2012rational]. The fact that the same form of modulation is at the heart of Bayesian inference for our relatively general generative model, that it is derived without any further assumptions, and is not a-priori defined is in our view an important contribution to the field of adaptive learning algorithms in computational neuroscience. Furthermore, the results of our three approximate methods (Particle Filtering, Variational SMiLe, and Message Passing with fixed $N$ number of messages) as well as some previously developed ones [@adams2007bayesian; @fearnhead2007line; @nassar2010approximately; @nassar2012rational] demonstrate that the surprise-based modulation of the learning rate is a generic phenomenon. Therefore, regardless of whether the brain uses Bayesian inference or an approximate algorithm [@mathys2011bayesian; @friston2010free; @nassar2010approximately; @nassar2012rational; @gershman2014statistical; @bogacz2017tutorial; @bogacz2019dopamine; @gershman2019does; @findling2019imprecise; @prat2020human], the notion of surprise and the way it modulates learning (i.e. and ) looks generic. Bayes Factor Surprise as a novel measure of surprise {#surp_disc .unnumbered} ---------------------------------------------------- In view of a potential application in the neurosciences, a definition of surprise should exhibit two properties: (i) surprise should reflect how unexpected an event is, and, (ii) surprise should modulate learning. Surprising events are indications that our belief is far from the real world, suggesting to update our model of the world, or, for large surprise, simply forget it. Forgetting is the same as returning to the prior belief. However, an observation $y_{t+1}$ can be unexpected under both the prior $\pi^{(0)}$ and the current beliefs $\pi^{(t)}$. In these situations, it is not obvious whether forgetting helps. Therefore, the modulation between forgetting or not should be based on a comparison between the probability of an event under the current belief $P(y_{t+1}; \pi^{(t)})$ and its probability under the prior belief $P(y_{t+1}; \pi^{(0)})$. The definition of the Bayes Factor Surprise $\textbf{S}_{\mathrm{BF}}$ as the ratio of $P(y_{t+1}; \pi^{(t)})$ and $P(y_{t+1}; \pi^{(0)})$ exploits this insight. The Bayes Factor Surprise appears as a modulation factor in the recursive form of the exact Bayesian update rule for a hierarchical generative model of the environment. When two events are equally probable under the prior belief, the one which is less expected under the current belief is more surprising - satisfying the first property. At the same time, when two events are equally probable under the current belief, the one which is more expected under the prior belief is more surprising - signaling that forgetting may be beneficial. $\textbf{S}_{\mathrm{BF}}$ can also be written (using ) in a more explicit way as $$\label{Eq:S_GM_ind_GM} \textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)}) = \frac{P(y_{t+1}; \pi^{(0)})}{P(y_{t+1}; \pi^{(t)})} = \frac{ \mathbb{E}_{\pi^{(0)}} \big[ P_Y(y_{t+1} | \Theta ) \big] }{ \mathbb{E}_{\pi^{(t)}} \big[ P_Y(y_{t+1} | \Theta ) \big] }.$$ Note that the definition by itself is independent of the specific form of the generative model. In other words, even in the cases where data is generated with another generative model (e.g. the real world), $\textbf{S}_{\mathrm{BF}}$ could be a candidate surprise measure in order to interpret brain activity or pupil dilation. We formally discussed the connections between the Bayes Factor Surprise and Shannon Surprise [@shannon1948mathematical], and show that they are closely linked. We showed that the modulated adaptation rate ($\gamma$) used in (approximate) Bayesian inference is a function of the difference between the Shannon Surprise under the current and the prior beliefs, but cannot be expressed solely by the Shannon Surprise under the current one. Our formal comparisons between these two different measures of surprise lead to specific experimentally testable predictions. The Bayesian Surprise $\textbf{S}_{\mathrm{Ba}}$ [@storck1995reinforcement; @schmidhuber2010formal; @itti2006bayesian] and the Confidence Corrected Surprise $\textbf{S}_{\mathrm{CC}}$ [@faraji2018balancing] are two other measures of surprise in neuroscience. The learning modulation derived in our generative model cannot be expressed as a function of $\textbf{S}_{\mathrm{Ba}}$ and $\textbf{S}_{\mathrm{CC}}$. However, one can hypothesize that $\textbf{S}_{\mathrm{Ba}}$ is computed after the update of the belief to measure the information gain of the observed event, and is therefore not a good candidate for online learning modulation. The Confidence Corrected surprise $\textbf{S}_{\mathrm{CC}}$ takes into account the shape of the belief, and therefore includes the effects of confidence, but it does not consider any information about the prior belief. Hence, a result of $ \Bar{M}_1({\hat{\theta}},\delta,s=+1,C) = \Bar{M}_1({\hat{\theta}},\delta,s=-1,C)$ in our first experimental prediction would be consistent with the corresponding behavioral or physiological indicator reflecting the $\textbf{S}_{\mathrm{CC}}$. Difference in Shannon Surprise, an alternative perspective {#difference-in-shannon-surprise-an-alternative-perspective .unnumbered} ---------------------------------------------------------- Following our formal comparison in the “Experimental prediction” section, $\textbf{S}_{\mathrm{BF}}$ can be expressed as a deterministic function of the difference in Shannon Surprise as $$\begin{aligned} \label{Sgm_Shannon} \textbf{S}_{\mathrm{BF}} &= \frac{(1-p_c) e^ {\Delta \textbf{S}_{\mathrm{Sh}} } }{1 - p_c e^ {\Delta \textbf{S}_{\mathrm{Sh}} }}. \end{aligned}$$ All of our theoretical results can be rewritten by replacing $\textbf{S}_{\mathrm{BF}}$ with this function of $\Delta \textbf{S}_{\mathrm{Sh}}$. Moreover, since there is a 1-to-1 mapping between $\textbf{S}_{\mathrm{BF}}$ and $\Delta \textbf{S}_{\mathrm{Sh}}$, from a systemic point of view, it is not possible to specify whether the brain computes the former or the latter by analysis of behavioral data and biological signals. This suggests an alternative interpretation of surprise-modulated learning as an approximation of Bayesian inference: What the brain computes and perceives as surprise or prediction error may be Shannon Surprise, but the modulating factor in a three-factor synaptic plasticity rule [@lisman2011neohebbian; @fremaux2016neuromodulated; @gerstner2018eligibility] may be implemented by comparing the Shannon Surprise values under the current and the prior beliefs. Future directions {#future-directions .unnumbered} ----------------- A natural continuation of our study is to test our experimental predictions in human behavior and physiological signals, in order to investigate which measures of surprise are used by the brain. Along a similar direction, our approximate learning algorithms can be evaluated on human behavioral data from experiments that use a similar generative model [@nassar2012rational; @nassar2010approximately; @wilson2013mixture; @behrens2007learning; @angela2005uncertainty; @glaze2015normative; @heilbron2019confidence]. in order to assess if our proposed algorithms achieve similar or better performance in explaining data. Finally, our methods can potentially be applied to model-based reinforcement learning in non-stationary environments. In recent years, there has been a growing interest in adaptive or continually learning agents in changing environments in the form of Continual learning and Meta-learning [@lomonaco2019continual; @traore2019discorl]. Many Continual learning model-based approaches make use of some procedure to detect changes [@lomonaco2019continual; @nagabandi2018learning]. Integrating $\textbf{S}_{\mathrm{BF}}$ and a learning rate $\gamma(\textbf{S}_{\mathrm{BF}})$ into a reinforcement learning agent would be an interesting future direction. Methods {#methods .unnumbered} ======= Generative model {#generative-model .unnumbered} ---------------- The graphical model corresponding to our generative model is shown in . Formally, we sample $\Theta_0$ from the prior $\pi^{(0)}$, and for $t \geq 1$ the generative model is $$\begin{aligned} & \Delta H_t \sim \text{Bernoulli}(p_c)\, , \label{Eq:GenModel} \\ & \textbf{P}(\Theta_t=\theta|\Delta H_t = \Delta h_t, \Theta_{t-1}=\theta')= \left\{ \begin{array}{lr} \delta (\theta - \theta') & \text{if} \quad \Delta h_t = 0\, ,\\ \pi^{(0)}(\theta) & \text{if} \quad \Delta h_t = 1\, , \end{array} \right. \label{Eq:GenModel2}\\ & \textbf{P}(Y_t=y|\Theta_{t}=\theta) = P_{Y}(y | \theta)\, . \label{Eq:GenModel3}\end{aligned}$$ Random variables are indicated by capital letters, and values by small letters. $\textbf{P}$ stands for either probability density function (for the continuous variables) or probability mass function (for the discrete variables), and $\delta$ is the Dirac or Kronecker delta distribution, respectively. $P_{Y}$ is the time-invariant likelihood function. Proof of the proposition {#proof-of-the-proposition .unnumbered} ------------------------ By definition $$\label{Eq:Belief_Def_1_App} \pi^{(t+1)}(\theta)\equiv \textbf{P}(\Theta_{t+1}=\theta|Y_{1:t+1} = y_{1:t+1}).$$ We exploit the Markov property of the generative model in equations \[Eq:GenModel\], \[Eq:GenModel2\], and \[Eq:GenModel3\], condition on the fixed past $y_{1:t}$ and rewrite $$\label{Eq:Belief_Def_2_App} \pi^{(t+1)}(\theta) = \frac{ P_{Y}(y_{t+1} | \theta) \textbf{P}(\Theta_{t+1}=\theta|Y_{1:t} = y_{1:t}) }{\textbf{P}(Y_{t+1} = y_{t+1}|Y_{1:t} = y_{1:t})}.$$ By marginalization over the hidden state $\Delta H_{t+1}$, the second factor in the numerator of can be written as $$\begin{aligned} \textbf{P}(\Theta_{t+1}=\theta|Y_{1:t}=y_{1:t}) &= (1-p_c) \pi^{(t)}(\theta) + p_c \pi^{(0)}(\theta). \end{aligned}$$ The denominator in can be written as $$\begin{aligned} \textbf{P}(Y_{t+1}=y_{t+1}|Y_{1:t}=y_{1:t}) &= \int P_Y(y_{t+1}|\theta) \textbf{P}(\Theta_{t+1}=\theta|Y_{1:t}=y_{1:t}) d \theta\\ &= (1-p_c) \int P_Y(y_{t+1}|\theta) \pi^{(t)}(\theta) d \theta + p_c \int P_Y(y_{t+1}|\theta) \pi^{(0)}(\theta) d \theta\\ &= (1-p_c) P(y_{t+1}; \pi^{(t)}) + p_c P(y_{t+1}; \pi^{(0)}). \end{aligned}$$ where we used the definition in . Using these two expanded forms, can be rewritten $$\begin{aligned} \label{Eq:pi_post_1} \pi^{(t+1)}(\theta) &= \frac{ P_Y(y_{t+1}|\theta) \Big( (1-p_c) \pi^{(t)}(\theta) + p_c \pi^{(0)}(\theta) \Big) }{ (1-p_c) P(y_{t+1}; \pi^{(t)}) + p_c P(y_{t+1}; \pi^{(0)}) }. \end{aligned}$$ We define $P(\theta|y_{t+1})$ as the posterior given a change in the environment as $$\begin{aligned} P(\theta|y_{t+1}) = \frac{P_Y(y_{t+1}|\theta) \pi^{(0)}(\theta)} {P(y_{t+1}; \pi^{(0)})}. \end{aligned}$$ Then, we can write as $$\begin{aligned} \label{Eq:pi_post_2} \pi^{(t+1)}(\theta) &= \frac{ (1-p_c) P(y_{t+1}; \pi^{(t)}) \pi_B^{(t+1)}(\theta) + p_c P(y_{t+1}; \pi^{(0)}) P(\theta|y_{t+1}) }{ (1-p_c) P(y_{t+1}; \pi^{(t)}) + p_c P(y_{t+1}; \pi^{(0)}) }\\ &= \frac{ \pi_B^{(t+1)}(\theta) + \frac{p_c}{1-p_c} \frac{P(y_{t+1}; \pi^{(0)})}{P(y_{t+1}; \pi^{(t)})} P(\theta|y_{t+1}) }{ 1 + \frac{p_c}{1-p_c} \frac{P(y_{t+1}; \pi^{(0)})}{P(y_{t+1}; \pi^{(t)})} }\\ &= (1 - \gamma_{t+1}) \pi^{(t+1)}_B(\theta) + \gamma_{t+1} P(\theta|y_{t+1}), \end{aligned}$$ where $\pi^{(t+1)}_B(\theta)$ is defined in , and $$\begin{aligned} \gamma_{t+1} &= \gamma \Big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)}), \frac{p_c}{1-p_c} \Big)\\ \end{aligned}$$ with $\textbf{S}_{\mathrm{BF}}$ defined in , and $\gamma(\text{S}, m)$ defined in . Thus our calculation yields a specific choice of surprise ($\text{S} = \textbf{S}_{\mathrm{BF}}$) and a specific value for the saturation parameter $m = \frac{p_c}{1-p_c}$. Derivation of the weight update for Particle Filtering (Algo. 1) {#derivation-of-the-weight-update-for-particle-filtering-algo.-1 .unnumbered} ---------------------------------------------------------------- We derive here the weight update for the particle filter. We start by defining the number of changes from beginning until time $t$ as the random variable $H_t = \sum_{k=1}^{t} \Delta H_k$. In this subsection, to shorten notation, we drop the explicit mentioning of the random variables, e.g. $Y_{1:t+1}$, and display only their values, e.g. $y_{1:t+1}$. The difference in our formalism from a standard derivation [@sarkka2013bayesian] is the absence of the Markov property of conditional observations (i.e $\textbf{P}(y_{t+1} | h_{1:t+1}, y_{1:t}) \neq \textbf{P}(y_{t+1} | h_{t+1})$). Our goal is to perform the approximation $$\label{Eq:pDh_pf_approx_method} P(\Delta h_{1:t+1} | y_{1:t+1}) \approx \sum_{i=1}^{N} w_{t+1}^{(i)} \delta (\Delta h_{1:t+1} - \Delta h_{1:t+1}^{(i)} ) \, .$$ Given a proposal sampling distribution $Q$, for the weight of particle $i$ at time $t+1$ we have $$\begin{aligned} w_{t+1}^{(i)} & \propto \frac{\textbf{P}(h_{1:t+1}^{(i)} | y_{1:t+1})}{Q(h_{1:t+1}^{(i)} | y_{1:t+1})} \propto \frac{\textbf{P}(h_{1:t+1}^{(i)}, y_{t+1} | y_{1:t})}{Q(h_{1:t+1}^{(i)} | y_{1:t+1})} \\ w_{t+1}^{(i)} & \propto \frac{\textbf{P}(y_{t+1} | h_{1:t+1}^{(i)}, y_{1:t}) \textbf{P}(h_{t+1}^{(i)} | h_{1:t}^{(i)}, y_{1:t}) \textbf{P}(h_{1:t}^{(i)} | y_{1:t}) } {Q(h_{t+1}^{(i)} | h_{1:t}^{(i)}, y_{1:t+1}) Q(h_{1:t}^{(i)} | y_{1:t})}\, , \end{aligned}$$ where the only requirement for the proposal distribution $Q$ is the essential property of the generative model (and hence $\textbf{P}$) that the previous hidden states $h_{1:t}^{(i)}$ are independent of the next observation $y_{t+1}$. Notice that $ w_{t}^{(i)} \propto \frac{\textbf{P}(h_{1:t}^{(i)} | y_{1:t})} {Q(h_{1:t}^{(i)} | y_{1:t})} $ are the weights calculated at the previous time step and that $\textbf{P}(h_{t+1}^{(i)} | h_{1:t}^{(i)}, y_{1:t}) = \textbf{P}(h_{t+1}^{(i)} | h_{t}^{(i)})$. Therefore $$\begin{aligned} \label{w_pf_1} w_{t+1}^{(i)} & \propto \frac{\textbf{P}(y_{t+1} | h_{1:t+1}^{(i)}, y_{1:t}) \textbf{P}(h_{t+1}^{(i)} | h_{t}^{(i)})} {Q(h_{t+1}^{(i)} | h_{1:t}^{(i)}, y_{1:t+1})} w_{t}^{(i)}\, . \end{aligned}$$ We use the optimal proposal function in terms of variance of the weights [@doucet2000sequential] $$\begin{aligned} \label{Q_pf_1} Q(h_{t+1}^{(i)} | h_{1:t}^{(i)}, y_{1:t+1}) = \textbf{P}(h_{t+1}^{(i)} | h_{1:t}^{(i)}, y_{1:t+1})\, . \end{aligned}$$ Using Bayes’ rule and equations \[w\_pf\_1\] and \[Q\_pf\_1\], and after a few steps of algebra, we have $$\begin{aligned} w_{t+1}^{(i)} & \propto \frac{\textbf{P}(y_{1:t+1} | h_{1:t}^{(i)})} {\textbf{P}(y_{1:t}| h_{1:t}^{(i)})} w_{t}^{(i)}\\ & \propto \frac{(1-p_c)\textbf{P}(y_{1:t+1} | h_{1:t}^{(i)}, \Delta h^{(i)}_{t+1}=0 ) + p_c \textbf{P}(y_{1:t+1} | h_{1:t}^{(i)}, \Delta h^{(i)}_{t+1}=1)} {\textbf{P}(y_{1:t}| h_{1:t}^{(i)})} w_{t}^{(i)}\, . \end{aligned}$$ To specify the stationary periods separated by change-points, we define two new variables $m^k_{(i)}$ and $n^k_{(i)}$. We define $m^k_{(i)}$ as the time point when a new hidden state $h_t^{(i)}=k$ started, i.e $m^k_{(i)} = \min \{ j \in \{1,...,t\} | h_j^{(i)} = k \}$, and similarly we define $n^k_{(i)} = \max \{ j \in \{1,...,t\} | h_j^{(i)} = k \}$ as the time point when a new hidden state $h_t^{(i)}=k$ stopped. By doing so, for each particle, the observations $y_{1:t}$ can be partitioned to $h_{t}^{(i)}$ stationary sections, i.e. $\{y_{m^k_{(i)}:n^k_{(i)}}: 1 \leq k \leq h_{t}^{(i)} \}$. Therefore, we group together the observations coming from the same hidden state and drop the conditioning on the hidden states since this information is incorporated in the $m^k_{(i)}, n^k_{(i)}$ variables. We have $$\begin{aligned} \textbf{P}(y_{1:t+1} | h_{1:t}^{(i)}, \Delta h^{(i)}_{t+1}=0) & = \prod_{k=1}^{h_{t-1}^{(i)}}\textbf{P}'(y_{m^k_{(i)}:n^k_{(i)}})\textbf{P}'(y_{m_{(i)}^{h_{t}^{(i)}}:t+1})\\ \textbf{P}(y_{1:t+1} | h_{1:t}^{(i)}, \Delta h^{(i)}_{t+1}=1) & = \prod_{k=1}^{h_{t}^{(i)}}\textbf{P}'(y_{m_{(i)}^k:n_{(i)}^k})\textbf{P}(y_{t+1})\\ \textbf{P}(y_{1:t}| h_{1:t}^{(i)}) & = \prod_{k=1}^{h_{t-1}^{(i)}}\textbf{P}'(y_{m_{(i)}^k:n_{(i)}^k})\textbf{P}'(y_{m_{(i)}^{h_{t}^{(i)}}:t})\, , \end{aligned}$$ where we use $\textbf{P}'$ instead of $\textbf{P}$ when we drop conditioning on hidden states, i.e. for the cases that the observations correspond to time points with the same hidden state. This gives us after a few steps of algebra $$\begin{aligned} \label{Eq:w_update_normalized} w_{t+1}^{(i)} & = \Big[(1-p_c) P(y_{t+1}; {\hat{\pi}}_i^{(t)}) + p_c P(y_{t+1}; \pi^{(0)}) \Big] w_{t}^{(i)} / Z\, , \end{aligned}$$ where $P(y_{t+1}; {\hat{\pi}}_i^{(t)})$ and $P(y_{t+1}; {\hat{\pi}}_i^{(0)})$ are defined as in , and $Z$ is the normalization factor $$\label{Eq:w_norm_factor} Z = (1-p_c) P(y_{t+1}; {\hat{\pi}}^{(t)}) + p_c P(y_{t+1}; \pi^{(0)})\, .$$ We now compute the weights corresponding to $\pi_{B}^{(t+1)}$ as defined in $$\label{Eq:wB_sup} w_{B, t+1}^{(i)} = \frac{P(y_{t+1}; {\hat{\pi}}_i^{(t)})}{P(y_{t+1}; {\hat{\pi}}^{(t)})} w_{t}^{(i)}\, .$$ Combining , and we can then re-write the weight update rule as $$\begin{aligned} \label{Eq:w_update_final} w_{t+1}^{(i)} &= (1-\gamma_{t+1})w_{B, t+1}^{(i)} + \gamma_{t+1}w_{t}^{(i)}\, , \end{aligned}$$ where $\gamma_{t+1} = \gamma \Big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}^{(t)}), \frac{p_c}{1-p_c} \Big)$ of . At every time step $t+1$ we sample each particle’s hidden state $h_{t+1}$ from the proposal distribution. We calculate the change probability $$\begin{aligned} \label{Eq:proposal_final_supp} Q(\Delta h_{t+1}^{(i)} = 1| h_{1:t}^{(i)}, y_{1:t+1}) &= \frac{\textbf{P}(y_{1:t+1} | h_{1:t}^{(i)}, \Delta h^{(i)}_{t+1}=1) \textbf{P}(\Delta h^{(i)}_{t+1}=1 | h_{1:t}^{(i)})} {\textbf{P}(y_{1:t}| h_{1:t}^{(i)})}\\ &= \gamma \Big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\pi}}_i^{(t)}), \frac{p_c}{1-p_c} \Big)\, . \end{aligned}$$ We implemented the Sequential Importance Resampling algorithm [@gordon1993novel], [@doucet2000sequential], where the particles are resampled when their effective number falls below a threshold. The effective number of the particles is defined as [@doucet2000sequential], [@sarkka2013bayesian] $$N_{\text{eff}} \approx \frac{1}{\sum_{i=1}^{N}(w_t^{(i)})^2}\, .$$ When $N_{\text{eff}}$ is below a critical threshold, the particles are resampled with replacement from the categorical distribution defined by their weights, and all their weights are set to $w_t^{(i)} = 1/N$. We did not optimize the parameter $N_{\text{eff}}$, and following [@doucet2009tutorial], we performed resampling when $N_{\text{eff}} \leq N/2$. Derivation of the optimization-based formulation of VarSMiLe (Algo. 2) {#derivation-of-the-optimization-based-formulation-of-varsmile-algo.-2 .unnumbered} ---------------------------------------------------------------------- To derive the optimization-based update rule for the Variational SMiLe rule and the relation of the bound $B_{t+1}$ with surprise, we used the same approach used in [@faraji2018balancing]. #### Derivation of the update rule Consider the general form of the following variational optimization problem: $$\begin{aligned} \label{Eq:Genral_Opt} q^*(\theta) &= \text{argmin }\textbf{D}_{KL} \big[ q(\theta) || p_1(\theta) \big]\\ & q(\theta) \text{ s.t. } \textbf{D}_{KL} \big[ q(\theta) || p_2(\theta) \big] < B \text{ and } \mathbb{E}_q[1]=1, \end{aligned}$$ where $B \in \big[ 0, \textbf{D}_{KL}[ p_1(\theta) || p_2(\theta)] \big]$. On the extremes of $B$, we will have trivial solutions $$\begin{aligned} q^*(\theta)= \left\{ \begin{array}{lr} p_2(\theta) & \text{if} \quad B=0\\ p_1(\theta) & \text{if} \quad B = \textbf{D}_{KL}[ p_1(\theta) || p_2(\theta)]. \end{array} \right.\\ \end{aligned}$$ Note that the Kullback–Leibler divergence is a convex function with respect to its first argument, i.e. $q$ in our setting. Therefore, both the objective function and the constraints of the optimization problem in are convex. For convenience, we assume that the parameter space for $\theta$ is discrete, but the final results can be generalized also to the continuous case with some considerations - see [@beal2003variational] and [@faraji2018balancing]. For the discrete setting, the optimization problem in can be rewritten as $$\begin{aligned} \label{Eq:Genral_Opt_Disc} q^*(\theta) &= \text{argmin } \sum_{\theta} q(\theta) \big( \text{log}(q(\theta)) - \text{log}(p_1(\theta)) \\ & q(\theta) \text{ s.t. } \sum_{\theta} q(\theta) \big( \text{log}(q(\theta)) - \text{log}(p_2(\theta)) < B \text{ and } \sum_{\theta} q(\theta) = 1. \end{aligned}$$ For solving the mentioned problem, one should find a $q$ which satisfies the Karush–Kuhn–Tucker (KKT) conditions [@boyd2004convex] for $$\mathcal{L} = \sum_{\theta} q(\theta) \text{log}\Big( \frac{q(\theta)}{p_1(\theta)} \Big) + \lambda \sum_{\theta} q(\theta) \text{log}\Big( \frac{q(\theta)}{p_2(\theta)} \Big) - \lambda B + \alpha - \alpha \sum_{\theta} q(\theta),$$ $$\begin{aligned} \frac{\partial \mathcal{L}}{\partial q(\theta)} &= \text{log}\Big( \frac{q(\theta)}{p_1(\theta)} \Big) + 1 + \lambda \text{log}\Big( \frac{q(\theta)}{p_2(\theta)} \Big) + \lambda - \alpha \\ &= (1+\lambda)\text{log}( q(\theta)) - \text{log}( p_1(\theta)) - \lambda \text{log}( p_2(\theta)) + 1 + \lambda - \alpha, \end{aligned}$$ where $\lambda$ and $\alpha$ are the parameters of the dual problem. Defining $\gamma = \frac{\lambda}{1+\lambda}$, and considering the partial derivative to be zero, we have $$\text{log}( q^*(\theta)) = (1-\gamma) \text{log}( p_1(\theta)) + \gamma \text{log}( p_2(\theta)) + \text{Const}(\alpha, \gamma),$$ where $\alpha$ is always specified in a way to have $\text{Const}(\alpha, \gamma)$ as the normalization factor $$\begin{aligned} \text{Const} & (\alpha, \gamma) = -\text{log}(Z(\gamma))\\ & \text{where } Z(\gamma) = \sum_{\theta} p_1^{1-\gamma}(\theta) p_2^{\gamma}(\theta). \end{aligned}$$ According to the KKT conditions, $\lambda \geq 0$, and as a result $\gamma \in [0,1]$. Therefore, considering $p_1(\theta)={\hat{\pi}}_B^{(t+1)}(\theta)$ and $p_2(\theta)=P(\theta|y_{t+1})$, the solution to the optimization problem of and is as . #### Proof of the claim that $B$ is a decreasing function of Surprise According to the KKT conditions $$\lambda \Big( \textbf{D}_{KL}[ q^{*}(\theta) || p_2(\theta)] - B \Big) = 0.$$ For the case that $\lambda \neq 0$ (i.e. $\gamma \neq 0$), we have $B$ as a function of $\gamma$ $$\begin{aligned} B(\gamma) &= \textbf{D}_{KL}[ q^{*}(\theta) || p_2(\theta)]\\ &= (1-\gamma) \mathbb{E}_{q^{*}} \Big[ \text{log} \big( \frac{p_1(\theta)}{p_2(\theta)} \big) \Big] - \text{log} \big( Z(\gamma) \big). \end{aligned}$$ Now, we show that the derivate of $B(\gamma)$ with respect to $\gamma$ is always non-positive. To do so, we first compute the derivative of $Z(\gamma)$ as $$\begin{aligned} \frac{\partial \text{log}(Z(\gamma))}{\partial \gamma} &= \frac{1}{Z(\gamma)} \frac{\partial}{\partial \gamma} \sum_{\theta} p_1^{1-\gamma}(\theta) p_2^{\gamma}(\theta)\\ &= \frac{1}{Z(\gamma)} \sum_{\theta} p_1^{1-\gamma}(\theta) p_2^{\gamma}(\theta) \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big)\\ &= \mathbb{E}_{q^{*}} \Big[ \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big)\Big], \end{aligned}$$ and the derivate of $\mathbb{E}_{q^{*}} [ h(\theta)]$ for an arbitrary $h(\theta)$ as $$\begin{aligned} \frac{\partial \mathbb{E}_{q^{*}} [ h(\theta)]}{\partial \gamma} &= \frac{\partial }{\partial \gamma} \sum_{\theta} q^{*}(\theta) h(\theta)\\ &= \sum_{\theta} q^{*}(\theta) h(\theta) \frac{\partial }{\partial \gamma} \text{log}(q^{*}(\theta))\\ &= \sum_{\theta} q^{*}(\theta) h(\theta) \frac{\partial }{\partial \gamma} \Big( (1-\gamma) \text{log}( p_1(\theta)) + \gamma \text{log}( p_2(\theta)) -\text{log}(Z(\gamma))\Big)\\ &= \mathbb{E}_{q^{*}} \Big[ h(\theta) \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big)\Big] - \mathbb{E}_{q^{*}} \big[ h(\theta) \big] \mathbb{E}_{q^{*}} \Big[ \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big)\Big]. \end{aligned}$$ Using the last three equations, we have $$\begin{aligned} \frac{\partial B(\gamma)}{\partial \gamma} &= -(1-\gamma) \Big( \mathbb{E}_{q^{*}} \Big[ \Big( \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big) \Big)^2 \Big] - \mathbb{E}_{q^{*}} \Big[ \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big)\Big]^2 \Big)\\ &= -(1-\gamma) \textbf{Var}_{q^{*}} \Big[ \text{log} \big( \frac{p_2(\theta)}{p_1(\theta)} \big) \Big] \leq 0, \end{aligned}$$ which means that $B$ is a decreasing function of $\gamma$. Since $\gamma$ is an increasing function of surprise, $B$ is also a decreasing function of surprise. Derivations of Message Passing $N$ (Algo. 3) {#derivations-of-message-passing-n-algo.-3 .unnumbered} -------------------------------------------- For the sake of clarity and coherence, we repeat here some steps performed in the Results section. Following the idea of [@adams2007bayesian] let us first define the random variable $R_t = \text{max} \{ n \in \mathbb{N} : H_{t-n+1} = H_{t} \}$. This is the time window from the last change point. Then the exact Bayesian form for $\pi^{(t)}(\theta)$ can be written as $$\begin{aligned} \pi^{(t)}(\theta) &= \textbf{P}(\Theta_{t+1} = \theta |Y_{1:t}=y_{1:t})\\ &= \sum_{r_t=1}^{t} \textbf{P}(R_t = r_t |Y_{1:t}=y_{1:t}) \textbf{P}(\Theta_{t+1} = \theta | R_t = r_t, Y_{1:t}=y_{1:t}). \end{aligned}$$ To have a formulation similar to the one of Particle Filtering we rewrite the belief as $$\begin{aligned} \label{Eq:MPN_particle_form} \pi^{(t)}(\theta) &= \sum_{k=0}^{t-1} w^{(k)}_t \textbf{P}(\Theta_{t+1} = \theta | R_t = t-k, Y_{1:t}=y_{1:t}), \end{aligned}$$ where $w^{(k)}_t = \textbf{P}(R_t = t-k |Y_{1:t}=y_{1:t})$ is the weight of the particle $k$ at time $t$. To update the belief after observing $Y_{t+1} = y_{t+1}$, one can use the exact Bayesian recursive formula , for which one needs to compute $\pi_B^{(t+1)}(\theta)$ as $$\begin{aligned} \pi_B^{(t+1)}(\theta) &= \frac{\pi^{(t)}(\theta) P_Y(y_{t+1}|\theta)} {P(y_{t+1}; \pi^{(t)})}\\ &= \frac{1}{P(y_{t+1}; \pi^{(t)})} \sum_{k=0}^{t-1} w^{(k)}_t \textbf{P}(\Theta_{t+1} = \theta | R_t = t-k, Y_{1:t}=y_{1:t}) P_Y(y_{t+1}|\theta).\\ \end{aligned}$$ Using Bayes’ rule and conditional independence of observations, we have $$\begin{aligned} \pi_B^{(t+1)}(\theta) &= \frac{1}{P(y_{t+1}; \pi^{(t)})} \sum_{k=0}^{t-1} w^{(k)}_t \frac{ \textbf{P}( Y_{k+1:t}=y_{k+1:t} | \Theta_{t+1} = \theta, R_t = t-k) \pi^{(0)}(\theta) } { \textbf{P}( Y_{k+1:t}=y_{k+1:t} | R_t = t-k) } P_Y(y_{t+1}|\theta) \\ &= \frac{1}{P(y_{t+1}; \pi^{(t)})} \sum_{k=0}^{t-1} w^{(k)}_t \frac{ \prod_{i=k+1}^{t+1} P_Y( y_i | \theta) \pi^{(0)}(\theta) } { \textbf{P}( Y_{k+1:t}=y_{k+1:t} | R_t = t-k) }\, , \end{aligned}$$ and once again, by using the Bayes’ rule and the conditional independence of observations, we find $$\begin{aligned} \pi_B^{(t+1)}(\theta) &= \frac{1}{P(y_{t+1}; \pi^{(t)})} \sum_{k=0}^{t-1} w^{(k)}_t \frac{ \textbf{P}( Y_{1:t+1}=y_{1:t+1} | R_{t+1} = t-k+1) } { \textbf{P}( Y_{1:t}=y_{1:t} | R_t = t-k) } \times\\ & \times \textbf{P}(\Theta_{t+1} = \theta | R_{t+1} = t-k+1, Y_{1:t+1}=y_{1:t+1}). \end{aligned}$$ This gives us $$\begin{aligned} \pi_B^{(t+1)}(\theta) = \sum_{k=0}^{t-1} & w^{(k)}_t \frac{ P(y_{t+1}; \pi_k^{(t)}) } { P(y_{t+1}; \pi^{(t)}) } \times\\ & \times \textbf{P}(\Theta_{t+1} = \theta | R_{t+1} = t-k+1, Y_{1:t+1}=y_{1:t+1}), \end{aligned}$$ and finally $$\begin{aligned} w^{(k)}_{B,t+1} = \frac{ P(y_{t+1}; \pi_k^{(t)}) } { P(y_{t+1}; \pi^{(t)}) } w^{(k)}_t. \end{aligned}$$ The update in the last equation is identical to the update of the Particle Filter weights that correspond to a Bayesian update defined in . Using the recursive formula, the update rule for the weights for $\quad 0 \leq k \leq t-1$ is $$\begin{aligned} \label{Eq:SMP_w_oldparticles} & w^{(k)}_{t+1} = (1-\gamma_{t+1}) w^{(k)}_{B,t+1} = (1-\gamma_{t+1}) \frac{ P(y_{t+1}; \pi_k^{(t)}) } { P(y_{t+1}; \pi^{(t)}) } w^{(k)}_t, \end{aligned}$$ and for the newly added particle $t$ $$\begin{aligned} \label{Eq:SMP_w_newparticle} & w^{(t)}_{t+1} = \gamma_{t+1}, \end{aligned}$$ where $\gamma_{t+1} = \gamma \big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)}) , m = \frac{p_c}{1-p_c} \big)$ of . The MP$N$ algorithm uses , , and for computing the belief for $t \leq N$ - which is same as the exact Bayesian inference. For $t > N$, it first updates the weights in the same fashion as and , keeps the greatest $N$ weights, and sets the rest weights equal to 0. After normalizing the new weights, it uses (but only over the particles with non-zero weights) to compute the belief ${\hat{\pi}}^{(t)}$. For the particular case of exponential family, see Algorithm \[Alg:MPN\] for the pseudocode. Surprise-modulation as a framework for other algorithms {#ext_int_related .unnumbered} ------------------------------------------------------- ### SMiLe Rule {#sec:smile .unnumbered} The Confidence Corrected Surprise [@faraji2018balancing] is $$\textbf{S}_{CC}(y_{t+1}; {\hat{\pi}}^{(t)}) = \textbf{D}_{KL} \big[ {\hat{\pi}}^{(t)}(\theta) || \tilde{P}(\theta|y_{t+1}) \big],$$ where $\tilde{P}(\theta|y_{t+1})$ is the scaled likelihood defined as $$\label{Eq:scaled_like} \tilde{P}(\theta|y_{t+1}) = \frac{P_Y(y_{t+1}|\theta)}{\int P_Y(y_{t+1}|\theta') d \theta'}.$$ Note that this becomes equal to $P(\theta|y_{t+1})$ if the prior belief $\pi^{(0)}$ is a uniform distribution; cf. . With the aim of minimizing the Confidence Corrected Surprise by updating the belief during time, [@faraji2018balancing] suggested an update rule solving the optimization problem $$\begin{aligned} \label{Eq:Mod_SMiLe_Min_1} {\hat{\pi}}^{(t+1)}(\theta) &= \arg \min_q \textbf{D}_{KL} \big[ q(\theta) || \tilde{P}(\theta|y_{t+1}) \big] \\ & \text{s.t. } \textbf{D}_{KL} \big[ q(\theta) || {\hat{\pi}}^{(t)}(\theta) \big] \leq B_{t+1}, \end{aligned}$$ where $B_{t+1} \in \big[ 0, \textbf{D}_{KL}[ P(\theta|y_{t+1}) || {\hat{\pi}}^{(t)}(\theta)] \big]$ is an arbitary bound. The authors showed that the solution to this optimization problem is $$\label{Eq:Mod_SMiLe_Log_Rec} \text{log} \big( {\hat{\pi}}^{(t+1)}(\theta) \big) = (1-\gamma_{t+1}) \text{ log} \big( {\hat{\pi}}^{(t)}(\theta) \big) + \gamma_{t+1} \text{ log} \big( \tilde{P}(\theta|y_{t+1}) \big) + \text{Const.},$$ where $\gamma_{t+1} \in [0,1]$ is specified so that it satisfies the constraint in . Although looks very similar to , it signifies a trade-off between the latest belief ${\hat{\pi}}^{(t)}$ and the belief updated by only the most recent observation $\tilde{P}(\theta|y_{t+1})$, i.e. a trade-off between adherence to the current belief and reset. While SMiLe adheres to the current belief ${\hat{\pi}}^{(t)}$, Variational SMiLe integrates the new observation with the current belief to get ${\hat{\pi}}_B^{(t)}$, which leads to a trade-off similar to the one of the exact Bayesian inference ( and ) and of Particle Filtering (). To modulate the learning rate by surprise, [@faraji2018balancing] considered the boundary $B_{t+1}$ as a function of the Confidence Corrected Surprise, i.e. $$\label{Eq:SMiLe_modulation} \begin{aligned} B_{t+1} &= B_{\text{max}} \gamma \Big(\textbf{S}_{CC}(y_{t+1}) , m \Big)\\ & \text{where } B_{\text{max}} = \textbf{D}_{KL}[ P(\theta|y_{t+1}) || {\hat{\pi}}^{(t)}(\theta)]\, , \end{aligned}$$ where $m$ is a free parameter. Then, $\gamma_{t+1}$ is found by satisfying the constraint of the optimization problem in using and . ### Nassar’s algorithm {#nassars-algorithm .unnumbered} For the particular case that observations are drawn from a Gaussian distribution with known variance and unknown mean, i.e. $y_{t+1}|\mu_{t+1} \sim \mathcal{N}(\mu_{t+1}, \sigma^2)$ and $\theta_t = \mu_t$, [@nassar2012rational; @nassar2010approximately] considered the problem of estimating the expected $\mu_t$ and its variance rather than a probability distribution (i.e. belief) over it, implicitly assuming that the belief is always a Gaussian distribution. The algorithms of [@nassar2012rational; @nassar2010approximately] were developed for the case that, whenever the environment changes, the mean $\mu_{t+1}$ is drawn from a uniform prior with a range of values much larger than the width of the Gaussian likelihood function. The authors showed that in this case, the expected $\mu_{t+1}$ (i.e. ${\hat{\mu}}_{t+1}$) estimated by the agent upon observing a new sample $y_{t+1}$ is $$\label{Eq:nassar_original} {\hat{\mu}}_{t+1} = {\hat{\mu}}_t + \alpha_{t+1} ( y_{t+1} - {\hat{\mu}}_t),$$ with $\alpha_{t+1}$ the adaptive learning rate given by $$\label{Eq:nassar_original_learningrate} \alpha_{t+1} = \frac{1 + \Omega_{t+1}{\hat{r_t}}}{1 + {\hat{r_t}}},$$ where ${\hat{r_t}}$ is the estimated time since the last change point (i.e. the estimated $r_t = \text{max} \{ n \in \mathbb{N} : H_{t-n+1} = H_{t}$) and $\Omega_{t+1} = P(\text{change}|y_{t+1})$ the probability of a change given the observation. Note that this quantity, i.e. the posterior change point probability, is the same as our adaptation rate $\gamma_{t+1}$ of . In the next section, we extend their approach to a more general case where the prior is a Gaussian distribution with arbitrary variance, i.e. $\mu_{t+1} \sim \mathcal{N}(\mu_0, \sigma_0^2)$. We then discuss the relation of this method to Particle Filtering. ### Nas10$^{*}$ and Nas12$^{*}$ algorithms {#sec:nassar_star .unnumbered} Let us consider that $y_{1:t}$ are observed, the time since the last change point $r_t$ is known, and the agent’s current estimation of $\mu_t$ is ${\hat{\mu}}_{t}$. It can be shown (see Supplementary Material for the derivation) that the expected $\mu_{t+1}$ (i.e. ${\hat{\mu}}_{t+1}$) upon observing the new sample $y_{t+1}$ is $$\label{Eq:Nassar_post_m_4b} {\hat{\mu}}_{t+1} = (1 - \gamma_{t+1})\Big( {\hat{\mu}}_t + \frac{1}{\rho + r_t + 1} (y_{t+1} - {\hat{\mu}}_t) \Big) + \gamma_{t+1} \Big( \mu_0 + \frac{1}{\rho + 1} (y_{t+1} - \mu_0) \Big)\, ,$$ where $\rho = \frac{\sigma^2}{\sigma^2_0} $, $\mu_0$ is the mean of the prior distribution and $\gamma_{t+1}$ is the adaptation rate of . We can see that the updated mean is a weighted average, with surprise-modulated weights, between integrating the new observation with the current mean ${\hat{\mu}}_t$ and integrating it with the prior mean $\mu_0$, in the same spirit as the other algorithm we considered here. can also be seen as a surprise-modulated weighted sum of two delta rules: one including a prediction error between the new observation and the current mean $(y_{t+1} - {\hat{\mu}}_t)$ and one including a prediction error between the observed sample and the prior mean $(y_{t+1} - \mu_0)$. In order to obtain a form similar to the one of [@nassar2012rational; @nassar2010approximately], we can rewrite the above formula as $$\label{Eq:Nassar_post_m_final_2} {\hat{\mu}}_{t+1} = \frac{\rho}{\rho + 1} \Big({\hat{\mu}}_t + \gamma_{t+1}(\mu_0 - {\hat{\mu}}_t) \Big) + \frac{1}{\rho + 1} \Big( {\hat{\mu}}_t + \alpha_{t+1} (y_{t+1} - {\hat{\mu}}_t) \Big)\, ,$$ where we have defined $\alpha_{t+1} = \frac{\rho + \gamma_{t+1} r_t +1}{\rho + r_t + 1}$. Hence the update rule takes the form of a weighted average, with fixed weights, between two delta rules: one including a prediction error between the prior mean and the current mean $(\mu_0 - {\hat{\mu}}_t)$ and one including a prediction error between the observed sample and the current mean $(y_{t+1} - {\hat{\mu}}_t)$, both with surprise-modulated learning rates. In [@nassar2012rational; @nassar2010approximately] the true new mean after a change point is drawn from a uniform distribution with a range of values much larger than the width of the Gaussian likelihood. Their derivations implicitly approximate the uniform distribution with a Gaussian distribution with $\sigma_0 \gg \sigma$. Note that if $\sigma_0 \gg \sigma$ then $\rho \rightarrow 0$, so that the first term of disappears, and $\alpha_{t+1} = \frac{1 + \gamma_{t+1} r_t}{1 + r_t}$. This results in the delta-rule of the original algorithm in and , with $\gamma_{t+1}=\Omega_{t+1}$. All of the calculations so far were done by assuming that $r_t$ is known. However, for the case of a non-stationary regime with a history of change points, the time interval $r_t$ is not known. [@nassar2012rational; @nassar2010approximately] used the expected time interval ${\hat{r}}_t$ as an estimate. We make a distinction here between [@nassar2012rational] and [@nassar2010approximately]: In [@nassar2010approximately] ${\hat{r}}_t$ is calculated recursively on each trial in the same spirit as : ${\hat{r}}_{t+1} = (1 - \gamma_{t+1})({\hat{r}}_t + 1) + \gamma_{t+1}$, i.e., at each time step, there is a probability $(1 - \gamma_{t+1})$ that ${\hat{r}}_t$ increments by 1 and a probability $\gamma_{t+1}$ that it is reset to 1. So ${\hat{r}}_{t+1}$ is the weighted sum of these two outcomes. Hence, combined with the expected time interval ${\hat{r}}_t$ constitutes a generalization of the update rule of [@nassar2010approximately] for the case of Gaussian prior $\mathcal{N}(\mu_0, \sigma_0^2)$. We call this algorithm Nas10$^{*}$ (see Alg. \[Alg:Nas10\] in Supplementary Material). In [@nassar2012rational], the variance ${\hat{\sigma}}^2_{t+1} = \text{Var}[\mu_{t+1} | y_{1:t+1}]$ is estimated given ${\hat{\mu}}_t$, ${\hat{r}}_t$, and ${\hat{\sigma}}^2_{t}$. Based on this variance, ${\hat{r}}_{t+1} = \frac{\sigma^2}{{\hat{\sigma}}^2_{t+1}} - \frac{\sigma^2}{\sigma^2_{0}}$ is computed. The derivation of the recursive computation of ${\hat{\sigma}}^2_{t+1}$ for the case of Gaussian priors can be found in the Supplementary Material. We call the combination of with this way of computing the expected time interval ${\hat{r}}_t$ Nas12$^{*}$ (see Alg. \[Alg:Nas12\] in Supplementary Material). These two versions of calculating ${\hat{r}}_t$ in [@nassar2010approximately] and [@nassar2012rational] give different results, and we compare our algorithms with both Nas10$^{*}$ and Nas12$^{*}$ in our simulations. Note that, as discussed in the section “Online Bayesian inference modulated by surprise” of the Results, the posterior belief at time $t+1$ does not generally belong to the same family of distributions as the belief of time $t$. However, we therefore approximate for both algorithms the posterior belief $P(\theta | y_{1:t+1})$ by a Gaussian. ### Nassar’s algorithm and Particle Filtering with one particle {#sec:nassar_pf1 .unnumbered} In the case of Particle Filtering (cf. ) with only one particle, at each time step we sample the particle’s hidden state with change probability $Q(\Delta h_{t+1}^{(1)} = 1| \Delta h_{1:t}^{(1)}, y_{1:t+1}) = \gamma_{t+1}$, generating a posterior belief that takes two possible values with probability (according to the proposal distribution) $$\begin{aligned} \label{Eq:pf1_postbelief} Q({\hat{\pi}}^{(t+1)}(\theta) = {\hat{\pi}}^{(t+1)}_B(\theta) | \ y_{t+1}) &= 1 - \gamma_{t+1},\\ Q({\hat{\pi}}^{(t+1)}(\theta) = P(\theta|y_{t+1}) | \ y_{t+1}) &= \gamma_{t+1}. \end{aligned}$$ So, *in expectation*, the updated belief will be $$\begin{aligned} \label{Eq:pf1_expectedpostbelief} \mathbb{E}_{Q}[{\hat{\pi}}^{(t+1)}(\theta)] = (1 - \gamma_{t+1}){\hat{\pi}}^{(t+1)}_B(\theta) + \gamma_{t+1} P(\theta|y_{t+1}). \end{aligned}$$ If we apply to ${\hat{\mu}}_{t+1}$, we find that $\mathbb{E}_{Q}[{\hat{\mu}}^{(t+1)}] $, is identical to the generalization of [@nassar2010approximately] (see ). Moreover, in Particle Filtering with a single particle, we sample the particle’s hidden state, which is equivalent to sampling the interval ${\hat{R}}_{t+1}$. Since ${\hat{R}}_{t+1}$ takes the value ${\hat{r}}_t + 1$ with $(1 - \gamma_{t+1})$ and the value of $1$ (=reset) with probability $\gamma_{t+1}$, the *expected value* of ${\hat{R}}_{t+1}$ is $$\label{Eq:pf1_expectedrt+1} \mathbb{E}_{Q}[{\hat{R}}_{t+1}] = (1 - \gamma_{t+1})({\hat{r}}_t + 1) + \gamma_{t+1}.$$ In other words, in [@nassar2010approximately], the belief is updated based on the *expected* ${\hat{r}}_t$, whereas in Particle Filtering with one particle, the belief is updated using the *sampled* ${\hat{r}}_t$. In summary, the two methods will give different estimates on a trial-per-trial basis, but the same result in expectation. The pseudocode for Particle Filtering with one particle for the particular case of the Gaussian estimation task can be found in Alg. \[Alg:PF\_Gaussian\] in the Supplementary Material. Application to the exponential family {#expfam .unnumbered} ------------------------------------- For our all three algorithms Particle Filtering, Variational SMiLe, and Message Passing with fixed number $N$ of particles, we derive compact update rules for ${\hat{\pi}}^{(t+1)}(\theta)$ when the likelihood function $P_Y(y|\theta)$ is in the exponential family and $\pi^{(0)}(\theta)$ is its conjugate prior. In that case, the likelihood function has the form $$\label{Eq:exp_fam_likelihood} P_Y(y|\theta) = h(y)\text{exp}\big(\theta^T \phi(y) - A(\theta)\big),$$ where $\theta$ is the vector of natural parameters, $h(y)$ is a positive function, $\phi(y)$ is the vector of sufficient statistics, and $A(\theta)$ is the normalization factor. Then, the conjugate prior $\pi^{(0)}$ has the form $$\label{Eq:exp_fam_conj_prior} \begin{aligned} \pi^{(0)}(\theta) &= \textbf{P}_{\pi}\big(\Theta = \theta; \chi^{(0)}, \nu^{(0)}\big)\\ &= \tilde{h}(\theta) f\big(\chi^{(0)}, \nu^{(0)}\big) \text{exp}\big(\theta^T \chi^{(0)} - \nu^{(0)} A(\theta)\big) \end{aligned}$$ where $\chi^{(0)}$ and $\nu^{(0)}$ are the distribution parameters, $\tilde{h}(\theta)$ is a positive function, and $f\big(\chi^{(0)}, \nu^{(0)}\big)$ is the normalization factor. For this setting and while $\pi^{(t)}=\textbf{P}_{\pi}\big(\Theta = \theta; \chi^{(t)}, \nu^{(t)}\big)$, the “Bayes Factor surprise” has the compact form $$\label{Eq:S_GM_exp} \textbf{S}_{\mathrm{BF}} \Big( y_{t+1};\textbf{P}_{\pi}\big(\Theta = \theta; \chi^{(t)}, \nu^{(t)}\big) \Big) = \frac{f\big(\chi^{(t)}+\phi(y_{t+1}), \nu^{(t)}+1\big)}{f\big(\chi^{(0)}+\phi(y_{t+1}), \nu^{(0)}+1\big)} \frac{f\big(\chi^{(0)}, \nu^{(0)}\big)}{f\big(\chi^{(t)}, \nu^{(t)}\big)}\, .$$ The pseudocode for Particle Filtering, Variational SMiLe, and MP$N$ can be seen in Algorithms \[Alg:PF\], \[Alg:Var\_SMiLe\], and \[Alg:MPN\], respectively. Simulation task {#simulation-task .unnumbered} --------------- In this subsection, we first argue why the mean squared error is a proper measure for comparing different algorithms with each other, and then we explain the version of Leaky integrator which we used for simulations. ### Mean squared error as an optimality measure {#mean-squared-error-as-an-optimality-measure .unnumbered} Consider the case that at each time point $t$, the goal of an agent is to have an estimation of the parameter $\Theta_t$ as a function of the observations $Y_{1:t}$, i.e. ${\hat{\Theta}}_t = f(Y_{1:t})$. The estimator which minimizes the mean squared error $\textbf{MSE} [ {\hat{\Theta}}_t ] = \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \big[ ( {\hat{\Theta}}_t - \Theta_t)^2 \big]$ is $${\hat{\Theta}}_t^{\text{Opt}} = \mathbb{E}_{\textbf{P}(\Theta_{t}|Y_{1:t} )} \big[ \Theta_t \big] = \mathbb{E}_{\pi^{(t)}} \big[ \Theta_t \big],$$ which is the expected value of $\Theta_t$ conditioned on the observations $Y_{1:t}$, or in other words under the Bayes-optimal current belief (see [@papoulis1989probability] for a proof). The MSE for any other estimator ${\hat{\Theta}}_t$ can be written as (see below for the proof) $$\begin{aligned} \label{eq:MSE_dMSE} &\textbf{MSE} [ {\hat{\Theta}}_t ] = \textbf{MSE} [ {\hat{\Theta}}_t^{\text{Opt}} ] + \Delta \textbf{MSE} [ {\hat{\Theta}}_t ],\\ & \text{ where } \Delta \textbf{MSE} [ {\hat{\Theta}}_t ] = \mathbb{E}_{\textbf{P}(Y_{1:t})} \Big[ ({\hat{\Theta}}_t - {\hat{\Theta}}^{\text{Opt}}_t)^2 \Big] \geq 0. \end{aligned}$$ This means that the MSE for any arbitrary estimator ${\hat{\Theta}}_t$ includes two terms: the optimal MSE and the mismatch of the actual estimator from the optimal estimator ${\hat{\Theta}}_t^{\text{Opt}}$. As a result, if the estimator we are interested in is the expected value of $\Theta_t$ under the approximate belief ${\hat{\pi}}^{(t)}$ computed by each of our algorithms (i.e. ${\hat{\Theta}}_t' = \mathbb{E}_{{\hat{\pi}}^{(t)}} \big[ \Theta_t \big]$), the second term in , i.e. the deviation from optimality, is a measure of how good the approximation is. **Proof for the algorithms without sampling:** Consider ${\hat{\Theta}}^{\text{Opt}}_t = f_{\text{Opt}}(Y_{1:t})$. Then, for any other arbitrary estimator ${\hat{\Theta}}_t = f(Y_{1:t})$ (except for the ones with sampling), we have $$\begin{aligned} \textbf{MSE} [ {\hat{\Theta}}_t ] &= \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \big[ ( {\hat{\Theta}}_t - \Theta_t)^2 \big]\\ &= \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \big[ ( f(Y_{1:t}) - \Theta_t)^2 \big]\\ &= \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \Big[ \big( (f(Y_{1:t}) - f_{\text{Opt}}(Y_{1:t})) + (f_{\text{Opt}}(Y_{1:t}) - \Theta_t) \big)^2 \Big]. \end{aligned}$$ The quadratic term in the last line can be expanded and written as $$\begin{aligned} \textbf{MSE} [ {\hat{\Theta}}_t ] = &\mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \Big[ (f(Y_{1:t}) - f_{\text{Opt}}(Y_{1:t}))^2 \Big] +\\ & \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \Big[ (f_{\text{Opt}}(Y_{1:t}) - \Theta_t)^2 \Big]+\\ & 2\mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)} \Big[ (f(Y_{1:t}) - f_{\text{Opt}}(Y_{1:t}))(f_{\text{Opt}}(Y_{1:t}) - \Theta_t) \Big]. \end{aligned}$$ The random variables in the expected value of the first line are not dependent on $\Theta_t$, so it can be computed over $Y_{1:t}$. The expected value of the second line is equal to $\textbf{MSE} [ {\hat{\Theta}}^{\text{Opt}}_t ]$. It can also be shown that the expected value of the third line is equal to 0, i.e. $$\begin{aligned} \text{3rd line } &= 2 \mathbb{E}_{\textbf{P}(Y_{1:t})} \Big[ \mathbb{E}_{\textbf{P}(\Theta_t|Y_{1:t})} \Big[(f(Y_{1:t}) - f_{\text{Opt}}(Y_{1:t}))(f_{\text{Opt}}(Y_{1:t}) - \Theta_t) \Big] \Big]\\ &= 2 \mathbb{E}_{\textbf{P}(Y_{1:t})} \Big[ (f(Y_{1:t}) - f_{\text{Opt}}(Y_{1:t}))(f_{\text{Opt}}(Y_{1:t}) - \mathbb{E}_{\textbf{P}(\Theta_t|Y_{1:t})} [\Theta_t]) \Big] = 0, \end{aligned}$$ where in the last line we used the definition of the optimal estimator. All together, we have $$\begin{aligned} \textbf{MSE} [ {\hat{\Theta}}_t ] = \textbf{MSE} [ {\hat{\Theta}}^{\text{Opt}}_t ] + \mathbb{E}_{\textbf{P}(Y_{1:t})} \Big[ ({\hat{\Theta}}_t - {\hat{\Theta}}^{\text{Opt}}_t)^2 \Big]. \end{aligned}$$ **Proof for the algorithms with sampling:** For particle filtering (and any kind of estimator with sampling), the estimator is not a deterministic function of observations $Y_{1:t}$. Rather, the estimator is a function of observations as well as a set of random variables (samples) which are drawn from a distribution which is also a function of observations $Y_{1:t}$. In our case, the samples are the sequence of hidden states $H_{1:t}$. The estimator can be written as $${\hat{\Theta}}^{\text{PF}}_t = f(Y_{1:t}, H_{1:t}^{(1:N)}),$$ where $H_{1:t}^{(1:N)}$ are $N$ iid samples drawn from the proposal distribution $Q(H_{1:t}|Y_{1:t})$. MSE for this estimator should also be averaged over the samples, which leads to $$\begin{aligned} \textbf{MSE} [ {\hat{\Theta}}^{\text{PF}}_t ] &= \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)Q(H_{1:t}^{(1:N)}|Y_{1:t})} \big[ ( {\hat{\Theta}}^{\text{PF}}_t - \Theta_t)^2 \big]\\ &= \mathbb{E}_{\textbf{P}(Y_{1:t},\Theta_t)Q(H_{1:t}^{(1:N)}|Y_{1:t})} \big[ ( f(Y_{1:t}, H_{1:t}^{(1:N)}) - \Theta_t)^2 \big]. \end{aligned}$$ Similar to what we did before, the MSE for particle filtering can be written as $$\begin{aligned} \textbf{MSE} [ {\hat{\Theta}}^{\text{PF}}_t ] &= \textbf{MSE} [ {\hat{\Theta}}^{\text{Opt}}_t ] + \mathbb{E}_{\textbf{P}(Y_{1:t})Q(H_{1:t}^{(1:N)}|Y_{1:t})} \Big[ ({\hat{\Theta}}^{\text{PF}}_t - {\hat{\Theta}}^{\text{Opt}}_t)^2 \Big]\\ &= \textbf{MSE} [ {\hat{\Theta}}^{\text{Opt}}_t ] + \mathbb{E}_{\textbf{P}(Y_{1:t})} \Big[ \mathbb{E}_{Q(H_{1:t}^{(1:N)}|Y_{1:t})} \big[ ({\hat{\Theta}}^{\text{PF}}_t - {\hat{\Theta}}^{\text{Opt}}_t)^2 \big] \Big] \end{aligned}$$ which can be written in terms of bias and variance over samples as $$\begin{aligned} \textbf{MSE} [ {\hat{\Theta}}^{\text{PF}}_t ] = &\textbf{MSE} [ {\hat{\Theta}}^{\text{Opt}}_t ]\\ &+ \mathbb{E}_{\textbf{P}(Y_{1:t})} \Big[ \text{Var}_{Q(H_{1:t}^{(1:N)}|Y_{1:t})}({\hat{\Theta}}^{\text{PF}}_t) + \text{Bias}_{Q(H_{1:t}^{(1:N)}|Y_{1:t})}({\hat{\Theta}}^{\text{PF}}_t,{\hat{\Theta}}^{\text{Opt}}_t)^2 \Big]. \end{aligned}$$ ### Leaky integration {#leaky-integration .unnumbered} Gaussian task: The goal is to have an estimation of the mean of the Gaussian distribution at each time $t$, denoted by ${\hat{\theta}}_t$. Given a leak parameter $\omega \in (0,1]$, the leaky integrator estimation is $$\begin{aligned} {\hat{\theta}}_t = \frac{\sum_{k=1}^{t} \omega^{t-k} y_k }{ \sum_{k=1}^{t} \omega^{t-k} }. \end{aligned}$$ Categorical task: The goal is to have an estimation of the parameters of the categorical distribution at each time $t$, denoted by ${\hat{\theta}}_t = [{\hat{\theta}}_{i,t}]_{i=1}^{N}$ for the case that there are $N$ categories. Given a leak parameter $\omega \in (0,1]$, the leaky integrator estimation is $$\begin{aligned} {\hat{\theta}}_{i,t} = \frac{\sum_{k=1}^{t} \omega^{t-k} \delta(y_k - i) }{ \sum_{k=1}^{t} \omega^{t-k} }, \end{aligned}$$ where $\delta$ is the Kronecker delta function. Derivation of the Formula Relating Shannon Surprise to the Modulated Learning Rate {#derivation-of-the-formula-relating-shannon-surprise-to-the-modulated-learning-rate .unnumbered} ---------------------------------------------------------------------------------- Given the defined generative model, the Shannon surprise upon observing $Y_{t+1} = y_{t+1}$ can be written as $$\label{Eq:S_Shannon} \begin{aligned} \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(t)}) &= \text{log} \Big( \frac{1}{\textbf{P}(Y_{t+1} = y_{t+1}|Y_{1:t} = y_{1:t})} \Big)\\ &= \text{log} \Big( \frac{1}{(1-p_c) P(y_{t+1}; \pi^{(t)}) + p_c P(y_{t+1}; \pi^{(0)})} \Big)\\ &= \text{log} \Big( \frac{1}{P(y_{t+1}; \pi^{(0)})} \Big) + \text{log} \Big( \frac{1}{ p_c } \frac{1}{1 + \frac{1}{m} \frac{1}{\textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)})}} \Big)\\ &= \textbf{S}_{\mathrm{Sh}}(y_{t+1}; \pi^{(0)}) + \text{log} \Big( \frac{\gamma_{t+1}}{ p_c } \Big), \end{aligned}$$ where $\gamma_{t+1} = \gamma \Big(\textbf{S}_{\mathrm{BF}}(y_{t+1}; \pi^{(t)}) , m = \frac{p_c}{1-p_c} \Big) $ of . As a result, the modulated adaptation rate can be written as in and the Bayes Factor Surprise as in . Experimental predictions {#experimental-predictions .unnumbered} ------------------------ ### Setting {#setting .unnumbered} Consider a Gaussian task where $Y_t$ can take values in $\mathbb{R}$. The likelihood function $P_Y(y|\theta)$ is defined as $$\begin{aligned} P_Y(y|\theta) = \mathcal{N}(y; \theta, \sigma^2), \end{aligned}$$ where $\sigma \in \mathbb{R}^{+}$ is the standard deviation, and $\theta \in \mathbb{R}$ is the mean of the distribution, i.e. the parameter of the likelihood. Whenever there is a change in the environment (with probability $p_c \in (0,1)$), the value $\theta$ is drawn from the prior distribution $\pi^{(0)}(\theta) = \mathcal{N}(\theta; 0, 1)$. ### Theoretical proofs for prediction 1 {#theoretical-proofs-for-prediction-1 .unnumbered} For our theoretical derivations for our first prediction, we consider the specific but relatively mild assumption that the subjects’ belief $\pi^{(t)}$ at each time is a Gaussian distribution $$\begin{aligned} \pi^{(t)}(\theta) = \mathcal{N}(\theta; {\hat{\theta}}_t, {\hat{\sigma}}_t^2), \end{aligned}$$ where ${\hat{\theta}}_t$ and ${\hat{\sigma}}_t$ are determined by the learning algorithm and the sequence of observations $y_{1:t}$. This is the case when the subjects use either VarSMiLe, Nas10$^{*}$, Nas12$^{*}$, pf1, MP1, or Leaky Integration as their learning rule. With such assumptions, the inferred probability distribution $P(y;\pi^{(t)})$ can be written as $$\begin{aligned} P(y;\pi^{(t)}) = \mathcal{N}(y ; {\hat{\theta}}_t, \sigma^2 + {\hat{\sigma}}_t^2). \end{aligned}$$ As mentioned in the Results section, we define, at time $t$, the prediction error as $\delta_{t+1} = y_{t+1} - {\hat{\theta}}_{t}$ and the “sign bias” as $s_{t+1} = \text{sign}(\delta_{t+1} {\hat{\theta}}_{t})$. Then, given an absolute prediction ${\hat{\theta}}>0$, an absolute prediction error $\delta>0$, a standard deviation $\sigma_C$, and a sign bias $s \in \{ -1, 1 \}$, the average Bayes Factor Surprise is computed as $$\begin{aligned} \Bar{\textbf{S}}_{\mathrm{BF}}&({\hat{\theta}}, \delta,s,\sigma_C) = \frac{1}{|\mathcal{T}|} \sum_{t \in \mathcal{T}} \textbf{S}_{\mathrm{BF}}(y_{t}; {\hat{\pi}}^{(t-1)}),\\ & \text{where } \mathcal{T} = \{ t: |{\hat{\theta}}_{t-1}| = {\hat{\theta}}, |\delta_t| = \delta, {\hat{\sigma}}_t = \sigma_C, s_t = s \}. \end{aligned}$$ It can easily be shown that the value $\textbf{S}_{\mathrm{BF}}(y_{t}; {\hat{\pi}}^{(t-1)})$ is same for all $t \in \mathcal{T}$, and hence the average surprise is same as the surprise for each time point. For example, the average surprise for $s=+1$ is equal to $$\begin{aligned} \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=+1,\sigma_C) = \frac{\mathcal{N}({\hat{\theta}}+ \delta ; 0, \sigma^2 + 1)} {\mathcal{N}(\delta ; 0, \sigma^2 + \sigma_C^2)}. \end{aligned}$$ Similar formulas can be computed for $\Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}}, \delta,s=-1,\sigma_C)$. Then, the difference $\Delta \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,\sigma_C) = \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=+1,\sigma_C) - \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=-1,\sigma_C)$ can be computed as $$\begin{aligned} \Delta \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,\sigma_C) = \frac{\mathcal{N}({\hat{\theta}}+ \delta ; 0, \sigma^2 + 1) - \mathcal{N}({\hat{\theta}} - \delta ; 0, \sigma^2 + 1)}{\mathcal{N}(\delta ; 0, \sigma^2 + \sigma_C^2)}. \end{aligned}$$ It can be shown that $$\begin{aligned} \Delta \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,\sigma_C) < 0 \text{ and } \frac{\partial}{\partial \delta} \Delta \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,\sigma_C) < 0, \end{aligned}$$ for all ${\hat{\theta}}>0$, $\delta>0$, and $\sigma_C>0$. The first inequality is trivial, and the proof for the second inequality is given below. The average Shannon Surprise can be computed in a similar way. For example, for $s=+1$, we have $$\begin{aligned} \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,s=+1,\sigma_C) &= - \log \Big( p_c \mathcal{N}({\hat{\theta}} + \delta ; 0, \sigma^2 + 1) + (1-p_c) \mathcal{N}(\delta ; 0, \sigma^2 + \sigma_C^2) \Big), \end{aligned}$$ and then the difference $\Delta \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,\sigma_C) = \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,s=+1,\sigma_C) - \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,s=-1,\sigma_C)$ can be computed as $$\begin{aligned} \label{Eq:Shannon_diff_experiment} \Delta \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,\sigma_C) = \log \Big( \frac{1 + m \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=-1,\sigma_C)}{1 + m \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=+1,\sigma_C)} \Big), \end{aligned}$$ where $m=\frac{p_c}{1-p_c}$. Then, using the results for the Bayes Factor Surprise, we have $$\begin{aligned} \Delta \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,\sigma_C) > 0 \text{ and } \frac{\partial}{\partial \delta} \Delta \Bar{\textbf{S}}_{Sh}({\hat{\theta}},\delta,\sigma_C) > 0, \end{aligned}$$ for all ${\hat{\theta}}>0$, $\delta>0$, and $\sigma_C>0$. See below for the proof of the second inequality. #### Proof of the 2nd inequality for $\textbf{S}_{\mathrm{BF}}$: Let us define the variables $$\begin{aligned} \sigma_d^2 &= \sigma^2 + \sigma_C^2 \quad \sigma_n^2 = \sigma^2 + 1,\\ \end{aligned}$$ as well as the functions $$\begin{aligned} f_1(\delta) &= \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=+1,\sigma_C) = \frac{\sigma_d}{\sigma_n} \text{exp}\big( \frac{\delta^2}{2\sigma_d^2} - \frac{(\delta + {\hat{\theta}})^2}{2\sigma_n^2} \big)\\ f_2(\delta) &= \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=-1,\sigma_C) = \frac{\sigma_d}{\sigma_n} \text{exp}\big( \frac{\delta^2}{2\sigma_d^2} - \frac{(\delta - {\hat{\theta}})^2}{2\sigma_n^2} \big)\\ f(\delta) &= \Delta \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,\sigma_C) = f_1(\delta) - f_2(\delta). \end{aligned}$$ The following inequalities hold true $$\begin{aligned} f(\delta) <0 & \Rightarrow f_1(\delta) < f_2(\delta)\\ \sigma_C^2 < \sigma^2_0 = 1 & \Rightarrow \sigma_d^2 < \sigma_n^2. \end{aligned}$$ Then, the derivative of $f(\delta)$ can be compute as $$\begin{aligned} \frac{d}{d\delta}f(\delta) &= f_1(\delta) \big( \frac{\delta}{\sigma_d^2} - \frac{\delta + {\hat{\theta}}}{\sigma_n^2} \big) - f_2(\delta) \big( \frac{\delta}{\sigma_d^2} - \frac{\delta - {\hat{\theta}}}{\sigma_n^2} \big) \\ &= - \frac{{\hat{\theta}}}{\sigma_n^2} \big( f_1(\delta) + f_2(\delta) \big) \Big( 1 - \frac{f_1(\delta) - f_2(\delta)}{f_1(\delta) + f_2(\delta)} \frac{\delta}{{\hat{\theta}}} \big( \frac{\sigma_n^2}{\sigma_d^2} -1 \big) \Big)\\ &< 0. \end{aligned}$$ Therefore we have $\frac{\partial}{\partial \delta} \Delta \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,\sigma_C) = \frac{d}{d\delta}f(\delta) <0$. #### Proof of the 2nd inequality for $\textbf{S}_{\mathrm{Sh}}$: Using the functions we defined for the previous proof, and after computing the partial derivative of , we have $$\ \begin{aligned} \frac{\partial}{\partial \delta} \Delta\Bar{\textbf{S}}_{\mathrm{Sh}}({\hat{\theta}},\delta,\sigma_C) &= \frac{\partial}{\partial \delta} \log \Big( \frac{1 + m \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=-1,\sigma_C)}{1 + m \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}},\delta,s=+1,\sigma_C)} \Big)\\ &= \frac{d}{d \delta} \log \Big( \frac{1 + m f_2 (\delta)}{1 + m f_1 (\delta)} \Big). \end{aligned}$$ The derivative of the last term can be written in terms of the derivates of $f_1$ and $f_2$, indicated by $f_1'$ and $f_2'$, respectively, $$\ \begin{aligned} \frac{\partial}{\partial \delta} \Delta\Bar{\textbf{S}}_{\mathrm{Sh}}({\hat{\theta}},\delta,\sigma_C) &= \frac{m f_2' (\delta)}{1 + m f_2 (\delta)}-\frac{m f_1' (\delta)}{1 + m f_1 (\delta)}\\ &= \frac{ - m f' (\delta)}{\big( 1 + m f_1 (\delta) \big) \big( 1 + m f_2 (\delta) \big)} + \frac{ m^2 (f_1 f_2'-f_1' f_2) (\delta)}{\big( 1 + m f_1 (\delta) \big) \big( 1 + m f_2 (\delta) \big)}. \end{aligned}$$ The 1st term is always positive based on the proof for $\textbf{S}_{\mathrm{BF}}$. The 2nd term is also always positive, since $$\ \begin{aligned} (f_1 f_2'-f_1' f_2) (\delta) &= f_1(\delta)f_2(\delta) \Big( \big( \frac{\delta}{\sigma_d^2} - \frac{\delta - {\hat{\theta}}}{\sigma_n^2} \big) - \big( \frac{\delta}{\sigma_d^2} - \frac{\delta + {\hat{\theta}}}{\sigma_n^2} \big) \Big) \\ &= f_1(\delta)f_2(\delta) \frac{2 {\hat{\theta}}}{\sigma_n^2} > 0. \end{aligned}$$ As a result, we have $\frac{\partial}{\partial \delta} \Delta\Bar{\textbf{S}}_{\mathrm{Sh}}({\hat{\theta}},\delta,\sigma_C)>0$. ### Simulation procedure for prediction 1 {#simulation-procedure-for-prediction-1 .unnumbered} In order to relax the main assumption of our theoretical proofs (i.e. the belief is always a Gaussian distribution), to include the practical difficulties of a real experiment (e.g. to use $|\delta_t| \approx \delta$ instead of $|\delta_t| = \delta$), and to have an estimation of the effect size, we also performed simulations for our first experimental prediction. For each simulated subject, the procedure of our simulation was as follows: 1. We fixed the hyper parameters $\sigma^2$ and $p_c$ for producing samples. 2. We selected a learning algorithm (e.g. pf20) and fixed its corresponding tuned parameters (based on our simulations in the Results section). 3. We applied the learning algorithm over a sequence of observations $y_{1:T}$. Note that in a real experiment, this step can be done through a few episodes, which makes it possible to have a long sequence of observations, i.e. large $T$. 4. At each time $t$, we saved the values $y_t$, ${\hat{\theta}}_t$, ${\hat{\sigma}}_t$, $\delta_t$, $s_t$, $\textbf{S}_{\mathrm{Sh}}(y_{t}; {\hat{\pi}}^{(t-1)})$, and $\textbf{S}_{\mathrm{BF}}(y_{t}; {\hat{\pi}}^{(t-1)})$. Then, given an absolute prediction ${\hat{\theta}}>0$, an absolute prediction error $\delta>0$, a standard deviation $\sigma_C>0$, and a sign bias $s \in \{ -1, 1 \}$, we defined the set of time points $$\ \begin{aligned} \mathcal{T} = \{ 1 < t \leq T: | |{\hat{\theta}}_{t-1}| - {\hat{\theta}}| < \Delta \theta, | |\delta_t| - \delta| < \Delta \delta, |{\hat{\sigma}}_t - \sigma_C| < \Delta \sigma_C, s_t = s \}, \end{aligned}$$ where $| |{\hat{\theta}}_{t-1}| - {\hat{\theta}}| < \Delta \theta$, $| |\delta_t| - \delta| < \Delta \delta$, and $|{\hat{\sigma}}_t - \sigma_C| < \Delta \sigma_C$ are equivalent to $|{\hat{\theta}}_{t-1}| \approx {\hat{\theta}}$, $|\delta_t| \approx \delta$, and ${\hat{\sigma}}_t \approx \sigma_C$, respectively. $\Delta \theta$, $\Delta \delta$, and $\Delta \sigma_C$ are positive real values that should be determined based on practical limitations (mainly the length of the observation sequence $T$). We then computed the average surprise values as $$\ \begin{aligned} \Bar{\textbf{S}}_{\mathrm{BF}}({\hat{\theta}}, \delta,s,\sigma_C) &= \frac{1}{|\mathcal{T}|} \sum_{t \in \mathcal{T}} \textbf{S}_{\mathrm{BF}}(y_{t}; {\hat{\pi}}^{(t-1)})\\ \Bar{\textbf{S}}_{Sh}({\hat{\theta}}, \delta,s,\sigma_C) &= \frac{1}{|\mathcal{T}|} \sum_{t \in \mathcal{T}} \textbf{S}_{\mathrm{Sh}}(y_{t}; {\hat{\pi}}^{(t-1)}). \end{aligned}$$ We repeated this procedure for $N$ different simulated subjects (with different random seeds). The average of $\Bar{\textbf{S}}_{\mathrm{BF}}$ and $\Bar{\textbf{S}}_{Sh}$ over $N=20$ subjects, for two learning algorithms (i.e. Nas12$^{*}$ and pf20), and for $T=500$, ${\hat{\theta}} = 1$, $\sigma_C = 0.5$, $\Delta \theta = 0.25$, $\Delta \delta = 0.1$, and $\Delta \sigma_C = 1$ is shown in .B. The results are the same as what was predicted by our theoretical analysis. ### Simulation procedure for prediction 2 {#simulation-procedure-for-prediction-2 .unnumbered} For our second prediction, the theoretical proof is trivial. However, in order to have a setting similar to a real experiment (e.g. to use $P(y_{t+1};{\hat{\pi}}^{(t)}) \approx p$ instead of $P(y_{t+1};{\hat{\pi}}^{(t)}) = p$), and to have an estimation of the effect size, we used simulations also for our second experimental predictions. We followed the same procedure as the one for the simulation of the first prediction. For each simulated subject, and at each time $t$, we saved the quantities $P(y_{t+1};{\hat{\pi}}^{(t)})$, $P(y_{t+1};{\hat{\pi}}^{(0)})$, $\textbf{S}_{\mathrm{Sh}}(y_{t}; {\hat{\pi}}^{(t-1)})$, and $\textbf{S}_{\mathrm{BF}}(y_{t}; {\hat{\pi}}^{(t-1)})$. Then, for a given a probability value $p>0$, we defined the set of time points $$\ \begin{aligned} \mathcal{T} = \{ 0 \leq t \leq T: | P(y_{t+1};{\hat{\pi}}^{(t)}) - p| < \Delta p, | P(y_{t+1};{\hat{\pi}}^{(0)}) - p| < \Delta p \}, \end{aligned}$$ where $| P(y_{t+1};{\hat{\pi}}^{(t)}) - p| < \Delta p$ and $| P(y_{t+1};{\hat{\pi}}^{(0)}) - p| < \Delta p$ are equivalent to $P(y_{t+1};{\hat{\pi}}^{(t)}) \approx p$ and $P(y_{t+1};{\hat{\pi}}^{(0)}) \approx p$, respectively. $\Delta p$ is a positive real value that should be determined based on practical limitations (mainly the length of the observation sequence $T$). We then computed the average surprise $\Bar{\textbf{S}}_{\mathrm{BF}}(p)$ and $\Bar{\textbf{S}}_{\mathrm{Sh}}(p)$ over $\mathcal{T}$ for each value of $p$. We repeated this procedure for $N$ different simulated subjects (with different random seeds). The average of $\Bar{\textbf{S}}_{\mathrm{BF}}$ and $\Bar{\textbf{S}}_{Sh}$ over $N=20$ subjects, for two learning algorithms (i.e. Nas12$^{*}$ and pf20), and for $T=500$ and $\Delta p = 0.0125$ is shown in .B. Supporting information {#supporting-information .unnumbered} ====================== Modified algorithm of [@nassar2012rational; @nassar2010approximately]: Adaptation for Gaussian prior {#adaptednassar .unnumbered} ---------------------------------------------------------------------------------------------------- ### Recursive update of the estimated mean for Gaussian prior {#recursive-update-of-the-estimated-mean-for-gaussian-prior .unnumbered} Let us first consider the case of a stationary regime (i.e. no change points) where observed samples are drawn from a Gaussian distribution with known variance, i.e. $y_{t+1}|\theta \sim \mathcal{N}(\theta, \sigma^2)$, and the parameter $\theta$ is also drawn from a Gaussian distribution $\theta \sim \mathcal{N}(\mu_0, \sigma_0^2)$. After having observed samples $y_1, ..., y_{t+1}$, it can be shown that, using Bayes’ rule, the posterior distribution $P(\theta | y_{1:t+1}) = \pi^{(t+1)}_B(\theta)$ is \[Eq:Nassar\_Gaussian\_mean\_update\] P(| y\_[1:t+1]{}) = ( ; \_[B, t+1]{} = ( + ), \^2\_[B, t+1]{} = ) . An estimate of $\theta$ is its expected value $\mathbb{E}(\theta | y_{1:t+1}) = \mu_{B, t+1}$. In a nonstationary regime where, after having observed $y_1, ..., y_t$ from the same hidden state, there is the possibility for a change point upon observing $y_{t+1}$, the posterior distribution is \[Eq:Nassar\_Bayesian\_posterior\_Rec\_Formula\_0\] P(| y\_[1:t+1]{}) = (1 - \_[t+1]{}) P(| y\_[1:t+1]{}, h\_[t+1]{} = 0) + \_[t+1]{} P(|y\_[t+1]{}, h\_[t+1]{} = 1) . To facilitate notation in this subsection we denote $\Delta h_{t+1} = 0$ as “stay” and $\Delta h_{t+1} = 1$ as “change” so that \[Eq:Nassar\_Bayesian\_posterior\_Rec\_Formula\_1\] P(| y\_[1:t+1]{}) = (1 - \_[t+1]{}) P(| y\_[1:t+1]{}, ) + \_[t+1]{} P(|y\_[t+1]{}, ) Note that the above is equivalent to Bayesian recursive formula () of the main text, where $\gamma_{t+1}$ is the adaptation rate we saw in of the main text, and is essentially the probability to change given the new observation, i.e. $P(\text{change} | y_{t+1})$. In [@nassar2010approximately] this quantity is denoted as $\Omega_{t+1}$. Taking into account we have \[Eq:SNassar\_Exp\_values\] & (| y\_[1:t+1]{}, ) = \_[B, t+1]{} = ( + ) ,\ & (| y\_[1:t+1]{}, ) = ( + ) , where $r_t$ is the time interval of observations coming from the same hidden state, calculated at time $t$. Taking the expectation of the estimated mean upon observing the new sample $Y_{t+1} = y_{t+1}$ is \[Eq:SNassar\_post\_m\_1\] \_[t+1]{} = (1 - ) ( + ) + ( + ) , where we dropped the subscript $t+1$ in $\gamma$ to simplify notations. We have \[Eq:SNassar\_post\_m\_2\] \_[t+1]{} = (1 - ) ( + + ) + ( + ) . Since ${\hat{\mu}}_{t} = \frac{1}{\frac{1}{\sigma^2_0} + \frac{r_t}{\sigma^2}} \Big(\frac{\mu_0}{\sigma^2_0} + \frac{\sum_{i=t+1 - r_t}^{t}y_i}{\sigma^2}\Big)$, after a few lines of algebra we have \[Eq:SNassar\_post\_m\_3\] \_[t+1]{} &= (1 - )\_t + \_0 + (1 - ) (y\_[t+1]{} - \_t) + (y\_[t+1]{} - \_0). We now define $\rho = \frac{\sigma^2}{\sigma^2_0} $ and find \[Eq:SNassar\_post\_m\_4\] \_[t+1]{} = (1 - )\_t + \_0 + (1 - ) (y\_[t+1]{} - \_t) + (y\_[t+1]{} - \_0) . A rearrangement of the terms and inclusion of the dependency of $\gamma$ on time yields \[Eq:SNassar\_post\_m\_4b\] \_[t+1]{} = (1 - \_[t+1]{})( \_t + (y\_[t+1]{} - \_t) ) + \_[t+1]{} ( \_0 + (y\_[t+1]{} - \_0) ) . In order to obtain a form similar to the one of [@nassar2012rational; @nassar2010approximately] we continue and we spell out the terms that include the quantities ${\hat{\mu}}_t, \mu_0$ and $y_{t+1}$ \[Eq:SNassar\_post\_m\_5\] \_[t+1]{} &= (1 - )\_t - (1 - ) \_t\ &+ \_0 - \_0\ &+ (1 - ) y\_[t+1]{} + y\_[t+1]{} Using that $\frac{1}{\rho + r_t + 1} = \frac{1}{\rho + 1} - \frac{r_t}{(\rho + 1)(\rho + r_t + 1)}$ we have \[Eq:SNassar\_post\_m\_6\] \_[t+1]{} &= (1 - )\_t - (1 - ) \_t + (1 - ) \_t\ &+ \_0 - \_0\ &+ (1 - ) y\_[t+1]{} - (1 - ) y\_[t+1]{} + y\_[t+1]{}. After a further step of algebra we arrive at \[Eq:SNassar\_post\_m\_7\] \_[t+1]{} &= ( (1 - )\_t + \_0 ) + ( (1 - ) (\_t - y\_[t+1]{}) + y\_[t+1]{} ) . If we define $1 - \alpha = (1 - \gamma) \frac{r_t}{\rho + r_t + 1} \Rightarrow \alpha = 1 - (1 - \gamma) \frac{r_t}{\rho + r_t + 1} \Rightarrow \alpha = \frac{\rho + \gamma r_t +1}{\rho + r_t + 1}$ and rearrange the terms, we have \[Eq:SNassar\_post\_m\_final\_1\] \_[t+1]{} &= ( (1 - )\_t + \_0 ) + ( (1 - )\_t + y\_[t+1]{} )\ \_[t+1]{} &= (\_t + (\_0 - \_t) ) + ( \_t + (y\_[t+1]{} - \_t) ) . Adding back the dependency of $\gamma$ and $\alpha$ on time we finally have \[Eq:SNassar\_post\_m\_final\_2\] \_[t+1]{} = (\_t + \_[t+1]{}(\_0 - \_t) ) + ( \_t + \_[t+1]{} (y\_[t+1]{} - \_t) ) . ### Recursive update of the the Estimated Variance for Gaussian Prior {#recursive-update-of-the-the-estimated-variance-for-gaussian-prior .unnumbered} In [@nassar2012rational] the authors calculate first the variance ${\hat{\sigma}}^2_{t+1} = Var(\theta | y_{1:t+1})$ and based on this compute then ${\hat{r}}_{t+1}$. We derive here these calculations for the case of Gaussian prior. We remind once again that P(| y\_[1:t+1]{}) = (1 - \_[t+1]{}) P(| y\_[1:t+1]{}, ) + \_[t+1]{} P(|y\_[t+1]{}, ) Then for the variance ${\hat{\sigma}}^2_{t+1} = Var(\theta | y_{1:t+1})$ we have \[SNassar\_variance\_1\] \^2\_[t+1]{} &= (1- )\^2\_[stay]{} + \^2\_[change]{} + (1- )(\_[stay]{} - \_[change]{})\^2\ &= (1- )\^2\_[B, t+1]{} + \^2\_[change]{} + (1- )(\_[B, t+1]{} - \_[change]{})\^2 where $\sigma^2_{B, t+1} = \frac{1}{\frac{1}{\sigma^2_0} + \frac{r_t+1}{\sigma^2}}$ and $\sigma^2_{change} = \frac{1}{\frac{1}{\sigma^2_0} + \frac{1}{\sigma^2}}$. We have defined earlier $\rho = \frac{\sigma^2}{\sigma^2_0} $ so that \[SNassar\_variance\_A\] A &= (1- )\^2\_[B, t+1]{} + \^2\_[change]{} = (1- ) + Using, as before, that $\frac{1}{\rho + r_t + 1} = \frac{1}{\rho + 1} - \frac{r_t}{(\rho + 1)(\rho + r_t + 1)}$ we have A &= ( 1 - (1 - ) ). We have defined earlier the learning rate $\alpha = 1 - (1 - \gamma) \frac{r_t}{\rho + r_t + 1}$, so we can write A &= Note that $\mu_{t} = \frac{1}{\frac{1}{\sigma^2_0} + \frac{r_t}{\sigma^2}} \Big(\frac{\mu_0}{\sigma^2_0} + \frac{\sum_{i=t+1 - r_t}^{t}y_i}{\sigma^2}\Big)$ so for the calculation of the last term we have \[Eq:SNassar\_variance\_B1\] B &= \_[B, t+1]{} - \_[change]{}\ &= ( + ) - ( + ). We now rearrange terms \[Eq:SNassar\_variance\_B2\] B &= \_t + ( - ) (y\_[t+1]{} - \_t) - \_0 - (y\_[t+1]{} - \_0),\ and finally we have \[Eq:SNassar\_variance\_final\] \^2\_[t+1]{} = & + ( 1- )B\^2. Implementation of Nas10$^{*}$, Nas12$^{*}$ and Particle Filtering with 1 particle for the Gaussian estimation task {#nassar_pseudocodes .unnumbered} ------------------------------------------------------------------------------------------------------------------ We provide here the pseudocode for the algorithms Nas10$^{*}$, Nas12$^{*}$ and Particle Filtering with 1 particle for the Gaussian estimation task. Observations are drawn from a Gaussian distribution with known variance and unknown mean, i.e. $y_{t+1}|\mu_{t+1} \sim \mathcal{N}(\mu_{t+1}, \sigma^2)$ and $\theta_t = \mu_t$. When there is a change, the parameter $\mu$ is also drawn from a Gaussian distribution $\mu \sim \mathcal{N}(\mu_0, \sigma_0^2)$. Algorithms \[Alg:Nas10\], \[Alg:Nas12\], and \[Alg:PF\_Gaussian\] estimate the expected $\mu_{t+1}$ (i.e. ${\hat{\mu}}_{t+1}$) upon observing a new sample $y_{t+1}$. After re-writing we have \[Eq:S\_GM\_gaussian\] **S**\_ ( y\_[t+1]{}; \_[t]{}, \_[t]{} ) = Note that Algorithm \[Alg:PF\_Gaussian\] is a translation of Algorithm \[Alg:PF\] to the case of a single particle (where there are no weights to calculate) and for the Gaussian distribution as a particular instance of the exponential family. Specify $m=p_c/(1-p_c)$, $\mu_{0}$, $\sigma_{0}$, $\sigma$ and $\rho = \sigma^2 / \sigma_{0}^2$. Initialize ${\hat{\mu}}_{0}$, ${\hat{\sigma}}_{0}$, ${\hat{r}}_0$ and $t \gets 0$. Observe $y_{t+1}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\mu}}_{t}, {\hat{\sigma}}_{t})$ using [ ]{} Compute $\gamma_{t+1} = \gamma \big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\mu}}_{t}, {\hat{\sigma}}_{t}) , m \big)$ as in [ ]{} Compute ${\hat{\mu}}_{t+1}$ using [ ]{} Compute ${\hat{r}}_{t+1} = (1 - \gamma_{t+1})({\hat{r}}_t + 1) + \gamma_{t+1}$ [ ]{} Compute ${\hat{\sigma}}_{t+1} = \big[ \frac{1}{\sigma_0^2} + \frac{{\hat{r}}_{t+1}}{\sigma^2} \big]^{-1}$ [ ]{} $t \gets t+1$ Specify $m=p_c/(1-p_c)$, $\mu_{0}$, $\sigma_{0}$, $\sigma$ and $\rho = \sigma^2 / \sigma_{0}^2$. Initialize ${\hat{\mu}}_{0}$, ${\hat{\sigma}}_{0}$, ${\hat{r}}_0$ and $t \gets 0$. Observe $y_{t+1}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\mu}}_{t}, {\hat{\sigma}}_{t})$ using [ ]{} Compute $\gamma_{t+1} = \gamma \big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\mu}}_{t}, {\hat{\sigma}}_{t}) , m \big)$ as in [ ]{} Compute ${\hat{\mu}}_{t+1}$ using [ ]{} Compute the expected variance ${\hat{\sigma}}_{t+1}$ using [ ]{} Compute the expected time interval ${\hat{r}}_{t+1} = \frac{\sigma^2}{{\hat{\sigma}}^2_{t+1}} - \frac{\sigma^2}{\sigma^2_{0}}$ [ ]{} $t \gets t+1$ Specify $m=p_c/(1-p_c)$, $\mu_{0}$, $\sigma_{0}$, $\sigma$ and $\rho = \sigma^2 / \sigma_{0}^2$. Initialize ${\hat{\mu}}_{0}$, ${\hat{\sigma}}_{0}$, ${\hat{r}}_0$ and $t \gets 0$. Observe $y_{t+1}$ [ ]{} Compute $\textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\mu}}_{t}, {\hat{\sigma}}_{t})$ using [ ]{} Compute $\gamma_{t+1} = \gamma \big( \textbf{S}_{\mathrm{BF}}(y_{t+1}; {\hat{\mu}}_{t}, {\hat{\sigma}}_{t}) , m \big)$ as in [ ]{} Sample $ \Delta h_{t+1}^{(1)} \sim \text{Bernoulli}\big(\gamma_{t+1}\big)$ [ ]{} ${\hat{\mu}}_{t+1} \gets {\hat{\mu}}_{t} + \frac{1}{\rho + {\hat{r}}_t + 1} (y_{t+1} - {\hat{\mu}}_{t})$ and ${\hat{r}}_{t+1} \gets {\hat{r}}_{t} + 1$ ${\hat{\mu}}_{t+1} \gets \mu_{0} + \frac{1}{\rho + 1} (y_{t+1} - \mu_{0})$ and ${\hat{r}}_{t+1} \gets 1$ [ ]{} Compute the expected variance ${\hat{\sigma}}_{t+1} = \frac{\sigma^2}{{\hat{r}}_{t+1} + \rho}$ [ ]{} $t \gets t+1$ Acknowledgments {#acknowledgments .unnumbered} =============== This research was supported by Swiss National Science Foundation No. 200020\_184615) and by the European Union Horizon 2020 Framework Program under grant agreement No.785907 (Human Brain Project, SGA2). \#1\#1\#1\#1 [10]{} Preuschoff K, t Hart BM, Einhauser W. Pupil dilation signals surprise: Evidence for noradrenaline’s role in decision making. Frontiers in neuroscience. 2011;5:115. Nassar MR, Rumsey KM, Wilson RC, Parikh K, Heasly B, Gold JI. Rational regulation of learning dynamics by pupil-linked arousal systems. Nature neuroscience. 2012;15:1040–1046. Modirshanechi A, Kiani MM, Aghajan H. Trial-by-trial surprise-decoding model for visual and auditory binary oddball tasks. NeuroImage. 2019;196:302–317. Ostwald D, Spitzer B, Guggenmos M, Schmidt TT, Kiebel SJ, Blankenburg F. Evidence for neural encoding of Bayesian surprise in human somatosensation. NeuroImage. 2012;62:177–188. Mars RB, Debener S, Gladwin TE, Harrison LM, Haggard P, Rothwell JC, et al. Trial-by-trial fluctuations in the event-related electroencephalogram reflect dynamic changes in the degree of surprise. Journal of Neuroscience. 2008;28:12539–12545. Yu AJ, Dayan P. Uncertainty, neuromodulation, and attention. Neuron. 2005;46:681–692. Gerstner W, Lehmann M, Liakoni V, Corneil D, Brea J. Eligibility traces and plasticity on behavioral time scales: experimental support of neohebbian three-factor learning rules. Frontiers in neural circuits. 2018;12. Nassar MR, Wilson RC, Heasly B, Gold JI. An approximately Bayesian delta-rule model explains the dynamics of belief updating in a changing environment. Journal of Neuroscience. 2010;30:12366–12378. Behrens TE, Woolrich MW, Walton ME, Rushworth MF. Learning the value of information in an uncertain world. Nature neuroscience. 2007;10:1214. Glaze CM, Kable JW, Gold JI. Normative evidence accumulation in unpredictable environments. Elife. 2015;4:e08825. Heilbron M, Meyniel F. Confidence resets reveal hierarchical adaptive learning in humans. PLoS computational biology. 2019;15:e1006972. Faraji M, Preuschoff K, Gerstner W. Balancing new against old information: the role of puzzlement surprise in learning. Neural computation. 2018;30:34–83. Friston K, FitzGerald T, Rigoli F, Schwartenbeck P, Pezzulo G. Active inference: a process theory. Neural computation. 2017;29:1–49. Schwartenbeck P, FitzGerald T, Dolan R, Friston K. Exploration, novelty, surprise, and free energy minimization. Frontiers in psychology. 2013;4:710. Friston K. The free-energy principle: a unified brain theory? Nature reviews neuroscience. 2010;11:127. Bogacz R. A tutorial on the free-energy framework for modelling perception and learning. Journal of mathematical psychology. 2017;76:198–211. Adams RP, MacKay DJ. Bayesian online changepoint detection. arXiv preprint arXiv:07103742. 2007;. Fearnhead P, Liu Z. On-line inference for multiple changepoint problems. Journal of the Royal Statistical Society: Series B (Statistical Methodology). 2007;69:589–605. Aminikhanghahi S, Cook DJ. A survey of methods for time series change point detection. Knowledge and information systems. 2017;51:339–367. Wilson RC, Nassar MR, Gold JI. Bayesian online learning of the hazard rate in change-point problems. Neural computation. 2010;22:2452–2476. Cummings R, Krehbiel S, Mei Y, Tuo R, Zhang W. Differentially private change-point detection. In: Advances in Neural Information Processing Systems; 2018. p. 10825–10834. Lin K, Sharpnack JL, Rinaldo A, Tibshirani RJ. A sharp error analysis for the fused lasso, with application to approximate changepoint screening. In: Advances in Neural Information Processing Systems; 2017. p. 6884–6893. Masegosa A, Nielsen TD, Langseth H, Ramos-L[ó]{}pez D, Salmer[ó]{}n A, Madsen AL. Bayesian models of data streams with hierarchical power priors. In: Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org; 2017. p. 2334–2343. Findling C, Chopin N, Koechlin E. Imprecise neural computations as source of human adaptive behavior in volatile environments. bioRxiv. 2019; p. 799239. Shannon C. A mathematical theory of communication. Bell System Technical Journal 27: 379-423 and 623–656. 1948;20. Gordon NJ, Salmond DJ, Smith AF. Novel approach to nonlinear/non-Gaussian Bayesian state estimation. In: IEE proceedings F (radar and signal processing). vol. 140. IET; 1993. p. 107–113. Kutschireiter A, Surace SC, Sprekeler H, Pfister JP. Nonlinear Bayesian filtering and learning: a neuronal dynamics for perception. Scientific reports. 2017;7:8722. Shi L, Griffiths TL. Neural implementation of hierarchical Bayesian inference by importance sampling. In: Advances in neural information processing systems; 2009. p. 1669–1677. Huang Y, Rao RP. Neurons as Monte Carlo Samplers: Bayesian? Inference and Learning in Spiking Networks. In: Advances in neural information processing systems; 2014. p. 1943–1951. Legenstein R, Maass W. Ensembles of spiking neurons with noise support optimal probabilistic inference in a dynamically changing environment. PLoS computational biology. 2014;10. Lisman J, Grace AA, Duzel E. A neoHebbian framework for episodic memory; role of dopamine-dependent late LTP. Trends in neurosciences. 2011;34:536–547. Efron B, Hastie T. Computer age statistical inference. vol. 5. Cambridge University Press; 2016. S[ä]{}rkk[ä]{} S. Bayesian filtering and smoothing. vol. 3. Cambridge University Press; 2013. zkan E, [Š]{}m[í]{}dl V, Saha S, Lundquist C, Gustafsson F. Marginalized adaptive particle filtering for nonlinear models with unknown time-varying noise parameters. Automatica. 2013;49:1566–1575. Maheu M, Dehaene S, Meyniel F. Brain signatures of a multiscale process of sequence learning in humans. Elife. 2019;8:e41541. George CP, Doss H. Principled Selection of Hyperparameters in the Latent Dirichlet Allocation Model. Journal of Machine Learning Research. 2017;18:162–1. Liu J, West M. Combined parameter and state estimation in simulation-based filtering. In: Sequential Monte Carlo methods in practice. Springer; 2001. p. 197–223. Doucet A, Tadi[ć]{} VB. Parameter estimation in general state-space models using particle methods. Annals of the institute of Statistical Mathematics. 2003;55:409–422. Joshi S, Gold JI. Pupil size as a window on neural substrates of cognition. PsyArXiv. 2019;. Lieder F, Daunizeau J, Garrido MI, Friston KJ, Stephan KE. Modelling trial-by-trial changes in the mismatch negativity. PLoS computational biology. 2013;9. Kopp B, Lange F. Electrophysiological indicators of surprise and entropy in dynamic task-switching environments. Frontiers in human neuroscience. 2013;7:300. Meyniel F, Maheu M, Dehaene S. Human inferences about sequences: A minimal transition probability model. PLoS computational biology. 2016;12:e1005260. Musiolek L, Blankenburg F, Ostwald D, Rabovsky M. Modeling the N400 brain potential as Semantic Bayesian Surprise. In: 2019 Conference on Cognitive Computational Neuroscience; 2019. Konovalov A, Krajbich I. Neurocomputational Dynamics of Sequence Learning. Neuron. 2018;98:1282–1293. Loued-Khenissi L, Pfeuffer A, Einh[ä]{}user W, Preuschoff K. Anterior insula reflects surprise in value-based decision-making and perception. NeuroImage. 2020; p. 116549. Huettel SA, Mack PB, McCarthy G. Perceiving patterns in random series: dynamic processing of sequence in prefrontal cortex. Nature neuroscience. 2002;5:485–490. Mathys C, Daunizeau J, Friston KJ, Stephan KE. A Bayesian foundation for individual learning under uncertainty. Frontiers in human neuroscience. 2011;5:39. Wilson RC, Nassar MR, Gold JI. A mixture of delta-rules approximation to Bayesian inference in change-point problems. PLoS computational biology. 2013;9:e1003150. Prat-Carrabin A, Wilson RC, Cohen JD, Da Silveira RA. Human Inference in Changing Environments with Temporal Structure. BioRxiv. 2020; p. 720516. Daw N, Courville A. The pigeon as particle filter. Advances in neural information processing systems. 2008;20:369–376. Fr[é]{}maux N, Gerstner W. Neuromodulated spike-timing-dependent plasticity, and theory of three-factor learning rules. Frontiers in neural circuits. 2016;9:85. Yu AJ. Change is in the eye of the beholder. Nature neuroscience. 2012;15:933. Yu AJ, Cohen JD. Sequential effects: superstition or rational behavior? In: Advances in neural information processing systems; 2009. p. 1873–1880. Gershman SJ, Radulescu A, Norman KA, Niv Y. Statistical computations underlying the dynamics of memory updating. PLoS computational biology. 2014;10:e1003939. Storck J, Hochreiter S, Schmidhuber J. Reinforcement driven information acquisition in non-deterministic environments. In: Proceedings of the international conference on artificial neural networks, Paris. vol. 2. Citeseer; 1995. p. 159–164. Schmidhuber J. Formal theory of creativity, fun, and intrinsic motivation (1990–2010). IEEE Transactions on Autonomous Mental Development. 2010;2:230–247. Itti L, Baldi PF. Bayesian surprise attracts human attention. In: Advances in neural information processing systems; 2006. p. 547–554. Bogacz R. Dopamine role in learning and action inference. BioRxiv. 2019; p. 837641. Gershman SJ. What does the free energy principle tell Us about the brain? arXiv preprint arXiv:190107945. 2019;. Lomonaco V, Desai K, Culurciello E, Maltoni D. Continual Reinforcement Learning in 3D Non-stationary Environments. arXiv preprint arXiv:190510112. 2019;. Traor[é]{} R, Caselles-Dupr[é]{} H, Lesort T, Sun T, Cai G, D[í]{}az-Rodr[í]{}guez N, et al. DisCoRL: Continual Reinforcement Learning via Policy Distillation. arXiv preprint arXiv:190705855. 2019;. Nagabandi A, Clavera I, Liu S, Fearing RS, Abbeel P, Levine S, et al. Learning to adapt in dynamic, real-world environments through meta-reinforcement learning. arXiv preprint arXiv:180311347. 2018;. Doucet A, Godsill S, Andrieu C. On sequential Monte Carlo sampling methods for Bayesian filtering. Statistics and computing. 2000;10:197–208. Doucet A, Johansen AM. A tutorial on particle filtering and smoothing: Fifteen years later. Handbook of nonlinear filtering. 2009;12:3. Beal MJ. Variational algorithms for approximate Bayesian inference. Ph.D. thesis, University College London; 2003. Boyd S, Vandenberghe L. Convex optimization. Cambridge university press; 2004. Papoulis A, Saunders H. Probability, random variables and stochastic processes. American Society of Mechanical Engineers Digital Collection; 1989.

ArXiv

--- abstract: 'Depinning of an interface from a random self–affine substrate with roughness exponent $\zeta_S$ is studied in systems with short–range interactions. In 2$D$ transfer matrix results show that for $\zeta_S<1/2$ depinning falls in the universality class of the flat case. When $\zeta_S$ exceeds the roughness ($\zeta_0=1/2$) of the interface in the bulk, geometrical disorder becomes relevant and, moreover, depinning becomes . The same unexpected scenario, and a precise location of the associated tricritical point, are obtained for a simplified hierarchical model. It is inferred that, in 3$D$, with $\zeta_0=0$, depinning turns first–order already for $\zeta_S>0$. Thus critical wetting may be impossible to observe on rough substrates.' address: - 'Dipartimento di Fisica e INFN (Gruppo Collegato di Udine, Sezione di Trieste), Universitá di Udine, 33100 Udine, Italy' - 'INFM – Dipartimento di Fisica e Sezione INFN, Universitá di Padova, 35131 Padova, Italy' author: - 'G. Giugliarelli' - 'A. L. Stella' title: Discontinuous Interface Depinning from a Rough Wall --- 15truept Wetting and depinning phenomena occur when the interface between two coexisting phases unbinds from an attractive substrate [@forgacs]. In 2$D$, where an Ising interface can be well represented by a directed path [@chui], critical wetting occurs as a rule, and first–order wetting is predicted only for special setups [@forgacs]. In 3$D$, on the other hand, numerical simulations have shown critical, first–order and tricritical wetting for the Ising model. The same phenomena in random media have attracted a lot of attention recently. Many works considered the effects of quenched disorder due to impurities on interface depinning. Other studies addressed disorder restricted to a geometrically smooth (chemical surface disorder) [@forgacs1]. Both bulk and chemical surface disorder may modify the universality class of the unbinding transition, but generally not its continuous, second–order character in 2$D$.[@forgacs] In the present Letter we address the role of disorder due to wall roughness in determining the nature of the wetting transition. Disordered geometry is most amenable to theoretical treatment when characterized by simple scaling laws, like in the case of self–affine or fractal substrates. So far, the effects of such roughness were seldom discussed, mostly in connection with complete wetting . In spite of this, real substrates with self–affine geometry are met in many situations and have been the object of recent experiments [@pfeifer]. For these substrates the average transverse width of the wall scales with the longitudinal length $X$ as $X^{\zeta_S}$ ($0<\zeta_S<1$). We show that in the 2$D$ Ising model the nature of the depinning transition changes drastically upon increasing the self–affine roughness exponent $\zeta_S$. For $\zeta_S\lesssim1/2$ depinning remains continuous as in the flat case, and disorder is irrelevant. For $\zeta_S\gtrsim 1/2$ an unusual depinning occurs. Since the intrinsic roughness of a 2$D$ interface is $\zeta_0=1/2$, it is natural to expect $\zeta_S=1/2$ as precise threshold. The emerging scenario is that of geometrical disorder determining a tricritical phenomenon for depinning. We are able to locate the tricritical point by accurate renormalization group (RG) calculations on a hierarchical model, which further suggest $\zeta_S=1/2$ as the tricritical threshold in 2$D$. Extrapolation of our results to the experimentally most relevant case of 3$D$, where $\zeta_0=0$ in ordered bulk, suggests that a minute disorder in substrate geometry could suppress critical wetting in favor of a first–order transition. 6.0truecm [**FIG.1.**]{} Example of rough substrate boundary (continuous) and interface (dotted) configuration. Let $x$ and $y$ be integer coordinates of points on a square lattice. A “[*wall*]{}” is given by a self–avoiding path directed according to the positive $x$ axis (Fig. 1). To each wall step parallel to the $x$ axis, a height $S_x$, equal to the ordinate of its left end, $x$, is associated. For simplicity we consider only wall configurations obeying the restriction: $s_x=S_x-S_{x+1}=\pm 1$. In order to generate wall geometries with a preassigned roughness exponent $\zeta_S$, we used a randomized version of an algorithm due to Mandelbrot [@mandelbrot]. Given a wall profile, the interface can assume configurations obeying $h_x-h_{x+1}=0,\pm 1$ and $h_x\geq S_x$, if $h_x$ is the height of the horizontal step at $x$ (Fig. 1). The interface is like a (partially) directed walk or polymer, and its Hamiltonian is $${\cal H}=\sum_x [\varepsilon(1+|z_x-z_{x+1}+s_x|)-u\delta_{z_x,0}] \eqnum{1}$$ where $z_x\equiv h_x-S_x$, and $\varepsilon, u>0$. According to (1), at temperature $T$, fugacities $\omega\equiv e^{-\varepsilon/T}$ and $k\equiv e^{u/T}$ are associated to each (horizontal or vertical) step of the walk, and to each horizontal step on the wall, respectively. With a wall profile covering a distance $X$ the partition function is $${\cal Z}_X=\sum_{all\ walks} \omega^L k^l \eqnum{2}$$ where the sum is restricted, e.g., to walks from the origin $(0,0)$ to any point $(X,y)$, with $y\geq S_X$. $L$ and $l$ indicate total length and number of horizontal steps on the wall, respectively. ${\cal Z}_X$ is a functional of the wall profile. In order to calculate it we use transfer matrices: $$({\bf T}_{s_x})_{m,n}=\omega [\delta_{n,m\!-\!s_x}\!+\!(\delta_{n,m\!-\!s_x\!-\!1}\!+\! \delta_{n,m\!-s_x\!+\!1})\omega] k^{\delta_{n,0}} \eqnum{3}$$ where $m$ and $n$ range on the allowed $z_x$ and $z_{x+1}$, respectively. The partition function thus becomes $${\cal Z}_X=\sum_{l,p} \left( \prod_{x=0}^{X-1} {\bf T}_{s_x}\right )_{l,p} \phi_0(p) \eqnum{4}$$ where, with the left end of the interface grafted at the origin, $\phi_0(p)=\delta_{p,0}$. A wall profile corresponds to a sequence of factors ${\bf T}_1$, ${\bf T}_{-1}$ in the product of eq.(4). ${\cal Z}_X$ is thus given asymptotically in terms of the largest Lyapunov eigenvalue [@crisanti] $$\lambda_{max}=\!\lim_{X\to\infty}\!\left [\!{{||\left(\prod_{x=0}^{X-1} {\bf T}_{s_x}\right) {\vec \phi}_0||} \over {|| {\vec \phi}_0||}}\! \right ]^{1\over X}\!\!=\! \lim_{X\to\infty} \!\left[\!{{||{\vec \phi}_X||}\over {||{\vec \phi}_0||}}\!\right] ^{1\over X} \eqnum{5}$$ We verified that different long ${\bf T}$ sequences, i.e. wall profiles, lead to the same eigenvalue within good accuracy. Thus, the quenched free energy is $\ln \lambda_{max}(\omega,k,\zeta_S)= \lim_{X\to\infty} \overline{\ln {\cal Z}_X}/X$, where the bar indicates quenched averaging over wall profiles. In the random context depinning is most efficiently detected by studying the behavior of quantities which can be directly related to the components of ${\vec{\phi}}_X$. Examples are the average probability, $\overline{P_0}(\omega,k,\zeta_S)=\lim_{X\to\infty}\overline{{1\over X} \sum_{x=0}^{X-1} \phi_x^2(0)/||\vec{\phi}_x||^2}$, that the horizontal step lies on the wall, and the average distance of the interface from the substrate, $\overline{\langle z\rangle}=\lim_{X\to\infty}\overline{ {1\over X} \sum_{x=0}^{X-1} \sum_z z\phi_x^2(z)/||\vec{\phi}_x||^2}$ [@forgacs]. For our determinations we used up to 50 independent profiles with $x\leq 2^{20}$ for which the components of $\vec{\phi}_x$ were computed up to a distance from the wall $z_{max}\simeq 2\cdot 10^4$. TABLE I. Values of $\psi$ and $k_c$ at $\omega=1/2$ from fits of $\overline{\langle z\rangle}$. ----------- ----------------- ------------------- $\zeta_S$ $\psi$ $k_c$ 0 1 4/3$^{\ a}$ 1/3 1.01 $\pm$ 0.01 1.772 $\pm$ 0.001 2/5 1.01 $\pm$ 0.02 1.828 $\pm$ 0.001 1/2 1.03 $\pm$ 0.04 1.908 $\pm$ 0.003 ----------- ----------------- ------------------- $^{\ a}\ $ Exact results for a flat substrate \[1\]. Rather than considering variations of $\overline{P_0}$ or $\overline{\langle z\rangle}$ along curves parametrized by temperature, we choose to follow $\omega=const$. lines. In the case $\zeta_S\lesssim 1/2$, e.g., upon approaching $k=k_c$ from above with $0<\omega<1$, $\overline{\langle z\rangle}$ is well fitted by $\overline{\langle z\rangle}=A(k-k_c)^{-\psi}+B$, with $\psi$ always compatible with the exactly known flat wall value of 1 [@forgacs]. $k_c$ of course depends on $\omega$ and $\zeta_S$. Some $\psi$ and $k_c$ determinations are reported in Table I for $\omega=1/2$. For $\zeta_S\lesssim 1/2$ disorder in the wall geometry does not appear to lead to a new universality class for depinning. By treating a different model with continuum many–body techniques, Li and Kardar found second–order depinning with $\psi=1$ for all $\zeta_S<1$. At variance with this conclusion we find here that the situation drastically changes for $\zeta_S\gtrsim 1/2$. In this range $\overline{\langle z\rangle}$ has a much steeper, abrupt rise at $k_c$, so that the previous fit becomes clearly inappropriate. Further strong evidence of a change of the nature of the transition at $\zeta_S\simeq 1/2$ comes from the behavior of $\overline{P_0}$ (Fig. 2). While for $\zeta_S\lesssim1/2$, $\overline{P_0}\sim (k-k_c)^\rho$, for $k\to k_c^+$, with $\rho\simeq 1$, as with flat substrate [@forgacs], for $\zeta_S\gtrsim 1/2$ a discontinuity in $\overline{P_0}$ shows up (Fig. 2), becoming sharper and sharper as finite–size effects are reduced. The amplitude of the discontinuous jump in $\overline{P_0}$ also increases with $\zeta_S$. 8.5truecm [FIG. 2.]{} The probability $\overline{P_0}$ as a function of $k$ at $\omega=1/2$ and for different $\zeta_S$ values. Thus, for $\zeta_S\gtrsim 1/2$ substrate roughness is relevant and, moreover, drives the depinning transition first–order. $\zeta_S=1/2$ is the natural candidate as border value between continuous and discontinuous regimes. Indeed, for $\zeta_S>1/2$ the wall roughness exceeds the roughness $\zeta_0=1/2$ [@lipowski] of the interface. In 2$D$, first–order depinning is quite unexpected in the context of interfacial phenomena. Only two special ways of obtaining it have been conceived so far, by introducing either an attractive defect line in the bulk [@forgacs2], or longitudinally fully correlated disorder [@nieu]. Here first order is caused by sufficiently strong geometrical surface disorder, which also reveals opposite in its effects to its chemical counterpart. Indeed, while higher roughness induces first–order, in the defect line case chemical surface disorder drives the transition back continuous [@forgacs3]. This latter effect is certainly what one would expect at first sight on the basis of experience with phase transitions [@hui]. Like in the special examples of refs. [@forgacs2; @nieu], our first order depinning needs not be accompanied by off-coexistence prewetting phenomena. Extrapolation to 3$D$ of our findings is natural and has remarkable and unexpected consequences. Systems with a dominance of short–range interactions are, e.g., metallic substrate–adsorbates or, even more, type–I superconductors [@okki]. Since typically interfaces in pure systems are only logarithmically rough in 3$D$ ($\zeta_0=0$), a minimum of substrate roughness should be sufficient to give first–order wetting. This offers a further possible explanation of the fact that critical wetting is so elusive from the experimental point of view. A most recent work predicts critical wetting for superconductor interfaces [@okki]. For such interfaces $\zeta_0$ is not known, unfortunately, but could be rather small, if not zero. This means that special care in using smooth substrates should be exerced, in order to observe the predicted phenomenon. 6.0truecm [FIG. 3.]{} Construction rule of the DHL (level $n=0$ to level $n=1$) and schematic picture of the lattice at level $n$, with the four $n-1$ level units. A wall configuration (heavy) crossing the left units, and two polymer configurations (dotted) are reported. With such wall configuration eq. (6) applies. By reinterpreting $\omega$ and $k$ as monomer fugacity and Boltzmann factor for contacts, respectively, our model describes polymer adsorption [@privman88]. Apart from a change in the ensemble (${\cal Z}=\sum_X {\cal Z}_X$), the transfer matrices are the same. Criticality ($\cal Z$ dominated by infinite length polymer configurations) implies $\lambda_{max}=1$. We find that $\lambda_{max}(1/2,k,\zeta_s)=1$ for all $k<k_c(1/2,\zeta_S)$, where $k_c$ was defined above. For $k>k_c$ criticality occurs at $\omega<1/2$, indicating that the polymer is adsorbed [@privman88]. Indeed $\omega=1/2$ marks criticality for the polymer in the bulk. The dependence of $k_c$ on $\zeta_S$ shows that adsorption becomes more difficult with increasing roughness. One can also show that the fraction of monomers adsorbed on the wall should have the same singular behavior as $\overline{P_0}$, when moving at $\omega=1/2$. Thus, like interface depinning, polymer adsorption undergoes a change from second to first order upon increasing $\zeta_S$. To gain more insight into this change, we stick to polymer language and consider a simplified model of adsorption defined on diamond hierarchical lattice (DHL) (Fig. 3). Self–avoiding paths on DHL have often been used to mimic directed polymers in 2$D$ [@derrida]. A wall joining the two ends is obtained as follows: at level $n=0$ of DHL construction the wall always coincides with the unique existing bond. For $n=n_{max}>0$, the wall is determined by backward iteration. Starting from $n_{max}$, at each level, $n$, we choose whether the wall passes through the left or right DHL units of level $n-1$, with probabilities $1-\Delta$ and $\Delta$, respectively, and so on. If, e.g., we put $\Delta=0$, the process is deterministic and we create a single wall coinciding with the left border of the lattice. For $\Delta=1/2$ we generate with equal probability all possible walls through the lattice. Given a wall, we consider a polymer, with partition ${\cal Z}_n$, joining the ends of the lattice and laying, e.g., to the right of the wall. As in eq. (2), $\omega$ and $k$ are step and wall contact fugacities, respectively. For $\Delta=0$ we deal with a polymer attracted ($k>1$) by the left DHL border, which plays the role of a “flat”, deterministic substrate. When $\Delta$ rises, the “roughness” of the now random wall increases. Transverse hills and valleys are felt more and more by the polymer. For a given wall, one can compute the polymer partition function iteratively, using ${\cal Z}_{b,n+1}=2{\cal Z}_{b,n}^2$ and $${\cal Z}_{n+1}={\cal Z}_{n,1}{\cal Z}_{n,2}+{\cal Z}_{b,n}^2 \eqnum{6}$$ or $${\cal Z}_{n+1}={\cal Z}_{n,1}{\cal Z}_{n,2} \eqnum{7}$$ Eq.(6) or (7) is chosen, according to whether, at level $n+1$, the $n$–th level diamonds, 1 and 2, crossed by the wall, are the left or the right ones, respectively. ${\cal Z}_{b,n}$ is clearly the “bulk” partition function on DHL at $n$–th level, in absence of wall. Initial conditions are ${\cal Z}_0=k\omega$ and ${\cal Z}_{b,0}=\omega$. Considering first $\Delta=0$, eq.(6) induces a two parameters RG mapping by putting ${\cal Z}_{b,1}=\omega'$ and ${\cal Z}_{1}={\omega}'k'$. Clearly the transformation of ${\cal Z}_{b,n}$ implies that the bulk criticality condition is $\omega=1/2$ as in the Euclidean case. The value $k=1$ is marginally unstable for $k\gtrsim1$ and separates adsorbed ($k>1$) from desorbed ($k<1$) regimes. Thus, for $\Delta=0$ a positive attraction is always sufficient to adsorb the polymer. Eq. (6) applies for all $n$ and the problem has a nontrivial fixed point with ${\cal Z}_b^*=1/2$ and ${\cal Z}^*=1/2$. $P_0(\omega=1/2, k,\Delta=0)$ can be extrapolated by iteration. Due to marginality, $P_0$ starts rising with zero slope, but continuously, for $k=1$ (Fig. 4). For $\Delta>0$ the ${\cal Z}_n$’s become random variables and we must iterate their probability distribution, ${\cal P}_n$. This cannot be done exactly. However, starting from ${\cal P}_0=\delta({\cal Z}-k\omega)$, we follow the distribution at level $n$ by iteratively sampling it. From a large number ($\simeq 10^6$) of ${\cal Z}$ values distributed according to ${\cal P}_{n-1}$, we generate a sample distributed according to ${\cal P}_n$ by choosing many pairs of ${\cal Z}$ values and obtaining from each pair a new value of ${\cal Z}$ according to eq.(6) or (7), with probabilities $1-\Delta$ and $\Delta$, respectively. The resulting ${\cal Z}$’s constitute a sample of ${\cal P}_n$. This procedure could be iterated for $n\leq 40$ with extremely high accuracy. 8.5truecm [FIG. 4.]{} $\overline{P_0}$ on the DHL at $\omega=1/2$ for various $\Delta$ values. The dashed lines mark the discontinuities for $\Delta>1/2$. $\overline{P_0}$ at $\omega=1/2$ is plotted as a function of $k$ in Fig. 4 for different $\Delta$’s. For $0<\Delta< 1/2$, $\overline{P_0}$ rises continuously from zero as $c(k-k_c)$. Thus, $\rho=1$ for $0<\Delta< 1/2$. The slope $c$ and $k_c$ are both increasing with $\Delta$. Apart from $c=0$ at $\Delta=0$, due to the accidental marginality, such behaviour reproduces what is observed in 2$D$ when $\zeta_S< 1/2$. For $\Delta<1/2$ the polymer is more “rough” than the wall. Only when $\Delta=1/2$ the latter has the same freedom to develop through the DHL as a polymer has within the “bulk”. On the other hand, $\Delta>1/2$ corresponds qualitatively to $\zeta_S> 1/2$, because the polymer feels more and more the wall limiting its options when developing through the DHL. For $\Delta=1/2$ we have evidence that $c=\infty$, with a still continuous transition. This infinite slope indeed anticipates a sharp discontinuity in $\overline{P_0}$ for $\Delta>1/2$. So, the hierarchical model contains ingredients reproducing, at least qualitatively, the scenario emerging for the Euclidean model, and gives a suggestive indication of the way in which continuous transitions switch to first order at the expected threshold $\zeta_S=1/2$. The dependence of $k_c$ on $\Delta$ mimics that on $\zeta_S$ in 2$D$, and further motivates the correspondence between $\Delta$ and $\zeta_S$ in the two cases. We stress that, since all paths have the same length, and there is no natural recipe for defining a transversal distance on DHL, the notion of roughness must always be mediated in some way: here we can link roughness to $\Delta$. Our hierarchical model provides a remarkable example of the tricritical transition we are dealing with in this paper, and allows an essentially exact determination of its location and properties. Moreover, the threshold we find at $\Delta=1/2$ is strongly suggestive of the precise location of the tricritical point in the Euclidean case. Summarizing, in 2$D$, interface depinning or directed polymer adsorption on rough substrate with $\zeta_S<1/2$ are continuous and in the same universality as in the flat case. For $\zeta_S>1/2$ roughness is relevant and, moreover, the transitions acquire an unusual, discontinuous nature. This result, not anticipated so far , warns that some continuum approaches may not be able to catch the correct physics of depinning from rough substrates. We expect similar properties, and the same threshold $\zeta_S=1/2$, for directed polymer adsorption on a self–affine surface in 3$D$. Indeed, directed polymers have no upper critical dimension, and for them $\zeta_0=1/2$ in all $d$. Although the single polymer adsorption regime is not easily accessible experimentally, we believe that our results should be relevant for stretched polymers [@ward]. In 3$D$ the interface of a pure system typically has $\zeta_0=0$ [@lipowski]. We thus conjecture that depinning occurs discontinuously as soon as $\zeta_S>0$, in cases when the interactions are predominantly short–range, like in metallic systems. Our results also bear on the observability of the recently predicted critical wetting in the case of type–I superconductors [@okki], which is perhaps the most strict physical example of short–range interface–substrate interactions. Numerical calculations have been partly supported by CNR within the CRAY project of Statistical Mechanics. We thank Joseph Indekeu for valuable criticism and suggestions and Mehran Kardar for stimulating discussions. G. Forgacs, R. Lipowsky and Th.M. Nieuwenhuizen, in “[*Phase Transitions and Critical Phenomena*]{}”, by C. Domb and J.L. Lebowitz, Vol. 14, Academic Press, 1991. S.T. Chui and J.D. Weeks, [*Phys. Rev.*]{} [**B23,**]{} 2438 (1981). D.A. Huse and C.L. Henley, [*Phys. Rev. Lett.*]{} [**54,**]{} 2708 (1985). M. Kardar, [*Phys. Rev. Lett.*]{} [**55,**]{} 2235 (1985); M. Kardar and D.R. Nelson, [*Phys. Rev. Lett.*]{} [**56,**]{} 472 (1986). R. Lipowski and M.E. Fisher, [*Phys. Rev. Lett.*]{} [**56,**]{} 472 (1986). G. Forgacs, J.M. Luck, Th.M. Nieuwenhuizen and H. Orland, [*Phys. Rev. Lett.*]{} [**57**]{} 2184 (1986). H. Li and M. Kardar, [*Phys. Rev.*]{} [**B42,**]{} 6546 (1990); M. Kardar and J.O. Indekeu, [*Europhys. Lett.*]{} [**12,**]{} 61 (1990). P. Pfeifer, Y.J. Wu, M.W. Cole and J. Krim, [sl Phys. Rev. Lett.]{} [**62,**]{} 1997 (1989); see also J. Krim, I. Heyvaert, C. Van Haesendonck and Y. Bruynseraede, [*Phys. Rev. Lett.*]{} [**70,**]{} 57 (1993). G. Giugliarelli and A.L. Stella, [*Physica A*]{} [**212,**]{} 12 (1994). B.B. Mandelbrot, [*Physica Scripta*]{} [**32,**]{} 257 (1985). A. Crisanti, G. Paladin and A. Vulpiani, in “[*Products of Random Matrices in Statistical Physics*]{}”, edited by H.K. Lotsch, Springer Series in Solid State Sciences, Vol.[**104**]{} (Springer, Berlin, 1993). G. Forgacs, M.N. Svrakic and V. Privman, [*Phys. Rev.*]{} [**E37,**]{} 3818 (1988). Th.M. Nieuwenhuizen, [*J. Phys.*]{} [**A21,**]{} L567 (1988). G. Forgacs and Th.M. Nieuwenhuizen, [*J. Phys.*]{} [**A21,**]{} 3871 (1988). K. Hui and N. Berker, [*Phys. Rev. Lett.*]{} [**62,**]{} 2507 (1989). J.O. Indekeu and J.M.J. van Leeuwen, [*Phys. Rev. Lett.*]{} [**75,**]{} 1618 (1995). V. Privman, G. Forgacs and H.L. Frisch, [*Phys. Rev.*]{} [**B37,**]{} 9897 (1988). B. Derrida and R.B. Griffiths, [*Europhys. Lett.*]{} [**8,**]{}111 (1989). I.M. Ward, in “[*Structure and Properties of Oriented Polymers*]{}”, (Wiley, New York, 1975).

ArXiv

--- abstract: 'Liquid democracy allows members of an electorate to either directly vote over alternatives, or delegate their voting rights to someone they trust. Most of the liquid democracy literature and implementations allow each voter to nominate only one delegate per election. However, if that delegate abstains, the voting rights assigned to her are left unused. To minimise the number of unused delegations, it has been suggested that each voter should declare a personal ranking over voters she trusts. In this paper, we show that even if personal rankings over voters are declared, the standard delegation method of liquid democracy remains problematic. More specifically, we show that when personal rankings over voters are declared, it could be undesirable to receive delegated voting rights, which is contrary to what liquid democracy fundamentally relies on. To solve this issue, we propose a new method to delegate voting rights in an election, called *breadth-first delegation*. Additionally, the proposed method prioritises assigning voting rights to individuals closely connected to the voters who delegate.' author: - Grammateia Kotsialou$^1$ - | Luke Riley$^2$ $^1$Department of Political Economy, King’s College London, UK\ $^2$Department of Informatics, King’s College London, UK {grammateia.kotsialou, luke.riley}@kcl.ac.uk. bibliography: - 'FinalBib.bib' title: 'Incentivising Participation in Liquid Democracy with Breadth-First Delegation[^1]' --- Introduction {#sec:intro} ============ Liquid democracy is a middle ground between direct and representative democracy, as it allows each member of the electorate to directly vote on a topic, or temporarily choose a representative by delegating her voting rights to another voter. Therefore individuals who are either apathetic for an election, or trust the knowledge of another voter more than their own, can still have an impact on the election result (through delegating). An individual who casts a vote for themselves and for others is known as a *guru* (@ChristoffG17 [-@ChristoffG17]). Liquid democracy has recently started gaining attention in a few domains which we discuss to show an overview of the general societal interest. In the political domain, local parties such as the Pirate Party in Germany, Demoex in Sweden, the Net Party in Argentina and Partido de Internet in Spain have been experimenting with liquid democracy implementations. Additionally, the local governments of the London Southwark borough and the Italian cities Turino and San Dona di Piave are working on integrating liquid democracy for community engagement processes (@BoellaFGKNNSSST18 [-@BoellaFGKNNSSST18]). In the technology domain, the online platform LiquidFeedback uses a liquid democracy system where a user selects a single guru for different topics (@LFBook [-@LFBook]; @KlingKHSS15 [-@KlingKHSS15]). Another prominent online example is GoogleVotes (@GoogleVotes [-@GoogleVotes]), where each user wishing to delegate can select a ranking over other voters. Regardless of the increasing interest in liquid democracy, there exists outstanding theoretical issues. This work focuses on liquid democracy systems where each individual wishing to delegate can select a ranking over other voters. In such systems, given the common interpretation that delegations of voting rights are multi-step and transitive[^2], we observe that: searching for a guru follows a depth-first search in a graph that illustrates all delegation preferences within an electorate, e.g. nodes represent the voters and directed edges the delegation choices for each voter. For this reason, we name this standard approach of delegating voting rights as *depth-first delegation*. Despite its common acceptance, we came across an important disadvantage for this rule. We show that when depth-first delegation is used, it could be undesirable to receive the voting rights of someone else. At this point, we emphasize that disincentivising voters to participate as gurus is in contrast to the ideology of liquid democracy. Motivated by this, we propose a new rule for delegating voting rights, the *breadth-first delegation*, which guarantees that casting voters (those who do not delegate or abstain) weakly prefer to receive delegated voting rights, i.e. to participate as a guru. We outline this paper as follows: Section \[sec:intro\] discusses the latest applied and theoretical developments in liquid democracy and gives our model’s preliminaries. In Sections \[sec:Delegation\] & \[sec:Participation\], we define delegation graphs, delegation rules and two types of participation. Section \[sec:delrulepref\] formally introduces a new delegation rule while Section \[sec:delrulesegs\] compares this rule with the standard one. Finally, Section \[sec:con\] concludes this work. Related work {#sec:RelatedWork} ------------ There currently exists a lack of theoretical analysis on liquid democracy. However, we summarise the main differences of our work to the main undertaking so far. As outlined by @Brill18 ([-@Brill18]), one of the main ongoing issues in liquid democracy is how to handle personal rankings over voters. His work discusses possible solutions around this issue without giving a formalised model, which this paper does. For two election alternatives and a ground truth on which the correct one is, @KahngMP18 ([-@KahngMP18]) find that: (a) there is no decentralised liquid democracy delegation rule that is guaranteed to outperform direct democracy and (b) there is a centralised liquid democracy delegation rule that is guaranteed to outperform direct democracy as long as voters are not completely misinformed or perfectly informed about the ground truth. In comparison, our model can be used in a wider variety of elections, as it allows for multiple alternatives and no ground truth. Additionally our delegation rules can be used in a central or decentralised manner, thus the negative result (a) does not apply to our paper. The work of @ChristoffG17 ([-@ChristoffG17]) focuses on the existence of delegation cycles and inconsistencies that can occur when there are several binary issues to be voted on with a different guru assigned for each issue. In comparison, we avoid delegation cycles by stating that a delegation chain (a path from a delegating voter to their assigned guru) cannot include the same voter more than once. Furthermore, individual rationality issues between multiple elections is out of scope for this work. Last, @BrillT18 ([-@BrillT18]) introduce a special case of @ChristoffG17’s model, which allows a single voter to be assigned several gurus. However, our model assigns one guru per voter. Similar to our work, GoogleVotes (@GoogleVotes [-@GoogleVotes]) allows a user to select a ranking over other voters and uses, what the authors describe as, a back-track breadth first search to assign a guru to a voter. We cannot complete a more comprehensive comparison to GoogleVotes as they have published only a general description of their system (without a formal model). However, we know that their delegation rule is different to our proposed breadth-first delegation rule as in the Tally/Coverage section of their video example (from minute 32 of @GoogleVotesPresentation [-@GoogleVotesPresentation]), their rule assigns guru $C$ to voter $F$, while our rule would assign guru $A$ to voter $F$. Preliminaries {#sec:Model} ------------- Consider a set of voters ${\mathcal{V}}$ and a set of alternatives or outcomes ${\mathcal{A}}$. The set of possible electorates is given by ${\cal E}({\mathcal{V}}) = 2^{\mathcal{V}}\backslash \{ \emptyset \}$, i.e. non-empty subsets of ${\mathcal{V}}$. In our model, for every election there are three sets of electorates $V^a, V^c, V^d \in {\cal E}({\mathcal{V}})$ such that $V^a \cap V^c \cap V^d = \emptyset$ and $V^a \cup V^c \cup V^d = {\mathcal{V}}$, where sets $V^a$, $V^c$, $V^d$ consist of those who abstain, cast a vote and wish to delegate their voting rights, respectively. A *preference relation over alternatives* for a voter $i \in {\mathcal{V}}$ is denoted by $\succ^{{\mathcal{A}}}_i$ and is a binary relation on ${\mathcal{A}}$, i.e.: for $x,y\in A$ with $x \neq y$, the expression $x \succ^{{\mathcal{A}}}_i y$ indicates that voter $i$ strictly prefers alternative $x$ over alternative $y$. A *preference relation over voters* for voter $i \in {\mathcal{V}}$ is denoted by $\succ^{{\mathcal{V}}}_i$ and is a binary relation on ${\mathcal{V}}$, i.e.: for $i, x,y \in {\mathcal{V}}$ with[^3] $i \neq x, y$ and $x \neq y$, the expression $x \succ^{{\mathcal{V}}}_i y$ indicates that voter $i$ strictly prefers to delegate her voting rights to voter $x$ instead of voter $y$. For both preference relations, we allow an index to identify ranking positions e.g. for any $i \in {\mathcal{V}}^d$, her $m$-th preferred voter is denoted by $\succ^{\mathcal{V}}_{i, m}$. Consider a binary relation on a set $W$ given by $\succ^{W}_i$. Then $\succ^{W}_i$ is: $(a)$ *complete* iff for every pair $x, y \in W$ either $x \succ^{W}_i y$ or $y \succ^{W}_i x$ holds, $(b)$ *antisymmetric* iff for every pair $x, y \in W$, if $x \succ^{W}_i y$ then $y \succ^{W}_i x$ does not hold, and $(c)$ *transitive* iff for all $x, y, z \in W$, if $x \succ^{W}_i y$ and $y \succ^{W}_i z$, then $x \succ^{W}_i z$. Both preference relations over alternatives and preference relations over voters are antisymmetric and transitive but not complete (we do not enforce voters to rank every other member of the electorate as we consider this an unrealistic scenario for large electorates). The set of all possible preference relations $\succ^{{\mathcal{A}}}_i$ and $\succ^{{\mathcal{V}}}_i$, for any $i \in {\mathcal{V}}$, are denoted by ${R^{\mathcal{A}}}$ and ${R^{\mathcal{V}}}$, respectively. A *preference profile over alternatives* is a function $P^{\mathcal{A}}:\cal E ({\mathcal{V}}) \rightarrow $ 2$^{R^{\mathcal{A}}}$, where $P^A(N)$ returns a set of preference relations over alternatives (maximum one for each voter in $N$). For example, given an electorate $N=\{i,j,k\}$, a preference profile $P^A(N)$ could return $\{(i, \succ^{{\mathcal{A}}}_i),(j, \succ^{\mathcal{A}}_j)\}$, meaning that agent $i$ and $j$ have been assigned a preference relation over alternatives but $k$ has not (as $k$ is either delegating or abstaining). Similarly, we define as a *preference profile over voters* a function $P^{\mathcal{V}}:\cal E ({\mathcal{V}}) \rightarrow $ 2$^{R^{\mathcal{V}}}$, where $P^{\mathcal{V}}(N)$ returns a set of preference relations over voters (maximum one for each voter in $N$). Given profiles $P^{\mathcal{A}}$ and $P^{\mathcal{V}}$, voters are assigned to the $V^a$, $V^c$ and $V^d$ electorates as follows. If $(i, \succ^{\mathcal{A}}_i) \in P^{\mathcal{A}}(N)$, we infer that voter $i$ casts a vote according to $\succ^{\mathcal{A}}_i$ and therefore becomes a member of the casting electorate $V^c$. If $(i, \succ^{\mathcal{A}}_i) \notin P^{\mathcal{A}}(N)$ and $(i, \succ^{\mathcal{V}}_i) \in P^{\mathcal{V}}(N)$, then $i$ becomes a member of the delegating electorate $V^d$. If $(i, \succ^{\mathcal{A}}_i) \notin P^{\mathcal{A}}(N)$ and $(i, \succ^{\mathcal{V}}_i) \notin P^{\mathcal{V}}(N)$, $i$ becomes a member of the abstaining electorate $V^a$. Given an electorate $N$, adding or removing a preference relation over alternatives (or over voters) from a preference profile over alternatives $P^{\mathcal{A}}(N)$ (or over voters $P^{\mathcal{V}}(N)$), is denoted as follows. For a tuple $(i, \succ^{\mathcal{A}}_i) \in P^{\mathcal{A}}(N)$, a voter $j \in {\mathcal{V}}~\backslash ~N$ and $j$’s assigned preference relation over alternatives $\succ^{\mathcal{A}}_j\in R^{\mathcal{A}}$: $$\begin{aligned} & P^{\mathcal{A}}_{-i}(N) := P^{\mathcal{A}}(N)~\backslash ~\{ (i, \succ^{\mathcal{A}}_i)\}, \nonumber\\ & P^{\mathcal{A}}_{+ (j, \succ^{\mathcal{A}}_j)} (N):= P^{\mathcal{A}}(N)~ \cup ~\{(j, \succ^{\mathcal{A}}_j)\}.\nonumber\end{aligned}$$ Similarly, for a tuple $(i, \succ^{\mathcal{V}}_i)$ $\in P^{\mathcal{V}}(N)$, a voter $j \in {\mathcal{V}}~ \backslash ~N$ and $j$’s assigned preference relation over voters $\succ^{\mathcal{V}}_j \in R^{\mathcal{V}}$: $$\begin{aligned} &P^{\mathcal{V}}_{-i}(N) := P^{\mathcal{V}}(N)~\backslash ~\{ (i, \succ^{\mathcal{V}}_i)\}, \nonumber \\ &P^{\mathcal{V}}_{+(j, \succ^{\mathcal{V}}_j)}(N) := P^{\mathcal{V}}(N)~ \cup ~\{(j, \succ^{\mathcal{V}}_j)\}. \nonumber\end{aligned}$$ To simplify the above, we will be using the notation $P^{\mathcal{A}}, P^{\mathcal{V}}, P^{\mathcal{A}}_{-i}$, $P^{\mathcal{A}}_{+ j}$, $P^{\mathcal{V}}_{-i}$ and $P^{\mathcal{V}}_{+j}$, accordingly. Delegation graph and delegation rules {#sec:Delegation} ===================================== We use a graph to model possible delegations between voters: A *delegation graph* is a weighted directed graph $G=({\mathcal{V}}, E, w)$ where: - ${\mathcal{V}}$ is the set of nodes representing the agents registered as voters; - $E$ is the set of directed edges representing delegations between voters; and - $w$ is the weight function $w : E \mapsto \mathbb{N}$ that assigns a value to an edge $(i, j)$ equal to $i$’s preference ranking of $j$. To generate a delegation graph, we use the following: Define as $g$ the *delegation graph function* which takes as input a preference profile over voters $P^{\mathcal{V}}$ and returns the related delegation graph $G=({\mathcal{V}}, E, w)$ with the following property: for every $ i, j \in {\mathcal{V}}$ and $i \neq j$, - if there exists $(i, \succ^{\mathcal{V}}_{i}) \in P^{\mathcal{V}}$ with $\succ^{\mathcal{V}}_{i,x} = j$; then there exists $(i, j) \in E$ where $w((i, j)) = x$. We can evaluate a delegation graph through the following: A *delegation rule function* $d$ takes as input a preference profile over alternatives $P^{\mathcal{A}}$ together with a delegation graph $G$, and returns another preference profile over alternatives $\hat{P}^{{\mathcal{A}}}$ that resolves delegations. More specifically, - if $(i, \succ^{\mathcal{A}}_i) \in \hat{P}^{{\mathcal{A}}}$, then $i$ casts her vote, - if $(i, \succ^{\mathcal{A}}_j) \in \hat{P}^{{\mathcal{A}}}$ for a voter $j \neq i$, then $j$ becomes $i$’s final delegate, i.e. her guru, - if $(i, \succ^{\mathcal{A}}_k) \notin \hat{P}^{{\mathcal{A}}}$ for any $k \in {\mathcal{V}}$, then $i$ abstains. \[ref:DR\] For each voter $i \in {\mathcal{V}}$, a delegation rule analyses the subtree of the delegation graph rooted at node $i$ and decides whether $i$ casts, delegates or abstains. If voter $i$ is found to delegate, the chosen delegation rule function will traverse $i$’s subtree to find $i$’s guru. To get the outcome of an election, we use a voting rule function $f$. In our model, $f$ takes as input the modified preference profile over alternatives $\hat{P}^{{\mathcal{A}}}$ (which incorporates delegations) and returns a single winner or a ranking over alternatives (depending on the voting rule used). In Section \[sec:delrulesegs\], we show that the output of the voting rule depends on the chosen delegation rule, meaning that we could get different election results when only the delegation rule function is different, i.e. $f(d(P^{\mathcal{A}}, g(P^{\mathcal{V}}))) \neq f(d'(P^{\mathcal{A}}, g(P^{\mathcal{V}})))$, when $d \neq d'$. Cast and guru participation {#sec:Participation} =========================== The key property that we investigate is participation. The participation property holds if a voter, by joining an electorate, is at least as satisfied as before joining. This property has been defined only in the context of vote casting (@Fishburn [-@Fishburn]; @Moulin [-@Moulin]). Due to the addition of delegations in our model, we establish two separate definitions of participation reflecting this new functionality[^4]. For both of the following definitions, note that for an electorate $N \in {\cal E}({\mathcal{V}})$, the set of all preference profiles over alternatives is given by $\cal{P}$$^{{\mathcal{A}},N}$, while the set of all preference profiles over delegates is given by $\cal{P}$$^{{\mathcal{V}},N}$. A voting rule $f$ satisfies the *cast participation* property when every voter $i \in {\mathcal{V}}$ weakly prefers joining any possible voting electorate $V^c$ compared to abstaining and regardless of who is in the delegating electorate $V^d$. The *Cast Participation* property holds for a voting rule $f$ iff: $$\begin{aligned} f(d(P^{\mathcal{A}}, g(P^{\mathcal{V}}))) \succeq^{{\mathcal{A}}}_i f(d(P^{\mathcal{A}}_{-i}, g(P^{\mathcal{V}}))),\end{aligned}$$ for every possible disjoint casting and delegating electorates $V^c, V^d \in {\cal E}({\mathcal{V}})$, where $i \in V^c$, and every possible preference profile for these electorates $P^A \in {\cal P}^{{\mathcal{A}}, V^c}$ and $P^{\mathcal{V}}\in {\cal P}^{{\mathcal{V}}, V^d}$. For any casting and delegating electorates $V^c$ and $V^d$, a voting rule $f$ satisfies the *guru participation* property when any voter $i \in V^c$ weakly benefits from receiving additional voting rights of any voter $j \in {\mathcal{V}}$. The *Guru Participation* property holds for a voting rule $f$ iff: $$\begin{aligned} \label{guru_p} f(d(P^{\mathcal{A}}, g(P^{\mathcal{V}}))) \succeq^{{\mathcal{A}}}_i f(d(P^{\mathcal{A}}, g(P^{\mathcal{V}}_{-j}))),\end{aligned}$$ for every possible disjoint casting and delegating electorates $V^c, V^d \in {\cal E}({\mathcal{V}})$, where $i \in V^c$ , $j \in V^d$, and every possible profile $P^A \in {\cal P}^{{\mathcal{A}}, V^c}$ and $P^{\mathcal{V}}\in {\cal P}^{{\mathcal{V}}, V^d}$ that assign $j$’s vote to guru $i$, i.e. $(j, \succ^{\mathcal{A}}_i) \in d(P^{\mathcal{A}}, g(P^{\mathcal{V}}))$. Let $F$ be the set of all voting rules. It is known that only a subset $\bar{F}\subset F$ satisfy (cast) participation. For example, @Fishburn ([-@Fishburn]) show that single transferable vote does not satisfy (cast) participation, while @Moulin ([-@Moulin]) shows there is no Condorcet-consistent voting rule satisfying this property given 25 or more voters. We explores guru participation for voting rules in $\bar{F}$, and our results build on the following observation, which we intuitively descibe: if a new voter $j$ joins the delegating electorate and only one voter $i$ from the casting electorate increases the number of times she is assigned as a guru, then guru participation is satisfied. Consider $i, j \in {\mathcal{V}}$, profile $\hat{P}^{\mathcal{A}}$ returned by $d(P^{\mathcal{A}}, g(P^{\mathcal{V}}))$ and profile $\hat{P}^{'{\mathcal{A}}}$ returned by $d(P^{\mathcal{A}}, g(P^{\mathcal{V}}_{+j}))$, where $i$ has been assigned as $j$’s guru, i.e. $(j, \succ^{\mathcal{V}}_i) \in \hat{P}^{'{\mathcal{A}}}$. Guru $i$ becomes weakly better off after $j$ delegates if the following holds. For every $k \in {\mathcal{V}}$: 1. $k$’s vote is assigned to guru $l \in {\mathcal{V}}$ by both returned preference profiles, i.e. $(k, \succ^{\mathcal{V}}_l) \in \hat{P}^{{\mathcal{A}}} \cap \hat{P}^{'{\mathcal{A}}}$, or 2. $k$’s vote is assigned to guru $i$ after $j$ joins the delegating electorate, i.e. $(k, \succ^{\mathcal{A}}_i) \in \hat{P}^{'{\mathcal{A}}}$, Introducing breadth-first delegation {#sec:delrulepref} ==================================== Recall that liquid democracy allows for multi-step delegations. Therefore, the guru of any $i \in V^d$ could be any voter $j \in V^c$ who is in the sub tree of the delegation graph with root $i$. In addition, the assigned guru $j$ may not be included in $i$’s preference relation $\succ^{\mathcal{V}}_i$, i.e. it could be that $\nexists ~x$ such that $\succ_{i,x}^{\mathcal{V}}= j$. In this case, there is at least one intermediate delegator between voter $i$ and the assigned guru $j$. To find the exact intermediate delegators, we introduce delegation chains. A *delegation chain* for a voter $i \in V^d$ starts with $i$, then lists the intermediate voters in $V^d$ who have further delegated $i$’s voting rights and ends with $i$’s assigned guru $j \in V^c$. These chains (see Definition \[deft:dc\]) must satisfy the following conditions: $(a)$ no voter occurs more than once in the chain (to avoid infinite delegation cycles that could otherwise occur) and $(b)$ each member of the chain must be linked to the next member through an edge in the delegation graph, which is generated from the given preference profile over voters. Given profiles $P^{\mathcal{A}}$ and $P^{\mathcal{V}}$, a voter $i \in V^d$ and her guru $j \in V^c$, we define a *Delegation Chain* for $i$ to be an ordered tuple $C_i = \langle i, \ldots, j \rangle$ with the following properties: 1. For an integer $x \in [1, |C_i|]$, let $C_{i, x}$ indicate the voter at the $x$-th position in $C_i$. Then for any pair of integers $x,y \in[1, |C_i|]$ with $x \neq y$, $$C_{i,x} \neq C_{i,y}.$$ 2. For each integer $z \in [1, |C_i|-1]$, there exists an edge $$(C_{i, z}, C_{i, z+1}) \in E \in g(P^{\mathcal{V}}).$$ \[deft:dc\] Observe that the $x$ in the expression $C_{i, x}$ also indicates how deep the voter $C_{i, x}$ is in the delegation graph subtree rooted with $i$. Thus sometimes we refer to $x$ as the depth of $C_{i, x}$ in $C_i$. The function $w$ takes as input a delegation chain and returns a list of the weights assigned to each edge among voters in $C_i$, that is, $w(C_i) = [w(C_i)_1, \ldots, w(C_i)_x, \ldots, w(C_i)_{n-1}]$, where $w(C_i)_x$ is the weight of edge $(C_{i, x}, C_{i, x+1})$ and $n= |C_i|$. Delegation chains can be used as a tool to find a guru for a voter $i \in V^d$. The standard interpretation of liquid democracy delegations prioritises all possible delegation chains involving $i$ and $i$’s most preferred voter $\succ_{i,1}^{\mathcal{V}}$ before all possible delegation chains involving $i$ and $i$’s second preferred voter $\succ_{i,2}^{\mathcal{V}}$ and so on. Note that this priority rule hold for the deeper levels of the delegation graph subtree rooted at $i$. In other words, we observe that the standard way to select $i$’ guru is to choose the first casting voter found through a depth first search in $i$’s subtree, which motivates the next definition. A *depth-first delegation* rule $d^D$ assigns guru $j$ to $i$ iff: $(a)$ there is a delegation chain $C_i$ that can be formed from $i$ to $j$, and $(b)$ there is no other delegation chain $C_i'$ leading to a different guru $k$ that has a smaller weight at the earliest depth after the root, compared to $C_i$. \[def:DFD\] For $i, j, k \in {\mathcal{V}}$, a *depth-first delegation rule* $d^{D}$ returns a profile $\hat{P}^{{\mathcal{A}}}$ with $(i, \succ^{\mathcal{A}}_j) \in \hat{P}^{{\mathcal{A}}}$ iff $(a)$ and $(b)$ hold: 1. $\exists~C_i$ with $C_{i, |C_i|} = j$, 2. $\nexists~C_i'$ such that: 1. $C'_{i, |C_i'|} = k$  for $k \neq j$, 2. - $\exists y$: $w(C_i')_y < w(C_i)_y$ and - $w(C_i')_x \leq w(C_i)_x$  for all $~0 < x < y$. ---------------------------------------- ---------------------------------------- {width=".5\linewidth"} {width=".5\linewidth"} $(a)$ $(b)$ ---------------------------------------- ---------------------------------------- \[fig:loop\] Consider the delegation graph in Figure \[fig:loop\] $(a)$. There are two delegation chains[^5] available for voter $p \in {\mathcal{V}}$: $C_p = \langle p, r\rangle$ and $C_p' = \langle p, q, s\rangle$ with weights $w(C_p) = [2]$ and $w(C_p') = [1, 2]$, respectively. The $d^D$ rule returns profile $\hat{P}^{{\mathcal{A}}}$ that assigns $s$ as the guru of $p$, i.e. $(p, \succ^{\mathcal{A}}_s) \in \hat{P}^{{\mathcal{A}}}$, due to inequality $w(C_p')_1 < w(C_p)_1$. Note that $C_p'$ satisfies Definition \[def:DFD\]. \[ex:dd\] In Example \[ex:dd\], $p$’s voting right is assigned to guru $s$, but why should $s$ (who is the second preference of $q$) outrank agent $p$’s explicit second preference $r$? This question gains even more importance the longer the depth first delegation chain is. Given this issue, we define a novel delegation rule that prioritises a voter’s explicit preferences as follows. A *breadth-first delegation* rule $d^B$ assigns guru $j$ to $i$ iff: $(a)$ there is a delegation chain $C_i$ that can be formed from $i$ to $j$; and $(b)$ there is no other delegation chain $C_i'$ leading to a different guru $k$ with: either a shorter length or, an equal length and a smaller weight at the earliest depth after the root, compared to $C_i$. \[def:BFD\] For $i,j,k \in {\mathcal{V}}$, a *breadth-first delegation rule* $d^{B}$ returns a profile $\hat{P}^{{\mathcal{A}}}$ with $(i, \succ^{\mathcal{A}}_j) \in \hat{P}^{{\mathcal{A}}}$ iff $(a)$ and $(b)$ hold: 1. $\exists~C_i$ with $C_{i, |C_i|} = j$, 2. $\nexists~C_i'$ such that $C'_{i, |C_i|} = k$, for $k \neq j$, and 1. $|C_i'| < |C_i|$, or 2. - $|C_i'| = |C_i|$ and - $\exists y$: $w(C_i')_y < w(C_i)_y$ and - $w(C_i')_x \leq w(C_i)_x$  for all $~0 < x < y$. \[deft:BFD\] Consider the delegation graph in Figure \[fig:loop\] $(a)$. There are two delegation chains available for voter $p \in {\mathcal{V}}$: $C_p = \langle p, r\rangle$ and $C_p' = \langle p, q, s\rangle$ with weights $w(C_p) = [2]$ and $w(C_p') = [1, 2]$, respectively. The $d^B$ rule returns profile $\hat{P}^{{\mathcal{A}}}$ that assigns $r$ as the guru of $p$, i.e. $(p, \succ^{\mathcal{A}}_r) \in \hat{P}^{{\mathcal{A}}}$, due to inequality $|C_p| < |C_p'|$. Note that $C_p$ satisfies Definition \[deft:BFD\]. \[eg:3\] Depth-first versus breadth-first delegation {#sec:delrulesegs} =========================================== Through the next two examples, we show that different delegation rules can have different properties. More specifically, we present an instance where the depth-first delegation rule cannot guarantee guru participation, while the breadth-first delegation rule does. <span style="font-variant:small-caps;">delegation graph</span> <span style="font-variant:small-caps;">delegation rule</span> Yes No ---------------------------------------------------------------- --------------------------------------------------------------- ----- ---- Figure \[fig:loop\] $(a)$ $d^D$ 1 3 Figure \[fig:loop\] $(b)$ $d^D$ 3 2 Figure \[fig:loop\] $(a)$ $d^B$ 2 2 Figure \[fig:loop\] $(b)$ $d^B$ 3 2 : Election results for Figure \[fig:loop\] when using either the depth-first or the breadth-first delegation rule.[]{data-label="tab:electionResults"} Consider the delegation graph in Figure \[fig:loop\](a) and all possible delegation chains available to each voter in $V^d$: $C_p = \langle p, r\rangle$, $C_p' = \langle p, q, s\rangle$ and $C_q = \langle q, s \rangle$. Using rule $d^{D}$, voter $p$ is assigned guru $s$ through chain $C_p'$ (see Example \[ex:dd\]), while voter $q$ is also assigned guru $s$ through chain $C_q$. Therefore $d^{D}$ returns the preference profile over alternatives $\{ (p, \succ^{\mathcal{A}}_s), (q, \succ^{\mathcal{A}}_s), (s, \succ^{\mathcal{A}}_s), (r, \succ^{\mathcal{A}}_r)\}$. Using rule $d^{B}$ instead, voter $p$ is assigned guru $r$ through $C_p$ (see Example \[eg:3\]), while $q$’s guru remains the same. Therefore $d^{B}$ returns another preference profile over alternatives: $\{(p, \succ^{\mathcal{A}}_r), (q, \succ^{\mathcal{A}}_s), (s, \succ^{\mathcal{A}}_s), (r, \succ^{\mathcal{A}}_r)\}$. \[eg:ExampleNoLoop\] In the next example we focus on the case where the previously abstaining voter $t$ decides to delegate and show that the election result is inversed only when $d^D$ is used (see Table \[tab:electionResults\]). \[eg:ExampleLoop\] Consider the delegation graph in Figure \[fig:loop\](b) and all possible delegation chains available to each voter in $V^d$ with their respective edge weights: ---------------------------------------------------------------- ------------------------------------------------------------ <span style="font-variant:small-caps;">delegation chain</span> <span style="font-variant:small-caps;">edge weights</span> $C_p= \langle p, r \rangle$ $w(C_p)= [2]$, $C_p'= \langle p, q, s \rangle$ $w(C_p')= [1, 2]$, $C_p''= \langle p, q, t, r \rangle$ $w(C_p'')= [1, 1, 2]$, $C_q= \langle q, s \rangle$ $w(C_q)= [2]$, $C_q'= \langle q, t, r \rangle$ $w(C_q')= [1, 2]$, $C_q''= \langle q, t, p, r \rangle$ $w(C_q'')= [1, 1, 2]$, $C_t= \langle t, r \rangle$ $w(C_t)= [2]$, $C_t'= \langle t, p, r \rangle$ $w(C_t')= [1, 2]$, $C_t''= \langle t, p, q, s \rangle$ $w(C_t'')= [1, 1, 2]$. ---------------------------------------------------------------- ------------------------------------------------------------ Using rule $d^{D}$, observe that voter $p$ is assigned guru $r$ through the chain $C_p''$  due to  $w(C_p'')_1 < w(C_p)_1$, $w(C_p'')_1 = w(C_p')_1$  and  $w(C_p'')_2 < w(C_p')_2$. Voter $q$ is also assigned guru $r$ through chain $C_q''$ since $w(C_q'')_1 < w(C_q)_1$, $w(C_q'')_1 = w(C_q')_1$ and $w(C_q'')_2 < w(C_q')_2$. Last, voter $t$ is assigned guru $s$ through chain $C_t''$ because $w(C_t'')_1 < w(C_t)_1$, $w(C_t'')_1 = w(C_t')_1$ and $w(C_t'')_2 < w(C_t')_2$. Therefore rule $d^{D}$ returns the preference profile over alternatives $\{ (p, \succ^{\mathcal{A}}_r), (q, \succ^{\mathcal{A}}_r), (s, \succ^{\mathcal{A}}_s), (r, \succ^{\mathcal{A}}_r), (t, \succ^{\mathcal{A}}_s)\}$. Using rule $d^{B}$ instead, voter $p$ is assigned guru $r$ through the chain $C_p$ due to inequalities $|C_p| < |C_p'| < |C_p''|$. Voter $q$ is assigned guru $s$ through $C_q$ due to $|C_q| < |C_q'| < |C_q''|$ and voter $t$ is assigned guru $r$ through $C_t$ because of $|C_t| < |C_t'| < |C_t''|$. Therefore, rule $d^{B}$ returns the profile over alternatives $\{(p, \succ^{\mathcal{A}}_r), (q, \succ^{\mathcal{A}}_s), (s, \succ^{\mathcal{A}}_s), (r, \succ^{\mathcal{A}}_r), (t, \succ^{\mathcal{A}}_r)\}$. Examples \[eg:ExampleNoLoop\] and \[eg:ExampleLoop\] show that guru participation may not hold for depth-first delegation when a cycle exists in the delegation graph. Due to this cycle, when $t$ joins the election, both $r$ and $s$ receive new delegated voting rights, thus Observation 1 does not occur[^6]. We summarise the above for the set of voting rules satisfying cast participation $\bar{F}$. Given a voting rule $f \in \bar{F}$, guru participation is not guaranteed to hold when using the depth-first delegation rule $d^D$. \[theorem:1\] Consider the preference profile over alternatives and the preference profile over voters of Figure \[fig:loop\]$(b)$, $$\begin{aligned} & P^{\mathcal{A}}= \{(r,\succ^{\mathcal{A}}_r), (s,\succ^{\mathcal{A}}_s)\} \\ & P^{\mathcal{V}}= \{(p,\succ^{\mathcal{V}}_p), (q,\succ^{\mathcal{V}}_q), (t,\succ^{\mathcal{V}}_t)\}, \end{aligned}$$ where the preferences over alternatives for $r$ and $s$ are: “Yes” $\succ^{\mathcal{A}}_r$ “No", “No"$\succ^{\mathcal{A}}_s$ “Yes” and the preferences over voters for $p,q,t$ are: $q \succ^{\mathcal{V}}_p r$,  $t \succ^{\mathcal{V}}_q s$ and $p \succ^{\mathcal{V}}_t r$. We prove this theorem using Examples \[eg:ExampleNoLoop\] and  \[eg:ExampleLoop\]. In Example \[eg:ExampleNoLoop\], where voter $t$ abstains, rule $d^D$ returns profile $$\begin{aligned} d^{D}(P^{\mathcal{A}},g(P^{\mathcal{V}}_{-t}))= \{ (p, \succ^{\mathcal{A}}_s), (q, \succ^{\mathcal{A}}_s), (s, \succ^{\mathcal{A}}_s), (r, \succ^{\mathcal{A}}_r)\},\end{aligned}$$ which gives three votes (via $s$) for alternative “No" and one vote (via $r$) for alternative “Yes" (see also Table \[tab:electionResults\]). From Example \[eg:ExampleLoop\] where voter $t$ delegates, rule $d^D$ returns profile $$\begin{aligned} & d^{D}(P^{\mathcal{A}},g(P^{\mathcal{V}}))= \\ & \{ (p, \succ^{\mathcal{A}}_r), (q, \succ^{\mathcal{A}}_r), (s, \succ^{\mathcal{A}}_s), (r, \succ^{\mathcal{A}}_r), (t, \succ^{\mathcal{A}}_s)\},\end{aligned}$$ which gives three votes for “Yes" and two votes for “No". Observe that the election result changes from “No" to “Yes" despite the fact that $t$ votes for “No" through her guru $s$, i.e. $(t, \succ^{\mathcal{A}}_s) \in d^{D}(P^{\mathcal{A}}, g(P^{\mathcal{V}}))$. Note that due to the preference “No"$\succ^{\mathcal{A}}_s$ “Yes”, we get $$\begin{aligned} \label{s} f(d^{D}(P^{\mathcal{A}},g(P^{\mathcal{V}}))) \prec^{{\mathcal{A}}}_s f(d^{D}(P^{\mathcal{A}},g(P^{\mathcal{V}}_{-t}))),\end{aligned}$$ where $f$ is a voting rule satisfying cast participation. However, the preference expressed by implies that guru $s$ becomes worst off after $t$ delegates to her, which violates the definition of guru participation , proving this theorem. We highlight that if a delegation graph has no cycle then guru participation is guaranteed to hold for the depth-first delegation rule, which show through Lemma \[lem:DFDLemma1\] and Theorem \[theorem:1b\]. When using depth-first delegation rule $d^D$, if there is no cycle in the delegation graph then Observation 1 holds. \[lem:DFDLemma1\] Assume there exists a delegation graph with no cycles where Observation 1 does not hold. We show that the only case where Observation 1 does not hold is when a cycle exists. Recall that, by Observation 1, guru participation is guaranteed to hold if whenever a voter $j$ joins the delegating electorate, there exists only one voter, say $i$, in the casting electorate who increases the number of times she becomes a guru. Consider another voter $k$ who changes her assigned guru to a voter $l$ after $j$ joins the delegating electorate, where $l \neq i$ and $k \neq j$. This means that, apart from $i$, voter $l$ also increases the times she becomes a guru. Next we describe that, when $d^D$ is used, this case can only arise through the following circumstance. Let guru $i$ be assigned to voter $j$ through delegation chain $C_j = \langle j, \ldots, i \rangle$ and guru $l$ be assigned to voter $k$ through delegation chain $C_k= \langle k,...,j,...,l \rangle$. Chain $C_k$ must pass through $j$ because all chains without $j$ are available before $j$ delegates. Note that even if both chains pass through voter $j$, they end at different gurus. For $d^D$, this only occurs if there exists a voter $h$ with $h \neq i, j, l$, such that $$\begin{aligned} &C_j= \langle j,\ldots, h, \ldots, i \rangle \quad \text{and} \label{cycle1} \\ &C_k=\langle k,\ldots, h, \ldots, j, \ldots, l \rangle. \label{cycle2}\end{aligned}$$ The reason for the above is that $k$’s delegation goes through $h$ to reach $j$, but then the preferred delegation from $j$ passes through $h$ (see chain $C_j$). As $k$’s delegation already includes $h$ before $j$, and an intermediator voter cannot be repeated (definition \[ref:DR\]), $k$ uses another route to guru $l$ (through a less preferred option of $j$). From and , observe that there exists a cycle in the graph, i.e. the cycle $\langle h, ..., j, ..., h \rangle$, which contradicts our assumption and proves the lemma. Given a voting rule $f \in \bar{F}$ and a delegation graph with no cycles, guru participation is guaranteed to hold when using the depth-first delegation rule $d^D$. \[theorem:1b\] We prove this using Lemma \[lem:DFDLemma1\] and Observation 1. We have previously shown that depth-first delegation does not guarantee guru participation when the delegation graph contains cycles. The next theorem states that breadth-first delegation always guarantees guru participation. To show this, we first introduce the following observation and lemma. \[ob:BFD2\] Consider two voters $j$ and $k$ in a delegating electorate. Using the breadth-first delegation rule $d^B$, if $k$ is assigned guru $l$ through a delegation chain $C_k$ with $j \notin C_k$, then $k$ is assigned guru $l$ even when $j$ abstains. This is because rule $d^B$ has used $C_k$ ahead of any possible delegation chain that includes $j$. Therefore chain $C_k$ will still be used by $d^B$ when $j$ is in the abstaining electorate and no possible delegation chain that includes $j$ can be formed. Consider two voters $j$ and $k$ in a delegating electorate. Using the breadth-first delegation rule $d^B$, if voter $k$ is assigned her guru through a delegation chain $C_k$ with $j \in C_k$, then $k$ is assigned the same guru as $j$. \[lem:BFD1\] Assume that, using $d^B$, voter $j$ is assigned guru $i$ through delegation chain $C_j = \langle j \dots,i \rangle$ and $k$ is assigned a different guru $l$ through a delegation chain that includes $j$, i.e. $C_k = \langle k, \dots, j, \dots, l \rangle$. Then either $(a)$ or $(b)$ occurs: 1. rule $d^B$ should use chain $C'_j = \langle j, \dots,l \rangle$, which contradicts the assumption that $C_j$ is used, 2. there exists a shared intermediate voter $e$ such that $$\begin{aligned} &C_j = \langle j, \dots,e,g, \dots,i \rangle \quad \text{and} \\ &C_k = \langle k, \dots, f, j, \dots, l \rangle,\end{aligned}$$ where $e \in \langle k, \dots,f\rangle$. Recall that $d^B$ prioritises shorter length delegation chains (see definition \[def:BFD\]). We show that voter $k$ has a shorter delegation chain available that does not include $j$, i.e. there exists a $C_k'$ such that $|C_k'| < |C_k|$ and $j \notin C'_k$. Let $C_k' =\langle k, \dots, e, g, \dots,i \rangle$. According to $d^B$, the delegation chain used to assign $j$’s guru, $\langle j \dots, e, g, \dots,i\rangle$, is shorter or equal in length to any other alternative, thus $|\langle j, \dots, e, g, \dots, i\rangle| \leq |\langle j, \dots, l\rangle|$. Observe that $$\begin{aligned} |\langle g, \dots, i\rangle| < |\langle j, \ldots, e, g, \dots, i\rangle| \leq |\langle j, \dots, l\rangle \Rightarrow \\ |\langle k, \dots, e\rangle| + |\langle g, \dots, i\rangle| < |\langle k, \dots, e\rangle| + |\langle j, \dots, l\rangle|. \end{aligned}$$ Since $e \in \langle k, \ldots, f \rangle$, we can rewrite the previous as $$\begin{aligned} |\langle k,...,e\rangle| + |\langle g,...,i\rangle| < |\langle k,...,f\rangle| + |\langle j,...,l\rangle|.\end{aligned}$$ Therefore, rule $d^B$ should use $C'_k$ to assign $k$’s guru. However, since $j \notin C'_k$, the assumption is contradicted. The contradictions of both $(a)$ and $(b)$ prove this lemma. Given a voting rule $f \in \bar{F}$, guru participation is guaranteed to hold when using the breadth-first delegation rule $d^B$. By Observation \[ob:BFD2\], given voters $j$ and $k$ in the delegating electorate, if a voter $k$ does not delegate through $j$, then $k$’s assigned guru (if any) is the same as if $j$ abstained. By Lemma \[lem:BFD1\], if a voter $k$ delegates through $j$, then the guru of $k$ is the same as the guru of $j$. Combining the above cases, we show that (regardless of $k$ delegating through $j$ or not), whenever a voter $j$ joins the delegating electorate and is assigned to a guru $i$, then $i$ is the only casting voter who increases the number of times she becomes a guru. Since also $f \in\bar{F}$, then Observation 1 holds, meaning that the breadth-first delegation rule $d^B$ is guaranteed to satisfy guru participation. Conclusion and future work {#sec:con} ========================== In this paper, we discuss the depth-first and the breadth-first delegation rule proving that only the latter has the desirable property that every guru weakly prefers receiving delegating voting rights. However, there could be other delegation rules satisfying the same or other interesting properties that improve the concept of liquid democracy. Towards this path, we note that the issue of current liquid democracy implementations suffering from a small subset of gurus representing a large part of the electorate (@KlingKHSS15 [-@KlingKHSS15]) could be counteracted by the breadth-first delegation rule, as this rule favours keeping delegated voting rights close to their origin. We strongly believe that this hypothesis should be investigated. Other interesting future work include investigating guru participation with voting rules that do not satisfy cast participation, relaxing the assumption of strict personal rankings over voters, and analysing other types of participation. [^1]: This work was supported by the EPSRC grant EP/P031811/1. [^2]: If a voter $i$ delegates to a voter $j$, then $i$ transfers to $j$ the voting rights of herself and all the others that had been delegated to $i$. [^3]: A voter cannot include herself in her preference relation over voters. [^4]: There could be other interesting participation properties for liquid democracy, such as incentivising deviation from delegating to casting. But this is out of the paper’s scope, as we focus on finding delegation rules that weakly benefit casting voters who become gurus. [^5]: Recall that $\langle p, q, t\rangle$ is not a valid delegation chain as $t \notin V^c$. [^6]: Observation 1 states how guru participation can be satisfied when a voting rule satisfying cast participation is used: when a voter joins the delegating electorate, if only one voter increase the number of times assigned as a guru, then this voter is weakly better off.

ArXiv

--- abstract: 'The excess adsorption $\Gamma $ in two-dimensional Ising strips $(\infty \times L)$ subject to identical boundary fields, at both one-dimensional surfaces decaying in the orthogonal direction $j$ as $-h_1j^{-p}$, is studied for various values of $p$ and along various thermodynamic paths below the critical point by means of the density-matrix renormalization-group method. The crossover behavior between the complete wetting and critical adsorption regimes, occurring in semi-infinite systems, are strongly influenced by confinement effects. Along isotherms $T=const$ the asymptotic power law dependences on the external bulk field, which characterize these two regimes, are undercut by capillary condensation. Along the pseudo first-order phase coexistence line of the strips, which varies with temperature, we find a broad crossover regime where both the thickness of the wetting film and $\Gamma$ increase as function of the reduced temperature $\tau$ but do not follow any power law. Above the wetting temperature the order parameter profiles are not slab-like but exhibit wide interfacial variations and pronounced tails. Inter alia, our explicit calculations demonstrate that, contrary to opposite claims by Kroll and Lipowsky \[Phys. Rev. B [**28**]{}, 5273 (1983)\], for $p=2$ critical wetting transitions do exist and we determine the corresponding wetting phase diagram in the $(h_1,T)$ plane.' author: - 'A. Drzewiński,$^{1}$ A. Maciołek,$^{2,3}$$^{,4}$ A. Barasiński,$^{1}$ and S. Dietrich$^{2,3}$' title: 'Interplay of complete wetting, critical adsorption, and capillary condensation.' --- Introduction {#sec:1} ============ Far away from phase boundaries, i.e., deep in the one-phase region, condensed matter is perturbed by confining walls only within a thin layer proportional to the bulk correlation length $\xi$, i.e., at most approximately 15 Å. This changes drastically if the thermodynamic state of the bulk system is moved towards the boundary of phase transitions between the bulk phases. If the phase transition is of [*second*]{} order the bulk correlation length diverges and so-called critical adsorption occurs, i.e., the perturbation due to the wall located at $y=0$ penetrates deeply into the bulk, resulting in an algebraic divergence of the thickness of the interfacial structure $\int_0^{\infty}\left(\rho(y)-\rho_b\right)dy/(\rho(0^+)-\rho_b)\equiv \Gamma/(\rho(0^+)-\rho_b)\to \infty$, where for a one-component fluid as a paradigmatic case $\rho_b$ is the bulk number density for a given temperature $T$ and chemical potential $\mu$, and $\rho(y) $ is the fluid number density profile [@binder; @diehl]. Moreover, the order-parameter profile $m(y)\equiv (\rho(y)-\rho_c)/\rho_c$, where $\rho_c$ is the bulk critical number density, is governed by a universal scaling function which decays exponentially on the length scale of $\xi$ in the direction $y$ normal to the surface [@binder; @diehl]. If the phase transition is of [*first*]{} order wetting phenomena occur as a result of a subtle interplay between the substrate potential, interactions among fluid particles, and entropic contributions, in particular interfacial fluctuations [@dietrich]. The thickness of the interfacial structure diverges too, but here with exponents which are determined by the decay exponents of the pair potentials between the fluid particles and of the substrate potential [@dietrich]. These two types of phenomena are of significant practical importance, in particular for the condensed phase being a fluidum near the gas-liquid transition, such as water near the thermodynamic coexistence with its vapor, or binary liquid mixtures such as water and hydrocarbons near liquid-liquid phase transitions. In this context various applications arise ranging from the use of colloidal suspensions [@xia] to petroleum recovery [@bertrand]. Wetting films are relevant in many types of liquid coating processes, such as lubrication and adhesion and also for microfluidics and nanoprinting [@deGennes; @tabeling]. Critical adsorption plays an important role, e.g., for heterogeneous nucleation in supercritical solvents [@supercritical] or in micro- and nanofluidic systems in order to achieve wetting of these small structures  [@austin]. ![Schematic drawing of the liquid-gas coexistence curve $\Delta\mu=0$ in the $(\Delta\mu,T)$ bulk phase diagram. The thick solid line indicates the bulk coexistence line. $T_w$ is the transition temperature for a critical wetting transition. Thermodynamic paths (I) and (II), along which the behavior of the adsorption is discussed in the main text, are displayed as well as various crossover lines. The hatched area is the complete wetting regime. Note that $3\nu\simeq 1.89$ is larger than $\Delta\simeq 1.56$.[]{data-label="fig:0"}](rys01.eps){width="7.0cm"} A characteristic property of liquids is that first-order gas-liquid or liquid-liquid phase transitions end at critical points. Therefore critical adsorption and wetting must emerge from each other upon moving along the first-order phase boundary. This creates a puzzle. Liquids governed by dispersion forces (decaying asymptotically $\sim r^{-6}$ with distance $r$) belong to the Ising universality class [@pfeuty], i.e., the critical exponents describing the singular behavior of various thermodynamic quantities and structural properties are those of the Ising universality class. This also holds at interfaces. Specifically, along bulk coexistence $\Delta \mu=(\mu-\mu_0(T))/(k_BT_c)=0$ one has asymptotically  [@binder; @diehl] $$\label{eq:I_1} \Gamma \sim |\tau|^{\beta-\nu}, \qquad \Delta \mu=0,$$ where $\tau\equiv (T-T_c)/T_c$ is the reduced deviation from the critical temperature $T_c$ and $\Delta \mu $ measures, in units of $k_BT_c$, the deviation of the chemical potential from its value $\mu_0(T)$ at two-phase coexistence, whereas along the path $\tau=0$ $$\label{eq:I_2} \Gamma \sim |\Delta \mu|^{(\beta-\nu)/\Delta}, \qquad \tau=0,$$ where the critical exponents $\nu, \beta$, and $\Delta$ take the values of the Ising model (in $d=3$: $\nu=0.6301(4), \beta=0.32653(10)$, and $\Delta=3\nu-\beta = 1.564 $ [@pelisseto]). The latter universal singularity $\Gamma \sim |\Delta \mu|^{-0.194}$ is weaker than the non-universal one $\Gamma \sim |\Delta \mu|^{-1/3}$ for complete wetting [@dietrich], although according to the general renormalization group (RG) arguments the dispersion forces, which are responsible for this non-universal behavior, at first glance should give rise only to corrections to scaling (i.e., subdominant power laws). Naturally the question arises how the system manages to restore universality (i.e., dominance by the weaker universal power law) upon moving from the non-universal complete wetting behavior to the universal critical adsorption behavior. This issue has been addressed in Refs. [@tarazona:87; @dietrich:88:0; @TelodaGama:91] and the proposed scenario is such that upon approaching bulk coexistence, i.e., for $\tau$ fixed and $\Delta \mu\to 0^-$ (see path (II) in Fig. \[fig:0\]), the system will [*always*]{} cross over ultimately to the complete wetting regime but such that upon approaching $T_c$ (i.e., for smaller $|\tau|$) this crossover occurs closer and closer to the coexistence curve. The proposed mechanism is based on the argument that long-ranged dispersion forces are relevant and can dominate if the thickness $\ell \sim |\Delta \mu|^{-1/3}$ of the wetting layer is much larger than the bulk correlation length $\xi\sim |\tau|^{-\nu}$. As described more closely in Sec. \[sec:2\], for dispersion forces in $d=3$, this occurs if $|\Delta \mu| \ll |\tau|^{3\nu}$ (see the hatched area in Fig. \[fig:0\]). Accordingly, upon approaching $T_c$ this complete wetting regime becomes vanishingly small. This expectation follows similar arguments as how dispersion forces become irrelevant for, e.g., two-point correlation functions in the bulk or for the critical adsorption profile itself, for which RG theory predicts exponential decays which, however, are dominated asymptotically by the algebraic decay due to the dispersion forces. Nonetheless the universal scaling functions, which capture the exponential decay, are expected to describe also the neighborhood of $T_c$ correctly, even in the presence of the irrelevant dispersion forces. So far this puzzle has not been resolved satisfactorily by explicit calculations. The main reason is that mean field theory (MFT) as the natural theoretical starting point does not capture the correct $d=3$ exponents, which is essential for disentangling the different contributions and to describe the crossover. Within MFT, $\nu=\beta$ and the adsorption diverges only $\sim \ln |\Delta \mu|$. In systems governed by short-ranged forces only, within MFT both critical adsorption and complete wetting follow a divergence $\sim \ln|\Delta \mu|$ and therefore they cannot be distinguished; but the divergences differ beyond MFT. However, going beyond MFT is a difficult theoretical challenge, due to the spatial inhomogeneity. There is also an urgent experimental need to resolve this issue. An ellipsometry study of gravity-thinned complete wetting layers in the binary-liquid mixture of cyclohexane and methanol [@fenistein:02] reports results which are [*not*]{} in accordance with the theoretical predictions. The data indicate the divergence of the film thickness upon approaching the bulk critical temperature characterized by a critical exponent which is distinct from the expected, critical exponent for the bulk correlation length. Scaling arguments have been put forward according to which the observed [*effective*]{} exponent is associated with a broad, intermediate scaling regime facilitating the crossover from complete wetting in the presence of dispersion forces to critical scaling [@fenistein:02]. Also recent neutron reflectometry data for the adsorption from alkane-perfluoroalkane mixtures at fluorophobic and fluorophilic surfaces [@bowers:04] are [*not*]{} in agreement with the theoretical predictions. These authors have found that the behavior of the adsorption, as a function of temperature in the one-phase region upon approaching liquid-liquid coexistence with the composition well removed from the critical composition, can be represented by a power law with an exponent which differs from both that for complete wetting and that for critical adsorption [@bowers:04]. The predictions described above for the crossover between the complete wetting regime of fluids with long-ranged interactions and the critical regime are based on general scaling arguments and on the analysis of simple models for the effective interface Hamiltonian which are supposed to describe the relevant physics at length scales much larger than the bulk correlation length [@dietrich:88:0]. Only few microscopic studies are available which test these ideas [@tarazona:87; @TelodaGama:91]. Those are based on density functional methods and are of mean-field character, which in $d=3$ provides a correct description of complete wetting in the presence of dispersion forces but fails in the regime dominated by critical fluctuations. This makes it even more important to thoroughly analyze a model system which takes into account both the bulk and surface critical fluctuations of semi-infinite systems near $T_c$ as well as the interfacial fluctuations associated with wetting phenomena. As a representative of the corresponding universality class, an Ising ferromagnet is well suited for such an analysis. In addition, addressing the issue in two spatial dimensions ($d=2$) allows one to gain deep insight because the “exact” solvability provides, inter alia, the correct exponents. There is a recent experimental interest in such genuine two-dimensional systems in the context of proteins immersed in a fluid two-component lipid membrane which is near phase separation, including a critical point belonging to the $d=2$ Ising universality class [@silvius; @honerkamp]. If two such proteins are close to each other the structural properties of the membrane between them can be described in terms of the strips studied here, giving even rise to interesting effective interactions between the proteins. In the present analysis we use the density-matrix renormalization group (DMRG) method [@white:92; @nishino; @peschel; @schollwoeck] to numerically investigate the behavior of the adsorption in a two-dimensional ($d=2$) Ising model with short-ranged as well as long-ranged surface fields. The DMRG method provides essentially exact numerical results for thermodynamic quantities and correlation functions, including the magnetization profiles. This allows us to study systematically the adsorption properties along various thermodynamic paths and thus to test the aforementioned predictions of the mesoscopic, effective interface Hamiltonian approaches. The DMRG method is based on the transfer matrix approach and it is a numerically very efficient iterative truncation algorithm for constructing the effective transfer matrices for large systems. The method was originally developed by White [@white:92] for the diagonalization of quantum spin chains. It was then adapted by Nishino [@nishino] to two-dimensional classical systems. The DMRG method allows one to study strips that are infinitely long with widths up to $L=700$ lattice constants with arbitrary surface and bulk fields. The comparison with exact results in the case of a vanishing bulk field and in the presence of contact surface fields shows that the DMRG method provides a very high accuracy for a broad range of temperatures. Because the method requires finite values of $L$ the influence of the distant surface on the structural properties near the surface under study cannot be neglected, in particular, near the critical point — even for strips as wide as $L=700$. On one side, this for the DMRG method unavoidable confinement of the system complicates the aforementioned crossover behavior due to finite-size effects and capillary condensation. On the other side, the interplay between complete wetting, critical adsorption, and capillary condensation do not appear to have been discussed before, although it might be of relevance for adsorption-induced colloidal aggregation in binary liquid mixtures [@beysens; @guo]. Our approach is naturally suited for investigating interesting and relevant aspects of these important phenomena. The paper is organized as follows. In Sec. \[sec:2\] theoretical predictions for semi-infinite systems are summarized. In Sec. \[sec:3\] we describe the microscopic model and the method. Numerical results for the phase diagram of the present model are described in Sec. \[sec:pd\]. In Sec. \[sec:wtemp\] we determine the wetting temperature for different ranges of the boundary fields. Results for the magnetization profiles, the thickness of the wetting layer, and the adsorption along the isotherms and along the line of the pseudo-phase coexistence are reported in Sec. \[sec:res\]. Section \[sec:5\] summarizes and concludes our results. Description of semi-infinite systems {#sec:2} ==================================== A useful global characterization of the interfacial structure near a single wall is provided by the coverage $\Gamma$ defined as the excess number of fluid particles per area adsorbed on the confining substrate: $$\label{eq:1} \Gamma=\int_0^{\infty}(\rho(y)-\rho_b)dy$$ where $\rho _b$ is the bulk number density for a given temperature $T$ and chemical potential $\mu$; here the fluid number density profile $\rho({\bf r})\equiv \rho(y)$ is assumed to vary only in the direction normal to the wall located at $y=0$. Complete wetting {#subsec:cw} ---------------- If the substrate potential is sufficiently strong there is a wetting transition temperature $T_w$ such that, if the bulk gas phase approaches gas-liquid coexistence $\mu_0(T)$ along isotherms at temperatures $T>T_w$, $\Gamma$ diverges due to complete wetting, i.e., a macroscopically thick wetting film is formed. The equilibrium thickness $\ell_0$ of the wetting film can be defined as $$\label{eq:el} \ell_0=\Gamma /(\rho_l-\rho_g),$$ where $\rho_l$ and $\rho_g$ are the bulk number densities of the liquid and gas phase, respectively, at coexistence. In magnetic language the fluid is an Ising ferromagnet, the gas phase corresponds to the spin down phase, the liquid phase to the spin up phase, and the difference between the substrate potential and its analogue for fluid-fluid interactions to a surface field; the undersaturation $\Delta \mu=(\mu-\mu_0(T))/(k_BT_c)$ is proportional to the bulk field $H$. In the [*c*]{}omplete wetting regime the increase of the adsorption upon approaching bulk coexistence can be described as [@dietrich] $$\label{eq:2} \Gamma (\Delta \mu \to 0,T)\sim |\Delta \mu|^{-\beta_s^{co}}, \quad T_w<T<T_c.$$ The exponent $\beta_s^{co}$ for this [*s*]{}urface quantity depends on the form of the fluid-fluid and substrate-fluid forces as well as on the spatial dimension $d$. In $d=3$, $\beta_s^{co}$ is non-universal; $\beta_s^{co}=0$ (i.e., $\sim \ln\Delta \mu$) for short-ranged forces whilst $\beta_s^{co}=1/p$ for wall-fluid and fluid-fluid pair potentials decaying as $r^{-(d+p)}$ ($p=3$ for non-retarded dispersion forces). Because the upper critical dimension for complete wetting with long-ranged forces is $d^{*}_s=3-4/(p+1)<3$ [@dietrich; @schick; @lipowsky] these mean-field exponents $\beta_s^{co}=1/p$ are not altered by interfacial fluctuations. For short-ranged forces, i.e., $p\to\infty$ one has $d^{*}_s=3$ so that fluctuations matter in $d=3$, but it turns that for complete wetting they change only the amplitude of the thickness of wetting film [@dietrich]. In $d=2$ interfacial fluctuations in the wetting films are much stronger and $2=d<d^{*}_s$ for both short-ranged and dispersion forces ($p=4$ for the latter in $d=2$). Accordingly the complete wetting exponent takes a universal value which turns out to be given by $\beta_s^{co} =1/3$, provided $p\geq 3$ so that fluctuations dominate [@schick; @lipowsky]. From the point of view of an effective interface Hamiltonian this latter universality of $\beta_s^{co}$ is due to the entropic effects of the unbinding interface which give rise to an effective repulsive interaction for the gas-liquid interface, taken to be located on average at $y=\ell$, which decays $\sim {\ell}^{-\kappa}$ with $\kappa=2(d-1)/(3-d)$; $\kappa(d=2)=2$. If this entropic repulsion dominates the effective interaction contribution $\sim {\ell}^{-(p-1)}$, i.e., if $p > \kappa +1$ one finds $\beta_s^{co}=1/3$ in $d=2$; this defines the so-called weak fluctuation regime for complete wetting. According to this argument, for $p<3$, one has $\beta_s^{co}=1/p$. The considerations leading to the above predictions are valid only if the equilibrium wetting film thickness $\ell_0 $ is much larger than the bulk correlation length $\xi $, i.e., $\ell_0 \gg \xi$. Critical adsorption {#subsec:ca} ------------------- Near a critical point $T_c$, a confining wall generically provides an effective surface field $h_1$ acting on the order parameter (OP) $m$ describing the continuous phase transition and leading to the so-called critical adsorption [@binder; @diehl; @fdg]. The ensuing decay of the OP profile $m(y)\equiv (\rho(y)-\rho_c)/\rho_c$, where $\rho_c$ is the critical density, follows the power law $$\label{eq:3} m(y)\sim y^{-\beta/\nu}, \qquad T=T_c,$$ where $\beta$ and $\nu$ are the known critical exponents of the bulk OP and of the correlation length $\xi$, respectively. Off the critical point a crossover to the exponential decay $$\label{eq:4} m(y)\sim \exp(-y/\xi)$$ takes place at the distance $y\sim \xi$ from the surface. This is described by universal scaling functions $P_{\pm}$ such that $m_{\pm}(y,\tau,\Delta \mu)=m_b^{(0)} P_{\pm}(y/\xi_{\pm}(\tau,\Delta\mu),\Delta \mu|\tau|^{-\Delta})$, where $m_b^{(0)}=m(y\to\infty,T<T_c,\Delta\mu=0^{-})=m_0|\tau|^{\beta}<0$, $\tau=(T-T_c)/T_c$, and $\Delta$ is the so-called gap exponent; $\pm$ corresponds to $\tau \gtrless 0$. Accordingly the adsorption $\Gamma_{\pm}$ diverges as $$\label{eq:5} \Gamma_{\pm} = m_0|\tau|^{\beta}\xi_{\pm}(\tau,\Delta\mu)K_{\pm}(\Delta\mu|\tau|^{-\Delta}),$$ where $K_{\pm}(x)=\int_0^{\infty}[P_{\pm}({\tilde y}_{\pm},x)-P_{\pm}(\infty,x)]d{\tilde y}_{\pm}$ [@floeter]. Near the critical point the bulk correlation length acquires the scaling form $$\label{eq:6} \xi_{\pm}(\tau, \Delta\mu) = |\tau|^{-\nu}\Xi_{\pm}(\Delta\mu |\tau|^{-\Delta}),$$ where $\Xi_{\pm}$ are scaling functions. Since $\Xi_{\pm}(0)= const =\xi_0^{\pm}$ one has \[eq:xitau\] \_(,0)= \_0\^||\^[-]{}, whereas $\Xi_{\pm}(x\to \pm \infty)\sim |x|^{-\nu/\Delta}$ and thus at $\tau=0$ \[eq:ximu\] (0,)=\_0\^[()]{}||\^[-/]{}. Together with $K_{\pm}(x\to 0)=const$ and $K_{\pm}(x\to \pm\infty)\sim |x|^{\beta/\Delta}$, this implies that upon approaching $(T_c, \mu_0(T_c))$ along the path $\Delta\mu=0$ the adsorption diverges according to Eq. (\[eq:I\_1\]) whereas it diverges according to Eq. (\[eq:I\_2\]) along the path $\tau=0$. We note that both Eq. (\[eq:I\_1\]) and Eq. (\[eq:I\_2\]) are consistent with the scaling behavior of $\Gamma\sim m\xi$ due to $m\sim |\tau|^{\beta}$ and $\xi\sim|\tau|^{-\nu}$ for $\Delta \mu=0$ and due to $m\sim |\Delta\mu|^{1/\delta}$, $\delta=\Delta/\beta$, and $\xi\sim |\Delta\mu|^{-\nu/\Delta}$ for $\tau=0$. Within MFT $\beta=\nu=1/2$ which results in a logarithmic divergence along both paths. Equation  (\[eq:I\_2\]) is expected to hold [@dietrich:88:0] also below $T_c$ for $|\Delta\mu|\gg |\tau|^{\Delta}$ on both the wetting and non-wetting side of the coexistence curve, i.e., if the argument of the scaling functions $\Xi_{\pm}$ approaches infinity so that the behavior of $\xi_{\pm}$ is governed by $\Delta\mu$ (Eq. (\[eq:ximu\])). This regime is referred to as the [*critical adsorption regime*]{}. Similarly, above $T_c$ Eq. (\[eq:I\_1\]) is expected to be valid within the regime $|\Delta\mu|\ll |\tau|^{\Delta}$, i.e., where the behavior of $\xi$ is governed by $\tau$ (Eq. (\[eq:xitau\])). The same should hold close to bulk coexistence below $T_c$. However, it has been argued [@dietrich:88:0; @tarazona:87; @fenistein:02] that even near the critical point long-ranged forces are important, in the sense that for $\tau$ fixed and $\Delta \mu\to 0^-$ the system will [*always*]{} cross over to the complete wetting regime (Eq. (\[eq:2\])) although upon approaching $T_c$ (i.e., for smaller $\tau$) this crossover will occur closer and closer to the coexistence curve. The expectation for this scenario to hold is based on the argument that long-ranged forces are relevant and thus dominant for $\ell_0 \gg \xi$. Since $\ell_0 \sim |\Delta \mu|^{-\beta_s^{co}}$ and $\xi \sim |\tau|^{-\nu}$, this leads to the condition $|\Delta \mu| \ll |\tau|^{\nu/\beta_s^{co}}$, i.e., $|\Delta \mu| \ll \tau^{3\nu}$ for dispersion forces in $d=3$. Thus on the gas side region between the coexistence curve $\Delta \mu=0$ and the curve $\Delta \mu =const\times |\tau|^{\nu/\beta_s^{co}}$ the adsorption should be governed by wetting phenomena whereas the divergence of the adsorption $|\tau|^{\beta-\nu}$ according to Eq. (\[eq:I\_1\]) should be limited to the range $|\Delta\mu|^{1/\Delta} \ll |\tau| \ll |\Delta \mu|^{\beta_s^{co}/\nu}$. Furthermore, the effective interface Hamiltonian approach predicts that contrary to the critical adsorption regime, in the wetting dominated region the divergence should depend on the choice of the thermodynamic path taken. In particular, for any isotherm within this regime one expects $\Gamma\sim |\Delta\mu|^{-\beta_s^{co}}$, but along a path $\Delta\mu=const|\tau|^x$ with $x>\nu/\beta_s^{co}$ one expects $\Gamma\sim |\tau|^{\beta-x\beta_s^{co}}$. This follows from Eq. (\[eq:el\]) with $\Delta \rho \sim |\tau|^{\beta}$ and $\ell_0\sim |\Delta \mu|^{-\beta_s^{co}}$. Crossover phenomena {#subsec:cross} ------------------- These regimes described above are expected to give rise to rich crossover phenomena for the adsorption $\Gamma (\Delta\mu,T)$ upon crossing boundary lines between them along various thermodynamic paths. Two of them are particularly relevant for the present work: (i) isotherms $T=const<T_c$ with $\Delta \mu\to 0^-$ (see path (II) in Fig. \[fig:0\]) and (ii) a path parallel to the coexistence curve on the gas side with a small undersaturation $\Delta \mu=const < 0$ and with $T\to T_c$ (path (I) in Fig. \[fig:0\]). Along an isotherm $T=const >T_w$ and below the curve $\Delta \mu \sim -|\tau|^{\Delta}$ (see Fig. \[fig:0\]) $\Gamma$ increases as $|\Delta \mu|^{(\beta-\nu)/\Delta}$ upon approaching bulk coexistence from the gas side, i.e., for decreasing $|\Delta \mu|$, until one enters the crossover region between the lines $\Delta\mu=-const |\tau|^{\Delta}$ and $\Delta\mu=-const |\tau|^{\nu/\beta_s^{co}}$ in which $\Gamma$ increases further; but therein no specific and well defined power law can be expected. Finally, when the path crosses the curve $\Delta\mu=-const|\tau|^{\nu/\beta_s^{co}}$ it enters the regime governed by wetting phenomena so that the adsorption should diverge as $|\Delta\mu|^{-\beta_s^{co}}$. For $\mu=\mu_0$ the wall is wet and the adsorption is infinite (provided $T_w<T\le T_c$). The behavior of the adsorption along a path parallel to the coexistence curve is expected to be equally rich. Let us consider the case that the wetting transition at coexistence is continuous at $T=T_w$. According to the theory of wetting phenomena [@dietrich] the adsorption $\Gamma$ increases smoothly to some finite value (because $|\Delta\mu| >0$) upon approaching the wetting temperature $T_w$. This increase is governed by the scaling laws with respect to $\Delta\mu$ and $\tau$ which are associated with critical wetting. Upon a further increase of the temperature the adsorption should slightly decrease because $\Gamma=\ell_0 \Delta \rho$, $\ell_0=const$ due to $\Delta\mu =const$, and $\Delta \rho=\rho_l-\rho_g\sim |\tau|^{\beta}$ until the crossover line $\Delta\mu=-const|\tau|^{\nu/\beta_s^{co}} $ to the critical adsorption regime is reached. There $\Gamma$ should increase $\sim |\tau|^{\beta-\nu}$ until the next crossover line $\Delta\mu=-const |\tau|^{\Delta}$ is encountered beyond which $\Gamma$ as a function of $\tau$ should saturate at a certain large value $\sim |\Delta\mu|^{(\beta-\nu)/\Delta}$. Finally, after passing the right branch of the crossover line $\Delta\mu \sim -\tau^{\Delta}$ (above $T_c$) the adsorption should decrease $\sim \tau^{\beta-\nu}$ upon increasing $\tau$. ![ Scaling behavior of the capillary condensation line $H_{ca}(\tau,L)$ near bulk criticality $(T=T_c,H=0)$. For $\tau=(T-T_c)/T_c\ne 0$ and $L\to \infty$, $H_{ca}(\tau,L)$ approaches bulk coexistence $H=0$ as $|\tau|^{\Delta-\nu}/L$ and the capillary critical point $(T_{c,L},H_{c,L})$ approaches $T_c$ and $H=0$ as $L^{-1/\nu}$ and $L^{-\Delta/\nu}$, respectively. The latter is valid for strong surface field and takes the form $L^{-(\Delta-\Delta_1)/\nu}$, where $\Delta_1$ is the surface gap exponent, for weak $h_1$  [@BLM]. Here $T_{c,L}<T_c$. Note the upward bent of the capillary condensation line; this feature is due to $\Delta-\nu<1$ (see, c.f., the text following Eq. (\[eq:capsc\])). However, this bent is weak: $\Delta-\nu=7/8=0.875$ in $d=2$ and $\Delta-\nu\simeq 0.936$ in $d=3$. []{data-label="fig:12"}](rys13.eps){width="7.5cm"} Microscopic model {#sec:3} ================= In this section we introduce the microscopic model within which we investigate quantitatively the crossover regimes described above. Specifically, we consider an Ising ferromagnet in a slit geometry subject to identical boundary fields. Contingent on the type of numerical approach we shall use, our results refer to $d=2$ strips defined on a square lattice of size $M\times L, M\to \infty$. The lattice consists of $L$ parallel rows at spacing $a$, so that the width of the strip is $La$; in the following we set $a=1$. Successive rows are labeled by an index $j$. At each site, labeled $(k,j)$, there is an Ising spin variable taking the value $\sigma_{k,j}=\pm 1$. The boundary surfaces are located in the rows $j=1$ and $j=L$ and periodic boundary conditions (PBCs) are assumed in the lateral $x$ direction. Our model Hamiltonian for the strip with PBCs and $M\to \infty$ is given by $$\begin{aligned} \label{eq:3_1} {\cal H} & = & -J\left( \sum_{\langle kj,k'j'\rangle}\sigma_{j,k}\sigma_{j',k'}\right. \nonumber \\ \mbox{}&\mbox{}&+\left.\sum_{j=1}^{L}V^{ext}_{j,L}\sum_k\sigma_{k,j}+H\sum_{k,j}\sigma_{k,j}\right),\end{aligned}$$ where the first sum is over all nearest-neighbor pairs and the external potential is measured in units of $J>0$. $V^{ext}_{j,L}=V^{s}_{j}+V^{s}_{L+1-j}$ is the total boundary field experienced by a spin in the row $j$; it is the sum of the two independent wall contributions. The [*s*]{}ingle-boundary field $V^s_{j}$ is assumed to have the form $$\label{eq:3_2} V^s_{j}=\frac{h_1}{j^p}$$ with $p>0$ and $h_1>0$. $H$ is a bulk magnetic field. According to Eq. (\[eq:3\_1\]) $h_1$ and $H$ are dimensionless. This model can be viewed as being obtained from a $2d$ lattice gas model mimicking a two-dimensional one-component fluid with a short-ranged interaction potential between the fluid particles and either short-ranged or long-ranged substrate potentials. The equivalence between the lattice gas and the Ising model implies the following relationships (see, e.g., Ref. [@pandit]): the bulk magnetic field $H$ in the former is proportional to the deviation of the chemical potential from the bulk phase boundary $\mu_0(T)$ in the latter, i.e., $H\sim \Delta \mu$. The lattice-gas analogue of the number density in the fluid is related to the magnetization by $\rho=(m+1)/2$, so that $\Delta \rho=\rho_l-\rho_g\sim 2m_b$, where $m_b$ is the spontaneous magnetization. Finally $4J$ corresponds to the strength of the attractive pair potential between the fluid particles, taken to be short-ranged so that in the lattice gas model it can be modelled by nearest neighbor interactions. $V^{ext}$ is a combination of the substrate potential and the liquid-liquid interaction. These relationships can be extended to binary liquid mixtures  [@getta]. We recall that although there is no longer any true phase transition for finite $L$, in two-dimensional Ising strips with large $L$ there is still a line of extremely weakly rounded first-order transitions ending at a pseudocritical point the location of which in the plane $(H,T)$ spanned by the bulk field and the temperature $T$ depends on the character of the surface fields [@fisherprivman; @binder:08]. For surfaces which prefer the same bulk phase, the phenomenon equivalent to capillary condensation takes place. The pseudo-phase coexistence between phases of spin up and spin down occurs along a line $H_{ca}(T,L)$, which is given approximately by the analogue of the Kelvin equation [@kelvin]: $$\label{eq:Kelvin} H_{ca}(T,L)\approx -\sigma(T)/\left(L|m_b^{(0)}(T)|\right),$$ where $\sigma(T)$ is the interfacial tension (divided by $J$) between the coexisting bulk phases and $m_b^{(0)}(T)<0$ is the spontaneous bulk magnetization for $H=0^{-}$. For the $d=2$ Ising model, the surface tension is given exactly $\beta J \sigma(T)=2(K-K^*)$, where $K=J/k_BT$ and $K^*$ satisfies $\sinh 2K\sinh 2K^*=1$. The occurrence of thick wetting films of $+$ spins at the two surfaces, for a thermodynamic bulk state corresponding to $-$ spins, gives rise to nontrivial corrections to Eq. (\[eq:Kelvin\]) which shift the condensation line to larger values of $|H|$ [@binder:08; @evans:90]. The pseudo-coexistence line ends at a pseudo-capillary critical point $(H_{c,L},T_{c,L})$ where $T_{c,L}(h_1,p)$ lies below the temperature $T_c$ of the bulk critical point. Its position as well as that of $H_{c,L}$ depends on $L$ and the strength and the range of the surface fields. For large $L$ and strong $h_1$ the shifts of the critical temperature and of the bulk field $H_{c,L}(h_1,p)$ are given by [@binder:08; @fisher] (see, c.f., Fig. \[fig:12\]) $$\label{eq:shifts} T_{c,L}-T_c\sim -L^{-1/\nu}, \quad H_{c,L}\sim -L^{-\Delta/\nu}$$ with $\nu=1$ and $\Delta=15/8$ for the two-dimensional Ising model. Phase diagram {#sec:pd} ============= In Fig. \[fig:1\] we show the phase diagram for the present model calculated by using the DMRG method for a strip of width 340 with $h_1=0.8$ and three choices of the parameter $p$ describing the decay of the boundary field: $p=50, 3$, and 2. The case $p=50$ is expected to resemble the behavior corresponding to short-ranged surface forces. In this figure we display the various crossover lines discussed in Sects. \[sec:1\] and \[sec:2\] and the thermodynamic paths along which we have calculated the adsorption $\Gamma$. The pseudo-phase coexistence line, the crossover lines, and the adsorption have been determined for $2d$ Ising strips within the DMRG method. ![Phase diagram for a $d=2$ Ising strip subject to identical boundary fields $V^{ext}_{j}$ (see Eq. (\[eq:3\_2\])) obtained by using DMRG for a strip width $L=340$ and the amplitude $h_1=0.8$ for the boundary fields. The thick solid line indicates the bulk coexistence line. The lines interpolating the symbols represent the pseudo-phase coexistence lines $H_{ca}(T,L;p)$ for three different values of the exponent $p$ governing the algebraic decay of the boundary fields: open circles correspond to $p=2$, open squares correspond to $p=3$, and stars correspond to $p=50$. Various crossover lines $H_{cr}^{(1)}$ and $H_{cr}^{(2)}$ with $d_1=1=d_2=1$ and $\nu=1$ and $\Delta=15/8$ in $d=2$, as discussed in the main text, are shown as well as the thermodynamic paths $(0)-(4)$ along which the adsorption has been calculated. $(1)-(4)$ are various isotherms and $(0)$ runs parallel to the pseudo-coexistence line. Thus path (0) resembles (in a certain sense, see later) a path of type (I) in Fig. \[fig:0\] whereas the paths $(1)-(4)$ do correspond to a path of type (II) in Fig. \[fig:0\].[]{data-label="fig:1"}](rys02.eps){width="7.5cm"} The thick solid line in Fig. \[fig:1\] indicates the bulk phase coexistence line $(H=0, T<T_c)$ terminating at the bulk critical point $(H=0,T=T_c=[\ln(1+\sqrt 2)]^{-1}J/k_B\simeq 2.269 J/k_B)$. The symbols (open circles for $p=2$, open squares for $p=3$, and stars for $p=50$) show the pseudo-phase coexistence. It turns out that the pseudo-phase coexistence line for long-ranged boundary fields is located slightly further away from the bulk coexistence line, especially at lower temperatures, than the pseudo-phase coexistence line for the short-ranged boundary fields ($p=50$). These pseudo-lines have been identified as those positions $(H,T)$ in the phase diagram where the total magnetization of the strip vanishes, i.e., $\sum_{j=1}^Lm_{j}=0$ with $m_{j}=\langle \sigma_{k,j} \rangle$. At fixed $T$ (fixed $H$) at those positions the free energy of the strip exhibits a maximum as a function of $H$ ($T$) which is the rounded remnant of the nonanalyticity (kink) of the free energy at the first-order phase transition in the bulk ($L=\infty$) system. As can be seen in Fig. \[fig:1\], for $T<T_c$ $(T>T_c)$ the line defined by the zeros of the total magnetization moves to less (more) negative values of $H$ upon increasing $T$. Determining the pseudocritical temperature $T_{c,L}$, where the pseudo-phase coexistence line ends, is difficult because in this quasi-one-dimensional system the critical point is not a sharp concept and one has to examine various criteria in order to estimate $T_{c,L}$. However, these different criteria, such as the maximum of the specific heat, or the erosion of the jumps in the adsorption and in the solvation force upon crossing the line $H_{ca}(T,L;p)$, provide somewhat different estimates for $T_{c,L}$. Determining $T_{c,L}$ for different values of $p$ and $L$ is beyond the scope of the present analysis. For the width $L=340$ and $h_1=0.8$ we approximately obtain $T_{c,L}\approx 2.2$; for more details concerning the short-ranged case see Ref. [@maciolek:01]. We have investigated how the location of the pseudo-phase coexistence line $H_{ca}(T,L;p)$ of capillary condensation changes as a function of $p$ for fixed width $L$, fixed amplitude $h_1$ of the surface field, and fixed temperature $T$. In Fig. \[fig:2\] we plot the logarithm of the difference $H_{ca}(T,L;p)-H_{ca}(T,L;\infty)$ as a function of $p$ calculated for $h_1=0.8$ and (i) for several values of the temperature at fixed $L=300$ (Fig. \[fig:2\](a)) as well as (ii) for several values of $L$ at fixed $T=1.8 J/k_B$ (Fig. \[fig:2\](b)). (Note that $p=\infty$ corresponds to a pure surface contact field at $j =1, L$ (see Eq. (\[eq:3\_2\])).) From these plots we can clearly distinguish two different regimes for the behavior of the pseudo-phase coexistence line, which are separated by a crossover region occurring for $1\lesssim p \lesssim 3$. In both of these two regimes the shift $H_{ca}(T,L;p)-H_{ca}(T,L;\infty)$ varies exponentially but with different decay constants. Moreover, for $p\lesssim 1 $ the shift relative to the short-ranged pseudo-phase coexistence line does not depend on the temperature; the dependence on the strip width $L$ becomes, for the range of $L$ considered here, very weak and finally negligible for $p\lesssim 0.25$, i.e., $H_{ca}(T,L;p)-H_{ca}(T,L;\infty)\approx B_0\exp (-Ap)$ with $A\simeq 5.2$ and $B_0\simeq 1.5$. For $p\gtrsim 3$ one has $H_{ca}(T,L;p)-H_{ca}(T,L;\infty)\simeq B(L,T)\exp (-Cp)$ with $C\simeq 0.723$. We note that, as expected, for $p$ fixed wider strips and higher temperatures give rise to smaller shifts of the pseudo-phase coexistence line. {width="7.5cm"} ![The relative shift of the pseudo-coexistence line for Ising strips as function of the range of the boundary fields characterized by the decay exponent $p$ (Eq. (\[eq:3\_2\])). The calculations are carried out by using the DMRG method for $h_1=0.8$ and (a) at fixed $L=300$ and for four temperatures $T^*=k_BT/J$; (b) at fixed temperature ($k_BT/J=T^*=1.8$) and for four thicknesses $L$ of the strip.[]{data-label="fig:2"}](rys03b.eps){width="7.5cm"} Determination of the wetting temperature for different ranges of the boundary fields {#sec:wtemp} ==================================================================================== It is important to assure that the range of temperatures, for which we perform our calculations of the adsorption, lies above the wetting temperature $T_w(p,h_1)$. For $h_1=0.8$ and a short range of the surface fields $(p=\infty)$ the wetting temperature in the semi-infinite $d=2$ Ising system equals $T_w(p=\infty,h_1=0.8)\simeq 1.41 J/k_B$ and the wetting transition is second order. Adding the long-ranged tail to the boundary fields shifts the wetting temperature. Wetting transitions in $d=2$ in the presence of long-ranged substrate potentials have been discussed by Kroll and Lipowsky [@kroll] within the continuum one-dimensional solid-on-solid model, with the conclusion that there is no wetting transition for fields which decay more slowly than $1/j^3$. In this case their prediction is that the interface remains pinned to the wall at all finite temperatures. This conclusion has been supported by providing upper and lower bounds for the ground-state energy $E_0$ of the corresponding Schrödinger equation. Within this approach the existence of the bound-state solution of the Schr" odinger equation ($E_0<0$) corresponds to a localized interface. When the bound state ceases to exists, $E_0\to 0$ signals the critical wetting transition. The construction of upper and lower bounds (based on using an exponentially decaying trial wave function) shows that the ground-state energy has a non-zero value (i.e., there is no transition) for a finite potential strength. The case $p=2$ has been studied by Privman and [Ŝ]{}vraki[ć]{} in Ref. [@privman] also within the SOS model; they argued that for attractive effective potentials decaying like $1/r$ the wetting transition is no longer sharp (becomes rounded). For the $d=2$ Ising ferromagnet the wetting phenomena for the case of the marginal range $p=3$ have been studied within the DMRG approach [@drzew] and by MC simulations [@albano; @devirgiliis]. These studies have shown that the presence of long-ranged tails in the boundary fields decreases the wetting temperature with respect to the case of the short-ranged surface fields and that $T_w(p=3,h_1=0.8)\simeq 0.75J/k_B $. Apart from Refs. [@kroll; @privman], the case $p<3$ has not been studied so far. In order to check the predictions by Kroll and Lipowsky, we investigate whether in the $d=2$ Ising ferromagnet with $p=2$ a wetting transition exists. In the finite systems we are studying, below the bulk critical temperature the wetting transition as a function of $h_1$ for any range $p$ of the boundary fields can be inferred from the so-called (weakly rounded in $d=2$) interface localization-delocalization (ILD) transition [@binder:08; @parryevans] which occurs in strips with antisymmetric boundary fields, i.e., for $V_{j}^{ext}=V_{j}^s-V_{L+1-j}^s$. This transition is the precursor of a wetting phase transition that occurs in the limit of infinite film thickness $(L\to\infty)$ at the critical curve $T_w(p,h_1)$. For $T<T_w(p,h_1)(T>T_w(p,h_1))$ such an interface is bound to (unbound from) the walls [@binder:08; @parryevans]. First, we note that $T_{w}(h_1\to 0,p)\to T_c$ for any $p$ (see Eq. (\[eq:3\_2\])). On the other hand $h_{1w}(T=0,p)$, which is the solution of the implicit equation $T_w(h_1,p)=0$ and denotes the critical surface field strength beyond which the system is wet even at $T=0$, shifts towards lower values upon decreasing $p$. There is no reason to expect a non-monotonic behavior of $T_w(h_1,p)$. Therefore, the gross features of the shape of the wetting transition line $T_w(h_1,p)$ for an arbitrary $p> 1$ can be inferred from localizing the position of the wetting transition $h_{1w}(T=0,p)$ in the ground state. The ground-state energy of the system (in units of $J$) can be found directly from the Hamiltonian with a vanishing bulk field. Because the system is translationally invariant along a strip, it is sufficient to consider only the configurations of a single column. In a [*p*]{}artial [*w*]{}etting regime with antisymmetric surface fields there are only two coexisting states (all spins up or all spins down) with the energy (per the number $M$ of columns) $$\begin{aligned} E_{pw}^{\pm} & = & -J(L-1)\mp h_{1}J \left( 1-\frac{1}{L^p}+\frac{1}{2^p}-\frac{1}{(L-1)^p}+\ldots \right. \mbox{}+\left.\frac{1}{(L-1)^p}-\frac{1}{2^p}+\frac{1}{L^p}-1 \right) \nonumber \\ &=&-J(L-1).\end{aligned}$$ In a wet regime at least one interface between spin up and spin down configurations has to be present. For even $L$ the lowest-energy configuration is that for the state with the interface located in the middle of the strip. For $p \to \infty$ (short-ranged case) $L-1$ degenerate states emerge with a single interface positioned at any of the rows but the ones closest to the surfaces. In general, the energy of a single column with an interface in the middle of a strip (spin up for $j=1,\ldots, L/2$ and spin down for $j=L/2+1,\ldots, L$) is $$\begin{aligned} \label{eq:gs_2} E_{wet} & = & -J(L-3)-h_{1}J \left( 1-\frac{1}{L^p}+\frac{1}{2^p}-\frac{1}{(L-1)^p}+\ldots \right. \mbox{}+\frac{1}{(L/2)^p}-\frac{1}{(L/2+1)^p} \nonumber \\ &-&\frac{1}{(L/2+1)^p}+\frac{1}{(L/2)^p}+\ldots \mbox{}-\left.\frac{1}{(L-1)^p}+\frac{1}{2^p}-\frac{1}{L^p}+1\right).\end{aligned}$$ In the ground state and at the ILD transition the energies $E_{pw}$ and $E_{wet}$ are equal. For a particular $L$ this determines the magnitude $h_{1w}^{ILD}(L)$: $$\begin{aligned} \label{eq:gs_3} h_{1w}^{ILD}(L)& = & 1/\left(1+\frac{1}{2^p}+\frac{1}{3^p}+\ldots+\frac{1}{(L/2)^p}-\frac{1}{(L/2+1)^p}-\ldots-\frac{1}{(L-1)^p}-\frac{1}{L^p} \right)\nonumber \\ & = & \left\{ \sum_{n=0}^{\infty}\frac{1}{n^p}-2\sum_{n=L/2+1}^{\infty}\frac{1}{n^p}+\sum_{n=L+1}^{\infty}\frac{1}{n^p} \right\}^{-1}.\end{aligned}$$ In order to find the critical wetting field $h_{1w}(L\to\infty)$ we take the limit $L\to\infty$. According to the second line in Eq. (\[eq:gs\_3\]) this leads to $$\label{eq:gs_4} h_{1w}=1/\sum_{n=0}^{\infty}\frac{1}{n^p}=1/\zeta(p)$$ where $\zeta(p)$ is the Riemann zeta function. Its values are known analytically only for certain even values: $\zeta(p=2)=\pi^2/6$, $\zeta(p=4)=\pi^4/90$ which gives $h_{1w}(p=2)\approx 0.6079$ and $h_{1w}(p=4)\approx 0.9239$. Other values can be found in tables of special functions, e.g., $\zeta(p=3)\approx 1.2021$ giving $h_{1w}(p=3)\approx 0.8319$. In the short-ranged limit ($p \to \infty$) the Riemann zeta function approaches $1$, confirming Abraham’s solution at $T=0$ [@abraham]. Even more interesting is the opposite limit $\zeta(p \to 1) \to \infty$ which results in $h_{1w}(p\to 1) \to 0$. This means that at $T=0$ the wetting transition does not exist for $p \leq 1$. ![Phase diagram at bulk coexistence $H=0$ for continuous wetting transitions in the $d=2$ Ising ferromagnet for (Eq. (\[eq:3\_2\])) $p=2$, $p=3$, and $p=50$ (quasi-short-ranged boundary field) obtained within the DMRG approach. For $p>1$ the values in the limit $T\to 0$ are known exactly for semi-infinite systems. The dashed line is the analytically known exact result for $p=\infty$ and for semi-infinite systems  [@abraham]. $T^*_w=k_BT_w/J$.[]{data-label="fig:3"}](rys04.eps){width="8.0cm"} In order to determine the location of a quasi-ILD transition between $T=0$ and $T=T_c$ for $p<3$ various criteria can be applied [@drzew]. Here we have adopted the approach involving the magnetic susceptibility $\chi$. The singularity (or a maximum) of the magnetic susceptibility $\chi$ is one of the most useful criteria for the localization of a phase transition (or of a pseudo-phase transition for a finite system). The magnetic susceptibility can be calculated as the second derivative of the free energy $f$ with respect to the bulk magnetic field $H$. This method is very convenient for the DMRG approach because the latter provides the free energy with a very high accuracy. Nevertheless the present case is somewhat special, because we want to determine $T_w(p,h_1)$ at $H=0$ (see Fig. \[fig:3\]), where in the partial wetting regime, i.e., for $T<T_w$, there is a first-order bulk transition. In the thermodynamic limit there is coexistence of phases with opposite magnetizations. Thus there is a discontinuity of the first derivative of the free energy $f$ (a jump of the magnetization $m=-\partial f/\partial H$) upon changing the sign of the bulk magnetic field. Accordingly, in order to calculate $\chi$ there, one has to calculate the derivatives for small nonzero bulk fields and then to consider the limit $H\to 0$. In the complete wetting regime, i.e., $T>T_w$, or equivalently above the ILD transition, the finite system exhibits only a single phase with an interface meandering freely between the walls so that there is no discontinuity of the free energy derivatives upon crossing $H=0$. For numerical calculations such as the ones within the DMRG method the necessity of performing an extra limiting procedure ($H \to 0$ in this case) is cumbersome. Therefore, instead of $\chi$, we have focused on another quantity $\chi_{0}$, which also corresponds to the second derivative of the free energy at fixed $T$ and $h_1$, but is calculated numerically in a symmetrical way with respect to $H=0$ by taking the free energy values at five equidistance points: $-2\Delta H$, $-\Delta H$, $0$, $\Delta H$ or $2\Delta H$; we typically used $\Delta H = 10^{-5}$. Because our calculations are always carried out for finite $L$, there is no discontinuity of the magnetization in the partial wetting regime. These discontinuities are replaced by functions which are rounded but steeply varying at $H=0$. In order to determine the ILD transition we have scanned the phase diagram at fixed $h_1$. The higher the temperature, the less steeply the magnetizations vary and the values of their derivative $\chi_{0}$ are smaller. Above the wetting temperature, where there is no discontinuity, $\chi_{0}(H=0,T)$ saturates for increasing $T$; here $\chi_{0}$ is equivalent to $\chi$. Therefore, at fixed $L$, the ILD transition can be identified by the maximal slope of $\chi_{0}$ or the minimum of its derivative with respect to temperature. Although all derivatives have been performed numerically, the high accuracy of the DMRG method guarantees very precise results. {width="7.5cm"} ![Magnetization profiles, relative to their bulk values $m_b(T,H)<0$, near one wall ($j >1$) along the isotherm $T^*=1.8$ (see path (1) in Fig. \[fig:1\]), calculated within the DMRG method for $d=2$ Ising strips of width $L=500$, for (a) short-ranged ($p=50$) and (b) for long-ranged ($p=2$) boundary fields of strength $h_1=0.8$. Along this isotherm the pseudo-capillary condensation transition occurs at $H_{ca}\simeq -0.00172$ for $p=50$ and at $H_{ca}\simeq -0.00195(4)$ for $p=2$. Thus both in (a) and (b) the full line for $H=-0.0017$ and $H=-0.0019$, respectively, corresponds to the capillary filled state with $m(j)-m_b(T,H)\simeq 2|m_b(T,H)|$.[]{data-label="fig:4"}](rys05b.eps){width="7.5cm"} Finally we have extrapolated $T_w(p,h_1;L)$ to the limit $L\to \infty$ in order to obtain the wetting temperatures $T_w(p,h_1)$. The obtained wetting phase diagram in the $(T,h_1)$ plane is shown in Fig. \[fig:3\] for $p=2, 3$, and $p=50$. The dashed line is the exact result by Abraham for $p=\infty$ [@abraham]. The close agreement of our data for $p=50$ with this exact result generates confidence in our numerical procedures. Our results show that contrary to the claim by Kroll and Lipowsky [@kroll] also for $p=2$ there are wetting transitions and that the wetting temperature of these continuous transitions can be well localized as a function of $h_1$. In view of the calculations of the adsorption, for our choice $h_1=0.8$ the surface is wetted for all temperatures. As already mentioned above, the claim by Kroll and Lipowsky is based on using a continuum planar solid-on-solid (SOS) approximation. SOS models ignore bulk-like fluctuations such as bubbles and the formation of “overhangs” of the line separating oppositely magnetized regions, while focusing on describing the long-wavelength behavior of the interface, expected to play the crucial role in wetting phenomena. In the present case of the $d=2$ Ising model the bulk fluctuations are particularly strong giving rise to a very diffuse interface. For the case of $p=2$ they might provide the mechanism for the unbinding of the interface from the wall, even though the long-wavelength interfacial fluctuations are strongly supressed by this very long-ranged boundary field. Numerical results and discussion {#sec:res} ================================ Using the DMRG method we have performed calculations of the magnetization profiles $m(j)$ from which we have obtained the thicknesses $\ell_0$ of the wetting layer and the adsorption $\Gamma$. The analysis of the shapes of the profiles $m(j)$ and of the scaling properties of $\ell_0$ gives a better understanding of the behavior of $\Gamma$ along different thermodynamic paths. We recall, that the predictions for the scaling behavior of the adsorption follows from the scaling behavior of $\ell_0$. In order to infer possible power laws governing the behavior of the thickness of the wetting layer and of the adsorption along different paths we have calculated [*local exponents*]{} of the quantities of interest as a function of $H$ or $\tau$. They are defined as $$z_i = \Bigl\lvert \frac{\ln Q (i+1) - \ln Q (i)}{\ln x_{i+1} - \ln x_{i}}\Bigr\rvert \,\, , \label{zN}$$ which is the discrete derivative of data $Q$ as a function of $x$ in a log-log plot; here $Q=\ell_0$ or $\Gamma$ and $x=H$ or $\tau$. Such a quantity probes the local slope at a given value $x_i$ of $x$ at a point $i$ along the path considered. It provides a better estimate of the leading exponent than a log-log plot itself. We have chosen the convention that if $Q$ decays algebraically as a function of $x$ then $Q\sim x^{-z}$ with $z>0$. The high quality of the DMRG data allows us to reliably carry out this numerical differentiation. Isotherms {#subsec:iso} --------- ### Magnetization profiles {#subsubsec:mp} In Figs. \[fig:4\] and \[fig:5\] we show a selection of magnetization profiles calculated for a strip width $L=500$ with $p=50$ and 2 along the two isotherms $T^*=1.8$ and $T^*=2.25$ indicated in Fig. \[fig:1\] as path (1) and (3), respectively. {width="7.5cm"} ![Magnetization profiles, relative to their bulk values $m_b(T,H)<0$, near one wall along the isotherm $T^*=2.25$ (see path (3) in Fig. \[fig:1\]), calculated within the DMRG method for $d=2$ Ising strips of width $L=500$, (a) for short-ranged ($p=50$) and (b) for long-ranged ($p=2$) boundary fields of strength $h_1=0.8$. Along this isotherm the pseudo-capillary condensation transition occurs at $H_{ca}\simeq -0.000191$ for $p=50$ and at $H_{ca}\simeq -0.00026(4)$ for $p=2$. Thus both in (a) and (b) the full line for $H=-0.0001$ corresponds to the capillary filled state with $m(j)-m_b(T,H)\simeq 2|m_b (T,H)|$.[]{data-label="fig:5"}](rys06b.eps){width="7.5cm"} From these plots one can see that within the accessible range of values $H<H_{ca}(T,L;p)$ the shapes of the profiles are [*not*]{} slab-like, even though the system size is very big so that $|H_{ca}(T,L)|$ is expected to be sufficiently small. They are characterized by a pronounced spatial variation at the emerging interface between the spin up and the spin down phase, where the profile varies quasi linearly with the distance from the wall, and by extended tails. For the short-ranged case ($p=50$) the profiles along the low-temperature isotherm ($T^*=1.8$) exhibit a rapid decay to their bulk values $m_b \equiv m_b(T,H)$ ($<0$ for $H<0$) with the bulk correlation length $\xi(T,H)= |\tau|^{-\nu}\Xi(H|\tau|^{\Delta})$ (see Eqs. (\[eq:6\])-(\[eq:ximu\])) as decay length which increases upon approaching the pseudo-phase coexistence. Only very close to the pseudo-phase coexistence line $H_{ca}(T,L;p)$ the emerging shape of the profile to a certain extent resembles that characteristic of the free interface between the spin up and the spin down bulk phases with a broad interfacial region and a rather narrow region where the magnetization stays somewhat close to the value $-m_b>0$ characteristic for the wetting phase corresponding to $H=0^+$; this latter region thickens as $H\to H_{ca}(T,L;p)$. For the long-ranged forces ($p=2$) the profiles along the low-temperature isotherm ($T^*=1.8$) exhibit an interface-like shape already away from the pseudo-phase coexistence line $H_{ca}(T,L;p)$. The interfacial part of these profiles is much more pronounced than for the short-ranged case ($p=50$) and their decay is much slower. The exponential decay of the profiles to the value $m_b(T,H)$ ultimately crosses over to the power law decay $\sim j^{-2}$. In general, algebraically decaying inter-particle and surface fields are known to generate algebraically decaying order parameter profiles [@SN]. For both $p=50$ and $p=2$ the profiles corresponding to $H\to 0$ along the high-temperature isotherm ($T^*=2.25$) decay much slower as compared to the low-temperature isotherm ($T^*=1.8$) and their shapes become quantitatively different (see Fig. \[fig:5\]). The magnetization in the first few layers near the walls decreases more rapidly so that there the profiles acquire a positive curvature. Further away from the wall there is an inflection point and the narrow pleateau occurring near the boundaries for the low-temperature isotherm (see Fig. \[fig:4\](b)) disappears. The interfacial region becomes very broad. ### Thickness of the wetting layer {#subsubsec:wl} In order to infer equilibrium thicknesses $\ell_0$ of the wetting layers from the magnetization profiles we have chosen the criterion of the change of sign of the curvature of the profile, i.e., we have assigned a thickness to a wetting layer which corresponds to the distance $j=\ell_0$ at which the profile exhibits its inflection point. Adopting other criteria for determining the wetting film thickness, e.g., the thickness which corresponds to the distance $j=j_0$ at which the magnetization vanishes, $m(j_0)=0$, leads, as it must be, to the same conclusions about physical observables. We have studied the behavior of $\ell_0$ along several isotherms. Representative data are shown and discussed below. They correspond to the isotherms $T^*= 1.8, 2.25$ (see the paths (1) and (3), respectively, in Fig. \[fig:1\]) and $T=T_c$ (see path (4) in Fig. \[fig:1\]) and have been obtained for several strip widths and ranges $p=50,5, 4, 3, 2, 1.5$ of the boundary fields. (The data for the isotherms $T^*=1.6$, 1.9, 2, and 2.2 are not shown.) We have chosen $h_1=0.8$ for which $T^*_w(p=\infty)\simeq 1.41$. All considered isotherms lie above the corresponding wetting temperatures $T_w(p)$. Along the isotherm $T^*=1.6$ the wetting layers are so thin that it is difficult to assign a suitable thickness $\ell_0$, especially for $p=50$. Along the isotherm at the higher temperature $T^*=1.8$ (path (1) in Fig. \[fig:1\]) this assignment is much clearer and the results for $\ell_0(H)$ and its local exponents are shown in Fig. \[fig:6\]. They were obtained for strips of width $L=500$ with $h_1=0.8$ and for various ranges $p$ of the boundary fields. {width="7.5cm"} ![ (a) Equilibrium wetting layer thickness $\ell_0$ in units of the lattice constant $a$ as a function of the bulk magnetic field $H$ calculated along the isotherm $T^*=1.8$ for $L=500$, $h_1=0.8$ and for the ranges $p=1.5, 2, 3, 4$, and 50 of the boundary fields. A rapid increase of $\ell_0$ upon approaching the pseudo-phase coexistence line $H_{ca}(T,L;p)$ supports the existence of the wetting transition for $p=2$ and 1.5. The values of $H_{ca}(T,L;p)$ are given by the points at which the various lines end on the right. The inset shows a log-log plot of these data. The thin solid lines indicate slopes (from top to bottom): 2/3, 1/2, and 1/3. (b) The local exponents $z$ for $\ell_0(H)$ (see Eq. (\[zN\])). For $H\to 0$, without the occurrence of capillary condensation, i.e., for $L\to \infty$, the expected exponents are $\beta_s^{co}=1/3$ $(\star)$ for $p=3, 4$, and 50, 1/2 $(\small{\blacksquare})$ for $p=2$, and 2/3 $(\bullet)$ for $p=1.5$. The dotted lines indicate the expected extrapolations $H\to 0$ for $L\to \infty$. The log-log plot supports these conclusions. We note that the local exponents do not always follow from naive visual impressions. For example, $z(p=50)>z(p=2,3,4)$, although in (a) the curves of $\ell_0(p=2,3,4)$ seem to increase somewhat stronger than the one for $\ell_0(p=50)$.[]{data-label="fig:6"}](rys07b.eps){width="7.5cm"} Path (1) lies entirely within the region between the bulk coexistence curve and the crossover boundary $H^{(1)}_{cr}= -d_1|\tau|^{3\nu}$ with $\nu =1$ for the $d=2$ Ising model, i.e., within the complete wetting regime (see Fig. \[fig:1\]). Indeed, for all values of $p$ we observe the increase of $\ell_0$ upon approaching the pseudo-phase coexistence line $H_{ca}(T,L;p)$. The lines in Figs. \[fig:6\]-\[fig:11\] end on the right when they reach the pseudo-phase coexistence line $H_{ca}(T,L;p)$ (see Fig. \[fig:1\]). Apart from this size dependence of the $H$–window with access for complete wetting, for $L=500$ and 300 (data for $L=300$ are not shown) the finite-size effects are negligible. As expected from the shape of the profiles (see Figs. \[fig:4\] and \[fig:5\]) for the same value of $H$ the wetting layer is much thicker for small values of the exponent $p$ than for short-ranged surface fields ($p=50$). The rate of increase of $\ell_0$, as characterized by the behavior of the local exponents $z$, depends on the range of the boundary field $p$ and varies along the isotherm. For $p=1.5$ the rate of increase is always largest, i.e., $z$ remains the largest along the entire isotherm (see Fig. \[fig:6\](b)). For $p=50, 4$, and 3 the local exponent $z$ decreases as $|H|\to 0$, i.e., the initial increase of the wetting layer is larger than the one observed very close to $H=0$ for which the plateau in the profile near the wall corresponding to the wetting phase at $H=0^+$ starts to form. For $p=2$ the very long-ranged nature of the boundary field enforces the occurrence of a narrow plateau in the profile near the wall even far away from $H=0$. Upon decreasing $|H|$ the width of this plateau increases and the emerging interfacial region broadens giving rise to the increase of $\ell_0$ which is initially smaller than for $p=50$. Only very close to $H=0$ this trend reverses and eventually the local exponents $z$ corresponding to $p=2$ become larger. For $p=3$ and 4 the rate of increase of $\ell_0$ remains smaller than for $p=2$ and 50 along the whole isotherm (for $|H|<0.008$). According to the predictions summarized in Sec. \[sec:2\] for semi-infinite systems $(L\to\infty)$, upon approaching the bulk coexistence line $H=0$ the thickness of the wetting film is expected to diverge for $p\ge 3$ with the exponent $\beta_s^{co}=1/3$ and for $p<3$ with the exponent $\beta_s^{co}=1/p$. As can be inferred from Fig. \[fig:6\] the local exponents $z$ calculated for $p=50, 4, 3$, and 2 tend towards their predicted values for $H\to 0$ but cannot reach them due to capillary condensation. For $p=1.5$ the local exponents have to behave non-monotonically in order to reach the expected value $1/p=2/3$. The behavior of $\ell_0$ along the isotherm $T^*=1.9$ is very similar (data not shown). Path (2) corresponds to $T^*=2.0$ and runs in between the crossover lines given by $H_{cr}^{(1)}= -d_1|\tau|^{3\nu}$ and $H_{cr}^{(2)}=-d_2|\tau|^{\Delta}$ with $\nu=1$ and $\Delta=15/8$ in $d=2$; it hits the pseudo-phase coexistence line of capillary condensation before it reaches the crossover boundary $H_{cr}^{(1)}=-d_1|\tau|^{3}$. Along that path the variation of $\ell_0$ is very similar to the one observed along the isotherm $T^*=1.8$ with the local exponents approaching the expected values of the complete wetting regime (data not shown). Again the asymptotic behavior is preempted by capillary condensation. {width="7.5cm"} ![ (a) Wetting layer thickness $\ell_0$ in units of the lattice constant $a$ as a function of the bulk magnetic field $H$ calculated along the isotherm $T^*=2.25$ (path (3) in Fig. \[fig:1\]) for $L=500$, $h_1=0.8$, and for the ranges $p=1.5, 2, 3, 4$, and 50 of the boundary fields. On the right the curves end at the corresponding capillary condensation point $H_{ca}(T,L;p)$. The inset shows a log-log plot of these data. The thin solid line indicates the slope 8/15. (b) The local exponents $z$ for $\ell_0(H)$ (see Eq. (\[zN\])). For $p\ge 2$ and $H\to 0$ they tend towards 0.48, $\ldots $, 0.52 whereas the expected value 8/15$ (\blacktriangle)$ is somewhat higher. According to Fig. \[fig:1\], along this isotherm capillary condensation prevents one to reach the ultimate complete wetting regime. For $p=1.5$ the behaviors of $\ell_0(H)$ and $z(H\to 0)$ are distinctly different from those for $p\ge 2$.[]{data-label="fig:7"}](rys08b.eps){width="7.5cm"} Results along the near-critical ($T^*=2.25$) and critical ($T^*=T^*_c$) isotherms (paths (3) and (4) in Fig. \[fig:1\]) are displayed in Fig. \[fig:7\] and Fig. \[fig:8\], respectively, for $L=500$. (For $L=300$ we observe a similar behavior; these data are not shown.) Both paths lie in the critical adsorption regime where one expects $\ell_0(H\to 0) \sim \xi \sim |\Delta H|^{-\nu/\Delta}$ with $\nu/\Delta=8/15$ in $d=2$. This universal behavior is expected to hold if the long-ranged boundary field decays sufficiently rapidly, i.e., if $p>(d+2-\eta)/2$. In this case the long-ranged part of the boundary field is irrelevant in the RG sense with respect to a pure contact surface field [@diehl]. Here $\eta$ is the critical exponent governing the algebraic decay of the two-point correlation function in the bulk and at $T_c$. In the present case of the $d=2$ Ising model $\eta(d=2)=1/4$ so that for $p> 15/8$ we expect to observe the power law $\ell_0 \sim |\Delta H|^{-8/15}$. For $p=1.5$ one has $p<(d+2-\eta)/2$. To our knowledge so far this case has not been studied in the literature in the context of critical adsorption. According to general scaling arguments [@diehl] the magnetization profile for the semi-infinite system under rescaling of distances by a factor $b$ behave as $$\begin{aligned} \label{eq:mlr} m_{\pm}(y,\tau,H,h_1,p) = &\mbox{} \nonumber \\ & b^{-\beta/\nu}m_{\pm}(yb^{-1},\tau b^{1/\nu},Hb^{\Delta/\nu},h_1b^{\omega_s}), \nonumber \\\end{aligned}$$ where $\omega_s\equiv (d+2-\eta)/2-p$. Because in the present case $\omega_s>0$ the range of such wall-fluid interactions is relevant in the RG sense at and near $T_c$, i.e., the profiles are affected by the decay of the boundary field even if $y$ is sufficiently large; in the critical region the profiles are no longer equivalent to those generated by a pure contact surface field ($p=\infty$). Choosing $b\sim y$ one obtains the scaling form of the magnetization $$\label{eq:mlr1} m_{\pm}(y,\tau,H,h_1,p)=y^{-\beta/\nu}{\cal M}_{\pm}(y/\xi_{\tau},y/\xi_{H},y^{\omega_s}h_1),$$ where $\xi_{\tau}\equiv \xi_{\pm}(\tau,0)$ is given by Eq. (\[eq:xitau\]) whereas $ \xi_{H}\equiv \xi(0,\Delta\mu)$ is given by Eq. (\[eq:ximu\]). For $\omega_s<0$ and large enough $y$ the scaling function ${\cal M}_{\pm}$ can be expanded in terms of their third argument demonstrating that in this case the long range of the wall-fluid interaction gives rise to contributions to the scaling functions of the magnetization profile which are subleading to the dominant universal $y^{-\beta/\nu}$ behavior valid for $y\lesssim \xi$ and $p=\infty$. Such an expansion cannot be performed for large $y$ if $\omega_s>0$ and in order to obtain a prediction in this case for the leading behavior of the magnetization profile and hence the thickness of the wetting layer $\ell_0$ and the adsorption, one has to calculate the full generalized scaling functions of the magnetization profiles. For $p<15/8$ such an analysis represents a necessary future research goal which, however, is beyond the scope of the present study. Nevertheless, we find it instructive and stimulating to show our numerical results for $p=1.5$ because they are strikingly different from those obtained for values of $p$ satisfying $p>(d+2-\eta)/2$. As along the noncritical isotherms $T^* \le 2.25$ we find that also along the isotherms $T^*=2.25$ and $T=T^*_c$ the increase of the wetting film thickness upon approaching $H=0$ is much more stronger for $p=1.5$ than for the other values of $p$. For $|H|\lesssim 0.01$ the local exponents attain $z(p=1.5)\simeq 0.64$, which indicates an algebraic increase of $\ell_0$ in this range of $H$ with $\ell_0\approx |H|^{-0.64}$. This effective exponent is close to the value $1/p=2/3=0.66(6)$ predicted for the complete wetting behavior for $p=1.5$. Closer to $H=0$ the functional form of the increase changes and $z$ increases strongly. For $p=50, 4$, and 3 the local exponents decrease upon approaching the pseudo-phase coexistence line $H_{ca}(T,L;p)$. This behavior is similar to the one occurring along the isotherm $T^*=1.8$; however, apart from the case $p=1.5$ the values of $z$ in Fig. \[fig:7\](b) are larger than the values in Fig. \[fig:6\](b), which indicates the stronger increase of the wetting film thickness. For $p=2$ the variation of the local exponents differs from that for the isotherm $T^*=1.8$: here $z$ is a decreasing function of $H$, almost identical to the one for $p=3$ (see Fig. \[fig:7\](b)). {width="7.5cm"} ![ (a) Wetting layer thickness $\ell_0$ in units of the lattice constant $a$ as a function of the bulk magnetic field $H$ calculated along the critical isotherm $T^*=T^*_c=2.269$ (path (4) in Fig. \[fig:1\]) for $L=500$, $h_1=0.8$, and for the ranges $p=1.5, 2, 3, 4$, and $50$ of the boundary fields. On the right the curves end at the corresponding capillary condensation point $H_{ca}(T,L;p)$. The inset shows a log-log plot of these data. The thin solid line indicates the slope 8/15. (b) The local exponents $z$ for $\ell_0(H)$ (see Eq. (\[zN\])). For $H\to 0$, without the occurrence of capillary condensation, i.e., for $L\to \infty$, the expected exponent is $\nu/\Delta=8/15$ $(\blacktriangle)$. The dotted lines indicate the expected extrapolation for $z(p\ge2,H\to 0$) with $L\to \infty$. For $p=1.5$ the behaviors of $\ell_0(H)$ and $z(H\to 0)$ are distinctly different from those for $p\ge 2$.[]{data-label="fig:8"}](rys09b.eps){width="7.5cm"} For all decay exponents $p\ge 2$ studied for $T^*=2.25$ the values of the local exponents $z(H\to 0)$ seem to attain values between 0.48 and 0.52. They would have to bend upwards for $H\to 0$ in order to reach the expected value of $8/15\simeq 0.533$. For the size $L=500$ capillary condensation does not allow to reach the ultimate complete wetting regime below the crossover line $H_{cr}^{(1)}$ (see Fig. \[fig:1\] for path (3)). For $p=3$ and 2, the variation of the local exponents $z$ along the critical isotherm $T=T_c$ is very similar to the one in Fig. \[fig:7\](b). In both cases, $z$ first decreases from the value $\approx 0.55$ above the expected one $8/15\simeq 0.53$ to the value $\approx 0.52$ at $|H|\approx 0.001$ and only then, for even smaller values of $H$, they seem to increase towards the expected value $ 8/15$. In the limit $H\to 0$, for $p=50$ and 4 the local exponents seem to approach the expected value 8/15 from above whereas for $p=3$ and 2 they will have to bend upwards do so from below. {width="7.5cm"} ![ Adsorption $\Gamma(H)$ (a) and the corresponding local exponents $z$ (b) along the isotherm $T^*=1.8$ (path (1) in Fig. \[fig:1\]) calculated within the DMRG approach for a $d=2$ Ising strip of width $L=500$ subject to symmetric boundary fields of strength $h_1=0.8$ and various ranges $p$ corresponding to power law decays $\sim j^{-p}$. For small $|H|$ all curves end at $H=H_{ca}(T,L;p)$ where capillary condensation sets in. Dotted lines indicate the expected extrapolations to $H=0$ for semi-infinite systems: $\beta_s^{co}=1/3$ $(\star)$ for $p\ge 3$, $\beta_s^{co}=1/2$ $(\small{\blacksquare})$ for $p=2$, and $\beta_s^{co}=2/3$ $(\bullet)$ for $p=1.5$.[]{data-label="fig:9"}](rys10b.eps){width="7.5cm"} ### Adsorption {#subsubsec:ads} For the present microscopic model in the strip geometry the adsorption $\Gamma $ is defined as $$\label{eq:adslatt} \Gamma =\sum_{j=1}^{L/2}(m_{j}-m_b(T,H)),$$ where $m_{j}\equiv \langle \sigma_{j}\rangle$ is the magnetization in row $j$ and $m_b(T,H)$ is the corresponding bulk magnetization. In the limit $L\to\infty$ this definition reduces to the semi-infinite quantity $\Gamma$ defined in Eq. (\[eq:1\]). Along the various paths indicated in Fig. (\[fig:1\]), at each point we have first calculated the bulk magnetization $m_b(T,H)$ using an equivalent system but with a vanishing surface field in order to minimize the influence of the boundary. {width="7.5cm"} ![The same as in Fig. \[fig:9\] along the isotherm $T^*=2.25$ (path (3) in Fig. \[fig:1\]). In (b) $(\nu-\beta)/\Delta=7/15$ is indicated $(\blacklozenge)$. The crossover to the ultimate complete wetting behavior is masked by the occurrence of capillary condensation. For $p=1.5$ the behaviors of $\Gamma(H)$ and $z(H\to 0)$ are distinctly different from those for $p\ge 2$.[]{data-label="fig:10"}](rys11b.eps){width="7.5cm"} Moreover, for the latter calculations as a representative spin the one in the middle of the strip was taken, as it is the one least affected by finite-size effects. In order to obtain the bulk magnetization $ m_b(T,H)$ we have extrapolated this midpoint value $m_{L/2}(T,H;h_1=0)$ to $L\to\infty$ from data calculated for $L=300, 400, 500, 600$, and $700$. The adsorption $\Gamma (H)$ and its local exponents calculated for a strip of width $L=500$ with $h_1=0.8$ and for various ranges $p$ of the boundary fields are shown in Figs. \[fig:9\], \[fig:10\], and \[fig:11\] for the paths (1), (3), and (4), respectively, as indicated in Fig. \[fig:1\]. {width="7.5cm"} ![ The same as in Fig. \[fig:10\] along the critical isotherm $T=T_c$ (path (4) in Fig. \[fig:1\]). In (b) $(\nu-\beta)/\Delta=7/15$ is indicated $(\blacklozenge)$. The crossover to the ultimate complete wetting behavior is masked by the occurrence of capillary condensation. For $p=1.5$ the behaviors of $\Gamma(H)$ and $z(H\to 0)$ are distinctly different from those for $p\ge 2$.[]{data-label="fig:11"}](rys12b.eps){width="7.5cm"} According to the predictions summarized in Sec. \[sec:2\], upon approaching the bulk coexistence line $H=0$ along path (1) $(T^*=1.8)$ the adsorption is expected to diverge for $p\ge 3$ with the exponent $\beta_s^{co}=1/3$. As can be seen in Fig. \[fig:9\](b) the local exponents calculated for $p=50, 4$, and 3 tend towards this predicted value for $H\to 0$ but do not reach it due to the occurrence of capillary condensation. For $p=2$ the local exponents as a function of the bulk field exhibit a negative curvature, consistent with the behavior of $z$ for the thickness $\ell_0$ of the wetting layer (see previous subsection), and seemingly tend to the exponent 1/2, which agrees with the predicted non-universal behavior for $p<3$, i.e., $\beta_s^{co}=1/p$. In order to reach the expected exponent $\beta^{co}_s=2/3$ for $p=1.5$, the local exponents have to vary in a non-monotonic way for $H\to 0$ (see Fig. \[fig:9\](b)). As expected from the analysis of the thickness of the wetting layer, the local exponents for the adsorption $\Gamma (H)$ along path (2) ($T^*=2.0$) behave in a similar way as along path (1) ($T^*=1.8$), i.e., for $p=50, 4$, and 3 they tend to the complete wetting exponent $\beta_s^{co}=1/3$, whereas for $p=2$ to the non-universal value $\beta_s^{co}=1/p=1/2$ (data not shown). As before, the asymptotic behavior is not yet completely reached when capillary condensation occurs. We note that the theoretical considerations in Sec. \[sec:2\] do not predict any specific power law behavior in terms of $H$ for the accessible $H$ values along this isotherm (2). Paths (3) and (4) lie in the critical adsorption regime and one expects (see Eq. (\[eq:I\_2\])) $\Gamma(H\to 0) \sim |H|^{-(\nu-\beta)/\Delta}$ as $H\to 0$ for $p\ge 15/8$. One can see from Figs. \[fig:10\](b) and  \[fig:11\](b) that for both $p=50$ and $p=4$ along both isotherms the local exponents seemingly tend to the predicted value $(\nu-\beta)/\Delta=7/15$ from above. In order to reach this value for $p=3$ and $2$, the local exponents have to bend upwards for $H\to 0$; whether for path (3) they ultimately tend to the complete wetting value $\beta_s^{co}=1/3$ and 1/2, respectively, after crossing both crossover lines $H_{cr}^{(1)}$ and $H_{cr}^{(2)}$ is unclear because for the finite values of $L$ considered here the asymptotic behavior is undercut by capillary condensation. (Note that for $\beta_s^{co}=1/2$ one has $H_{cr}^{(1)}=d_1|\tau|^{2\nu}$.) According to Fig. \[fig:1\] for $T^*=2.25$ this complete wetting regime is expected to be very narrow. For $p=1.5$ $(<15/8)$ the behavior of the critical adsorption is expected to be non-universal. Indeed, as compared to the other values of $p$ we observe a distinctly different behavior of the local exponents along both paths (3) and (4) with $z(H\to 0)$ strongly increasing. (Note that for $p=1.5$ one has $H_{cr}^{(1)}=d_1|\tau|^{3\nu/2}$.) For $\tau\ne 0$ the capillary condensation line $H_{ca}(\tau,L)$ is expected [@parry:92] to approach the bulk coexistence line $H=0$ as $1/L$ for $L\to \infty$ in accordance with the Kelvin equation (Eq. (\[eq:Kelvin\])). Near $T_c$, i.e., for $|\tau|\ll 1$ scaling arguments lead to the scaling behavior (see Eqs. (\[eq:6\]) and (\[eq:xitau\])): $$\label{eq:Kelvsc} H_{ca}(\tau,L)= |\tau|^{\Delta}{\tilde g}(L/\xi,h_1|\tau|^{-\Delta_1}),$$ where $\Delta_1$ is the surface gap exponent $(\Delta_1(d=2)=1/2)$ due to the presence of surface fields  [@diehl]; note that ${\tilde g}<0$. (In Eq. (\[eq:Kelvsc\]) the scaling of $h_1$ holds if $p$ is sufficiently large, i.e., $p> 15/8$.) Since $\xi$ depends on $\Delta\mu$ (see Eq. (\[eq:6\])), in Eq. (\[eq:Kelvsc\]) the first scaling variable of ${\tilde g}$ depends also on $H_{ca}(\tau,L)$, rendering an implicit equation for $H_{ca}(\tau,L)$. To leading order $H_{ca}(\tau,L)$ is given by Eq. (\[eq:Kelvsc\]) with $\xi$ from Eq. (\[eq:xitau\]). In the following $h_1$ is considered to be large enough so that ${\tilde g}(x,u\to\infty)=g(x)<0$ becomes independent of the second scaling variable. The analytic property $g(x\to\infty)\sim 1/x$ renders $H_{ca}(\tau,L\to \infty)\sim |\tau|^{\Delta-\nu}/L $ for $\tau\ne 0$ fixed and $L\to \infty$, in accordance with the Kelvin equation. On the other hand, for $L$ fixed and $\tau\to 0$ one has $g(x\to 0)\sim x^{-\Delta/\nu}$ (see below), which ensures that the limit $\tau \to 0$ renders a nontrivial function of $L$: $H_{ca}(\tau\to 0,L)\sim L^{-\Delta/\nu}$ with $\Delta/\nu=15/8$ in $d=2$. For weak surface fields $H_{ca}(\tau,L)\sim h_1|\tau|^{\Delta-\Delta_1}{\bar g}(L/\xi)$ with ${\bar g}(x\to 0)\sim x^{-(\Delta-\Delta_1)/\nu}$, ${\bar g}<0$, so that $H_{ca}(\tau\to 0,L)\sim L^{-(\Delta-\Delta_1)/\nu}$ [@BLM]. The capillary condensation line ends at a capillary critical point, at which the free energy of the confined system expressed in terms of the scaling variables $(L/\xi,H|\tau|^{-\Delta})$ is singular at a point $(x_0,y_0)$ implying Eq. (\[eq:shifts\]), which is consistent with the scaling behavior $H_{ca}(\tau,L\to\infty)\sim |\tau|^{\Delta-\nu}/L$. This scaling behavior of the capillary condensation line is sketched in Fig. \[fig:12\]. ![The log-log plot of the scaling function ${\hat g}(x=L/\xi)$ (see Eq. (\[eq:capsc\])) of the capillary condensation line $H_{ca}(\tau,L)$ calculated for several large values of $L$ and for a short-ranged ($p=50$) and a long-ranged ($p=3$) boundary field. In the limit $x\equiv L|\tau|^{\nu}\to 0$ the scaling function saturates at a constant nonzero value whereas ${\hat g}(x\to \infty) \sim x^w$ with $w\simeq 0.75$. The expected value $(\Delta-\nu)/\nu=7/8=0.875$ is not yet reached.[]{data-label="fig:13"}](rys14.eps){width="8.0cm"} With $\xi \sim |\tau|^{-\nu}$ (see Eq. (\[eq:6\])) the scaling form for the capillary condensation line can be written equivalently as $$\label{eq:capsc} H_{ca}(\tau,L)= L^{-\Delta/\nu}{\hat g}(L/\xi),$$ where ${\hat g}(x)=x^{\Delta/\nu}g(x)<0$. We test to which extent scaling holds for the present model with $p=50$ and $p=3$ by plotting $|H_{ca}|L^{\Delta/\nu}$ versus $L|\tau|^{\nu}$ for several strip widths $L$. Although we neglect the $H$ dependence of $\xi$ (see above and compare Eq. (\[eq:6\])) the data collapse is very good. From the log-log plot (see Fig. \[fig:13\]) one can see that ${\hat g}(x\to 0)= const$ which implies $g(x\to 0)\sim x^{-\Delta/\nu}$ (see above) and $H_{ca}(0,L)\sim L^{-\Delta/\nu}$. In the limit $x\gg 1$ our data fit very well to a power law ${\hat g}(x \to \infty )\sim x^{w}$ with the effective exponent $w\simeq 0.75$. $w$ is smaller than the value $(\Delta-\nu)/\nu=7/8=0.875$ which would lead to the predicted behavior $H_{ca}(\tau,L\to \infty)\sim |\tau|^{\Delta-\nu}/L $. One might be inclined to put the blame for the fact that $w$ has not yet reached this expected asymptotic value on corrections caused by the aforementioned $H$ dependence of the bulk correlation length, according to which Eq. (\[eq:capsc\]) reads $H_{ca}(\tau,L) L^{\Delta/\nu}={\hat g}(L|\tau|^{\nu}(\Xi(H_{ca}|\tau|^{-\Delta})^{-1})$ (see Eq. (\[eq:6\])); this additional dependence actually spoils the scaling in terms of the scaling variable $x=L|\tau|^{\nu}$. Since, however, scaling is observed, the latter dependence must be weak and the too small value of the exponent $w$ must be due to other corrections. \[subsec:alon\] {width="7.0cm"} ![Magnetization profiles near one wall, relative to their bulk values $m_b(T,H)$, along the thermodynamic path (0) (see Fig. \[fig:1\]) shifted by a constant $-10^{-6}$ away from the pseudo-phase coexistence line $H_{ca}(T,L)$, calculated within the DMRG method for $d=2$ Ising strips of width $L=700$ for short-ranged ((a) $p=50$) and of width $L=340$ for long-ranged ((b) $p=2$) boundary fields of strength $h_1=0.8$. Due to the slow decay, the profiles for $p=2$ are presented relative to their values $m(L/2)\ne m_b(T,H)$ at the midpoint of the film.[]{data-label="fig:14"}](rys14b.eps){width="7.0cm"} Due to the scaling behavior $H_{ca}(\tau,L)\sim |\tau|^{\Delta-\nu}/L$, for $L\gg 1$ but fixed, the pseudo-capillary condensation line appears to approach the bulk critical point (but without reaching it actually due to $L<\infty$) as $|\tau|^{\Delta-\nu}=|\tau|^{7/8}$ which is a weaker power law than $|H_{cr}^{(1)}|\sim |\tau|^{\nu/\beta_s^{co}}=|\tau|^3$ for the crossover line to complete wetting. This explains the observation that in $d=2$, upon approaching bulk coexistence sufficiently close to the critical point, the asymptotic complete wetting regime is always preempted by capillary condensation. More generally, we expect that in [*confined*]{} systems the asymptotic complete wetting divergence of the adsorption can be observed only if $\Delta-\nu=2-\alpha-\beta-\nu > \nu/\beta_s^{co}$. This condition is fulfilled neither in $d=2$ with $\Delta-\nu=0.875$ and $\nu/\beta_s^{co}=3$ nor in $d=3$ with $\Delta-\nu=0.936$ and $\nu/\beta_s^{co}=3\nu=1.89$. Thus we conclude that along an isotherm close to $T_c$ the ultimate crossover to the complete wetting behavior can only be observed in systems with macroscopically large transverse extensions. In the present confined system the analogue of the thermodynamic path parallel to the bulk coexistence line on the gas side with a small undersaturation $\Delta \mu=const>0$ and with $T\to T_c$ is the thermodynamic path $H_{(0)}(T,L)=H_{ca}(T,L)+\delta H$ along the pseudo-phase coexistence (capillary condensation) line $H_{ca}(T,L)$ shifted slightly by $\delta H$ to the spin down ($\sigma=-1$) side (see path (0) in Fig. \[fig:1\]). In view of the large computational effort required for the determination of the curve $H_{ca}(T,L)$ and of the value of the bulk magnetization $m_b(T,H)$ at each point along this line we have calculated the adsorption along this path only for a short-ranged boundary field ($p=50$), for the marginal case $p=3$ of long-ranged surface fields, and for $p=2$. We have considered three constant shifts away from the pseudo-phase coexistence line: $\delta H=-10^{-5}, -10^{-6}$, and $-10^{-7}$. ### Magnetization profiles {#subsubsec:mp1} In Fig. \[fig:14\] we have plotted a selection of magnetization profiles $m(z)$ along the thermodynamic path shifted by a constant $-10^{-6}$ away from the pseudo-coexistence line $H_{ca}(T,L)$ for $p=50$ (Fig. \[fig:14\](a)) and for $p=2$ (Fig. \[fig:14\](b)). For $p=3$ the profiles look qualitatively the same as for $p=2$; they are both characterized by a much broader interfacial region and larger tails than the ones for $p=50$. The shape of the profiles changes upon approaching bulk coexistence $H=0$ in a way similar to the case along the near-critical isotherm $T^*=2.25$. One observes a broadening of the interfacial region together with a rather small region where the magnetization stays close to the value $m_b(T,H)>0$ characteristic for being slightly on the upper side of the bulk coexistence curve, i.e., $H = 0^{+}$. Thermodynamic path along the pseudo – phase coexistence line ------------------------------------------------------------ ### Thickness of the wetting layer. {#subsubsec:wl1} For three undersaturations $\delta H$ relative to the pseudo-phase coexistence line Fig. \[fig:15\](a) shows the thickness of the wetting layers as a function of temperature calculated for a fixed strip width $L=340$ with strength $h_1=0.8$ and ranges $p=50, 3, 2$ of the boundary fields. The increase of the thickness of the wetting layer along these paths is very similar for all considered values of the decay exponent $p$. {width="7.0cm"} ![(a) Wetting layer thickness $\ell_0$ (defined as in Figs. \[fig:7\] and \[fig:8\]) as a function of temperature $T^*$ calculated for $L=340$, $h_1=0.8$, and various values of the exponent $p$, characterizing the range of the boundary fields, along the path (0) in Fig. \[fig:1\] with constant shift $\delta H= -10^{-6}$ away from the pseudo-phase coexisting line $H_{ca}(T,L;p)$ (see main text). For $p=50$, in addition the temperature dependence of $\ell_0$ for shifts $\delta H= -10^{-5}$ and $-10^{-7}$ is shown (dotted and thick solid line, respectively). Symbols denote predictions from the effective interface Hamiltonian (EIH) for $p=50$ (circle) and for $p=3$ (square) (see main text). (b) The local exponents $z$ for $\ell_0$ calculated from the data with constant shift $\delta H= -10^{-6}$ shown in (a) as a function of $\tau=(T-T_c)/T_c$. At the dotted and dashed vertical line the thermodynamic paths pass through the crossover line $H_{cr}^{(1)}$ and $H_{cr}^{(2)}$, respectively, shown in Fig. \[fig:1\] (and de facto independent of $\delta H$). Due to the finite peak of $\ell_0$ at $T_c$, below $H_{cr}^{(2)}$, i.e., for $|\tau|\lesssim 0.04$, the concept of the local exponents turns out to be less useful so that the lines in (b) have not been continued into this region.[]{data-label="fig:15"}](rys16b.eps){width="7.5cm"} At low temperatures, for $p=2$ the wetting layer is about three times thicker than for the short-ranged boundary field. Except for the close vicinity of the critical temperature, i.e., for $-0.98\lesssim T^*-T^*_c \lesssim 1.01$, where the finite-size effects are very pronounced, results obtained along paths with different shifts $\delta H$ but the same $p$ collapse onto a common curve. We note that for $p=50$ this common curve splits into three different lines just above the pseudocritical temperature $T^*_{c,L}(L=340,h_1=0.8) \approx 2.2$, i.e., when the pseudo-coexistence line $H_{ca}(T,L)$ intersects the crossover line $H_{cr}^{(2)}=-d_2|\tau|^{\Delta}$. For each $p$ the thickness of the wetting film attains a certain $L$-dependent value at the bulk critical temperature $T_c$, which also depends on the value $\delta H$ of the shift of the path (0). In order to calculate $\ell_0(T=T_c)$ for large $L$ we have performed an extrapolation scheme for the curves $\ell_0(T)$ to $T=T_c$, not taking into account data for temperatures higher than $T^*=2.24$ because those exhibit pronounced finite-size effects beyond the leading $L$–behavior. For $p=50$ we have found (data not shown) that the thickness of the wetting layer at $T=T_c$ is proportional to $L$, which is expected because $\ell_0$ cannot grow larger than $L/2$. For $p\ge 2$ it should diverge for $L\to \infty$. Above $T_c$ we observe a rapid decrease of $\ell_0$ as function of $T$. For increasing $T\lesssim T_c$ in Fig. \[fig:15\](a) we observe a continuous increase of the thickness of the wetting layer. We recall that, according to effective interface Hamiltonian studies, sufficiently far away from criticality, i.e., for $\ell_0 \gg \xi$, the wetting layer thickness should be independent of temperature along the path of [*constant*]{} undersaturation $\Delta \mu$, i.e., $ H=const$ with respect to bulk coexistence (see the last paragraph of Sec. \[sec:2\]). The present results are calculated along $H=H_{(0)}(L,T)=H_{ca}(T,L)+\delta H$ with $\delta H=const<0$ such that $|H|$ decreases for increasing temperatures, which should result in thicker films. This is in accordance with the data in Fig. \[fig:15\](a). In order to check whether this increase is captured by effective interfacial models we adopt a simple effective interface potential $W(\ell)$ valid for quasi short-ranged ($p=50$) boundary fields as well as for long-ranged boundary fields with a decay exponent $p=3$ or 2. Taking an undersaturation $H$ into account, we make the ansatz [@dietrich] (in units of $k_BT_c/a^2$) $$\label{eq:interfpot} W(\ell)= A\ell^{-(p-1)}+ B\ell^{-\kappa}+|H|\Delta m\ell$$ where an effective repulsive interaction $\sim B$, which accounts for the gain in entropy of the unbinding interface, has to be added to the leading long-ranged potential energy term $\sim A>0$ (see Subsec. \[subsec:cw\]) [@lipowsky]. One has $B=B_0 |k_BT/\Sigma(T)|>0$ [@lipowsky], where $\Sigma(T)$ is the surface stiffness of the interface, $B_0=const$ and $\Delta m=2|m_b^{(0)}(T)|$. The exponent $\kappa$ describing thermal wandering is given by $\kappa=2(d-1)/(3-d)$, which equals 2 in $d=2$ (see Subsec. \[subsec:cw\]) [@lipowsky]. In Eq. (\[eq:interfpot\]) the second (entropic) term dominates in $d=2$ for $p>3$. Thus for short-ranged forces $(p\gg 1$) the equilibrium wetting film thickness varies as function of $T$ along a thermodynamic path $H(T)$ as $$\label{eq:lsrf} \ell_0=(B_0k_BT/(\Sigma(T)|m_b^{(0)}(T)|))^{1/3}|H(T)|^{-1/3}, \quad p>3.$$ For $p=3$ the entropic term competes with the interaction term. This leads to the equilibrium thickness of the wetting film $$\begin{aligned} \label{eq:llr} \ell_0 = \;\;\;\;\;\;\;\;\;\;\:\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;& \nonumber \\ ((A+B_0k_BT/\Sigma(T))/|m_b^{(0)}(T)|)^{1/3}|H(T)|^{-1/3}, & \quad p=3. \nonumber \\\end{aligned}$$ $A$ is proportional to $ \Delta m=2|m_b^{(0)}(T)|$ so that $A/|m_b^{(0)}(T)|$ only weakly depends on $T$. For $p=2$ the first term in Eq. (\[eq:interfpot\]) dominates and the temperature dependence of the film thickness follows that of $H(T)\simeq H_{ca}(T,L)$: $$\label{eq:ellscl} \ell_0=(A/(2|m_b^{(0)}(T)|))^{1/2}|H_{ca}(T,L)|^{-1/2}, \quad p=2.$$ For the $d=2$ Ising model the surface stiffness is known exactly as $\Sigma(T)/(k_BT)=\sinh 2(K-K^*)$ where $K=J/(k_BT)$ and $\tanh K^*=\exp(-2K)$. Both $\Sigma(T)$ and $\Delta m(T)\equiv 2|m_b^{(0)}(T)| $ decrease with increasing temperature. Therefore, for all values of $p$, $\ell_0$ is expected to increase along the pseudo-coexistence line $H_{ca}(T,L)$. Sufficiently close to the bulk critical point the surface stiffness vanishes with the same power law as the surface tension of the interface $\sigma$, i.e., $\Sigma(\tau\to 0)\sim \sigma\sim |\tau|^{(d-1)\nu}=|\tau|$ for the $d=2$ Ising model. Since $\Delta m \sim |\tau|^{\beta}$ it follows that in the critical region for short-ranged forces $$\begin{aligned} \label{eq:ellsc} \ell_0\sim |\tau|^{-(\beta+(d-1)\nu)/3}|H_{ca}(\tau,L)|^{-1/3}=& \nonumber \\ |\tau|^{-3/8}|H_{ca}(\tau,L)|^{-1/3}, & \quad p=3. \nonumber \\\end{aligned}$$ with $H_{ca}(\tau,L)$ given by Eq. (\[eq:capsc\]). Close to $T_c$ and in the temperature range for which the pseudo-coexistence line $H_{ca}(\tau,L)$ lies in between the crossover lines $H_{cr}^{(1)}=-d_1|\tau|^{3\nu}$ and $H_{cr}^{(2)}=-d_2|\tau|^{\Delta}$ we have found (see Fig. \[fig:13\]), for both $p=50$ and $p=3$, that the absolute value of the scaling function ${\hat g}(x=L/\xi)$ of the pseudo-coexistence line (see Eq. (\[eq:capsc\])) decreases and crosses over from the power law behavior ${\hat g}(x\to \infty)\sim x^w$ with $w\simeq 0.75$ for $|\tau|\ne 0$ to a constant value ${\hat g}(x=0)$ at $\tau=0$. This implies that, for decreasing $|\tau|$, $\ell_0(\tau)$ crosses over from an increase $\sim |\tau|^{-5/8}$ to one $\sim |\tau|^{-3/8}$ (for $L$ fixed). Thus the effective interface model predicts that, along the path (0), $\ell_0(\tau\to 0)$ is an increasing function of temperature. But in view of the aforementioned crossover one cannot expect a purely algebraic behavior in this region. For $12 \lesssim L|\tau|^{\nu}\lesssim 50$, the variation of the pseudo-coexistence line $H_{ca}(T,L)$ at fixed $L$ is approximately algebraic with an effective exponent $w\simeq 0.75$. Inserting $H=H_{ca}(T,L)\sim |\tau|^{0.75}$ into the expression for $\ell_0$ (Eq. (\[eq:ellsc\])) gives $\ell_0 \sim |\tau|^{-0.625}$ for both $p=50$ and $p=3$. For $L|\tau|^{\nu}\ll 1$ one has $H_{ca}\to const$, so that in this limit Eq. (\[eq:ellsc\]) predicts $\ell_0(\tau\to 0)\sim |\tau|^{-3/8}$. For $p=2$ the pseudo-coexistence line $H_{ca}(T,L)$ seems to follow a power law with the same effective exponent as for $p=50$ and 3 for reduced temperatures smaller than the crossover line $H_{cr}(1)$ (see Fig. \[fig:1\]). Therefore on the basis of Eq. (\[eq:ellscl\]) for $12\lesssim L|\tau|^{\nu}\lesssim 50$ one expects $\ell_0 \sim |\tau|^{-0.375}$. For $p=50$ the behavior of the local exponents $z$ determined from the numerical data obtained for $L=340$ and shown in Fig. \[fig:15\](a) is consistent with the above predictions (see Fig. \[fig:15\](b)). We find that in between the crossover lines $H_{cr}^{(1)}=-d_1|\tau|^{3\nu}$ and $H_{cr}^{(2)}=-d_2|\tau|^{\Delta}$ the local exponents $z$ decrease continuously upon decreasing $|\tau|$. For $0.12 \lesssim L|\tau|^{\nu}\lesssim 50$, which with $L=340$ and $\nu=1$ corresponds to $0.04 \lesssim |\tau|\lesssim 0.15$, the value of $z$ for $p=50$ changes from $\simeq 0.80$ to $\simeq 0.52$. No saturation at the value 0.625, which would reflect the power law ${\hat g}(x)\sim x^{0.75}$ (see Eq. (\[eq:capsc\]) and the discussion above), is observed. However, the above temperature range lies outside the range of validity of the asymptotic power law behaviors of $\Sigma(T)$ and $\Delta m(T)$ leading to deviations from the algebraic variation $\ell_0\sim |\tau|^{-0.625}$ discussed above. In order to test the validity of the effective interface Hamiltonian approach we compare our full DMRG data with the predictions for the increase of the wetting film thickness along the line of pseudo-phase coexistence $H_{ca}(T,L)$ stemming from Eq. (\[eq:lsrf\]) by substituting our numerical data for $H(T)=H_{ca}(T,L)$ and $\Delta m(T)=2|m_b(T)|$ and by using the analytic expression for $\Sigma/(k_BT)$ given above. Treating $B_0$ as a fitting parameter we find very good agreement between these two approaches for $B_0=0.729$ (see Fig. \[fig:15\](a)) in the wide temperature range below $T^*\simeq 2.215$. We note, that for higher temperatures the system crosses over to the critical adsorption regime (beyond the crossover line $H_{cr}^{(2)}$). For $p=3$ and 2 we observe that at $H_{cr}^{(1)}$ the local exponents $z$ saturate at the values $\simeq 0.55$ and $\simeq 0.44$, respectively. In the case $p=3$ the exponent is slightly smaller than the prediction 0.625 based on the scaling behavior obtained within the effective interface model, whereas for $p=2$ it is slightly larger than the value 0.375 expected also within the effective interface model. We repeat the procedure described above to check the agreement of the effective interface Hamiltonian prediction in Eq. (\[eq:llr\]) with our DMRG data for $p=3$. We adopt $B_0=0.729$, i.e., the value determined from the fitting of Eq. (\[eq:lsrf\]) to the DMRG results for $p=50$, and treat $A/|m_b(T)|\approx const\equiv C$ as a free parameter. Very good agreement is obtained for $C\approx 1.5$ and temperatures $T^*\lesssim 2.15$ (see Fig. \[fig:15\](a)). However, the assumption $A/|m_b(T)|\approx const$ does not lead to an agreement of the DMRG data for $p=2$ (not shown) with the curve predicted by Eq. (\[eq:ellscl\]) with $H(T)=H_{ca}(T,L)$. Thus we conclude that in the case $p=2$ (not shown) the simple effective interface Hamiltonian given by Eq. (\[eq:interfpot\]) fails to describe the full DMRG data. ### Adsorption {#subsubsec:ads1} Results for the adsorption obtained for $L=340$, $h_1=0.8$, and $p=50, 3, 2$ are shown as a function of $T$ in Fig. \[fig:16\](a). {width="7.0cm"} ![ (a) Adsorption $\Gamma$ as a function of $T^*$ along the path (0) in Fig. \[fig:1\] with a constant shift $\delta H= -10^{-6}$ away from the pseudo-phase coexisting line $H_{ca}(T,L;p)$ (see main text), calculated within the DMRG method for $d=2$ Ising strips of width $L=340$ for short-ranged ($p=50$) and long-ranged ($p=3,2$) boundary fields of strength $h_1=0.8$. For $p=50$ the temperature dependence of $\Gamma$ for shifts $\delta H= -10^{-5}$ and $-10^{-7}$ is also shown (dotted and thick solid line, respectively). (b) The local exponents $z$ for $\Gamma$ calculated from the data shown in (a) for the constant shift $\delta H= -10^{-6}$. At the dotted and dashed vertical line the thermodynamic path passes through the crossover lines $H_{cr}^{(1)}$ and $H_{cr}^{(2)}$, respectively, shown in Fig. \[fig:1\] (and de facto independent of $\delta H$). As in Fig. \[fig:15\](b), in (b) the lines do not continue into the region $|\tau|\lesssim 0.04$.[]{data-label="fig:16"}](rys17b.eps){width="7.5cm"} In accordance with the expectations, the behavior of $\Gamma (T)$ upon increasing the temperature is qualitatively similar to the behavior of $\ell_0(T)$ in Fig. \[fig:15\](a). Results for different shifts $\delta H$ form a common curve above $T_c$ and for $T^*\lesssim 2.2$. Close to $T_c$ the adsorption exhibits a maximum which signals the vicinity of the end of the pseudo-phase coexistence line, i.e., the pseudo-critical point of a finite system. The position of this maximum shifts towards $T_c$ for wider strips. At $T_c$, $\Gamma$ attains a finite value $\Gamma (\tau=0,L)$. As expected, for long-ranged boundary fields the adsorption is slightly stronger than for short-ranged forces. According to the discussion in the last paragraph of Sec. \[sec:2\], for $p\ge 3$ and in between the two crossover lines $H_{cr}^{(1)}=-d_1|\tau|^{3\nu}$ and $H_{cr}^{(2)}=-d_2|\tau|^{\Delta}$ (see Fig. \[fig:1\]) $\Gamma$ is expected to increase $\sim |\tau|^{-(\nu-\beta)}$, which for $d=2$ corresponds to an exponent -0.875, provided the variation of $\xi$ is dominated by its dependence on $\tau$ (Eq. (\[eq:xitau\])). This is the case along paths of constant undersaturation $\Delta \mu \equiv H=const$. Along the pseudo-coexistence line and for $p=50$ the local exponents of $\Gamma$ vary strongly with temperature, also below the crossover line $H_{cr}^{(1)}$ (see Fig. \[fig:16\](b)). This is not compatible with the aforementioned algebraic variation of $\Gamma$ for $H=const$. In view of the discussion of the corresponding behavior of the wetting film thickness in the previous subsection, this difference can be traced back to the fact that the pseudo-coexistence line does not run in parallel to the bulk coexistence line. The decrease of the absolute value of the bulk field $H_{ca}(T,L)$ upon increasing temperature matters and in the scaling analysis of $\Gamma$ the whole scaling form of $\xi$ (Eq. (\[eq:6\])) has to be taken into account. According to Eqs. (\[eq:5\]) and (\[eq:6\]) this is encoded in the product of the scaling functions $\Xi_{\pm}$ and $K_{\pm}$ as function of $H|\tau|^{-\Delta}$, which alters the temperature dependence $\sim |\tau|^{-(\nu-\beta)}$. Due to $\Gamma \sim \ell_0\Delta m$ and the fact that $\Delta m$ decreases upon decreasing the reduced temperature $|\tau|$, one expects that the increase of $\Gamma $ is weaker than the increase of $\ell_0$. This is consistent with our numerical data for the local exponents $z$. At each point along the path (0) the value of $z$ obtained for $\Gamma$ is smaller than the corresponding value of $z$ obtained for $\ell_0$ (see Figs. \[fig:15\](b) and  \[fig:16\](b)). For $p=3$ and $p=2$ the variation of the local exponents of the adsorption in the temperature range in between the crossover lines $H_{cr}^{(1)}$ and $H_{cr}^{(2)}$ is less pronounced than that for $p=50$. Near $H_{cr}^{(1)}$ we find saturation for the local exponents $z$ at values which are smaller than the saturation values of $z$ of the corresponding thickness $\ell_0$ of the wetting film, i.e., $z\simeq 0.49$ and $z\simeq 0.36$ for $p=3$ and $p=2$, respectively (see Fig. \[fig:16\](b)). This is consistent with the relation $\Gamma \sim \ell_0\Delta m$ and can be explained by the same arguments as used in the previous subsection. We note, that $\Delta m\sim |\tau|^{\beta}$, with $\beta=1/8=0.125$ for the $d=2$ Ising model, leads to bigger differences in the local exponents for $\ell_0$ and $\Gamma$ than the aforementioned ones observed in the range of their saturation \[$p=3$: $z(\ell_0)-z(\Gamma)=0.55-0.49=0.06<0.125$; $p=2$: $z(\ell_0)-z(\Gamma)=0.44-0.36=0.08<0.125$\]. However, the above power laws are satisfied only sufficiently close to $T_c$, which is not reached along $H_{ca}(T,L)$ for the finite values of $L$ studied here. Summary and conclusions {#sec:5} ======================= We have studied two-dimensional Ising ferromagnets in strip geometries of width $L$ and with long-ranged boundary fields ($V^s_{j}=\frac{h_1}{j^p}, h_1>0$) (Eqs. (\[eq:3\_1\]) and (\[eq:3\_2\])). Based on scaling theory and the density-matrix renormalization-group method we have obtained the following main results: \(1) In Sects. \[sec:1\] and \[sec:2\] we have discussed the theoretical framework and the expectations due to scaling theory for the interplay between critical adsorption and complete wetting, including the crossover between them. For semi-infinite systems the corresponding scaling arguments lead to the schematic phase diagram shown in Fig. \[fig:0\]. \(2) The location of the pseudo-phase coexistence line $H_{ca}(T,L;p)$ of capillary condensation has been determined for various ranges $p$ of the boundary fields (Fig. \[fig:1\]). For positive and parallel surface fields capillary condensation occurs at negative values $H_{ca}(T,L;p)$ of the bulk field $H$. Increasing the range of the boundary fields shifts the pseudo-phase coexistence line towards more negative values of the bulk field $H$ (see Figs. \[fig:1\] and \[fig:2\]). Relative to the pseudo-capillary condensation line in the case that the surface fields act only on the two boundary layers $j=1$ and $j=L$, i.e., for $p=\infty$, $H_{ca}(T,L;p)$ varies exponentially as function of the range $p$ (Fig. \[fig:2\]). \(3) A critical wetting transition has been identified for $p=2$, i.e., for a boundary field decaying slower than the marginal one ($p=3$) (see Fig. \[fig:3\]). These findings disprove previous claims by Kroll and Lipowsky [@kroll] that in $d=2$ for $p<3$ there is no critical wetting transition at finite values of $h_1$. \(4) Within the accessible range of values for $T$ and $H$ the magnetization profiles both along isotherms (see Figs. \[fig:4\] and \[fig:5\]) and along the pseudo-capillary condensation line (see Fig. \[fig:1\]) are not slab-like. They exhibit a wide interfacial region and significant tails. Primarily the profiles approach their plateau values in the middle of the strip exponentially as function of the distance from the wall. The width of the emerging interfacial region and the decay length of the tails are proportional to the bulk correlation length and thus grow upon approaching $T_c$ or the pseudo-capillary condensation line. In the presence of long-ranged boundary fields the exponential decay of the profiles towards their bulk values $m_b(H,T)$ is followed by an algebraic decay $j^{-p}$ which finally is distorted by the presence of the distant wall. Features similar to those of these order parameter profiles have been inferred from neutron reflectrometry for the composition profile of a wetting film in a binary liquid mixture of n-hexane and perfluoro-n-hexane [@bowers:07]. \(5) The variation of the thickness $\ell_0$ of the wetting layer along various isotherms (Figs. \[fig:6\]-\[fig:8\]) and along the pseudo-phase coexistence line (Fig. \[fig:15\]) has been analyzed for wide strips and different ranges $p$ of the boundary field. Along both types of path we have found a gradual increase of $\ell_0$ upon approaching $T_c$. The asymptotic divergence of $\ell_0$ along the isotherms is preempted by capillary condensation. As discussed on general grounds at the end of Subsec. \[subsubsec:ads\], both in $d=2$ and $d=3$ along the isotherm close to $T_c$ the ultimate crossover to the complete wetting behavior (see hatched region and path II in Fig. \[fig:0\]) can only be observed in systems with macroscopically large transverse extensions. Along the isotherms and within the complete wetting regime the suitably defined local exponents of $\ell_0$ (Eq. (\[zN\])) tend to approach the values predicted by the corresponding effective interface Hamiltonian, i.e., $\beta_s^{co}=1/3$ for $p\ge 3$ and $\beta_s^{co}=1/p$ for $p<3$ (Figs. \[fig:6\] and \[fig:7\]). These results further support our findings that a critical wetting transition exists also for the values $p=2$ and $1.5$ of the decay exponent of the boundary field, i.e., smaller than the marginal case $p=3$. Along the pseudo-phase coexistence line $H_{ca}(T,L)$ the variation of the thickness of the wetting film is determined by the functional form of the temperature dependence of $H_{ca}(T,L)$ and for $p>2$ it agrees with the predictions of an effective interface Hamiltonian (Eq. (\[eq:lsrf\]) and Fig. \[fig:15\](a)). For long-ranged boundary fields with $p=3$ and $p=2$, the increase of $\ell_0$ within a certain range of temperature can be described by a power law with an effective exponent which is, however, not universal. It varies from $ca$ $ 0.55$ for $p=3$ to $ca$ $0.44$ for $p=2$ (Fig. \[fig:15\](b)). \(6) For $p=50, 4, 3$, and $2$ the adsorption $\Gamma$ (Eqs. (\[eq:1\]) and (\[eq:adslatt\])) has been calculated along various isotherms (see Figs. \[fig:9\]-\[fig:11\]). Along these thermodynamic paths, over a wide range of the bulk field $H$ the adsorption $\Gamma$ exhibits a continuous increase upon approaching the pseudo-phase coexistence line $H_{ca}(T,L;p)$. Within the accessible range of values for $H$ this increase cannot be described by simple power laws. Only very close to $H_{ca}(T,L;p)$ the local exponents of $\Gamma$ start to approach their predicted values, i.e., $\beta_s^{co}(p\ge 3)=1/3$ and $\beta_s^{co}(p=2)=1/2$ for complete wetting (see Subsec. \[subsec:cw\] and Figs. \[fig:9\] and \[fig:10\])) and $(\nu-\beta)/\Delta=7/15$ for critical adsorption (see Eq. (\[eq:I\_2\]), Subsec. \[subsec:ca\], and Fig. \[fig:11\]). However, these results are not entirely conclusive because the asymptotic regimes are preempted by capillary condensation so that one is unable to detect the crossover from critical adsorption to complete wetting which is expected to occur sufficiently close to bulk coexistence for isotherms with $T$ close to $T_c$. For the low temperature isotherms, which lie entirely within the complete wetting regime, the cases $p=1.5, 2$, i.e., for $p$ smaller than the marginal value $p=3$, differ distinctly from the cases $p\ge 3$ (see Subsec. \[subsec:cw\] and Fig. \[fig:9\]), which is a clear indication of the non-universality of complete wetting for these cases. For $p=1.5$ the long-ranged part of the boundary field is relevant in the RG sense and the behavior of $\Gamma$ is non-universal along the critical isotherm (see Fig. \[fig:11\](b)). \(7) The finite-size scaling predictions for the capillary condensation line (see Eqs. (\[eq:Kelvsc\]) and (\[eq:capsc\]) as well as Figs. \[fig:12\] and \[fig:13\]) are satisfied for both short-ranged ($p=50$) and long-ranged ($p=3$) boundary fields. \(8) For $p=50$, 3, and 2 the magnetization profiles (Fig. \[fig:14\]) and the adsorption $\Gamma$ (Fig. \[fig:16\]) have been calculated along the pseudo-capillary condensation line $H_{ca}(T,L)$. Over a wide range of temperatures below $T_c$, $\Gamma$ increases gradually (see Fig. \[fig:16\]). Near criticality its increase cannot, however, be described by a simple power law behavior. The reason is that along this thermodynamic path the temperature dependence of $\Gamma$ is determined also by the temperature variation of $H_{ca}(T,L)$, similarly to the thickness $\ell_0$ of the wetting film which is approximately related to $\Gamma$ as $\Gamma \simeq \ell_0 \Delta m$ where $\Delta m(T)$ is the temperature dependent difference between the magnetization of the spin up and the spin down bulk phases. Our theoretical results cannot be directly compared with the experimental findings in Refs. [@fenistein:02] and [@bowers:04], in which the thickness of the wetting layer [@fenistein:02] and the adsorption [@bowers:04] were measured along the paths corresponding to the path (1) in Fig. \[fig:0\], i.e., in these experiments the capillary condensation was absent. The data obtained from both experiments show that the behavior of $\ell_0$ and $\Gamma$ can be described by a power law but with an effective exponent which differs from both that for complete wetting and that for critical adsorption. Similarly, in our theoretical analysis of two-dimensional Ising strips, for long-ranged boundary fields with $p=3$ and $p=2$ the increase of $\Gamma$, for still sufficiently large reduced temperatures $|\tau|$, can be described by a power law with an effective exponent, here with a value of $ca.$ $0.44$ for $p=3$ and of $ca.$ $ 0.36$ for $p=2$ (Fig. \[fig:16\]). \(9) Our results indicate that the asymptotic behavior for $H\to 0$ occurs in a much narrower regime than one would expect and therefore requires much larger system sizes to suppress capillary condensation sufficiently in order to reach the asymptotic regime. One indication for having entered the asymptotic regime could be that a fully developed wetting layer has been formed. For the system sizes $L$ and thus the values of $H$ accessible in our calculations this does not yet fully occur. A. D. thanks the Wrocław Centre for Networking and Computing (grant No. 82) and the Computing Center of the Institute of Low Temperature and Structure Research PAS for access to their computing facilities. [99]{} K. Binder, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic, London, 1983), Vol. 8, p. 1. H. W. Diehl, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic, London, 1986), Vol. 10, p. 76. S. Dietrich, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic, London, 1988), Vol. 12, p. 1. Y. Xia, B. Gates, Y. Yin, and Y. Lu, Adv. Mater. [**12**]{}, 693 (2000). E. Bertrand and J. Petrol. Sci. Eng. [**33**]{}, 217 (2002). P. G. de Gennes, [*Rev. Mod. Phys.*]{} [**57**]{}, 827 (1985). P. Tabeling, [*Microfluidics*]{} (EDP Sciences, Paris, 2004). , edited by E. Kiran, P. G. Debenedetti and C. J. Peters, Proceedings of the NATO Advanced Study Institute, NATO Science Series E, Vol. 366, (Kluwer, Dordrecht, 2000). R. Riehn and R. H. Austin, Anal. Chem. [**78**]{}, 5933 (2006). P. Pfeuty and G. Toulouse, [*Introduction to the Renormalization Group and to Critical Phenomena*]{} (Wiley, London, 1977). A. Pelissetto and E. Vicari, Phys. Rep. [**368**]{}, 549 (2002). P. Tarazona, M. M. Telo da Gama and M. Robert, J. Chem. Phys. [**86**]{}, 1521 (1987). S. Dietrich, in [*Phase Transitions in Surface Films 2*]{}, edited by H. Taub, G. Torzo, H. J. Lauter, and S. C. Jr. Fain, Proceedings of the NATO Science Series B (Plenum, New York, 1991), Vol. 267, p. 391, and references therein. M. M. Telo da Gama and U. Marini Bettolo Marconi, Physica A [**171**]{}, 69 (1991). D. Fenistein, D. Bonn, S. Rafa[ï]{}, G. H. Wegdam, J. Meunier, A. O. Parry, and M. M. Telo da Gama, Phys. Rev. Lett. [**89**]{}, 096101 (2002). J. Bowers, A. Zarbakhsh, A. Querol, H. K. Christenson, I. A. McLure, and R. Cubitt, J. Chem. Phys. [**121**]{}, 9058 (2004). J. R. Silvius, [*Thermotropic phase transitions of pure lipids in model membranes and their modifications by membrane proteins*]{} (John Wiley, New York, 1982). A. R. Honerkamp-Smith, P. Cicuta, M. D. Collins, S. L. Veatch, M. den Nijs, M. Schick, and S. L. Keller, Biophys. J. [**95**]{}, 236 (2008). S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992); S. R. White, Phys. Rev. B [**48**]{}, 10345 (1993). T. Nishino, J. Phys. Soc. Jpn. [**64**]{}, 3598 (1995). , edited by I. Peschel, X. Wang, M. Kaulke, and K. Hallberg (Springer, Berlin, 1999), Vol. 528. U. Schollwoeck, Rev. Mod. Phys. [**77**]{}, 259 (2005). D. Beysens and D. Estève, Phys. Rev. Lett. [**54**]{}, 2123 (1985); B. M. Law, J.-M. Petit, and D. Beysens, Phys. Rev. E [**57**]{}, 5782 (1998). H. Guo, T. Narayanan, M. Sztuchi, P. Schall, and G. H. Wegdam, Phys. Rev. Lett. [**100**]{}, 188303 (2008). M. Schick, in [*Liquids at Interfaces*]{}, Proceedings of the Les Houches Summer School in Theoretical Physics, Session XLVIII, edited by J. Charvolin, J. F. Joanny, and J. Zinn-Justin (North-Holland, Amsterdam, 1990), p. 415. R. Lipowsky, Phys. Rev. B [**32**]{}, 1731 (1985). M. E. Fisher and P. G. de Gennes, C. R. Acad. Sci. Ser. B [**287**]{}, 207 (1978). G. Fl[" o]{}ter and S. Dietrich, Z. Phys. B [**97**]{}, 213 (1995). R. Pandit, M. Schick, and M. Wortis, Phys. Rev. B [**26**]{}, 5112 (1982). T. Getta and S. Dietrich, Phys. Rev. E [**47**]{}, 1856 (1993). V. Privman and M. E. Fisher, J. Stat. Phys. [**33**]{}, 385 (1983), and references therein. K. Binder, Annu. Rev. Mater. Res. [**38**]{}, 123 (2008). W. Thomson, Phil. Mag. [**42**]{}, 448 (1871). R. Evans, J. Phys.: Condens. Matter [**2**]{}, 8989 (1990), and references therein. M. E. Fisher and H. Nakanishi, J. Chem. Phys. [**75**]{}, 5857 (1981); H. Nakanishi and M. E. Fisher, [*ibid*]{} [**78**]{}, 3279 (1983). A. Maciołek, A. Drzewiński, and R. Evans, Phys. Rev. E [**64**]{}, 056137 (2001). D. M. Kroll and R. Lipowsky, Phys. Rev. B [**28**]{}, 5273 (1983). V. Privman and N. M [Ŝ]{}vraki[ĉ]{}, Phys. Rev. B [**37**]{}, 5974 (1988). A. Drzewiński and K. Szota, Phys. Rev. E [**71**]{}, 056110 (2005). E. V. Albano, K. Binder, and W. Paul, J. Phys.: Condens. Matter [**12**]{}, 2701 (2000). A. De Virgiliis, E. V. Albano, M. Müller, and K. Binder, J. Phys.: Condens. Matter [**17**]{}, 4579 (2005). A. O. Parry and R. Evans, Phys. Rev. Lett. [**64**]{}, 439 (1990). D. B. Abraham, Phys. Rev. Lett. [**44**]{}, 1165 (1980). M. Napiórkowski, W. Koch, and S. Dietrich, Phys. Rev. A [**45**]{}, 5760 (1992). A. O. Parry and R. Evans, Phys. Rev. A [**46**]{}, 5282 (1992). K. Binder, D. Landau, and M. M[ü]{}ller, J. Stat. Phys. [**110**]{}, 1411 (2003). J. Bowers, A. Zarbakhsh, I. A. McLure, J. R. P. Webster, R. Steitz, and H. K. Christenson, J. Phys. Chem. C [**111**]{}, 5569 (2007).

ArXiv

--- author: - 'Eric Dow[^1]' - 'Qiqi Wang[^2]' bibliography: - 'main\_sisc.bib' title: OPTIMIZATION OF GAUSSIAN RANDOM FIELDS --- random fields, pathwise sensitivity method, optimization 49N45, 60G15, 60G60 [^1]: Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, MA, 02139. (). [^2]: Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, MA, 02139. ().

ArXiv

--- abstract: | Scalar perturbations can grow during a phantomic cosmological phase as the big rip is approached, in spite of the high accelerated expansion regime, if the equation of state is such that $\frac{p}{\rho} = \alpha < - \frac{5}{3}$. It is shown that such result is independent of the spatial curvature. The perturbed equations are exactly solved for any value of the curvature parameter $k$ and of the equation of state parameter $\alpha$. Growing modes are found asymptotically under the condition $\alpha < - \frac{5}{3}$. Since the Hubble radius decreases in a phantom universe, such result indicates that a phantom scenario may not survive longtime due to gravitational instability. --- [**INSTABILITY OF SCALAR\ PERTURBATIONS IN A PHANTOMIC\ COSMOLOGICAL SCENARIO\ **]{} J.C. Fabris[^1]\ Institut d’Astrophysique de Paris\ 98bis, Boulevard Arago, 75014 Paris, France\ D.F. Jardim [^2] and S.V.B. Gonçalves[^3]\ Departamento de Física, Universidade Federal do Espírito Santo, 29060-900, Vitória, Espírito Santo, Brazil PACS: 98.80.-k, 95.36.+x In the ordinary theory of cosmological perturbation [@weinberg; @padmanabhan; @brand] there are two fundamental regimes defined from the notion of [*Jean’s length*]{}, $\lambda_J$. Perturbations whose scales are such that $\lambda > \lambda_J$ suffer gravitational instability since the gravitational attraction dominates over the pressure reaction to contraction; on the other hand, perturbations whose scales satisfy the condition $\lambda < \lambda_J$, tends to oscillates due to the effectiveness of the pressure opposition to the gravitational collapse. In the relativistic version of the theory of cosmological perturbation we must add a new relevant scale, given by the Hubble radius. The Hubble radius define, in some sense, the effective causal region, and consequently the region where the effects of the microphysics phenomenon (responsible for the pressure) may play a relevant rôle in the process of gravitational collapse. Perturbations whose scales are much large than the Hubble radius tends to become frozen, while for those smaller than the Hubble radius, the pressure tends to produce a damping effect. Moreover, if the spatial section of the metric representing the universe is not flat a new scale appears which is connected to the curvature parameter. In cosmology, concomitant with the interplay between the gravitational attraction and the pressure resistance, there is a supplementary relevant effect due to the expansion of the universe, which acts as a friction term in the fundamental equations of the perturbed quantities: the expansion leads to a damping in the evolution of perturbations. In general, for very large perturbations, that are connected with the large structures existing in the universe, there are two modes: a growing or a constant mode and a decreasing mode. However, if the pressure is negative, and the expansion becomes accelerated, gravity may become repulsive and we generally find only decreasing modes: the perturbations are always damped and structures can not be formed. There are quite strong evidences that the universe today is in an accelerated expansion phase [@spergel]. If this is true, the material content of the universe must be dominated by an exotic fluid whose pressure is negative. There are many candidates for this exotic fluid, like cosmological constant, quintessence, K-essence, Chaplygin gas, etc., each of them having its advantages and disadvantages from the theoretical point of view. We address the reader to recent reviews on the many dark energy models existing in the literature [@sahni; @durrer]. Such negative pressure fluid may generate repulsive effects driving the accelerated expansion. There are some claims that the observational data favors a phantom fluid, that is, a fluid whose negative pressure violates the dominant energy condition $p + \rho \geq 0$ [@caldwell]. Representing such fluid by a self-interacting scalar field, a phantom fluid requires a “wrong” sign in the kinetic term. This may imply instability at quantum level. However, interesting propositions have been made in what concerns such quite exotic fluid, mainly in the context of ghost condensation existing in some string configurations [@piazza]. Classically, one of the most striking feature of phantom fluids is the fact that its energy density grows with the expansion. This is consequence of the violation of the strong energy condition, and it may lead to a future singularity in a finite future time, the so-called [*big rip*]{}. This is, of course, a very undesirable feature. But it has been shown [@sergio1] that in a single fluid approximation, for a spatially flat universe, the scalar perturbations can grow when the scales of the perturbation are greater than the Hubble radius. Since, the Hubble radius decreases, for a phantom dominated universe, the isotropy and homogenous condition would not be satisfied anymore as the big rip is approached, leading perhaps to the avoidance of this future singularity, leaving a very inhomogeneous universe. This situation may occur if the pressure is negative enough in order to satisfy the condition $\frac{p}{\rho} = \alpha < - \frac{5}{3}$. In the present work we extend those result showing that phenomenon of enhancing of the inhomogeneities in large scales is independent of the spatial curvature, and that the critical point $\alpha = - \frac{5}{3}$ is present in any class of homogeneous and isotropic universe. In order to do so, we will solve the perturbed equations for scalar modes for any value of $k$ and $\alpha$. An asymptotic analysis will reveal the existence of critical behavior associated to $\alpha = - \frac{5}{3}$. For a universe dominated by a fluid whose equation of state is given by $p = \alpha\rho$, the relevant equation is $$\frac{a'^2}{a^2} + k = \frac{8\pi G}{3}\rho\,a^2 \quad , \quad \rho = \rho_0\,a^{-3(1 + \alpha)} \quad .$$ The primes indicate derivations with respect to the conformal time $\eta$, and $k$ is the spatial curvature parameter, $k = \pm1,0$. The solution for this equation may be written in a unified form as $$a(\eta) = a_0\biggr[\frac{1}{\sqrt{k}}\sin\biggr(\frac{1 + 3\alpha}{2}\sqrt{k}\eta\biggl)\biggl]^\frac{2}{1 + 3\alpha} \quad .$$ For $k = 1$, this solution represents a universe that begins and end at a singularity at $a = 0$ for $1 \leq \alpha < - \frac{1}{3}$, while for $\alpha < - \frac{1}{3}$ it represents a bouncing universe. On the other hand, for $k = - 1, 0$, it represents an ever expanding universe for any value of $\alpha$, with the following characteristics: for $1 \geq \alpha > - \frac{1}{3}$, the expansion implies $0 \leq \eta < \infty$ corresponding, in terms of the cosmic time $t$, to $0 \leq t < \infty$; for $- \frac{1}{3} > \alpha$, the expansion implies $- \infty < \eta \leq 0$, corresponding to $0 \leq t < \infty$ if $- \frac{1}{3} > \alpha \geq - 1$ and $- \infty < t \leq 0$ if $- 1 > \alpha$. This last feature leads to the notion of big rip. To study the evolution of scalar perturbations, we use the gauge invariant formalism. For a perfect fluid the evolution of the scalar perturbation is given by a single equation for the gravitational potential $\Phi$ [@brand; @mukhanov]: $$\label{perturbed} \Phi'' + 3(1 + \alpha)H\Phi' + \biggr\{\alpha\,n^2 + 2H' + (1 + 3\alpha)(H^2 - k)\biggl\}\Phi = 0 \quad ,$$ where $H = \frac{a'}{a}$ and $n^2$ is the eigenvalue of the Laplacian operator $\nabla^2\Phi = - n^2 \Phi$. The perturbed equation can be recast under the following form $$\label{equation} (1 - z)z\Phi'' + \frac{7 + 9\alpha}{2(1 + 3\alpha)}(1 - 2z)\Phi' + \tilde n^2\Phi = 0 \quad ,$$ where $$z = \frac{1 + \cos(\sqrt{k}\theta)}{2} \quad , \quad \tilde n^2 = \frac{4k}{(1 + 3\alpha)^2}\biggr[\alpha n^2 - 2(1 + 3\alpha)k\biggl] \quad , \quad \theta = \frac{1 + 3\alpha}{2}\eta \quad .$$ The solution of (\[equation\]) can be represented under the form of hypergeometric functions. It reads in general, for any value of $k$ and $\alpha$, as follows: $$\begin{aligned} \Phi_n(\eta) = c\, {}_2F_1\biggr[A_+,A_-;B;z\biggl] + \nonumber\\ \bar c\,z^{1 - B}{ }_2F_1\biggr[A_+ - B + 1,A_- - B + 1;2 - B;z\biggl] \quad ,\end{aligned}$$ where $$A_\pm = \frac{1}{2}\biggr\{6\frac{1 + \alpha}{1 + 3\alpha} \pm \sqrt{36\frac{(1 + \alpha)^2}{(1 + 3\alpha)^2} + 4\tilde n^2}\biggl\} \quad , \quad B = \frac{7 + 9\alpha}{2(1 + 3\alpha)} \quad ,$$ and $c,\bar c$ are constants. The asymptotic analysis for the flat case $k = 0$ was made in reference [@sergio1]. In this case, the hypergeometric’s function reduces to Bessel’s functions. It has been shown that one of the two modes remains constant, in the long wavelength limit, for any value of $\alpha$. In the same limit, however, the other mode decreases when $\alpha > - \frac{5}{3}$, but grows with time when $\alpha < - \frac{5}{3}$. For $\alpha = - \frac{5}{3}$, both modes are constants. In the case $k \neq 0$, such analysis, in terms of large or small wavelength limit, is more involved since, contrarily to the flat case, we have a scale given by the Hubble radius and another scale given by the curvature. It becomes easier, in this sense, and for our purpose more relevant, to consider the behavior in the extremes of the time interval. In order to perform this analysis, we must take into account some convenient transformation properties of the hypergeometric functions. In special, the following transformations will be useful (see reference [@grad]): $$\begin{aligned} {}_2F_1(A,B;C;z) = \frac{\Gamma(C)\Gamma(C-A-B)}{\Gamma(C-A)\Gamma(C-B)}{}_2F_1(A,B;A+B-C+1;1-z) &+&\nonumber\\ (1-z)^{C-A-B}\frac{\Gamma(C)\Gamma(A+B-C)}{\Gamma(A)\Gamma(B)}{}_2F_1(C-A,C-B;C-A-B+1;1-z)\quad &;&\\ {}_2F_1(A,B;C;z) = \frac{\Gamma(C)\Gamma(B-A)}{\Gamma(B)\Gamma(C-A)}(-1)^Az^{-A}{}_2F_1\biggr(A,A+1-C;A+1-B;\frac{1}{z}\biggl) &+& \nonumber\\ \frac{\Gamma(C)\Gamma(A-B)}{\Gamma(A)\Gamma(C- B)}(-1)^Bz^{-B}{}_2F_1\biggr(B,B+1-C;B+1-A;\frac{1}{z}\biggl) \quad &.&\end{aligned}$$ Hence, we have the following asymptotic behaviors: $$\begin{aligned} z \rightarrow 0 \quad \Rightarrow \quad {}_2F_1(A,B;C;z) &\sim& z^{1 - C} \quad ,\\ z \rightarrow 1 \quad \Rightarrow \quad {}_2F_1(A,B;C;z) &\sim& \frac{\Gamma(C)\Gamma(C-A-B)}{\Gamma(C - A)\Gamma(C - B)}(1 - z)^{C-A-B} \nonumber\\ &+& \frac{\Gamma(C)\Gamma(A+B-C)}{\Gamma(A)\Gamma(B)} \quad ,\\ z \rightarrow \infty \quad \Rightarrow \quad {}_2F_1(A,B;C;z) &\sim& \frac{\Gamma(C)\Gamma(B-A)}{\Gamma(B)\Gamma(C-B)}(-1)^A\,z^{-B}\nonumber\\ &+&\frac{\Gamma(C)\Gamma(A-B)}{\Gamma(A)\Gamma(C-B)}(-1)^B\,z^{-A} \quad .\end{aligned}$$ Using these expressions, we can determine the behavior of the perturbations in the two different extremities of the time interval, for each value of $k$. - $k = 1$. In this case the conformal time interval is $0 \leq \eta \leq \frac{2\pi}{1 + 3\alpha}$ for $\alpha > - \frac{1}{3}$ and $\frac{2\pi}{1 + 3\alpha} \leq \eta \leq 0$ for $\alpha < - \frac{1}{3}$. Using the asymptotic expressions written above, we find the following behaviors: for $\alpha > - \frac{1}{3}$, there is initially two decreasing modes and, as the universe approaches the big crunch at $\eta = \frac{2\pi}{1 + 3\alpha}$, there are a constant mode and a growing mode; for $- \frac{1}{3} > \alpha > - \frac{5}{3}$ there is initially, during the contraction phase, a growing mode and a constant mode, and as the scale factor diverges in the other asymptotic, there is a constant mode and a decreasing mode; for $\alpha = - \frac{5}{3}$, both modes are constant at the beginning of the contraction phase and at the end of the expansion phase; for $\alpha < - \frac{5}{3}$, there is a constant mode and a decreasing mode at the universe begins to contract, and there are two increasing modes as the universe approaches the big rip. - $k = - 1$. The range of the conformal time is $0 \leq \eta < \infty$ for $\alpha < - \frac{1}{3}$ and $- \infty < \eta \leq 0$ for $\alpha < - \frac{1}{3}$. The open case is more involved because, in opposition to the closed universe, the asymptotic behavior of the modes depends on the scale of the perturbation. Let us consider the situation where the eigenvalues of the Laplacian operator is null. Hence, when $\alpha > - \frac{1}{3}$ there is initially and in future infinity two decreasing modes. For $- \frac{1}{3} < \alpha < - \frac{5}{3}$ there is initially two decreasing modes, but in the future infinity there is a constant mode besides a decreasing mode; the case $\alpha = - \frac{5}{3}$ differs from the preceding one by the fact that in future infinity both modes are constant. Finally, when $\alpha < - \frac{5}{3}$, both modes are initially decreasing but they become growing modes as the big rip is approached. When $\alpha = - \frac{1}{3}$, in all cases, the same features observed in the flat universe are reproduced here, since for this particular equation of state the matter density scales as the curvature parameter in the Friedmann’s equation. The main conclusion of the previous analysis is that a phantom cosmological scenario is highly unstable against scalar perturbations under the condition of an isotropic and homogeneous background universe if the pressure is negative enough, that is $\frac{p}{\rho} < - \frac{5}{3}$: the scalar perturbations grows as the big rip is approached. As in the flat case, the Hubble radius shrinks with time in the phantomic case. This means that the large scale approximation becomes essentially valid asymptotically for all perturbation scales in the phantomic case. The analysis was made using a perfect fluid material content. However, at large scales, it is expected that a more fundamental representation, using for example, scalar fields, must give the same results as the perfect fluid case [@sergio2]. A curious feature of the above results concerns the critical point $\alpha = - \frac{5}{3}$. As far as we know, it does not correspond to any energy condition (contrarily to $\alpha = - \frac{1}{3}$ and $\alpha = - 1$). Also, it seems not related neither to the Hubble parameter, nor to deceleration parameter, or even to the statefinder parameters [@starobinsky]. However, it seems to be a general critical point for the analysis of perturbation, that is not reflected in the kinematical quantities like those we have quoted above. A dynamical system analysis of the background does not reveal any particular feature at $\alpha = - \frac{5}{3}$ [@joel]: it is not a critical point for the background. Moreover, in the perturbed equation (\[perturbed\]) there is no explicit special structure for that particular value of the parameter $\alpha$. The nature of this critical point must still be cleared up. We thank Jérôme Martin for his criticism and suggestions. We thank also CNPq (Brazil) and the French-Brazilian scientific cooperation CAPES/COFECUB (project number 506/05) for partial financial support. [90]{} S. Weinberg, [**Gravitation and cosmology**]{}, Wiley, New York (1972). T. Padmanabhan, [**Structure formation in the universe**]{}, Cambridge University Press, Cambridge (1993). R. Brandenberger, H. Feldman and V. Mukhanov, Phys. Rep. [**215**]{}, 203 (1992). D.N. Spergel et al, Astrophys. J. Suppl. [**170**]{}, 377 (2007). V. Sahni, Lect. Notes Phys. [**653**]{}, 141 (2004). R. Dürrer and R. Maartens, [*Dark energy and dark gravity*]{}, arXiv:0711.0077. R. Caldwell, Phys. Lett. [**B545**]{}, 23 (2002). J.C. Fabris and S.V.B. Gonçalves, Phys. Rev. [**D74**]{}, 027301 (2006). F. Piazza and S. Tsujikawa, JCAP [**0407**]{}, 004 (2007). V. Mukhanov, [**Physical foundations of cosmology**]{}, Cambridge University Press, Cambridge (2005). I.S. Gradshteyn and I.M. Rhyzik, [**Tables of integrals, series and products**]{}, Academic Press, London (1994). J.C. Fabris, S.V.B. Gonçalves and N.A. Tomimura, Class. Quant. Grav. [**17**]{}, 2983 (2000). V. Sahni, M. Sami, A.A. Starobinsky and U. Alam, JETP Lett. [**77**]{}, 201 (2003). A.B. Batista, J.C. Fabris, S.V.B. Goncalves and J. Tossa, Int. J. Mod. Phys. [**A16**]{}, 4527 (2001). [^1]: e-mail: [email protected]. On leave of absence of Departamento de Física, Universidade Federal do Espírito Santo, 29060-900, Vitória, Espírito Santo, Brazil [^2]: e-mail: [email protected] [^3]: e-mail: [email protected]

ArXiv

--- abstract: 'The censorship of toxic comments is often left to the judgment of imperfect models. Perspective API, a creation of Google technology incubator Jigsaw, is perhaps the most widely used toxicity classifier in industry; the model is employed by several online communities including *The New York Times* to identify and filter out toxic comments with the goal of preserving online safety. Unfortunately, Google’s model tends to unfairly assign higher toxicity scores to comments containing words referring to the identities of commonly targeted groups (e.g., “woman,” “gay,” etc.) because these identities are frequently referenced in a disrespectful manner in the training data. As a result, comments generated by marginalized groups referencing their identities are often mistakenly censored. It is important to be cognizant of this unintended bias and strive to mitigate its effects. To address this issue, we have constructed several toxicity classifiers with the intention of reducing unintended bias while maintaining strong classification performance.' author: - | Elizabeth Reichert[^1]\ `[email protected]` - | Helen Qiu\ `[email protected]` - | Jasmine Bayrooti\ `[email protected]` date: 'December 13, 2019' title: | Reading Between the Demographic Lines:\ Resolving Sources of Bias in Toxicity Classifiers --- Introduction ============ **Summary of approach.** In this paper, we describe a variety of toxic comment detection models crafted with the goal of lessening identity-driven bias while preserving high classification performance. To balance our dataset, we used over-sampling and under-sampling techniques in combination with natural text generation. We created logistic regression, neural network, and LSTM (Long Short-Term Memory) models to perform the toxicity classification of comments. We explored different features (including TF-IDF and GloVe embeddings), hyper-parameter configurations, and model architectures. We fed into our models the original training data as well as our re-balanced dataset and compared their performances both in terms of classification and mitigation of unintended bias. We also compared the performance of our best model to Perspective API by examining the prevalence of false positives in identity-related tweets by members of the U.S. House of Representatives that are assumed to be non-toxic. **Social impact.** We hope that the techniques explored in this paper will help to improve the fairness of toxic classifiers. Mistakenly flagging non-toxic content, especially content produced by marginalized voices, is an important social issue; it is critical to foster a safe online environment while also protecting the freedom of expression of all people. Related Work ============ Prior work in the realm of toxic speech detection (e.g. detecting offensive speech in tweets) has used bag-of-words approaches for feature representation along with models in logistic regression, decision trees, and linear SVMs to classify public social comments \[1, 2\]. More recent work has explored the application of CNN (Convolutional Neural Network) and GRU (Gated Recurrent Neural Network) in detecting hateful language on Twitter \[3\]. However, a common challenge with these models is the misclassification of content associated with frequently targeted communities \[4\]. To address Perspective API’s unintended bias problem, Jigsaw has proposed bias mitigation methods including mining assumed “non-toxic” data from Wikipedia articles to achieve a greater balance between toxic and non-toxic content referencing identity terms \[5\]. The optimal method to eliminate unintended bias resulting from imperfect training sets is not yet clear. According to a study on racial bias in abusive language detection, a one-size-fits-all approach to defining and detecting abusive language is not viable due to different speech norms in different communities, and providing more context, such as the author’s profile and the content that the author is responding to, is key to improving model performance \[6\]. Data ==== Datasets -------- **Toxicity-labeled comment data.** Our consists of approximately $1.8$ million comments from the currently-inactive online social platform Civil Comments. Each comment is labeled with a target value[^2] between $0$ and $1$ (inclusive) as well as several additional toxicity sub-type attributes (i.e., obscene, threat, insult, severe toxicity, identity attack, sexually explicit). Approximately $400,000$ comments are also marked with identity attributes representing the identities mentioned in the comment (e.g., “Muslim,” “bisexual,” etc.). We assign a binary toxicity label to each comment in our dataset, where the label “non-toxic” ($0$) corresponds to a toxicity score of below $0.5$ and the label “toxic” ($1$) corresponds to a toxicity score of between $0.5$ and $1$. The following are examples of comments from our dataset including toxicity scores and labels: - “This bitch is nuts. Who would read a book by a woman.”; Toxicity Score: $0.83$; Toxicity Label: $1$ (“toxic”) - “Why would you assume that the nurses in this story were women?”; Toxicity Score: $0.0$; Toxicity Label: $0$ (“non-toxic”) To speed up training, we randomly sampled $50\%$ of the $1.8$ million comments in our dataset to construct the training and test sets based on an $80:20$ split. **Identity-related tweets of politicians.** To measure the degree of identity-driven bias mitigation achieved by our top-performing model relative to Perspective API, we utilized a of tweets made by members of the U.S. House of Representatives. The dataset consists of the latest $200$ tweets (as of May 17th, 2018) from all members of the House of Representatives. A total of $84,502$ tweets are included along with the tweeter’s name and political party affiliation[^3] (i.e., Democrat or Republican). After filtering for tweets referencing at least one identity, $1,984$ tweets remained. We used this set of approximately $2,000$ identity-related tweets as an additional dataset to compare our best model’s level of unintended bias relative to that of Google’s Perspective API model. Features -------- **Term-frequency times inverse document-frequency (TF-IDF).** We use scikit-learn’s TF-IDF vectorizer, which is a normalized bag-of-words transformation of sentences, to construct features. The resulting vector for each comment is of length $SIZE_{corpus}$ (i.e., $198,234$). Each element in the vector is the word frequency in the comment multiplied by the inverse of the word frequency in the dataset. The goal of using TF-IDF instead of the raw frequencies of a word occurrences is to lessen the impact of words that occur very often in a given corpus (e.g., stop words such as “the,” “a,” “and,” etc.) that have empirically less informative than words that occur less frequently in the corpus. For the features generated via TF-IDF, our original training input was of shape ($721950, 198234$) and the shape of our test set was ($180487, 198234$). **GloVe embeddings.** We experimented with pre-trained GloVe embeddings as another way to extract features from the training data, since prior work suggests that GLoVe outperforms related models on similarity tasks and named entity recognition \[7\]. We used the $25$-dimension GloVe vectors trained from $2$ billion tweets on Twitter. We chose to use this specific GloVe embeddings model because the twitter data that the model was trained on is similar to the comment data on Civil Comments. For the features generated via GloVe embeddings, our original training input was of shape ($721950, 25$) and the shape of our test set was ($180487, 25$). Data Exploration ================ It is important to note that our dataset is imbalanced. Using a toxicity score threshold of $0.5$, there are roughly $8\%$ toxic examples and $92\%$ non-toxic examples. Out of the approximately $1.8$ million comments in our dataset, $405,130$ of those comments were randomly selected and marked with identity attributes representing the identities mentioned in the comment (e.g., “black,” “Christian,” etc.). We found that, out of the comments that were labeled with identity attributes, approximately $44\%$ did not reference any identities while around $56\%$ referenced at least one identity—roughly an even split. We found that comments that referenced at least one identity were more likely to be toxic than comments that did not reference any identities. The distribution of comments based on their toxicity and reference of identity terms is given in Figure 1.[^4] {width="35.00000%"} We notice that there is a higher incidence of toxicity in comments that reference identity terms than comments that do not reference any identity terms. Specifically, we observe that toxic comments are $14.9\%$ of comments that reference identities and only $6.8\%$ of comments that do not reference identities. This difference is highly statistically significant with $p < 0.00001$. Since we observe such a large difference in the prevalence of toxicity in identity versus non-identity comments, it is important that, when re-sampling comments to balance the split of toxic versus non-toxic examples in our dataset, we control for the split of identity versus non-identity examples within the toxic and non-toxic examples. To do so, we sample our data equally from each the following comment categories: toxic identity, toxic non-identity, non-toxic identity, and non-toxic non-identity. Methods ======= Baseline vs. Oracle ------------------- Our baseline is a Naive Bayes Classifier (NBC) whose input is a bag-of-words representation of a pre-processed comment. Due to limitations in computing power, we randomly sampled $25\%$ of our dataset (approximately $450,000$ comments) to construct our training and test sets. Performing an $80:20$ random split of the data, we trained with $360,976$ comments and ran the NBC on a test set of size $90,244$. We performed data pre-processing and feature extraction in scikit-learn. We observe that the Naive Bayes binary classification model performs poorly, with precision of approximately $15\%$ and recall of around $5\%$. This means that out of all of the toxic comments, our model only predicts $5\%$ of them as being toxic, and out of all of the comments our model predicts as toxic, only $15\%$ of them are actually toxic. The F1 score of this model is roughly $7.5\%$. The oracle is the human labeling of the toxicity of the comments. Humans are better able to comprehend the intended meaning of the comment and the contextual use of identity terms. The bag-of-words approach de-contextualizes comments by eliminating their sequential relationships, and the Naive Bayes model is not sophisticated enough to capture the meaning of the text—rather it attempts to discern the underlying distribution that produced the model’s inputs under the assumption that the inputs are independent. Logistic Regression ------------------- Logistic regression is commonly applied to classification problems and is a simple, yet powerful model. We implemented two logistic regression models: one with features generated via TF-IDF and one with features generated via GloVe embeddings. The output for the model was binary—either $0$ (non-toxic) or $1$ (toxic). Neural Network -------------- Neural networks have strong representational power; neural networks with at least one hidden layer are “universal approximators” and can approximate any continuous function \[9\]. We first trained a two-layer fully-connected neural network that has the fully-connected—ReLU—fully-connected architecture. The score for each output class is calculated with $f = \max(0, x W_1 + bias))W_2 + bias$. Here, the input $x$ is a $N \times D$ matrix where $N$ represents the number of training examples in the training set, $D$ represents the number of features per training example, and $W_1$ is a $D \times SIZE_{hidden}$ matrix that transforms the training data into a $N \times SIZE_{hidden}$ intermediate matrix. The ReLU activation function $\max(0, xW_1)$ is a non-linearity that is applied element-wise. Finally, $W_2$ is a matrix of size $SIZE_{hidden}\times C$, and thus $f$ has shape $N \times C$. The model outputs $C$ numbers interpreted as class scores for each training example. The parameters $W_1$ and $W_2$ are learned with stochastic gradient descent where the gradients are derived with chain rule and computed with back propagation. We used scikit-learn to create the neural network architecture. We first trained the two-layer neural net on scikit-learn’s TFIDF-vectorized features. The output layer produced classification scores for two classes—non-toxic represented by $0$, and toxic represented by $1$. The size of the hidden layer was $100$ and the learning rate was $1 \times e^{-5}$. We also experimented with transforming the comments into GloVe embeddings, summing up the GloVe vectors for each words in a comment and feeding in the sum vector as input. Once again, the output for the model was binary—either $0$ (non-toxic) or $1$ (toxic). {width="35.00000%"} To explore if a slightly deeper neural architecture would result in better model performance, we also trained a three-layer fully-connected neural network, with hidden layer sizes of $75$ and $50$ and a learning rate of $1 \times e^{-5}$. The input and output were of the same shape as the two-layer model. Two-Layer Bidirectional LSTM ---------------------------- Words that appear near the beginning of a sentence often influence the meaning of subsequent words, ultimately affecting the interpretation of that sentence. Therefore, to distinguish between appropriate versus inappropriate usage of identity terms, it is critical to process the context in which these identity words are used in each comment. We implemented a two-layer bidirectional LSTM (Long Short-Term Memory) model to capture word dependencies and sequential information. We chose a bidirectional architecture since it encapsulates information in both the forward and backward sequences of words in sentences. {width="50.00000%"} The model takes in a padded list of GloVe vectors as input, where each vector is a $25$-dimensional GloVe representation of a word in the comment. We dropped or padded each comment so that all comments had a length of $200$. The input to the first-layer LSTM has shape $(batch$ $size$, $comment$ $length$, $embed$ $length)$. The output of the bidirectional processing by the first-layer LSTM $lstm1$ is concatenated and fed into the second-layer LSTM $lstm2$. The output from $lstm2$ is then concatenated and put through a max pooling layer and an average pooling layer. We used both max pooling and average pooling and concatenated their results with the goal of extracting the most extreme word information as well as preserving the overall semantic information of each comment. Finally, with two linear layers we reduce the dimension of the output associated with each comment to $1$, which is then converted via sigmoid function to a value between $0$ to $1$ representing the probability of the comment being toxic. To prevent overfitting, we experimented with spatial dropout regularization on the input to $lstm1$. Using hyperparameter tuning, we selected $128$ as the number of LSTM hidden units, $512$ as the batch size, $1 \times e^{-5}$ as the learning rate, and $0.3$ as the dropout rate. After each epoch in training, we utilized our validation set of $40,000$ comments to evaluate the model’s F1 score and update our best model, which was later used for prediction on the test set. Our implementation of LSTM drew inspiration from . To determine the appropriate number of epochs, we plotted training loss against epoch value. For both the LSTM model trained on the original training set and the LSTM model trained on the balanced training set, after approximately 10 epochs, we observed that training loss converged, thus suggesting that a local optimum had been reached. {width="40.00000%"} {width="40.00000%"} Data Rebalancing ---------------- We used natural text generating techniques to produce additional comments to aid in balancing our training dataset. More specifically, we created our own comment templates for the following comment categories: toxic non-identity, toxic identity, non-toxic identity.[^5] We then filled in these templates with different identity terms (as well as identity slurs when the comments were designed to be offensive). Using these comments as source text, we utilized a bidirectional LSTM encoder and LSTM decoder to generate variations (or paraphrases) of these comments with temperature 0.8 \[11\]. The following are comments generated by the paraphraser: - I am black and I feel that this issue is important to me. - I’m black and I think this problem is significant to me. We then incorporated these comments into our training set and used over-sampling and under-sampling techniques to sample $50,000$ comments from each of the comment categories (i.e. toxic identity, toxic non-identity, non-toxic identity, and non-toxic non-identity). Thus, our rebalanced training set consisted of a total of $200,000$ comments. We ran the logistic regression, neural network, and LSTM models mentioned above again on the rebalanced training set, and tested the model performances on the original test set so that we were able to directly compare the performance of the models trained on both the original and rebalanced training sets. Results ======= Evaluation Metrics ------------------ **Classification performance.** We used area under the receiver operating characteristic curve (AUC) and F1 score[^6] on the test set as measures of model performance. Both metrics are commonly used to evaluate model performance on imbalanced data sets. **Mitigation of identity-driven unintended bias.** To measure the bias of our classifier, we compared the false positive rate ($FPR$) for comments that reference identities to that of comments that do not reference identities. We also compared the false negative rate ($FNR$) for identity versus non-identity comments. By the Equality of Odds, these rates should not differ if the model is fair \[11\]. **Comparison to Perspective API.** To evaluate the ability of our model to mitigate identity-driven unintended bias compared to Google’s toxicity classifier Perspective API, we examined tweets of U.S. Representatives referencing identities (e.g., “transgender,” “Jewish”). We assume that politicians would not tweet toxic content, and thus we used the prevalence of false positives in this dataset as a measurement of unintended bias. Model Performance ----------------- **Models trained on original training set.** We begin by evaluating the performance of our models trained on the original (unbalanced) training set. {width="50.00000%"} Before rebalancing, the two-layer neural network with TF-IDF features achieved superior classification performance with an AUC of $0.76$ and an F1 score of $0.79$. For all of our models, compared to TF-IDF vectorization, feature extraction using GLoVe embeddings resulted in reduced classification performance. This is likely because summing up the GLoVe-represented word vectors in a sentence leads to a loss of semantic contributions from individual words. GLoVe representations were more suitable for our LSTM model to learn word dependencies and sequential relationships, as evidenced by strong model performance; the two-layer bidirectional LSTM achieved an AUC of $0.75$ and an F1 score of $0.78$. Surprisingly, our three-layer neural network performed worse than our two-layer neural network. To determine if our three-layer neural network was overfit, we compared both the AUC and F1 score on the training set versus test set. We found that the AUCs and F1 scores are essentially equivalent for both datasets, suggesting that the worse performance is not due to overfitting. It is important to note that our logistic regression model, the simplest of all of our models, achieved comparable performance to our best model, with an AUC of $0.74$ and an F1 score of $0.78$, when both models used TF-IDF features. For each of our models, we notice that the false positive rate ($FPR$) for comments that reference identity terms is $2$ to $3$ times greater than the $FPR$ for comments that do not reference identity terms. Similarly, we see that, for each of our models, the false negative rate ($FNR$) for “identity” comments is $2$ to $3$ times greater than the $FNR$ for “non-identity” comments. Thus, since the $FPR$ and $FNR$ differ between the two groups (“identity” and “non-identity” comments), by the definition of fairness outlined in the Equality of Odds, we observe bias \[12\]. **Models trained on rebalanced training set.** Next, we evaluate the performance of our models trained on the rebalanced training set. {width="50.00000%"} After rebalancing, the two-layer neural network with TF-IDF features achieved the highest AUC ($0.82$), while the logistic regression model trained on TF-IDF features gave the highest F1 Score ($0.76$). Since GloVe embeddings did not perform well for our logistic regression and neural network models, we decided that when training on the rebalanced dataset, we would not use features via GloVe embeddings for those models. All of our models were similar in terms of classification performance, with AUCs ranging from $0.81$ to $0.82$. The F1 scores of our models are also very much alike (ranging from $0.72$ to $0.76$) with the exception of the LSTM model. Our lower F1 score for this model may be the result of a sub-optimal classification threshold[^7] since F1 score is dependent on the binary classification cutoff while AUC is not. We believe that the LSTM model could have outperformed our other models if it was supplied more training data. Since LSTM models have more hyperparameters, training on a larger dataset may have resulted in better tuning of those parameters. It makes sense that the $FPR$s are higher for our models trained on the rebalanced dataset than the $FPR$s for our models trained on the original training set, since, in the process of rebalancing, we are adding more toxic examples. We notice that, when comparing the $FPR$ and $FNR$ of the “identity” comments versus “non-identity” comments for each of our models, the ratio of the two values has decreased (instead of $2$ to $3$ times greater, the ratios of the rates are $1$ to $2.5$ times greater), suggesting unintended bias has been reduced. It is important to note that the increase in $FPR$s is larger in magnitude than the decrease in $FNR$s for our models trained on the balanced training set; ideally, rebalancing would have resulted in a smaller ratio between $FPR$s and $FNR$s for identity and non-identity comments in addition to reducing the magnitude of these rates. Our models trained on the rebalanced dataset perform better in terms of AUC on the test set than our models trained on the original training set. This suggests that rebalancing training data is an effective way to improve model classification performance. When comparing the top-performing models in terms of AUC, we observe an increase from $0.76$ to $0.82$. {width="50.00000%"} **Effects of incremental data rebalancing.** We examined the effects of incrementally rebalancing the training set on the classification performance and mitigation of unintended bias on our top-performing model (i.e., the two-layer neural network). We incrementally rebalanced each of the four comment categories (i.e., non-toxic identity, non-toxic non-identity, toxic identity, and toxic non-identity) in batches of $10,000$ comments until the dataset was fully balanced.[^8] We observe that, for all rebalancings, the AUC on the test set increases as the dataset becomes more balanced. When adding more toxic examples, unsurprisingly, we notice that $FPR$ increases while $FNR$ decreases; this is because increasing the number of toxic examples in our training set makes the model more likely to predict examples in the test set as toxic. Using the definition of fairness given by Equality of Odds, we visualize the mitigation of bias as the convergence of the $FPR$s and the convergence of the $FNR$s for identity and non-identity comments \[11\]. {width="20.00000%"} When rebalancing the number of non-toxic identity comment examples, we observe an increase in the test AUC with F1 score remaining constant. We notice that the $FPR$s slightly diverge while the $FNR$s slightly converge. {width="50.00000%"} When rebalancing the number of non-toxic non-identity comment examples, we observe an increase in the test AUC and F1 score. We see that the $FPR$s marginally converge while the $FNR$s marginally diverge. {width="50.00000%"} We find that the two-layer neural network performs worst in terms of AUC when there are no toxic identity comment examples. This makes sense since toxic comments commonly reference identity terms in a threatening or disrespectful manner. The test AUC increases dramatically as we rebalance the number of toxic identity comments while the F1 score slightly increases. The $FPR$s converge, intersecting at $10,000$ toxic identity comments, and then diverge while the $FNR$s substantially converge. As expected, the $FPR$s increase and the $FNR$s decrease. {width="50.00000%"} The test AUC increases while the F1 score remains constant when rebalancing the number of toxic non-identity comments. The $FPR$s converge while increasing, and the $FNR$s essentially remain equidistant while decreasing. {width="50.00000%"} **Bias mitigation comparison to Perspective API.** In the dataset of $1,984$ tweets made by U.S. politicians referencing identities that are assumed to be non-toxic, Perspective API labeled $10.3\%$ as toxic, while our best model (i.e., the two-layer neural network) only labeled $6.9\%$ as toxic. The lower prevalence of false positives achieved by our model indicates that we have outperformed Perspective API in terms of reducing identity-driven bias. We do not claim that our model performs generally better in terms of classification performance than Perspective API (it is doubtful since Google’s toxicity classifier has been trained on many more examples); however, this result does suggest that controlling for the reference of identity terms when rebalancing the training set to include an equal number of toxic and non-toxic examples is an effective means to mitigate identity-driven bias. Out of the $1,984$ tweets in our test set, $145$ were incorrectly marked as toxic by Perspective API but correctly predicted to be non-toxic by our top-performing model. These tweets included the following: - We must hold hearings & finally address that \#terrorism inflicted by white supremacy extremists is destroying \#USA. - When bias drives discipline, black girls miss out on the chance to learn. - \#WeRemember the victims of the Holocaust, including the millions of Jewish men, women, and children who were massacred. After examining these tweets, it appears that Perspective API had difficulty distinguishing toxicity from negative sentiment in tweets that referenced identity terms. Tweets referencing identity terms commonly describe inequities between or acts of hate against identity groups. Although these tweets may be calling out unfairness or violence (topics that have negative sentiment), the content of these messages is not toxic. Our model miscategorized $78$ tweets that Google’s model correctly predicted. Selected examples are given below: - Today we recognize Transgender Day of Visibility, seeing transgender people for who they are where they are. - To Jewish friends and family in NJ and around the world, I wish you all a Happy \#YomHaatzmaut. - To all my Muslim brothers and sisters and those observing in the \#CA13, I wish you a blessed, peaceful and happy month of Ramadan. The majority of tweets that our top-performing model incorrectly categorized had positive sentiment but referenced identity terms that were less common in our training data. Most of the identities in our training set related to gender (specifically women) and race, not sexuality and religion. We believe that having fewer examples of these identities hurt our model’s ability to correctly identify these tweets as non-toxic. To achieve better performance, it would be beneficial to train our model on a more comprehensive set of identity comment examples. Although when generating identity comment examples we included a wide range of identity terms, since our text generation was limited to paraphrasings, we were not able to achieve a breadth of positive contexts in which these identity terms are referenced. Future Work =========== We hope to explore models and techniques that will allow us to further reduce unintended bias while improving classification performance. We would like to explore more sophisticated models such as BERT in combination with LSTM, which incorporates self-attention mechanisms to make connections between words and relevant context and identify dependencies of words that are far away from each other \[13\]. We also wish to investigate more complex text generation techniques to increase the size and variety of our training set examples since having a greater volume of training examples will be useful when training more complex models. Acknowledgements ================ We would like to acknowledge Hancheng Cao for his superb mentorship throughout the research process; we thank him for his expert advice and encouragement. We would also like to recognize Victor Suthichai, whose paraphraser we customized to generate additional comment examples when balancing our training set and Benjamin Minixhofer, whose code we referenced when designing the architecture of our LSTM model. References ========== \[1\] Chatzakou, Despoina, et al. “Mean birds: Detecting aggression and bullying on twitter.” Proceedings of the 2017 ACM on Web Science Conference. ACM, 2017. \[2\] Davidson, Thomas et al. “Automated Hate Speech Detection and the Problem of Offensive Language.” ICWSM (2017). \[3\] Zhang, Ziqi et al. “Detecting Hate Speech on Twitter Using a Convolution-GRU Based Deep Neural Network.” ESWC (2018). \[4\] Burnap, Peter and Matthew Leighton Williams. “Cyber Hate Speech on Twitter: An Application of Machine Classification and Statistical Modeling for Policy and Decision Making.” (2015). \[5\] Dixon, Lucas et al. “Measuring and Mitigating Unintended Bias in Text Classification.” AIES (2018). \[6\] Davidson, Thomas et al. “Racial Bias in Hate Speech and Abusive Language Detection Datasets.” ArXiv abs/1905.12516 (2019). \[7\] Pennington, Jeffrey et al. “Glove: Global Vectors for Word Representation.” EMNLP (2014). \[8\] Cybenko, G. “Approximation by Superpositions of a Sigmoidal Function.” Math. Control Signals Systems (1989)2: 303-314. \[9\] Sap, Maarten et al. “The Risk of Racial Bias in Hate Speech Detection.” ACL (2019). \[10\] Minixhofer, Ben. “Simple LSTM - PyTorch.” Kaggle, 2019. https://www.kaggle.com/bminixhofer/simple-lstm-pytorch-version \[11\] Suthichai, Victor. “Paraphraser.” Github, 2017. https://github.com/vsuthichai/paraphraser \[12\] Moritz Hardt, Eric Price, and Nathan Srebro. 2016. “Equality of Opportunity in Supervised Learning.” CoRR abs/1610.02413 (2016). http://arxiv.org/abs/1610. 02413 \[13\] Jacob Devlin Ming-Wei Chang Kenton Lee Kristina Toutanova. 2019. “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding” arXiv:1810.04805v2 \[cs.CL\] 24 May 2019 [^1]: Equal Contribution, Stanford University [^2]: The target is the fraction of human labelers that found the comment “very toxic” (meaning severely hateful or disrespectful to the point of making one very likely to leave the platform) or “toxic” (meaning unreasonable or rude to the point of making one somewhat likely to leave the platform). [^3]: Approximately $51\%$ of tweets are made by Republicans while around $49\%$ of tweets are authored by Democrats. [^4]: “Toxic” refers to comments with a toxicity score of greater than or equal to $0.5$ while “non-toxic” refers to comments with a toxicity score less than $0.5$. “Identity” refers to comments that reference at least one identity term and “non-identity” refers to comments that do not reference any identity terms. An identity term is referenced if the fraction of human labelers that believe an identity term is referenced is greater than $0$. [^5]: Non-toxic non-identity comments covered too broad of a topic range for their generation to be worthwhile. [^6]: F1 score is defined as the harmonic mean of precision and recall. [^7]: The toxicity classification threshold is $0.5$. [^8]: The dataset is fully balanced when each of the comment categories has $50,000$ comments. The number of non-toxic comments was incrementally decreased and the number of toxic comments was incrementally increased to reflect the directionality of the rebalancing we performed on our original (imbalanced) dataset.

ArXiv

--- abstract: 'The work of Jaffe, Jenkins and Kimchi \[Phys. Rev. [**D**]{}79, 065014 (2009)\] is revisited to see if indeed the region of congeniality found in their analysis survives further restrictions from nucleosynthesis. It is observed that much of their congenial region disappears when imposing conditions required to produce the correct and required abundances of the primordial elements as well as ensure that stars can continue to burn hydrogen nuclei to form helium as the first step in forming heavier elements in stellar nucleosynthesis. The remaining region is a very narrow slit reduced in width from around 29 MeV found by Jaffe [*et al.*]{} to only about 2.2 MeV in the difference of the nucleon/quark masses. Further bounds on $\delta m_q /m_q$ seem to reduce even this narrow slit to the physical point itself.' author: - 'M. Hossain Ali' - 'M. Jakir Hossain' - Abdullah Shams Bin Tariq bibliography: - 'CongenialityNucleosynth.bib' title: Congeniality Bounds on Quark Masses from Nucleosynthesis --- A reasonably contemporary approach is to study, even without going into anthropic arguments, the nature of alternative universes as one changes the values of physical parameters. In the parameter space, one then looks for regions that could be similar to our universe and may possibly be congenial to the creation and sustenance of intelligent life [@Wei89; @Teg97; @Teg98a; @Teg98b; @Hog00; @Teg06; @Jaf09; @Don10]. Bounds thus obtained may be referred to as congeniality bounds. In a recent work, Jaffe, Jenkins and Kimchi [@Jaf09] studied how sensitive our universe would be to variations of quark masses. For this they chose to study the variations of masses of the three lightest quarks $u$, $d$ and $s$, under the constraint that the sum of these masses, $m_T$ remained fixed. They also studied variations of $m_T$. Their basic idea was to find the two lightest baryons for any quark mass combination and consider them to play the roles of the proton and neutron in forming nuclei. In this process they also considered $\Lambda_{\mathrm{QCD}}$ to be an adjusted free parameter that they tuned to keep the average nucleon mass at 940 MeV. They then studied the variation of nuclear stability and, in light of this, tried to obtain the regions of the parameter space where nuclear chemistry in a somewhat familiar form could be sustained. The starting point is that the three light quark masses would be changed keeping their sum $m_T$ fixed. This parameter space can be neatly shown in the form of an equilateral triangle \[Fig. \[mtriangle\]\] where the distances of a point from the base and right and left sides are, respectively, the masses of the up, down and strange quarks. ![\[mtriangle\]The model space of light quark masses for a fixed $m_T$ shown in the form of a triangle where the distance from the three sides give the three masses. Figure reproduced from [@Jaf09].](mass_triangle.eps){width="0.8\columnwidth"} In this manner they identified congenial regions in a triangle parametrized in terms of $x_3$ and $x_8$ \[Fig. \[x3x8\]\] defined as $$\begin{aligned} x_3 &=& \frac{2m_3}{\sqrt{3}}\frac{100}{m_T^\oplus}=\frac{100\left(m_u-m_d\right)}{\sqrt{3}m_T^\oplus}\\ x_8 &=& \frac{2m_8}{\sqrt{3}}\frac{100}{m_T^\oplus}=\frac{100\left(m_u+m_d-2m_s\right)}{\sqrt{3}m_T^\oplus}\end{aligned}$$ ![\[x3x8\]The model space of light quark masses parametrized in terms of $x_3$ and $x_8$ reproduced from [@Jaf09]. The point labeled ‘us’ points to the physical value in our present universe and therefore has coordinates $(x_3^\oplus, x_8^\oplus)$.](x3x8.eps){width="0.8\columnwidth"} Here $m_T^\oplus$ is the sum of the light ($u$, $d$ and $s$) quark masses in our universe. It is obvious that $x_3$ basically gives the isospin splitting while $x_8$ is related to the breaking of $SU(3)_f$ due to the mass of the strange quark. Their results can be summarized in Fig. \[cong\], where the congenial regions are indicated in green [^1]. This triangle is for $m_T^{\oplus}$, [*i.e.*]{} with $m_T$ as it is in the present universe. They have also studied variations in $m_T$, but our work is limited to commenting on the case of $m_T^\oplus$, understanding that the same arguments qualitatively extend to other values of $m_T$. This is further justified in the discussions near the end of this report.  Figure reproduced from [@Jaf09] identifying congenial regions in the quark mass triangle with green bands. The red and white regions are uncongenial and uncertain, respectively.](congenialtriangle.eps){width="0.8\columnwidth"} It is increasingly being understood that if there is complexity, fine-tuning is inevitable [@Brad11]. Even if one is not happy with anthropic arguments, we simply cannot get away from fine-tuning. With this in mind, the first impression that one has from Fig. \[cong\], is that the congenial region seems to be surprisingly large - allowing, around one order of magnitude variations in the quark masses. Though this already involves intricate compensating adjustments in $\Lambda_{\mathrm{QCD}}$ to keep the average ‘nucleon’ mass fixed. However, one should appreciate the difficulty in setting up a new framework in which a problem can be studied. From this perspective the authors of [@Jaf09] should be commended for presenting, literally from scratch, a setup for studying the congeniality bounds on quark masses. This setup can be extended removing some of the constraints used in any further work. Indeed, considering the significance of the work it was chosen first for a Viewpoint article in Physics [@Per09] and then went into a cover story in Scientific American [@Jen10]. In our work, we remain within the provided setup, but extend the analysis to bounds provided by nucleosynthesis. It should be noted here that whereas, on the one hand, nuclear masses and stability expectedly vary comparatively slowly with quark masses; on the other hand, the observed abundances of the lightest nuclei hydrogen and helium provide much more stringent bounds on the variation of nucleon masses. We report below how the congeniality triangle of Fig. \[cong\] is modified by the application of these constraints. At the outset, the variation of the octet baryon masses as one traverses along the borders of the triangle (Figs. 10 and 11 of [@Jaf09]) were reproduced to gain confidence in our code and our understanding of the framework. The fitted parameter values of $c_T$, $c_3$ and $c_8$ from Table III of Ref. [@Jaf09] were used in the equation $$M_B = C_0 + c_Tx_T + c_8x_8 + c_3x_3 + \left\langle B \left|H_{\mathrm{EM}}\right| B \right\rangle .$$ This leads to $$\begin{aligned} M_p = C_0 + 3.68\: x_T + 3.53\: x_8 + 1.24\: x_3 + 0.63\\ M_n = C_0 + 3.68\: x_T + 3.53\: x_8 - 1.24\: x_3 - 0.13.\end{aligned}$$ The quantity occurring most in the analysis below being $$M_n-M_p = -2.48\: x_3-0.76.$$ It may be noted here that using updated values of baryon masses from the Particle Data Book changes the parameters very slightly and this is neglected considering the qualitative nature of this work. After a clarification on the adjustment of $C_0$ (corresponding to an adjustment of $\Lambda_{\mathrm{QCD}}$) from the authors [@JJK] it was possible to reproduce the figures. Then the issue of further bounds from nucleosynthesis were studied. It is well known that the observed abundances of the primordial nuclei hydrogen (protons), helium (alpha particles) etc. are sensitively tied to the masses of the nucleons [@barrowtipler; @Hog00]. The slight difference in the masses of the proton and neutron are responsible for the survival of protons with the observed abundance. The (un)congenial regions of the triangle is explored further under these constraints. There are three cases that arise here: Case I: $x_3 > x_3^\oplus$ {#case-i-x_3-x_3oplus .unnumbered} -------------------------- Let us concentrate on the region on the upper right of the triangle with $x_3$ values greater than at the point labeled ‘us’ on the right hand side of the triangle. Of the two nucleons, the neutron is heavier by about 1.3 MeV. If it was just 0.8 MeV less, that would bring it below the electron capture threshold for protons, i.e. it would become energetically favourable for protons to capture electrons and become neutrons. All the protons would have been converted to neutrons in the Big Bang. The Universe would be full of neutrons and nothing else. We would not be here. In words of Barrow and Tipler [@barrowtipler] > Without electrostatic forces to support them, solid bodies would collapse rapidly into neutron stars or black holes. Thus, the coincidence that allows protons to partake in nuclear reactions in the early universe also prevents them decaying by weak interactions. It also, of course, prevents the 75% of the Universe which emerges from nucelosynthesis in the form of protons from simply decaying away into neutrons. If that were to happen no atoms would ever have formed and we would not be here to know it. \[Ref. [@barrowtipler], p. 400\] The same issue is also discussed by Hogan [@Hog00] > The $u$-$d$ mass difference in particular attracts attention because the $d$ is just heavier enough than $u$ to overcome the electromagnetic energy difference to make the proton ($uud$) lighter than the neutron ($udd$) and therefore stable. On the other hand, if it were a little heavier still, the deuteron would be unstable and it would be difficult to assemble any nuclei heavier than hydrogen. Therefore, it is necessary to have $$M_n-M_p \geq 0.5\:\mathrm{MeV}.$$ This reduces the congenial region on the upper right of the physical point to $$x_3 \leq -0.51.$$ Case II: $x_3 < x_3^\oplus$ {#case-ii-x_3-x_3oplus .unnumbered} --------------------------- Now let us move to the bottem-left side of the triangle (left of the $x_8$ axis) where $x_3$ values are smaller than that at the ‘us’-labeled point, again concentrating on the right hand side of the triangle. The key reaction by which hydrogen ‘burns’ in stars such as the sun involves the reaction $$\begin{aligned} p+p&\rightarrow& d+e^++\nu+0.42\,\mathrm{MeV}\\ e^++e^-&\rightarrow& 1\:\mathrm{MeV}\end{aligned}$$ So the total amount of energy released in this reaction is 1.42 MeV. If the neutron mass was 1.42 MeV (0.15%) more than it is, this reaction would not happen at all. It would need energy to make it go, rather than producing energy. Deuterons are a key step in burning hydrogen to helium. Without them, hydrogen would not burn, and there would be no long-lived stars and no stellar nucleosynthesis to produce the remaining elements. Therefore, it is necessary that $$M_n-M_p \leq 2.72\:\mathrm{MeV}.$$ This reduces the congenial region on the lower left of the physical point to $$x_3 \geq -1.4.$$ These two conditions, thus, significantly reduce the congenial corridor from $$-12.9\leq x_3 \leq 4.1$$ to $$-1.4\leq x_3 \leq -0.5.$$ It may be noted here that the width of this region is of the same order as the uncertainty in $x_3^\oplus$ itself due to uncertainties in the light quark masses given by $x_3^\oplus =-1.17\pm 0.43$. In fact it was a pleasant surprise to realise, rather late into our work, that Hogan [@Hog00] reached essentially similar conclusions which were expressed in terms of the up-down quark mass difference, $\delta m_{d-u}$ and Section IV of his review [@Hog00] is a recommended read for anybody interested in this issue. The approximately 1.4 + 0.8 = 2.2 MeV window of variation that we find is in agreement with the allowed region in Fig. 1 of the same [@Hog00]. Case III: Left half of the triangle {#case-iii-left-half-of-the-triangle .unnumbered} ----------------------------------- If we move to the left half of the triangle, we essentially replace the down quark with a strange quark. We know that the $s$-quark is, in some ways, like a heavy $d$-quark. In the left half of the triangle the $s$-quark is light and the $d$-quark is heavy. As if they simply interchange positions. That is why Jaffe [*e*t al.]{} [@Jaf09] seem to find a symmetric congenial region in the left of the triangle. The discussions for Cases I and II narrow it down, but do not remove it. However, let us now turn towards the coupling between $u$-$d$ and $u$-$s$. The $u$-$d$ coupling is much stronger, whereas the $u$-$s$ coupling is suppressed. This is described by the well-known Cabibbo angle $\theta_C$. Where the $u$-$d$ coupling carries a factor $cos\,\theta_C$ and the $u$-$s$ coupling carries a factor of $sin\,\theta_C$, the Cabibbo angle being about 13 degrees. This is like the present world with a much weaker weak interaction. This the case where the weak decay rate of neutrons is not strong enough to produce the primordial neutron-proton abundance ratio of 1:6. Without this we are left without enough protons, [*i.e.*]{} without enough hydrogen, which is key to both stellar burning and biological life itself. Therefore we are left with only a narrow region on the right \[Fig. \[newcong\]\]. ![\[newcong\] (color online) Fig. \[cong\] adapted by the further restrictions imposed leaving only a very narrow congenial slit in the bottom-right region of the triangle.](newcongenial.eps){width="0.8\columnwidth"} The only remaining question is probably regarding the length of this narrow region extending nearly up to the centre of the triangle. As one moves up this narrow slit towards the centre, away from the physical point, the up-down quarks become heavier keeping the down quark slightly heavier than the up. Meanwhile the strange quark becomes lighter to keep $m_T$ fixed. The physics considered here is probably not very sensitive to the strange quark mass. The increase in the up-down masses is offset by the compensating adjustment in $\Lambda_{\mathrm{QCD}}$ to keep the nucleon masses fixed. Therefore, the length of this region could probably be an artifact of the simultaneous and compensating tuning of quark masses and $\Lambda_{\mathrm{QCD}}$. This indeed has been one of the conclusions in [@Jaf09] as summarized more elegantly in [@Jen10]; as well as [@Brad11] where, reviewing the alternative universe landscapes studied by [@Agu01; @Har06; @Ada08; @Jaf09; @Jen10] it has been observed that if one is prepared to adjust another parameter in a compensating manner, it might be possible to find other regions in the parameter space that are also congenial. However, that does not remove the fine-tuning problem, as the alternative values are still finely tuned and this is inevitable to produce complexity as observed in our present universe. Here most of the alternatives are removed and the narrow region remains as a result of the compensating adjustments of $\Lambda_{\mathrm{QCD}}$. Indeed along the narrow region the sum of the two lightest quarks vary with the strange quark mass going in the opposite direction to keep $m_T$ fixed. If the effect of this could be quantified, it would probably be possible to restrict even the length of the narrow region \[See the Addendum\]. For example, as noted by Hogan [@Hog00], > ... the sum of the (up and down) quark masses controls the pion mass, so changing them alters the range of the nuclear potential and significantly changes nuclear structure and energy levels. Even a small change radically alters the history of nuclear astrophysics, for example, by eliminating critical resonances of nucleosynthesis needed to produce abundant carbon. [^2] Here it should be added that a more up-to-date view is that the strongest effect on the scalar scattering lengths and deuteron binding energy seem to be due to the “sigma-resonance” exchange (or correlated two-pion scalar-isoscalar exchange) dependence on $m_\pi$ [@Han08; @Pel10]. As mentioned at the outset, the analysis here has been limited to the case of $m_T = m_T^\oplus$. It has been noted by Jaffe [*et al.*]{} [@Jaf09] that the widths of the two major congenial bands on the bottom-left and bottom-right of the triangle are independent of $m_T$. Therefore, naturally the further exclusions for $x_3 > x_3^\oplus$, $x_3 < x_3^\oplus$ reducing the width of the band should also apply to other values of $m_T$. The exclusion of the left half of the triangle should also extend to other values of $m_T$. Therefore, in summary, it can be expected that for all values of $m_T$, after applying constraints from nucleosynthesis, there will only remain a similar very narrow congenial band at the bottom-right of the triangle. An additional comment is due here on the possibilities of universes with deuterons, sigma-hydrogen, or delta-helium playing the roles of hydrogen as listed in [@Jen10] as a summary of [@Jaf09]. The point made here does not contradict that these could be stable lightest elements. It is only pointed out that stability alone is not enough to produce and sustain nuclear chemistry in a manner familiar to us. Correct primordial abundances and conditions for sustained stellar burning provide constraints that are much more difficult to satisfy. This probably calls for a closer analysis of the other half in [@Jen10] related to possible universes without any weak interaction of [@Har06] where there indeed has been a detailed discussion of these issues pertaining to nucleosynthesis. However, that would have to be another project; whereas, this work is focused on [@Jaf09]. In summary, it can be observed that, primordial nuclear abundances and processes of stellar nucleosynthesis provide much more stringent constraints on quark masses than nuclear stability. Using these constraints it is possible to significantly reduce the congenial region in the space of light quark masses. Addendum {#addendum .unnumbered} ======== Our attention has been drawn through referee comments to studies of the bounds from nucleosynthesis [@Bed10; @Ber13], the latter appearing after the initial submission of this paper, on $\delta m_q /m_q$, where $m_q$ is the average of the light (up and down) quark mass and $\delta m_q$ is the change in $m_q$ keeping $m_u/m_d$ fixed. Coincidentally, along the length of the remaining narrow congenial region $m_u/m_d$ is approximately constant. The latest value is $\left|\delta m_q /m_q\right|<0.009$ [@Ber13]. There are other values in literature, but they are generally of the same order. Let us try to do a crude estimate of the effect of this constraint. For $m_q\approx 3.8$ MeV, $\delta m_q\approx 0.035$ MeV. From eqs. 8 and 9, we get $$M_N=\left(M_n+M_p\right)/2=C_0+3.68\: x_T+3.53\: x_8+2.5,$$ which given that $x_T$ is kept fixed, leads to $$\delta M_N=\delta C_0+3.53\: \delta x_8.$$ Now, $x_8$ as defined in eq. 2 can be re-expressed in terms of $m_q$ as $$x_8 =\frac{200\left(m_q-m_s\right)}{\sqrt{3}m_T^\oplus}.$$ If $m_T$ is kept fixed then $\delta m_s=-\delta m_q$, leading to $$\delta x_8 =\frac{400\left(\delta m_q\right)}{\sqrt{3}m_T^\oplus}=0.08,$$ where we have used $\delta m_q\approx 0.035$ MeV and $m_T^\oplus\approx 100$ MeV. The remaining congenial region is too small to show on a figure of this scale. In fact the region $x_8=x_8^\oplus\pm 0.08$ is too small to show on a plot of this scale and is also very small compared to the uncertainty in the value of $x_8^\oplus$ itself $x_8^\oplus =-59.5\pm 1.1$ due to the uncertainties in the determination of the light quark masses. However, one should remember our estimate is rather crude without appropriate consideration of the uncertainties in $m_q$. Taking these into account will increase the region, but keep it within the same order as the uncertainty in $x_8^\oplus$ itself. In short there is practically no congenial region outside $\left( x_3^\oplus ,x_8^\oplus\right)$. The authors are grateful to Robert L. Jaffe, Alejandro Jenkins and Itamar Kimchi for their kind replies to our queries and helpful suggestions. MHA and ASBT would also like to acknowledge the support of the Abdus Salam International Centre for Theoretical Physics (ICTP), Trieste, Italy through a Regular Associateship and a Junior Associateship, respectively. [^1]: If you are reading a black and white print, then green appears as lightly shaded and red as deep shaded [^2]: An interesting additional note is that, here Hogan cites Hoyle, F., D. N. F. Dunbar, W. A. Wenzel, and W. Whaling, 1953, Phys. Rev. [**92**]{}, 1095. This has been cited several times in different papers, sometimes with Phys. Rev. Lett. as the source, and a few times with the title “A state in C12 predicted from astrophysical evidence”. However, we failed to find any such article and would appreciate any information on this reference. The nearest match was D. N. F. Dunbar, R. E. Pixley, W. A. Wenzel, and W. Whaling, 1953, Phys. Rev. [**92**]{}, 649, an article on the resonance in $^{12}$C often dubbed the ‘Hoyle resonance’.

ArXiv

--- abstract: 'In this paper we consider in a bounded domain $\Omega \subset \mathbb{R}^N$ with smooth boundary an eigenvalue problem for the negative $(p,q)$-Laplacian with a Steklov type boundary condition, where $p\in (1,\infty)$, $q\in (2,\infty)$ and $p\neq q$. A full description of the set of eigenvalues of this problem is provided, thus essentially extending a recent result by Abreu and Madeira \[1\] related to the $(p,2)$-Laplacian.' address: - | Ovidius University, Faculty of Mathematics and Computer Science,\ 124 Mamaia Blvd, 900527 Constanţa, Romania - | Central European University, Department of Mathematics,\ Nador u. 9, 1051 Budapest, Hungary author: - Luminiţa Barbu - Gheorghe Moroşanu title: 'Eigenvalues of the negative $(p,q)$-Laplacian under a Steklov-like boundary condition' --- Introduction ============= In this paper we investigate the eigenvalue problem $$\label{eq:1.1} \left\{\begin{array}{l} A u:=-\Delta_p u-\Delta_q u=\lambda a(x) \mid u\mid ^{q-2}u\ \ \mbox{in} ~ \Omega,\\[1mm] \frac{\partial u}{\partial\nu_A}=\lambda b(x) \mid u\mid ^{q-2}u \hspace*{2.6cm}~ \mbox{on} ~ \partial\Omega, \end{array}\right. \tag{1.1}$$ under the following hypotheses $(H_{pq}) \ \ \ \ \ \ p\in (1, \infty),~ q\in (2, \infty),~ p\neq q$; $(H_{\Omega}) \ \ \ \ \ \ \Omega\subset \mathbb{R}^N, ~N\geq 2$, is a bounded domain with smooth boundary $\partial \Omega$; $(H_{ab}) \ \ \ \ \ \ a,b\in L^{\infty}(\Omega)$ are given nonnegative functions satisfying $$\label{eq:1.2} \int_\Omega a(x)~dx+\int_{\partial\Omega} b(x)~d\sigma >0. \tag{1.2}$$ We have denoted $$\frac{\partial u}{\partial\nu_A}:=\big(\mid \nabla u\mid ^{p-2}+\mid \nabla u\mid ^{q-2}\big)\frac{\partial u}{\partial\nu},$$ where $\nu$ is the unit outward normal to $\partial\Omega$. As usual $\Delta_p$ denotes the $p$-Laplacian, i.e., $\Delta_pu=div \, (|\nabla u|^{p-2}\nabla u)$. The operator $\Delta_p + \Delta q$, called $(p,q)$-Laplacian, occurs in quantum field theory. The solution $u$ of is understood in a weak sense, as an element of the Sobolev space $W:=W^{1,\max\{p,q\}}(\Omega)$ satisfying equation $\eqref{eq:1.1}_1$ in the sense of distributions and $\eqref{eq:1.1}_2$ in the sense of traces. Using a Green type formula (see [@CF], p. 71) we can define the concept of an eigenvalue of our problem as follows: \[def1\] $\lambda\in \mathbb{R}$ is an eigenvalue of problem if there exists $u_\lambda\in W \setminus \{0\}$ such that $$\label{eq:1.3} \begin{split} \int_\Omega \Big(\mid \nabla u_\lambda\mid ^{p-2}+&\mid \nabla u_\lambda\mid ^{q-2}\Big)\nabla u_\lambda \cdot \nabla v~dx \\ &=\lambda\Big(\int_\Omega a\mid u_\lambda\mid ^{q-2} u_\lambda v~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q-2} u_\lambda v~d\sigma\Big)~\forall~v\in W. \end{split} \tag{1.3}$$ Indeed, according to the mentioned Green type formula, $u\in W$ is a solution of if and only if it satisfies . Our goal is to determine the set of all eigenvalues of problem . Fortunately we are able to offer a complete description of this set (see Theorem 3.1 below). It is worth pointing out that this nice result is due to the fact that operator $A$ is nonhomogeneous ($p\neq q$). The homogeneous case ($p=q$) is more delicate. For example, if $p=q$, $a\equiv 1$ and $b\equiv 0$, then the eigenvalue set of the corresponding (Neumann type) problem is fully known only if $p=q=2$; otherwise, i.e. if $p=q\in (1, \infty)\setminus \{ 2\}$, then it is only known that, as a consequence of the Ljusternik-Schnirelman theory, there exists a sequence of positive eigenvalues of problem with $A=-2\Delta_p$ (see, e.g., [@Le]), but this sequence may not constitute the whole eigenvalue set. Note that the (nonhomogeneous) case $$p\in (1, \infty), \ q=2, \ p \neq q$$ has been considered recently by Abreu and Madeira in [@AM] where the reader can also find some useful historical comments. They assume weaker conditions on $a$ and $b$. In this paper we extend their result to the case $q>2$ but we restrict ourselves to functions $a \in L^{\infty}(\Omega), \ b\in L^{\infty}(\partial \Omega)$ since assuming weaker regularity for these functions leads to similar results without essential changes. Note that the case $$p\in (1,\infty), \ q\ge 2, \ p\neq q, \ a\equiv 1, \ b\equiv 0$$ has been solved in three previous papers, [@MMih], [@FMS], [@MM]. All these previous contributions are particular cases of the main result of this paper (Theorem 3.1). Preliminary results =================== Our hypotheses $H_{pq}$, $(H_{\Omega})$, $(H_{ab})$ will be assumed throughout this paper. If we choose $v=u_\lambda$ in (see Definition \[def1\]) we observe that the eigenvalues of problem cannot be negative numbers. It is also obvious that $\lambda_0=0$ is an eigenvalue of this problem and the corresponding eigenfunctions are the nonzero constant functions. So any other eigenvalue belongs to $(0,\infty)$. If we assume that $\lambda>0$ is an eigenvalue of problem and choose $v\equiv 1$ in we deduce that every eigenfunction $u_{\lambda}$ corresponding to $\lambda$ satisfies the equation $$\label{eq:2.1} \int_\Omega a\mid u_\lambda\mid ^{q-2} u_\lambda ~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q-2} u_\lambda ~d\sigma=0. \tag{2.1}$$ So all eigenfunctions corresponding to positive eigenvalues necessarily belong to the set $$\mathcal{C}:=\Big\{ u\in W;~\int_\Omega a\mid u\mid ^{q-2} u ~dx+\int_{\partial\Omega} b \mid u\mid ^{q-2} u ~d\sigma=0\Big\}.$$ This is a symmetric cone and using the Lebesgue Dominated Convergence Theorem (see also ) we can see that $\mathcal{C}$ is a weakly closed subset of $W$. In addition, $\mathcal{C}$ has nonzero elements. To show this, we first note that implis that either $|\{x\in \Omega;~a(x)> 0\}|_N>0$ or $a=0$ a.e. in $\Omega$ and $|\{x\in \partial\Omega;~b(x)> 0\}|_{N-1}>0$, where $|\cdot |_N$ and $|\cdot |_{N-1}$ denote the Lebesgue measures of the two sets. In the former case we choose $x_1, x_2\in \Omega,~ x_1\neq x_2$, $r>0$, such that $B_r(x_1)\cap B_r(x_2)=\emptyset,~B_r(x_k)\subset \Omega$, $|\{x\in B_r(x_k);~a(x)> 0\}|_N>0,\, k=1,2$, and consider the test functions $u_k: \Omega\rightarrow \mathbb{R}, \, k=1,2,$ $$u_k(x)= \left\{\begin{array}{l} e^{-\frac{1}{r^2-\mid x-x_k\mid^2}},~\mbox{if}~x\in B_r(x_k),\\[1mm] 0, \hspace*{1.8cm}\mbox{otherwise.} \end{array}\right.$$ Clearly $u_k\in W$, $k=1,2$. Denote $$\theta_k=\int_\Omega a u_k ^{q-1} ~dx+\int_{\partial\Omega} b u_k ^{q-1} ~d\sigma.$$ Obviously $\theta_k>0$, $k=1,2$. Define $\sigma_k=\theta_k^{\frac{-1}{q-1}},~~ k=1,2$. It is then easily seen that the function $v=\sigma_1 u_1-\sigma_2 u_2$ belongs to $\mathcal{C}\setminus \{0\}$. Of course, $tv\in \mathcal{C}$ for all $t\in \mathbb{R}$. A similar construction can be used in the later case, where restrictions of similar test functions to $B_r(x_k)\cap \partial \Omega$, $x_k\in \partial \Omega$, $k=1,2$, can be considered. \[remark11\] If for some $\lambda>0$ $u\in W\setminus\{0\}$ satisfies the equation $$\int_\Omega \Big(\mid \nabla u\mid ^{p}+\mid \nabla u\mid ^{q}\Big)~dx=\lambda\Big(\int_\Omega a\mid u\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u\mid ^{q} ~d\sigma\Big),$$ then $u$ cannot be a constant function (see ) and so $$\int_\Omega a\mid u\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u\mid ^{q} ~d\sigma> 0.$$ Therefore, denoting $\Gamma_1(u):=\{x\in \Omega;~a(x)u(x)\neq 0\},~\Gamma_2(u):=\{x\in \partial\Omega;~b(x)u(x)\neq 0\}$, we see that either $|\Gamma_1(u)|_N>0$ or $|\Gamma_2(u)|_{N-1}>0$. Obviously $u_{\lambda}$ corresponding to any eigenvalue $\lambda >0$ cannot be a constant function (see with $v=u_\lambda$ and ). Now, for $r>1$ define the set $$\mathcal{C}_r:=\Big\{ u\in W^{1,r}(\Omega);~\int_\Omega a\mid u\mid ^{r-2} u ~dx+\int_{\partial\Omega} b \mid u\mid ^{r-2} u ~d\sigma=0\Big\}.$$ Arguing as before, we infer that for all $r>0$ $\mathcal{C}_r$ is a symmetric, weakly closed (in $W^{1,r}(\Omega)$) cone, containing infinitely many nonzero elements. Note also that $\mathcal{C}=\mathcal{C}_q$ if $q>p$, otherwise (i.e., if $q<p$) $\mathcal{C}$ is a proper subset of $\mathcal{C}_q.$ Now let us define, $$\mathcal{C}_{1q}:=\mathcal{C}_{q}\cap\Big\{ u\in W^{1,q}(\Omega);\int_\Omega a\mid u\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u\mid ^{q} ~d\sigma=1\Big\}.$$ This set is nonempty. Indeed, let us suppose that $|\{ x\in \Omega; \, a(x)>0 \}|_N >0$ and choose $v= \sigma_1u_1 - \sigma_2u_2$ as before. We have $v\in C_q$ and $\int_{\Omega}a|v|^q \, dx >0$ so there exists a $t_* >0$ such that $$t_*^q \Big( \int_\Omega a\mid v\mid ^{q} ~dx+\int_{\partial\Omega} b \mid v\mid ^{q} ~d\sigma \Big)=1 \, .$$ Therefore $t_*v\in \mathcal{C}_{1q}$. A similar conclusion is obtained if $a=0$ a.e. in $\Omega$ but $| \{ x\in \partial \Omega; \, b(x)>0\} |_{N-1}>0$. Consider the minimization problem $$\label{eq:2.2} \underset{w\in\mathcal{C}_{1q}}{\inf }~J(w) \, , \tag{2.2}$$ where $J: W^{1,q}(\Omega)\rightarrow \mathbb{R}$ is defined by $J(w):=\int_\Omega \mid \nabla w\mid ^{q} ~dx.$ Functional $J$ is positively homogeneous of order $q$, convex and weakly lower semicontinuous. The next result states that $J$ attains its minimal value and this value is positive. \[lema1\] For each $q>1$ there exists $u^{*}\in \mathcal{C}_{1q}$ such that $J(u^{*})=\underset{w\in\mathcal{C}_{1q}}{\inf }~J(w)>0.$ It is well-known that functional $J$ is of class $C^1$ on $W^{1,q}(\Omega)$ and obviously $J$ is bounded below. Let $(u_n)\subset \mathcal{C}_{1q}$ be a minimizing sequence for $J$, i. e., $$J(u_n)\rightarrow \underset{w\in\mathcal{C}_{1q}}{\inf }~J(w):=\sigma.$$ We can prove that $(u_n)$ is bounded in $W^{1,q}(\Omega)$. Assume the contrary, that there exists a subsequence of $(u_n)$, again denoted $(u_n)$, such that $\parallel u_n\parallel_{L^q(\Omega)}\rightarrow\infty$ as $n\rightarrow\infty.$ Define $$v_n=\frac{u_n}{\parallel u_n\parallel_{L^q(\Omega)}} \ \ \ \forall~n\in \mathbb{N}\, .$$ Clearly sequence $(v_n)$ is bounded in $ W^{1,q}(\Omega)$ so there exist a $v\in W^{1,q}(\Omega)$ and a subsequence of $(v_n)$, again denoted $(v_n)$, such that $$v_n\rightharpoonup v ~\mbox{in} ~W^{1,q}(\Omega),$$ $$v_n\rightarrow v ~\mbox{in} ~L^{q}(\Omega),~v_n\rightarrow v ~\mbox{in} ~L^{q}(\partial\Omega).$$ As $\parallel v_n\parallel_{L^q(\Omega)}=1~\forall ~n\in \mathbb{N}$ we have $\parallel v\parallel_{L^q(\Omega)}=1$, and $$\int_\Omega \mid \nabla v\mid ^{q}~dx\leq \underset{n\rightarrow \infty}{\liminf}~\int_\Omega \mid \nabla v_n\mid ^{q}~dx=\underset{n\rightarrow \infty}{\liminf}\frac{1}{\parallel u_n\parallel^q_{L^q(\Omega)}}J(u_n)=0,$$ which shows that $v$ is a constant function. On the other hand, since $(v_n)\subset \mathcal{C}_{q}$ and $\mathcal{C}_{q}$ is weakly closed in $W^{1,q}(\Omega)$, we infer that $v\in \mathcal{C}_{q}$, hence $v\equiv 0$. But this contradicts the fact that $\parallel v\parallel_{L^q(\Omega)}=1$. Therefore, $(u_n)$ is indeed bounded in $W^{1,q}(\Omega)$, hence there exist $u^*\in W^{1,q}(\Omega)$ and a subsequence of $(u_n)$, which is also denoted $(u_n)$, such that $$u_n\rightharpoonup u^* ~\mbox{in} ~W^{1,q}(\Omega),$$ $$u_n\rightarrow u^* ~\mbox{in} ~L^{q}(\Omega),~u_n\rightarrow u^* ~\mbox{in} ~L^{q}(\partial\Omega).$$ By Lebesgue’s Dominated Convergence Theorem we obtain $u^*\in \mathcal{C}_{1q}$, so the weak lower semicontinuity of $J$ leads to $\sigma=J(u^*).$ In addition $J(u^*)>0$. Indeed, assuming by contradiction that $J(u^*)=0$ would imply that $u^* \equiv Const.$, which is impossible because $u^*\in \mathcal{C}_{1q}$. \[remarca\] For $p, \, q, \, \Omega$ satisfying our assumptions define $$\label{eq:2.3} \lambda_1:=\underset{w\in\mathcal{C}\setminus\{0\}}{\inf }~\frac{\int_\Omega\mid\nabla w\mid^q~dx}{\int_{\Omega}a\mid w\mid^q~dx+ \int_{\partial\Omega}b\mid w\mid^q~d\sigma}, \tag{2.3}$$ and $$\label{eq:2.4} \widetilde{\lambda}_1:=\underset{w\in\mathcal{C}\setminus\{0\}}{\inf }~\frac{\frac{1}{q}\int_\Omega\mid\nabla w\mid^q~dx+\frac{1}{p}\int_\Omega\mid\nabla w\mid^p~dx}{\frac{1}{q}\big(\int_{\Omega}a\mid w\mid^q~dx+\int_{\partial\Omega}b\mid w\mid^q~d\sigma\big)}. \tag{2.4}$$ Note that the denominators of the above fractions may equal zero for some $w$’s in $\mathcal{C}\setminus \{ 0\}$ and in such cases the corresponding numerators are obviously $\neq 0$ thus the values of those fractions are considered $\infty$ so they do not contribute to $\lambda_1$ or $\widetilde{\lambda}_1$. In fact $\lambda_1=\widetilde{\lambda}_1.$ Indeed, it is obvious that $\lambda_1\leq\widetilde{\lambda}_1$ and for the converse inequality we note that $\forall v\in \mathcal{C}\setminus\{0\}$, $t>0,$ we have $tv\in \mathcal{C}\setminus\{0\}$ and $$\begin{split} \widetilde{\lambda}_1=&\underset{w\in\mathcal{C}\setminus\{0\}}{\inf }~\frac{\frac{1}{q}\int_\Omega\mid\nabla w\mid^q~dx+ \frac{1}{p}\int_\Omega\mid\nabla w\mid^p~dx}{\frac{1}{q}\big(\int_{\Omega}a\mid w\mid^q~dx+\int_{\partial\Omega}b\mid w\mid^q~d\sigma\big)}\leq\\ &\frac{\int_\Omega\mid\nabla v\mid^q~dx}{\int_{\Omega}a\mid v\mid^q~dx+\int_{\partial\Omega}b\mid v\mid^q~d\sigma}+ t^{p-q}\frac{q\int_\Omega\mid\nabla v\mid^p~dx}{p\big(\int_{\Omega}a\mid w\mid^q~dx+\int_{\partial\Omega}b\mid v\mid^q~d\sigma\big)}. \end{split}$$ Now letting $t\rightarrow\infty$ if $q>p$, and $t\rightarrow 0$ if $q<p$, then passing to infimum for $v\in \mathcal{C}\setminus\{0\}$ we get the desired inequality. Therefore $\lambda_1$ can be expressed in two different ways (see and ). \[remark2\] As a consequence of Lemma \[lema1\] we have $\lambda_1>0$. Indeed, $$\lambda_1:=\underset{w\in\mathcal{\widetilde{C}}}{\inf }~\int_{\Omega}\mid\nabla w\mid^q~dx \, ,$$ where $\mathcal{\widetilde{C}}=\{v\in \mathcal{C}; \int_{\Omega}a\mid\nabla v\mid^q~dx+\int_{\partial\Omega}b\mid\nabla v\mid^q~d\sigma=1\}$. So $\lambda_1=J(u^*)$ for $p\leq q$ and $\lambda_1\geq J(u^*)$ if $p > q.$ Thus in both cases $\lambda_1>0.$ Now we recall a result which is known as the Lagrange multiplier rule (see, e.g., [@papa Thm. 2.2.18, p. 78]). \[lema2\] Let $X, Y$ be real Banach spaces and let $f:D\rightarrow \mathbb{R}$ be Fréchet differentiable , $g\in C^1(D,Y)$, where $D\subseteq X$ is a nonempty open set. If $v_0$ is a local minimizer of the constraint problem $${\min }~f(v), \ \ \ g(v)=0,$$ and $\mathcal{R}(g'(v_0))$ (the range of $g'(v_0)$) is closed, then there exist $\lambda^*\in \mathbb{R}$ and $y^{*}\in Y^{*}$ not both equal to zero such that $$\label{eq:2.5} \lambda^*f'(v_0)+y^{*}\circ g'(v_0)=0, \tag{2.5}$$ where $Y^{*}$ stands for the dual of $Y.$ \[remark13\] Define $$\label{eq:2.6} \parallel u\parallel_{ab}:=\parallel \nabla u\parallel_{L^p(\Omega)}+\Big(\parallel a^{1/q} u\parallel^q_{L^q(\Omega)}+ \parallel b^{1/q} u\parallel^q_{L^q(\partial\Omega)}\Big)^{1/q} \ \ \forall u\in W^{1,p}(\Omega) \tag{2.6}$$ If $p>q$ and $a, \, b$ satisfy $(H_{ab})$ then is a norm in $W^{1,q}(\Omega)$ equivalent with the usual norm of this space. This fact follows from [@DMP Proposition 3.9.55]. Indeed, the seminorm $$w(u):= \Big(\parallel a^{1/q} u\parallel^q_{L^q(\Omega)}+ \parallel b^{1/q} u\parallel^q_{L^q(\partial\Omega)}\Big)^{1/q} \ \ \forall u\in W^{1,p}(\Omega),$$ satisfies the two requirements of that proposition\ (i)  $\exists d>0$ such that $w(u)\leq d \parallel u\parallel_{W^{1,p}(\Omega)} \ \ \forall u\in W^{1,p}(\Omega),$ and\ (ii) if $u=\mbox{constant}$, then $w(u)=0$ implies $u\equiv 0.$ The main result ================ Let us state the main result of this paper: \[teorema1\] Assume that $(H_{pq})$, $(H_{\Omega})$ and $(H_{ab})$ above are fulfilled. Then the set of eigenvalues of problem is $\{0\} \cup (\lambda_1, \infty)$, where $\lambda_1$ is the positive constant defined by . We have alredy said that $\lambda_0=0$ is an eigenvalue of problem and any other eigenvalue of this problem belongs to $(0,\infty)$. Let us first prove that there is no eigenvalue of problem in $(0, \lambda_1]$. Assume by contradiction that there exists an eigenvalue $\lambda\in (0, \lambda_1]$ and let $u_\lambda\in \mathcal{C}\setminus\{0\}$ be a corresponding eigenfunction. Choosing $v=u_\lambda$ in yields $$\label{eq:3.1} \int_\Omega \big(\mid \nabla u_\lambda\mid ^{p}+\mid \nabla u_\lambda\mid ^{q}\big)~dx=\lambda\Big( \int_\Omega a\mid u_\lambda\mid ^{q}~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q} ~d\sigma \Big). \tag{3.1}$$ Note that $\int_\Omega a\mid u_\lambda\mid ^{q}~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q} ~d\sigma \neq 0$, otherwise $u_{\lambda}\equiv Const.$ (cf. ) which is impossible (see Remark \[remark11\]). On the other hand, as $u_\lambda\in \mathcal{C}\setminus\{0\}$, we derive from and $$\begin{split} \lambda\leq \lambda_1\leq &\frac{\int_\Omega\mid\nabla u_\lambda\mid^q~dx}{\int_{\Omega}a\mid u_\lambda \mid^q~dx+\int_{\partial\Omega}b\mid u_\lambda\mid^q~d\sigma}\\ &=\frac{\lambda\Big(\int_{\Omega}a\mid u_\lambda \mid^q~dx+\int_{\partial\Omega}b\mid u_\lambda\mid^q~d\sigma\Big)- \int_\Omega\mid \nabla u_\lambda\mid^p~dx}{\int_{\Omega}a\mid u_\lambda \mid^q~dx+\int_{\partial\Omega}b\mid u_\lambda\mid^q~d\sigma}\\ &= \lambda-\frac{\int_\Omega\mid\nabla u_\lambda\mid^p~dx}{\int_{\Omega}a\mid u_\lambda \mid^q~dx+ \int_{\partial\Omega}b\mid u_\lambda\mid^q~d\sigma}< \lambda, \end{split}$$ which is clearly impossible. In what follows we shall prove that every $\lambda>\lambda_1$ is an eigenvalue of problem . To this purpose we fix such a $\lambda$ and define the functional $\mathcal{J}_\lambda: W\rightarrow \mathbb{R},$ $$\mathcal{J}_\lambda(u)=\frac{1}{p}\int_\Omega \mid \nabla u\mid ^{p}~dx+\frac{1}{q}\int_\Omega\mid \nabla u\mid ^{q}~dx- \frac{\lambda}{q}\Big(\int_\Omega a\mid u\mid ^{q}~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q} ~d\sigma\Big).$$ It is easily seen that functional $\mathcal{J}_\lambda\in C^1(W\setminus\{0\};\mathbb{R})$ (even more, $\mathcal{J}_\lambda\in C^1(W;\mathbb{R})$ if $2<q<p$) and $$\begin{split} \langle\mathcal{J}'_\lambda(u),v\rangle&=\int_\Omega \mid \nabla u\mid ^{p-2}\nabla u\cdot\nabla v~dx+ \int_\Omega\mid \nabla u\mid ^{q-2}\nabla u\cdot\nabla v~dx\\ &-\lambda\Big(\int_\Omega a\mid u\mid ^{q-2}uv~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q-2}uv ~d\sigma\Big)\ \ \forall v\in W,~ u\in W\setminus\{0\}. \end{split}$$ So, according to Definition \[def1\], $\lambda>\lambda_1$ is an eigenvalue of problem if and only if there exists a critical point $u_\lambda\in W\setminus\{0\}$ of $\mathcal{J}_\lambda$, i. e. $\mathcal{J}'_\lambda(u_\lambda)=0$. We shall discuss two cases which are complementary to each other. **Case 1: $2<q<p$**. We shall prove that in this case functional $\mathcal{J}_\lambda$ is coercive on $\mathcal{C}\subset W=W^{1,p}(\Omega)$, i. e., $$\underset{\parallel u\parallel_{W^{1,p}(\Omega)}\rightarrow\infty, u\in\mathcal{C}}{\lim}\mathcal{J}_\lambda(u)=\infty.$$ To this purpose we define $T_1, T_2, T_3: \mathcal{C}\rightarrow\mathbb{R}$ as follows $$T_1(u)=\int_\Omega \mid \nabla u\mid ^{p}~dx,~~ T_2(u)=\int_\Omega\mid \nabla u\mid ^{q}~dx, ~~T_3(u)= \int_\Omega a\mid u\mid ^{q}~dx+\int_{\partial\Omega} b \mid u_\lambda\mid ^{q} ~d\sigma,$$ so $\mathcal{J}_\lambda(u)=\frac{1}{p}T_1(u)+\frac{1}{q}T_2(u)-\frac{\lambda}{q}T_3(u).$ We know from Remark \[remark13\] that the usual norm of $W^{1,p}(\Omega),$ denoted $\parallel \cdot \parallel_{W^{1,p}(\Omega)}$, is equivalent with the norm $\parallel \cdot \parallel_{ab}$ defined in . Thus $\parallel u\parallel_{W^{1,p}(\Omega)}\rightarrow\infty$ if and only if $\parallel u\parallel_{ab}=T_1(u)^{1/p}+T_3(u)^{1/q}\rightarrow\infty$. From we then have $$\label{eq:3.2} \lambda_1 T_3(u)\leq T_2(u)~~\forall~ u\in \mathcal{C}, \tag{3.2}$$ hence $$\frac{1}{p}T_1(u)+\frac{1}{q}T_2(u)\geq \frac{1}{p}\big(T_1(u)+T_2(u)\big)\geq \frac{1}{p}(1+\lambda_1)\big(T_1(u)+T_3(u)\big),$$ which implies $$\label{eq:3.3} \underset{\parallel u\parallel_{W^{1,p}(\Omega)}\rightarrow\infty, u\in\mathcal{C}}{\lim}\biggl(\frac{1}{p}T_1(u)+\frac{1}{q}T_2(u)\biggr)=\infty. \tag{3.3}$$ By Hölder’s inequality we have, $$T_2(u)\leq\mid\Omega\mid_N^{(p-q)/p} T_1(u)^{q/p}~~\forall~u\in W^{1,p}(\Omega),$$ so it follows from $$\label{eq:3.4} \underset{\parallel u\parallel_{W^{1,p}(\Omega)}\rightarrow\infty, u\in\mathcal{C}}{\lim}T_1(u)=\infty. \tag{3.4}$$ So, we obtin from and Hölder’s inequality $$\mathcal{J}_\lambda(u)\geq \frac{1}{p}T_1(u)+\frac{1}{q}T_2(u)-\frac{\lambda}{\lambda_1 q}T_2(u)\geq \frac{1}{p}T_1(u)- \frac{\lambda}{\lambda_1 q}\mid\Omega\mid^{(p-q)/p} T_1(u)^{q/p}.$$ Since $q<p$ the right-hand side of the above inequality tends to $\infty$ as $\parallel u\parallel_{W^{1,p}(\Omega)}\rightarrow\infty$ (cf. ) so $\mathcal{J}_\lambda$ is indeed coercive on $\mathcal{C}$. We note that $\mathcal{C}$ is a weakly closed subset of the reflexive Banach space $W=W^{1,p}(\Omega),$ and functional $\mathcal{J}_\lambda$ is weakly lower semicontinuous on $\mathcal{C}$ with respect to the norm of $W^{1,p}(\Omega).$ So $\mathcal{J}_\lambda$ has a global minimizer $u_*\in \mathcal{C}$, i.e., $\mathcal{J}_\lambda(u_*)=\min_{\mathcal{C}}\mathcal{J}_\lambda$ (see, e.g., ). From Remark \[remarca\] we know that $\lambda_1=\widetilde{\lambda}_1$, hence ) as $\lambda>\lambda_1=\widetilde{\lambda}_1$. Then (by ) there exists $u_{0\lambda}\in \mathcal{C}$ such that $\mathcal{J}_\lambda(u_{0\lambda})<0.$ It follows that $$\mathcal{J}_\lambda(u_*)\leq \mathcal{J}_\lambda(u_{0\lambda})<0,$$ which shows that $u_*\neq 0.$ In fact $u_*$ is a solution of the minimization problem $$\min_{v\in W} \mathcal{J}_\lambda(v),$$ under the restriction $$g(v):=\int_\Omega a\mid v\mid ^{q-2} v ~dx+\int_{\partial\Omega} b \mid v\mid ^{q-2} v ~d\sigma=0.$$ We can apply Lemma \[lema2\] with $X=W,~ D=W,~Y=\mathbb{R}, f=\mathcal{J}_\lambda,$ $g:W\rightarrow\mathbb{R}$ being the function just defined above, and $v_0=u_*,$ on the condition that $\mathcal{R}(g'(u_*))$ is a closed set. In fact we can show that $g'(u_*)$ is surjective, i.e., $\forall~~ c\in \mathbb{R}$ there exists a $w\in W$ such that $$\langle g'(u_*),w\rangle=c.$$ We seek $w$ of the form $w= u_*+\beta,~\beta \in \mathbb{\mathbb{R}} $. Thus we obtain from the above equation (using $u_*\in \mathcal{C}$) $$\beta(q-1)\Big(\int_\Omega a\mid u_*\mid ^{q-2} ~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} ~d\sigma\Big)=c,$$ which has a unique solution $\beta$ since $$\int_{\Omega} a\mid u_*{\mid}^{q-2}~dx + \int_{\partial \Omega}b \mid u_*{\mid}^{q-2} ~ d\sigma \neq 0 \, ,$$ otherwise $\mathcal{J}_{\lambda}(u_*)=p^{-1}\int_{\Omega}\mid \nabla u_*{\mid}^p +q^{-1}\int_{\Omega} \mid \nabla u_*{\mid}^q dx$ which contradicts $\mathcal{J}_{\lambda}(u_*) <0$. Thus $g'(u_*)$ is surjective, as claimed. By Lemma \[lema2\] there exist $\lambda^*, \mu\in \mathbb{R}$, not both equal to zero, such that $$\lambda^*\langle \mathcal{{J}}'_\lambda(u_*), v\rangle+\mu\langle g'(u_*), v\rangle=0,~~\forall~v\in W=W^{1,p}(\Omega),$$ or, equivalently, $$\begin{split} \lambda^*\biggl(\int_\Omega \Big(\mid \nabla u_*\mid ^{p-2}+&\mid \nabla u_*\mid ^{q-2}\Big)\nabla u_* \cdot \nabla v~dx, \\ &-\lambda\Big(\int_\Omega a\mid u_*\mid ^{q-2} u_* v~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} u_* v~d\sigma\Big)\biggr)\\ &+\mu (q-1)\Big(\int_\Omega a\mid u_*\mid ^{q-2} v~dx+\int_{\partial\Omega} b \mid u^*\mid ^{q-2} v~d\sigma\Big)=0 \ \ \forall v\in W. \end{split}$$ Choosing $v\equiv 1$ in the above equality and taking into account the fact that $u_*\in \mathcal{C}$ we get $$\mu \Big(\int_\Omega a\mid u_*\mid ^{q-2} ~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} ~d\sigma\Big)=0,$$ which implies $\mu=0$. Threfore $\lambda^*\neq 0$ and so $$\begin{split} \int_\Omega \Big(\mid \nabla u_*\mid ^{p-2}+&\mid \nabla u_*\mid ^{q-2}\Big)\nabla u_* \cdot \nabla v~dx, \\ &-\lambda\Big(\int_\Omega a\mid u_*\mid ^{q-2} u_* v~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} u_* v~d\sigma\Big)=0~\forall~v\in W, \end{split}$$ i. e., $\lambda$ is an eigenvalue of problem **Case 2: $q> 2$, $1<p<q$**. In this case $W=W^{1,q}(\Omega).$ Let $\lambda>\lambda_1$ be a fixed number. In this case we cannot expect coercivity on $W$ for functional $\mathcal{J}_\lambda$ which obviously belongs to $C^1(W\setminus\{0\};\mathbb{R}).$ We shall prove that $\mathcal{J}_\lambda$ has a critical point in $\mathcal{C}\setminus\{0\}$. To this purpose we consider a Nehari type manifold (see [@SW]): $$\begin{split} \mathcal{N}_\lambda&=\{v\in \mathcal{C}\setminus\{0\}; \langle \mathcal{J}'_\lambda(v),v\rangle=0\}\\ &=\Big\{v\in \mathcal{C}\setminus\{0\}; \int_\Omega \big(\mid \nabla v\mid ^{p}+\mid \nabla v\mid ^{q}\big)~dx =\lambda\Big(\int_\Omega a\mid v\mid ^{q} ~dx+\int_{\partial\Omega} b \mid v\mid ^{q} ~d\sigma\Big)\Big\}. \end{split}$$ It is natural to consider the restriction of $\mathcal{J}_\lambda$ to $\mathcal{N}_\lambda$ as any possible eigenfunction corresponding to $\lambda$ belongs to $\mathcal{N}_\lambda$. Note that on $\mathcal{N}_\lambda$ functional $\mathcal{J}_\lambda$ has the form $$\begin{split} \mathcal{J}_\lambda(u)&=\frac{1}{p}\int_\Omega \mid \nabla u\mid ^{p}~dx+\frac{1}{q}\int_\Omega\mid \nabla u\mid ^{q}~dx-\frac{\lambda}{q}\Big(\int_\Omega a\mid u\mid ^{q}~dx+\int_{\partial\Omega} b \mid u\mid ^{q} ~d\sigma\Big)\\ &=\frac{1}{p}\int_\Omega \mid \nabla u\mid ^{p}~dx-\frac{1}{q}\int_\Omega\mid \nabla u\mid ^{q}~dx=\frac{q-p}{qp}\int_\Omega \mid \nabla u\mid ^{p}~dx>0. \end{split}$$ We shall prove that there exists a point $u_*\in \mathcal{N}_\lambda$ where $\mathcal{J}_\lambda$ attains its minimal value, $m_\lambda:= \underset{w\in\mathcal{{N}}_\lambda}{\inf }{\mathcal{{J}}_\lambda (w) }$ and $\mathcal{J}'_\lambda(u^*)=0.$ The proof relies on essentially known and new arguments, and is divided into several steps as follows. **Step 1.** $\mathcal{{N}}_\lambda \neq \emptyset.$ Indeed, since $\lambda>\lambda_1,$ we deduce from that there exists a $v_0\in \mathcal{C}\setminus\{0\} $ such that $$\int_\Omega\mid\nabla v_0\mid^q~dx < \lambda \Big(\int_{\Omega}a\mid v_0\mid^q~dx+\int_{\partial\Omega}b\mid v_0\mid^q~d\sigma\Big).$$ We claim that for a convenient $t>0$, $tv_0\in \mathcal{{N}}_\lambda$. Since $\mathcal{{C}}$ is a cone, $tv_0\in \mathcal{{C}}$ for all $t\in \mathbb{R}$. So the condition $tv_0\in \mathcal{N}_\lambda$, $t>0$, reads $$t^p\int_\Omega \mid \nabla v_0\mid ^{p}~dx+t^q\int_\Omega\mid \nabla v_0\mid ^{q}~dx =\lambda t^q\Big(\int_\Omega a\mid v_0\mid ^{q} ~dx+\int_{\partial\Omega} b \mid v_0\mid ^{q} ~d\sigma\Big).$$ This equation can be solved for $t$, $$\label{eq:3.5} t=\Biggl(\frac{\int_\Omega \mid \nabla v_0\mid ^{p}~dx }{\lambda \big(\int_{\Omega}a\mid v_0\mid^q~dx+\int_{\partial\Omega}b\mid v_0\mid^q~d\sigma\big)-\int_\Omega\mid\nabla v_0\mid^q~dx}\Biggr)^{1/(q-p)}, \tag{3.5}$$ and hence for this $t$ we have $tv_0\in \mathcal{N}_\lambda$. **Step 2.** Every minimizing sequence $(u_n)\subset\mathcal{{N}}_\lambda$ for $\mathcal{{J}}_\lambda$ is bounded in $W=W^{1,q}(\Omega). $ Let $(u_n)\subset\mathcal{{N}}_\lambda$ be such a minimizing sequence for $\mathcal{{J}}_\lambda$. Since $u_n\in \mathcal{{N}}_\lambda$ for all $n$, we have $$\label{eq:3.6} \mathcal{J}_\lambda(u_n)=\frac{q-p}{qp}\int_\Omega \mid \nabla u_n\mid ^{p}~dx\rightarrow m_\lambda,~\mbox{as}~n\rightarrow\infty, \tag{3.6}$$ and $$\label{eq:3.7} \begin{split} 0&<\lambda\Big( \int_\Omega a\mid u_n\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u_n\mid ^{q} ~d\sigma\Big)-\int_\Omega\mid \nabla u_n\mid ^{q}~dx\\ &=\int_\Omega \mid \nabla u_n\mid ^{p}~dx\rightarrow \frac{qp}{q-p} m_\lambda,~\mbox{as}~n\rightarrow\infty. \end{split} \tag{3.7}$$ Assume by contradiction that $(u_n)$ is unbounded in $W^{1,q}(\Omega).$ Then, on a subsequence, again denoted $(u_n)$, we have $\Vert u_n\Vert_{ab} \rightarrow \infty$ (for details on $\Vert \cdot \Vert_{ab}$ see Remark \[remark13\]). It follows from that (on a subsequence) $$c_{n}:=\big(\parallel a^{1/q}u_n\parallel_{L^q(\Omega)}^q+\parallel b^{1/q} u_n\parallel_{L^q(\partial\Omega)}^q\big)^{1/q} \rightarrow \infty .$$ Denote $v_n=u_n/c_{n},~n\in\mathbb{N}.$ From we have $\int_\Omega\mid \nabla v_n\mid ^{q}~dx\leq \lambda $ for all $n$, so $(v_n)$ is bounded with respect to the norm $\parallel\cdot\parallel_{ab},$ which is equivalent to the usual norm of $W^{1,q}(\Omega)$. So there exists a $v_0\in W^{1,q}(\Omega)$ such that $v_n\rightharpoonup v_0$ in $W^{1,q}(\Omega)$ (hence also in $W^{1,p}(\Omega)$). Obviously, $v_n\rightarrow v_0$ in $L^{q}(\Omega)$ and also in $L^q(\partial\Omega).$ As $\mathcal{C}$ is weakly closed in $W$ and $(v_n)\subset \mathcal{C}$ we also have $v_0\in \mathcal{C}.$ Now, from we deduce $\int_\Omega\mid \nabla v_n\mid ^{p}~dx\rightarrow 0,$ and so $$\int_\Omega\mid \nabla v_0\mid ^{p}~dx\leq \underset{n\rightarrow\infty}{\liminf}\int_\Omega\mid \nabla v_n\mid ^{p}~dx=0.$$ Therefore $v_0$ is a constant function. In fact $v_0\equiv 0$ since $v_0\in \mathcal{C}$. It follows that $v_n\rightarrow 0$ in $L^{q}(\Omega)$ and in $L^{q}(\partial\Omega),$ which contradicts the fact that $$\parallel a^{1/q}v_n\parallel_{L^q(\Omega)}^q+\parallel b^{1/q}v_n\parallel_{L^q(\partial\Omega)}^q=1~\forall ~n\in \mathbb{N}.$$ **Step 3.** $m_\lambda:= \underset{w\in\mathcal{{N}}_\lambda}{\inf }{\mathcal{{J}}_\lambda (w) }>0.$ Assume that, on the contrary, $m_\lambda=0.$ Let $(u_n)\subset\mathcal{{N}}_\lambda$ be a minimizing sequence for $\mathcal{{J}}_\lambda.$ We have (see ) $$\label{eq:3.8} 0<\lambda\Big( \int_\Omega a\mid u_n\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u_n\mid ^{q} ~d\sigma\Big)-\int_\Omega\mid \nabla u_n\mid ^{q}~dx=\int_\Omega \mid \nabla u_n\mid ^{p}~dx\rightarrow 0~\mbox{as}~n\rightarrow\infty. \tag{3.8}$$ We know from Step 2 that $(u_n)$ is bounded in $W^{1,q}(\Omega),$ so there exists $u_0\in W^{1,q}(\Omega)$ such that, on a subsequence denoted again $(u_n)$, $u_n\rightharpoonup u_0$ in $W^{1,q}(\Omega)$ (hence also in $W^{1,p}(\Omega)$), and $u_n\rightarrow u_0$ in $L^q(\Omega)$, $u_n\rightarrow u_0$ in $L^q(\partial\Omega).$ Clearly $u_0\in \mathcal{C}$ and from we deduce that $u_0$ is a constant function, so $u_0\equiv 0.$ Summarizing, we have proved that $u_n\rightharpoonup 0$ in $W^{1,q}(\Omega)$. As in the previous step, we define $v_n=u_n/c_{n},~n\in \mathbb{N}$. Note that $c_n>0$ for all $n$ (otherwise, by all the $u_n$’s will be constant functions, which is impossible since they belong to $\mathcal{C}\setminus \{ 0\}$). By we see that $$\int_{\Omega} \mid \nabla v_n {\mid}^qdx < \lambda \ \ \ \forall n\in \mathbb{N} \, ,$$ so $(v_n)$ is bounded in $W^{1,q}(\Omega)$. As $(v_n)$ is a sequence in $\mathcal{C}$ which is weakly closed in $W^{1,q}(\Omega)$, it follows that there exists a $v_0\in \mathcal{C}$ such that, on a subsequence, $v_n \rightharpoonup v_0$ in $W^{1,q}(\Omega)$ and $v_n \rightarrow v_0$ in $L^q(\Omega)$ as well as in $L^q(\partial \Omega)$. Now, we divide by $c_n^q$ to obtain $$\int_{\Omega}\mid \nabla v_n \, {\mid}^p dx = c_n^{q-p}\big[ \lambda - \int_{\Omega} \mid \nabla v_n \, {\mid}^q dx \big] \rightarrow 0 \, .$$ Next, since $v_n \rightharpoonup v_0$ in $W^{1,q}(\Omega)$ (hence also in $W^{1,p}(\Omega)$), we infer that $$\int_{\Omega}\mid \nabla v_0 \, {\mid}^p dx \leq \underset{n\rightarrow\infty}{\liminf}\int_\Omega\mid \nabla v_n\mid ^{p}~dx=0.$$ Therefore $v_0$ is a constant function and in fact $v_0 \equiv 0$ since $v_0\in \mathcal{C}$. Thus, $v_n \rightarrow 0$ in both $L^q(\Omega)$ and $L^q(\partial \Omega)$. But this contradicts the fact that $$\parallel a^{1/q}v_n\parallel_{L^q(\Omega)}^q+\parallel b^{1/q}v_n\parallel_{L^q(\partial\Omega)}^q=1~\forall ~n\in \mathbb{N}.$$ This contradiction shows that $m_\lambda >0$. **Step 4.** There exists $u_*\in \mathcal{N}_\lambda$ such that $\mathcal{J}_\lambda(u_*)=m_\lambda.$ Let $(u_n)\subset\mathcal{{N}}_\lambda$ be a minimizing sequence: $\mathcal{{J}}_\lambda(u_n)\rightarrow m_\lambda.$ By Step 3 $(u_n)$ is bounded in $W^{1,q}(\Omega)$. Thus, on a subsequenc, $(u_n)$ converges weakly in $W^{1,q}(\Omega)$ to some $u_* \in W^{1,q}(\Omega)$ and strongly in both $L^{q}(\Omega)$ and $L^{q}(\partial\Omega)$ (to the same $u_*$). Thus, $$\label{eq:3.9} \mathcal{{J}}_\lambda(u_*)\leq \underset{n\rightarrow\infty}{\liminf} \mathcal{{J}}_\lambda(u_n)=m_\lambda. \tag{3.9}$$ As $(u_n)\subset\mathcal{{N}}_\lambda$ we have $$\label{eq:3.10} \int_\Omega \big(\mid \nabla u_n\mid ^{p}+\mid \nabla u_n\mid ^{q}\big)~dx =\lambda\big(\int_\Omega a\mid u_n\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u_n\mid ^{q} ~d\sigma\big), \tag{3.10}$$ $$\label{eq:3.11} \int_\Omega a\mid u_n\mid ^{q-2} u_n ~dx+\int_{\partial\Omega} b \mid u_n\mid ^{q-2} u_n ~d\sigma=0~\forall~\in \mathbb{N}. \tag{3.11}$$ It is easily seen that $u_*$ is not the null function. Indeed, assuming that $u_*\equiv 0$, we infer by that $(u_n)$ converges strongly to $0$ in $W^{1,q}(\Omega)$, hence also in $W^{1,p}(\Omega)$. Then will give $m_\lambda=0$ thus contradicting the statement of Step 3. Obviously $u_*\in \mathcal{{C}}\setminus \{0\}.$ Letting $n\rightarrow \infty$ in yields $$\label{eq:3.12} \int_\Omega \big(\mid \nabla u_*\mid ^{p}+\mid \nabla u_*\mid ^{q}\big)~dx \leq\lambda\big(\int_\Omega a\mid u_*\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q} ~d\sigma\big). \tag{3.12}$$ If holds with equality then we are done (cf. ). We shall prove that assuming strict inequality in leads to a contradiction. Thus, let us assume that $$\label{eq:3.13} \int_\Omega \big(\mid \nabla u_*\mid ^{p}+\mid \nabla u_*\mid ^{q}\big)~dx <\lambda\Big(\int_\Omega a\mid u_*\mid ^{q} ~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q} ~d\sigma\Big). \tag{3.13}$$ If we choose $t$ as in with $u_*$ instead of $v_0$, we have $t u_*\in \mathcal{N}_\lambda$ with $t\in (0,1)$. Next, using the form of $\mathcal{J}_\lambda$ on the Nehari manifold $\mathcal{{N}}_\lambda$, we get $$\label{eq:3.14} \mathcal{J}_\lambda(t u_*)=\frac{q-p}{qp}\int_\Omega \mid \nabla t u_*\mid ^{p}~dx=t^p\frac{q-p}{qp}\int_\Omega \mid \nabla u_*\mid ^{p}~dx. \tag{3.14}$$ In addition, $$\mathcal{J}_\lambda( u_n)=\frac{q-p}{qp}\int_\Omega \mid \nabla u_n\mid ^{p}~dx\Rightarrow m_\lambda=\underset{n\rightarrow\infty}{\lim}{\mathcal{{J}}_\lambda (u_n)}\geq\frac{q-p}{qp}\int_\Omega \mid \nabla u_*\mid ^{p}~dx.$$ Therefore, $$0<m_\lambda\leq \mathcal{J}_\lambda( t u_*) =t^p\frac{q-p}{qp}\int_\Omega \mid \nabla u_*\mid ^{p}~dx\leq t^p \underset{n\rightarrow\infty}{\lim}{\mathcal{{J}}_\lambda (u_n)}=t^p m_\lambda<m_\lambda,$$ which is impossible. **Step 5.** If $u_*\in \mathcal{N}_\lambda$ is the minimizer determined in Step 4, then $\mathcal{J}'_\lambda( u_*)=0.$ In fact $u_*$ is a solution of the minimization problem $$\min_{v\in W} \mathcal{J}_\lambda(v),$$ with the restrictions $$\label{eq:3.15} g_1(v):=\int_\Omega \big(\mid \nabla v\mid ^{p}+\mid \nabla v\mid ^{q}\big)~dx -\lambda\Big(\int_\Omega a\mid v\mid ^{q} ~dx+\int_{\partial\Omega} b \mid v\mid ^{q} ~d\sigma\Big)=0, \tag{3.15}$$ $$\label{eq:3.16} g_2(v):=\int_\Omega a\mid v\mid ^{q-2} v ~dx+\int_{\partial\Omega} b \mid v\mid ^{q-2} v ~d\sigma=0. \tag{3.16}$$ We shall use again Lemma \[lema2\], this time with $X=W, \ Y=\mathbb{R}^2, D=W\setminus \{0\}, \ f=\mathcal{J}_\lambda,$ $g=(g_1,g_2)$ where $g_1, \, g_2$ are defined above, $x_0=u_*$. All the assumptions of Lemma \[lema2\] can be checked easily, except the fact that $g'(u_*)$ has closed range. In fact we shall prove more, that $g'(u_*)$ is surjective, i.e., $\forall~~ (c_1, c_2)\in \mathbb{R}^2$ there exists a $w\in W$ such that $$\langle g_1'(u_*),w\rangle=c_1, ~~\langle g_2'(u_*),w\rangle=c_2.$$ If we choose in the above equations $w$ of the form $w=\alpha u_*+\beta,~ \alpha,\beta \in \mathbb{\mathbb{R}} $ and take into account the fact that $u_*\in \mathcal{N}_\lambda$, we obtain the following algebraic system $$\alpha(p-q)\int_\Omega \mid \nabla u_*\mid ^{p}dx=c_1,~~\beta(q-1)\Big(\int_\Omega a\mid u_*\mid ^{q-2} ~dx+ \int_{\partial\Omega} b \mid u_*\mid ^{q-2} ~d\sigma\Big)=c_2,$$ which has a unique solution $(\alpha , \beta)$ (from Remark \[remark11\] and $u_*\in \mathcal{N}_\lambda$ we see that the coefficients of $\alpha$ and $\beta$ are $\neq 0$). Thus $g'(u_*) $ is indeed surjective and so Lemma \[lema2\] is applicable to the above constraint minimization problem. Therefore there exist $\lambda^* \in \mathbb{R}$ and $(\mu_1, \mu_2)\in \mathbb{R}^2$, not both equal to zero, such that $$\lambda^*\langle \mathcal{{J}}'_\lambda(u_*), v\rangle+\mu_1\langle g'_1(u_*), v\rangle+\mu_2\langle g'_2(u_*), v\rangle=0,~~\forall~v\in W^{1,q}(\Omega),$$ or, equivalently, $$\begin{split} \lambda^*\biggl[\int_\Omega \Big(\mid \nabla u_*\mid ^{p-2}&+\mid \nabla u_*\mid ^{q-2}\Big)\nabla u_* \cdot \nabla v~dx - \lambda\Big(\int_\Omega a\mid u_*\mid ^{q-2} u_* v~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} u_* v~d\sigma\Big)\biggr]\\ &+\mu_1\biggl[p\int_\Omega \mid \nabla u_*\mid ^{p-2}\nabla u_* \cdot \nabla v~dx+q \int_\Omega\mid\nabla u_*\mid ^{q-2}\nabla u_* \cdot \nabla v~dx \\ &-q\lambda\Big(\int_\Omega a\mid u_*\mid ^{q-2} u_* v~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} u_* v~d\sigma\Big)\biggr]\\ &+\mu_2(q-1)\biggl[\int_\Omega a\mid u_*\mid ^{q-2} v~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} v~d\sigma\biggr]=0~~\forall v\in W. \end{split}$$ Testing with $v\equiv 1$ in the above equation and taking into account the fact that $u_*\in \mathcal{N}_\lambda$ we find $$\mu_2(q-1)\biggl[(\int_\Omega a\mid u_*\mid ^{q-2} ~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} ~d\sigma\bigr]=0,$$ which implies $\mu_2=0$ (since the coefficient of $\mu_2$ in the above equation $\neq 0$; see with $v=u_*$ and Remark \[remark11\]). Next, we test with $v=u_*$ and use with $v=u_*$ to obtain $$\mu_1(p-q)\int_\Omega \mid \nabla u_*\mid ^{p}~dx=0,$$ which implies $\mu_1=0$. Therefore, $\lambda^*\neq 0$, hence $$\int_\Omega \Big(\mid \nabla u_*\mid ^{p-2}+\mid \nabla u_*\mid ^{q-2}\Big)\nabla u_* \cdot \nabla v~dx =\lambda\Big(\int_\Omega a\mid u_*\mid ^{q-2} u_* v~dx+\int_{\partial\Omega} b \mid u_*\mid ^{q-2} u_* v~d\sigma\Big),$$ $\forall~v\in W$, i. e. $\lambda$ is indeed an eigenvalue of problem . This completes the proof of the theorem. Assume that $(H_{\Omega})$, $(H_{ab})$ are fulfilled and $q\ge 2$. We can show that, if in addition $1<p<q$, then $\lambda_1$ (defined in ) is in fact the first positive eigenvalue of the eigenvalue problem $$\label{eq:3.17} \left\{\begin{array}{l} -\Delta_q u=\lambda a(x) \mid u\mid ^{q-2}u\hspace*{1cm} \ \mbox{in} ~ \Omega,\\[1mm] \mid \nabla u\mid ^{q-2}\frac{\partial u}{\partial\nu}=\lambda b(x) \mid u\mid ^{q-2}u ~~~~~~ \mbox{on} ~ \partial\Omega. \end{array}\right. \tag{3.17}$$ As in the case of problem , $\lambda \in \mathbb{R}$ is called an eigenvalue of problem if there exists $u_\lambda \in W^{1,q}(\Omega) \setminus \{0\}$ such that $$\label{eq:3.18} \int_\Omega \mid \nabla u_\lambda\mid ^{q-2}\nabla u_\lambda \nabla v~dx =\lambda \Big(\int_\Omega a\mid u_\lambda\mid ^{q-2} u_\lambda v~dx+ \int_{\partial\Omega} b \mid u_\mu\mid ^{q-2} u_\mu v~d\sigma\Big)~~\forall v\in W^{1,q}(\Omega). \tag{3.18}$$ Obviously, $\lambda_0=0$ is an eigenvalue of and any other eigenvalue of this problem is positive (cf. with $v=u_\lambda$). For $q\geq 2,$ we can use Lemma \[lema2\] to show that the first positive eigenvalue of is given by $$\label{eq:3.19} \lambda_{1q}:=\underset{v\in\mathcal{C}_q\setminus\{0\}}{\inf }~\frac{\int_\Omega\mid\nabla v\mid^q~dx}{\int_{\Omega}a\mid v\mid^q~dx+ \int_{\partial\Omega}b\mid v\mid^q~d\sigma}. \tag{3.19}$$ First of all we see that there is no eigenvalue of in the interval $(0,\lambda_{1q})$. Assume the contrary, that there exists a $\lambda \in (0, \lambda_{1q})$ which is an eigenvalue and let $u_\lambda\in \mathcal{C}_q\setminus\{0\}$ be a correspunding eigenfunction. If we choose in $v=u_{\lambda}$ we get $$\label{eq:3.20} \int_\Omega \mid \nabla u_\lambda\mid ^{q}~dx=\mu\Big(\int_\Omega a\mid u_\lambda \mid ^{q}~dx+ \int_{\partial\Omega} b \mid u_\lambda \mid ^{q} ~d\sigma\Big). \tag{3.20}$$ As $u_\lambda \in \mathcal{C}_q\setminus\{0\}$, we have (see ) $$\lambda< \lambda_{1q}\leq \frac{\int_\Omega\mid\nabla u_\lambda\mid^q~dx}{\int_{\Omega}a\mid u_\lambda \mid^q~dx+\int_{\partial\Omega}b\mid u_\lambda\mid^q~d\sigma}=\lambda,$$ contradiction. Now, let us prove that $\lambda_{1q}$ is an eigenvalue of . We know from Lemma \[lema1\] that there exists $u^*\in \mathcal{C}_{1q}\setminus \{ 0\}$ such that $$\lambda_{1q}=J(u^*)=\underset{v\in\mathcal{C}_{1q}}{\min }~J(v).$$ We can prove that $J'(u^*)=0$. To this purpose we apply Lemma  \[lema2\] to problem with the constraints: $$\label{eq:3.21} h_1(v)=\int_\Omega a\mid v\mid ^{q} ~dx+\int_{\partial\Omega} b \mid v\mid ^{q} ~d\sigma-1=0, \tag{3.21}$$ $$\label{eq:3.22} h_2(v)=\int_\Omega a\mid v\mid ^{q-2} v ~dx+\int_{\partial\Omega} b \mid v\mid ^{q-2} v ~d\sigma=0. \tag{3.22}$$ Choose $X=W^{1,q}(\Omega), \ Y=\mathbb{R}^2, \ D=X, \ f=J, \ g=(h_1, h_2), \ x_0=u^*$. One can show by arguments similar to those used before that $g'(u^*)$ is surjective, so all the requirements of Lemma \[lema2\] are fulfilled. So there exist $\lambda^* \in \mathbb{R}$, $(\mu_1, \mu_2) \in \mathbb{R}^2$, not both equal to zero such that $$\label{eq:3.23} \begin{split} \lambda^*q\int_\Omega &\mid \nabla u^*\mid ^{q-2}\nabla u^* \nabla v~dx+\mu_{1}q \Big(\int_\Omega a\mid u^*\mid ^{q-2} u^* v~dx +\int_{\partial\Omega} b \mid u^*\mid ^{q-2} u^* v~d\sigma\Big) \\ &+\mu_2(q-1)\Big(\int_\Omega a\mid u^*\mid ^{q-2} v~dx+\int_{\partial\Omega} b \mid u^*\mid ^{q-2} v~d\sigma\Big)=0~~\forall v\in W^{1,q}(\Omega), \end{split} \tag{3.23}$$ Testing with $v=1$ in and observing that $u^*$ satisfies , we deduce that $\mu_2=0.$ Finally, chosing $v=u^*$ in and noting that $u^*$ satisfies we find $\lambda^* \lambda_{1q}+\mu_1=0$, where $\mu_1 \neq 0$, $\lambda^* \neq 0$. Replacing $\mu_1=-\lambda^* \lambda_{1q},~\mu_2=0$ in , we get $$\int_\Omega \mid \nabla u^*\mid ^{q-2}\nabla u^* \nabla v~dx - \lambda_{1q}\Big(\int_\Omega a\mid u^*\mid ^{q-2} u^* v~dx +\int_{\partial\Omega} b \mid u^*\mid ^{q-2} u^* v~d\sigma\Big) \ \ \forall v \in W^{1,q}(\Omega),$$ i. e., $(\lambda_{1q}, u^*)$ is an eigenpair of problem . Thus, if $q\ge 2$ and $1<p<q$ then $\lambda_1 = \lambda_{1q}$, so the eigenvalue set of problem is $\{ 0\}\cup (\lambda_{1q}, \infty)$, which is independent of $p$. If $2\le q<p$ then $\lambda_1\ge \lambda_{1q}$. J. Abreu, G. Madeira, Generalized eigenvalues of the $(p,2)$-Laplacian under a parametric boundary condition, ArXiv:1507.03299v2 \[math.AP\] 23 Mar 2016. H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equations, Springer, 2011. E. Casas, L. A. Fernández, A Green’s formula for quasilinear elliptic operators, J. Math. Anal. Appl., 142, 62-73, 1989. Z. Denkowski, S. Migórski, N. S. Papageorgiou, An Introduction to Nonlinear Analysis: Theory, Springer, New York, 2003. M. Fărcăşeanu, M. Mihăilescu, D. Stancu-Dumitru, On the set of eigen- values of some PDEs with homogeneous Neumann boundary condition, Nonlinear Analysis, 116, 19-25, 2015. A. Lê, Eigenvalue problems for p-Laplacian, Nonlinear Analysis, 64,1057-1099, 2006. L. Gasinski, N. S. Papageorgiou, Nonlinear Analysis, Chapman and Hall/CRC Taylor and Francis Group, Boca Raton, 2005. M. Mihăilescu, An eigenvalue problem possesing a continuous family of eigenvalues plus an isolated eigenvale, Commun. Pure Appl. Anal. 10, 701-708, 2011. M. Mihăilescu, G. Moroşanu, Eigenvalues of $-\triangle_p-\triangle_q$ under Neumann boundary condition, Canadian Math. Bull., 59(3), 606-616, 2016. N. S. Papageorgiou, S. Th. Kyritsi-Yiallourou, Handbook on Applied Analysis. Advances in Mechanics and Mathematics, 19, Springer, New York, 2009. M. Struwe, Variational Methods: Applications to Nonlinear Partial Differential Equations and Hamiltonian Systems, Springer, 1996. A. Szulkin, T. Weth, The Method of Nehary Manifold, Handbook of Nonconvex Analysis and Applications, Int. Press, Somerville, MA, 597-632, 2010.

ArXiv

--- abstract: 'models have achieved encouraging performance on the dialogue response generation task. However, existing -based response generation methods suffer from a *low-diversity problem*: they frequently generate generic responses, which make the conversation less interesting. In this paper, we address the low-diversity problem by investigating its connection with *model over-confidence* reflected in predicted distributions. Specifically, we first analyze the influence of the commonly used loss function, and find that the loss function prefers high-frequency tokens, which results in low-diversity responses. We then propose a loss function that improves over the loss function by incorporating a weighting mechanism conditioned on token frequency. Extensive experiments on benchmark datasets show that the loss function is able to substantially improve the diversity of existing state-of-the-art response generation methods, in terms of both automatic and human evaluations.' author: - Shaojie Jiang - Pengjie Ren - Christof Monz - Maarten de Rijke bibliography: - 'chatbot.bib' title: 'Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss' --- =1 &lt;ccs2012&gt; &lt;concept&gt; &lt;concept\_id&gt;10002951.10003317.10003331&lt;/concept\_id&gt; &lt;concept\_desc&gt;Information systems Users and interactive retrieval&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;/ccs2012&gt; Acknowledgments {#acknowledgments .unnumbered} =============== This research was supported by the China Scholarship Council, Ahold Delhaize, the Association of Universities in the Netherlands (VSNU), and the Innovation Center for Artificial Intelligence (ICAI). All content represents the opinion of the authors, which is not necessarily shared or endorsed by their respective employers and/or sponsors.

ArXiv

--- abstract: | We have performed a joint spectral and timing analysis of the outburst of GRS 1739-278 in 2014 based on Swift and INTEGRAL data. We show that during this outburst the system exhibited both intermediate states: hard and soft. Peaks of quasi-periodic oscillations (QPOs) in the frequency range 0.1-5 Hz classified as type-C QPOs have been detected from the system. Using Swift/BAT data we show that after the 2014 outburst the system passed to the regime of mini-outburst activity: apart from the three mini-outbursts mentioned in the literature, we have detected four more mini-outbursts with a comparable ($\sim$20 mCrab) flux in the hard energy band (15-50 keV). We have investigated the influence of the accretion history on the outburst characteristics: the dependence of the peak flux in the hard energy band in the low/hard state on the time interval between the current and previous peaks has been found (for the outbursts during which the system passed to the high/soft state).\ [**Keywords:**]{} X-ray novae, black holes, accretion, GRS 1739-278. address: - '1,2' - 1 - 1 - 1 - '1,2' author: - | **S. D.  Bykov** , E. V.  Filippova , I. A.  Mereminskiy , A. N.  Semena, A. A.  Lutovinov\ [*[$^1$]{}[Space Research Institute, Russian Academy of Sciences, Profsoyuznaya ul. 84/32, Moscow, 117997 Russia]{}\ [$^2$]{}[Higher School of Economics, Myasnitskaya 20, 101000 Moscow, Russia]{}*]{} title: 'Investigation of the Outburst Activity of the Black Hole Candidate GRS 1739-278' --- INTRODUCTION ============ During outbursts the transient systems with black hole candidates exhibit several characteristic states. The hardness-intensity and hardness-rms diagrams, on which the systems, as a rule, exhibit characteristic dependences, are used for their classification (Grebenev et al. 1993; Tanaka and Shibazaki 1996; Remillard and McClintock 2006; Belloni 2010; Belloni and Motta 2016). Initially, two states of such systems were detected: low/hard and high/soft (see Remillard and McClintock (2006) and references therein). According to the most popular accretion flow model, in such systems it is believed that the accretion disk in the low/hard state is truncated at a large inner radius, while the region between the disk and the compact object is filled with an optically thin hot plasma, a corona, which makes a major contribution to the source’s emission in the form of a power law with a high-energy cutoff. As the outburst develops, the inner radius of the accretion disk decreases and in the high/soft state the accretion disk makes a major contribution to the system’s emission (Grebenev et al. 1997; Gilfanov 2010). Subsequently, it was found that between these well-defined states the system could be in transition ones, whose established classification is currently absent (Remillard and McClintock 2006; Belloni and Motta 2016). In this paper we used the classification from Belloni and Motta (2016), according to which the source exhibits hard and soft intermediate states. These states are characterized by both thermal and power-law components in the source’s energy spectrum, but they do not differ greatly from the viewpoint of their spectral characteristics. Nevertheless, there are several distinctive features of these states. In the hard intermediate (and low/hard) state peaks of type C quasi-periodic oscillations (QPOs) and broadband noise at low frequencies with a fractional rms of both components of tens of percent are often observed in the source’s power spectrum. Type-B QPOs, which, as a rule, have a fractional rms of a few percent, and weak (rms &lt; 10%), frequency-dependent noise at low frequencies are detected in the soft intermediate state (Belloni and Motta 2016). The QPO frequency has inverse (type C) and direct (type B) dependences on the flux in the power-law component (Motta et al. 2011). The most popular QPO model is based on the Lense-Thirring precession of a hot corona near the compact object (Ingram et al. 2009), but there is no full physical picture for the formation of QPOs of different types. In the low/hard and hard intermediate states the systems with black hole candidates are observed in the radio band. During the transition to the soft intermediate state the system crosses the so-called “jet line”$\,$ and ceases to radiate in the radio band (Remillard and McClintock 2006; Belloni 2010). Thus, an investigation of the intermediate states is required for a more detailed study of the physical processes responsible for the formation of QPOs and jets. It is believed that during an outburst the system must pass from the low/hard to high/soft state and back, exhibiting the intermediate states and a characteristic “q”$\,$ shape on the hardness-intensity diagram (Belloni and Motta 2016). However, many sources with black hole candidates exhibit the so-called failed outbursts, when the system does not reach the high/soft state (Ferrigno et al. 2012, 2014a; Del Santo et al. 2015; Mereminskiy et al. 2017). The same system can exhibit both types of outbursts (Motta et al. 2010; Furst et al. 2015; Mereminskiy et al. 2017), with the same type of outbursts occurring at peak luminosities differing by tens of times. For example, in GRS 1739-278 the transition to the high/soft state was observed both during bright outbursts, when the peak flux reached 300 mCrab (15-50 keV), and mini-outbursts with a peak flux of 40 mCrab (15-50 keV) (Yan and Yu 2017). Different types of outbursts occur at close peak luminosities - a failed outburst at a peak flux of 30 mCrab in the 15- 50 keV energy band was detected in the same system GRS 1739-278 (Mereminskiy et al. 2017). At present, there is no full understanding of what physical conditions are necessary for the system’s transition from the low/hard to high/soft state, but it is clear that not only the change in accretion rate is responsible for this process. A hysteresis behavior is observed on the hardness-intensity diagram even within one full-fledged outburst of the source, i.e., the transition from the hard state to the soft one occurs at luminosities greater than the luminosity during the transition from the soft state to the hard one by several (or even tens of) times. The corona size (Homan et al. 2001), the accretion disk size (Smith et al. 2001), the evolution history of the inner accretion disk radius (Zdziarski et al. 2004), and the accretion disk mass (Yu et al. 2004; Yu and Yan 2009) are considered as additional parameters. Yu et al. (2007) and Wu et al. (2010) found a correlation between the luminosity at which the transition from the low/hard to high/soft state occurs and the peak luminosity in the high/soft state, based on which they put forward the idea about the influence of the accretion disk mass on the outburst evolution, and a correlation between the peak flux in the low/hard state and the time interval between the current and previous peak fluxes in the low/hard state. Based on these dependences, the authors hypothesized that the peak luminosity in the low/hard state is also determined by the mass of the accretion disk accumulated between outbursts. Yu et al. (2004, 2009) used mostly low-mass systems with neutron stars, which accounted for about 70% of the samples, to construct the relationship between the luminosity of the transition from the low/hard to high/soft state and the peak luminosity in the high/soft state. In contrast, the correlation between the peak flux in the low/hard state and the time interval between the current and previous peak fluxes in the low/hard state was found only for one system, GX 339-4 (based on eight outbursts detected from 1991 to 2006), and requires a further confirmation based on data for the outbursts from 2006 to 2018. Thus, an investigation of transient systems exhibiting several outbursts is required to develop a physical model for the outbursts of binary systems with black hole candidates, which allows the evolution of binary parameters from outburst to outburst to be measured. GRS 1739-278 belongs to such sources. [GRS 1739-278 ]{} ----------------- The X-ray source GRS 1739-278 was discovered by the SIGMA coded-mask telescope onboard the GRANAT observatory on March 18, 1996 (Paul et al. 1996). The peak flux from the source during the first recorded outburst was $\sim 800 - 1000$ mCrab in the 2-10 keV energy band (Borozdin et al. 1998). The corresponding radio source was detected by Durouchoux et al. (1996). During the 1996 outburst the system passed from the low/hard to high/soft state (Borozdin et al. 1998). QPO peaks at 5 Hz were detected in the source (Borozdin and Trudolyubov 2000). From the shape of the power spectrum and the total fractional rms it can be concluded that the system was recorded in the soft intermediate state and the recorded QPOs are type-B ones. The second outburst was recorded by the Swift/BAT monitor on March 9, 2014 (Krimm et al. 2014). A spectral analysis of this outburst based on Swift/XRT data was performed by Yu and Yan (2017) and Wang et al. (2018). They showed that during the outburst the system passed from the low/hard to high/soft state through the intermediate one, but no detailed analysis of the intermediate state was made. At the outburst onset (March 19, 2014) the system was also recorded by the INTEGRAL observatory. According to the analysis of these data, the source was recorded up to energies of $\sim 200$ keV, while the energy spectrum was fitted by a power law with a cutoff with the following parameters: a photon index $\Gamma=1.4\pm 2$ and a cutoff energy $E_{cut}=90^{+40}_{-20}$ keV (Filippova et al. 2014b). The source was observed by the NuSTAR observatory (Miller et al. 2015) 17 days after the outburst onset (March 26, 2014). A spectral analysis of the NuSTAR data showed that the system continued to be in the low/hard state. The reflected component of the hard emission presumably associated with an accretion disk whose inner radius must reach the innermost stable orbit was observed in the source’s spectrum, but the disk component itself was not recorded in the spectrum. Mereminskiy et al. (2019) performed a detailed timing analysis of the NuSTAR data, based on which QPOs at frequencies 0.3-0.7 Hz were detected in the variability power spectrum for GRS 1739-278. Based on data from the MAXI monitor, Wang et al. (2018) showed that the system passed back from the high/soft to low/hard state in November-December 2014. Two mini-outbursts (the peak flux in the 15-50 keV energy band was $\sim 40$ mCrab) were detected from the system 200 days after this outburst. A spectral analysis of the Swift/XRT data showed that during these mini-outbursts the system passed from the low/hard to high/soft state and returned back to the low/hard one (Yu and Yan 2017). The next outburst from the source was detected in September 2016 (Mereminskiy et al. 2016). During this outburst the peak flux in the 20-60 keV energy band was $\sim 30$ mCrab. Based on our analysis of the INTEGRAL and Swift data, it was shown that during the outburst the system was in the low/hard state and exhibited no transition to the high state. i.e., the outburst turned out to be a failed one (Mereminskiy et al. 2017). Some softening of the spectrum was recorded, i.e., the photon index increased from 1.73 at the outburst onset to 1.86 at the peak flux, with the power law having been observed up to energies $\sim 150$ keV without cutoffs; no contribution to the emission from the accretion disk was recorded. In this paper for the first time we have performed a simultaneous study of the spectral evolution and temporal variability of the system over the entire 2014 outburst, which has allowed the source’s intermediate states to be classified in detail, made a comparative analysis of the system’s behavior in all of the outbursts detected to date, and investigated the system’s behavior between outbursts. OBSERVATIONS ============ Swift Data ---------- GRS 1739-278 during the 2014 outburst was observed by the Swift/XRT telescope in the windowed timing mode (Burrows et al. 2005) from March 20, 2014 (MJD 56736) to November 1, 2014 (MJD 56962), i.e., the observations were begun on the 11th day after the outburst onset. A total of 39 observations with ObsID 000332030XY were carried out; below we will use the last two digits to denote the observation number. The Swift/XRT light curves and spectra of the source were obtained using an online repository (Evans et al. 2007). Events with grade 0 were selected to analyze the energy spectra. The source’s spectra were investigated in the 0.8-10 keV energy band, because at energies below 0.8 keV the response matrix is known inaccurately due to instrumental effects. The spectra obtained through the online repository by adding the counts in different bins were brought to the form in which there were at least 100 counts per energy channel with the grppha utility. This allowed the $\chi^2$ statistic to be used in fitting the spectra in the Xspec package (Arnaud 1996). Instrumental features at energies of 1.8 and 2.3 keV were detected in several spectra (marked in Table 1 by $\dagger$)[^1]. Therefore, when fitting these data, we, following the recommendations given at the mentioned link, used the gain command in the Xspec package, which modifies the response matrix by shifting the energies at which it was determined. The offset parameters are given in Table 2. Deviations of the data from the model at energies below 1 keV were also observed in several spectra (marked in Table 1 by $\star$). This feature of the spectrum is related to the position of the source’s image on the telescope’s detector[^2]. Using the telescope’s response matrix dependent on the source’s position on the detector (swxwt0s6psf1\_20131212v001.rmf) allowed the quality of the data fit at low energies to be improved. To construct the light curves, we used the 0.5-10 keV energy band, events with grade 0-2, the data were averaged over one observation. The typical exposure time for XRT observations was $\sim 1-2$ ks (see Table 1). The light curves used for our Fourier analysis were constructed in the 0.5-10, 0.5-3, and 3-10 keV energy bands with a time resolution of 0.01 s. The Swift/BAT light curve of the source was retrieved from the online database of light curves (Krimm et al. 2013). INTEGRAL Data ------------- We also used the data from the JEM-X and ISGRI/IBIS telescopes onboard the INTEGRAL observatory (Winkler et al. 2003) processed with the HEAVENS service (Walter et al. 2010). These data in combination with the quasi-simultaneous Swift/XRT spectral data were used to construct the source’s broadband spectra in the energy range 0.8-200 keV. When fitting the spectra, we took into account the systematic errors of 1 and 3% for the ISGRI[^3] and JEM - X[^4] instruments, respectively. The times and exposures of observations for the broadband spectra are given in Table 3. [ccc|cc|cc|c|c|c]{} $ID^a$ & ------------ MJD- -56000$^b$ ------------ & -------------- Exp. XRT$^c$ s -------------- & -------------------- $N_H^d$, $10^{22}$cm$^{-2}$ -------------------- & $\Gamma^e$ & ----------- $T_{in}$, keV $^f$ ----------- & ------------------------- $R_{in} cos^{-1/2}(i)$, km $^g$ ------------------------- & Flux$^h$ & $f_{po}^j$ & $\chi_N^2/dof$\ $01$ & 736 & $79$ & $1.51_{-0.63}^{+0.67}$ & $1.24_{-0.36}^{+0.37}$ & - & - & $1.57_{-0.29}^{+0.13}$ & 100 & $0.31/7$\ $02$ & $741$ & $2067$ & $1.64\pm0.04$ & $1.38\pm0.03$ & - & - & $3.49\pm0.04$ & 100 & $1.09/459$\ & & & $1.65\pm0.03$ & $1.38\pm0.02$ & - & - & $4.53\pm0.04$ & 100 & $1.06/496$\ & & & $2.08_{-0.1}^{+0.11}$ & $1.5$ & $0.24_{-0.03}^{+0.02}$ & $196.05_{-52.37}^{+70.17}$ & $4.44_{-0.04}^{+0.03}$ & $89\pm5$ & $1.11/495$\ & & & $2.96\pm0.09$ & $1.8$ & $0.18\pm0.01$ & $1378.65_{-198.48}^{+235.78}$ & $4.2\pm0.03$ & $51\pm4$ & $1.64/495$\ & & & $1.82_{-0.06}^{+0.08}$ & $2.05_{-0.04}^{+0.08}$ & - & - & $7.29_{-0.08}^{+0.07}$ & $100$ & $1.3/598$\ & & & $1.52_{-0.02}^{+0.08}$ & $1.5$ & $0.79_{-0.06}^{+0.04}$ & $15.7_{-0.73}^{+3.94}$ & $7.87_{-0.11}^{+0.06}$ & $82\pm4$ & $1.28/597$\ & & & $1.74_{-0.06}^{+0.08}$ & $1.8$ & $0.65_{-0.09}^{+0.05}$ & $19.22_{-3.96}^{+6.33}$ & $7.62_{-0.1}^{+0.03}$ & $90\pm5$ & $1.27/597$\ & & & $1.9_{-0.07}^{+0.09}$ & $2$ & $0.51_{-0.06}^{+0.05}$ & $25.76_{-10.15}^{+12.89}$ & $7.41_{-0.23}^{+0.03}$ & $94\pm4$ & $1.28/597$\ & & & $1.57_{-0.07}^{+0.08}$ & $1.88\pm0.05$ & - & - & $6.76_{-0.05}^{+0.06}$ & 100 & $1.17/552$\ & & & $1.37\pm0.06$ & $1.5$ & $0.76_{-0.06}^{+0.1}$ & $12.72_{-1.51}^{+1.98}$ & $7.17_{-0.14}^{+0.05}$ & $88\pm5$ & $1.18/551$\ & & & $1.62_{-0.1}^{+0.07}$ & $1.8$ & $0.55_{-0.1}^{+0.15}$ & $18.49_{-6.13}^{+13.25}$ & $6.87_{-0.2}^{+0.03}$ & $96\pm4$ & $1.17/551$\ & & & $1.85_{-0.07}^{+0.1}$ & $2$ & $0.26_{-0.03}^{+0.1}$ & $143.63_{-71.44}^{+135.22}$ & $6.62_{-0.08}^{+0.04}$ & $93\pm7$ & $1.18/551$\ & & & $2.26\pm0.03$ & $2.25\pm0.02$ & - & - & $10.67_{-0.08}^{+0.07}$ & 100 & $1.43/496$\ & & &$1.74\pm0.03$ & $1.5$ & $1.12\pm0.04$ & $14.24_{-0.85}^{+0.94}$ & $10.56_{-0.11}^{+0.07}$ & $53\pm4$ & $1.28/495$\ & & &$1.85\pm0.03$ & $1.8$ & $1.13\pm0.05$ & $12.09_{-0.84}^{+0.97}$ & $10.51\pm0.1$ & $66\pm6$ & $1.25/495$\ & & & $1.95\pm0.03$ & $2$ & $1.21_{-0.07}^{+0.06}$ & $9.21_{-0.68}^{+0.81}$ & $10.46_{-0.10}^{+0.09}$ & $75\pm8$ & $1.25/495$\ & & & $2.15\pm0.03$ & $2.11\pm0.02$ & - & - & $8.35_{-0.05}^{+0.06}$ & & $1.53/514$\ & & & $2.15_{-0.04}^{+0.05}$ & $1.5$ & $0.61\pm0.02$ & $41.96_{-4.21}^{+4.7}$ & $8.9_{-0.1}^{+0.05}$ & $70\pm2$ & $1.13/513$\ & & & $2.31_{-0.05}^{+0.06}$ & $1.8$ & $0.49\pm0.02$ & $62.57_{-9.27}^{+10.74}$ & $8.68_{-0.08}^{+0.05}$ & $78\pm3$ & $1.1/513$\ & & & $2.43\pm0.07$ & $2$ & $0.4_{-0.02}^{+0.03}$ & $97.64_{-20.61}^{+24.37}$ & $8.47_{-0.1}^{+0.03}$ & $82\pm4$ & $1.21/513$\ & & & $2.27\pm0.05$ & $2.28\pm0.03$ & - & - & $10.11_{-0.10}^{+0.11}$ & 1 00 & $1.36/366$\ & & & $1.69\pm0.04$ & $1.5$ & $1.17\pm0.06$ & $13.15_{-1.07}^{+1.23}$ & $9.91_{-0.2}^{+0.11}$ & $48\pm7$ & $1.16/365$\ & & & $1.79\pm0.05$ & $1.8$ & $1.2\pm0.07$ & $11.27_{-0.94}^{+1.15}$ & $9.86_{-0.21}^{+0.16}$ & $60\pm10$ & $1.16/365$\ & & & $1.89\pm0.05$ & $2$ & $1.28_{-0.09}^{+0.08}$ & $9.13_{-0.73}^{+0.89}$ & $9.80_{-0.14}^{+0.17}$ & $67\pm12$ & $1.16/365$\ & & & $1.75_{-0.11}^{+0.1}$ & $2.08\pm0.08$ & - & - & $6.14_{-0.11}^{+0.07}$ & 100 & $1.14/504$\ & & & $1.4_{-0.09}^{+0.07}$ & $1.5$ & $0.84_{-0.07}^{+0.12}$ & $12.87_{-2.89}^{+2.9}$ & $6.57_{-0.14}^{+0.09}$ & $81\pm7$ & $1.14/503$\ & & & $1.58_{-0.09}^{+0.1}$ & $1.8$ & $0.73_{-0.1}^{+0.16}$ & $12.53_{-4.37}^{+6.32}$ & $6.39_{-0.18}^{+0.07}$ & $90\pm8$ & $1.14/503$\ & & & $1.74_{-0.08}^{+0.1}$ & $2$ & $0.63_{-0.15}^{+0.29}$ & $11.57_{-8.42}^{+13.53}$ & $6.21_{-0.19}^{+0.04}$ & $96\pm7$ & $1.14/503$\ & & & $2.73\pm0.04$ & $2.22\pm0.02$ & - & - & $16.94_{-0.12}^{+0.1}$ & 100 & $1.37/525$\ & & & $2.32\pm0.04$ & $2$ & $1.42\pm0.06$ & $9.53_{-0.52}^{+0.57}$ & $16.45_{-0.14}^{+0.16}$ & $68\pm9$ & $1.09/524$\ & & & $2.51_{-0.06}^{+0.05}$ & $2.4$ & $1.85_{-0.05}^{+0.06}$ & $6.05_{-0.6}^{+0.58}$ & $16.26_{-0.26}^{+0.1}$ & $66\pm9$ & $1.12/524$\ & & & $1.8\pm0.02$ & - & $1.72\pm0.02$ & $10.09_{-0.23}^{+0.24}$ & $15.53_{-0.13}^{+0.09}$ & 0 & $1.64/525$\ & & & $2.37_{-0.09}^{+0.13}$ & $1.72_{-0.03}^{+0.08}$ & - & - & $27.3_{-0.41}^{+0.34}$ & 100 & $1.74/525$\ & & & $1.92_{-0.05}^{+0.09}$ & $2$ & $1.45_{-0.06}^{+0.03}$ & $15.95_{-1.49}^{+1.6}$ & $19.76_{-0.3}^{+0.22}$ & $11\pm7$ & $1.11/524$\ & & &$2.02_{-0.18}^{+0.11}$ & $2.4$ & $1.46_{-0.08}^{+0.07}$ & $16.07_{-1.91}^{+1.39}$ & $19.63_{-0.32}^{+0.27}$ & $11\pm8$ & $1.11/524$\ & & & $1.81_{-0.06}^{+0.12}$ & - & $1.49_{-0.07}^{+0.05}$ & $15.58_{-1.24}^{+1.87}$ & $19.55_{-0.76}^{+0.18}$ & 0 & $1.12/525$\ & & & $2.38\pm0.03$ & $2.3\pm0.02$ & - & - & $8.64_{-0.05}^{+0.06}$ & 100 & $1.18/495$\ & & & $2.04\pm0.03$ & $2$ & $1.14\pm0.06$ & $9.73_{-0.72}^{+0.86}$ & $8.49\pm0.08$ & $74\pm6$ & $0.99/494$\ & & & $2.25\pm0.04$ & $2.4$ & $1.70_{-0.06}^{+0.08}$ & $4.2_{-0.58}^{+0.54}$ & $8.36_{-0.13}^{+0.04}$ & $77\pm9$ & $1.04/494$\ & & & $1.53\pm0.02$ & - & $1.58\pm0.02$ & $8.4_{-0.19}^{+0.2}$ & $7.83_{-0.06}^{+0.05}$ & 0 & $1.97/495$ [ccc|cc|cc|c|c|c]{} $ID^a$ & ------------ MJD- -56000$^b$ ------------ & -------------- Exp. XRT$^c$ s -------------- & -------------------- $N_H^d$, $10^{22}$cm$^{-2}$ -------------------- & $\Gamma^e$ & ----------- $T_{in}$, keV $^f$ ----------- & ------------------------- $R_{in} cos^{-1/2}(i)$, km $^g$ ------------------------- & Flux$^h$ & $f_{po}^j$ & $\chi_N^2/dof$\ & & & $2.34\pm0.03$ & $2.34\pm0.02$ & - & - & $7.67_{-0.07}^{+0.05}$ & 100 & $1.16/497$\ & & & $1.96\pm0.03$ & $2$ & $1.13\pm0.05$ & $9.76_{-0.66}^{+0.77}$ & $7.55_{-0.08}^{+0.07}$ & $71\pm6$ & $0.99/496$\ & & & $2.12\pm0.04$ & $2.4$ & $1.58\pm0.05$ & $4.88^{+0.47}_{-0.50}$ & $7.37^{+0.04}_{-0.08}$ & $78\pm9$ & $1.02/496$\ & & & $1.48\pm0.02$ & - & $1.55\pm0.02$ & $8.27\pm0.19$ & $6.96_{-0.07}^{+0.04}$ & 0 & $1.99/497$\ & & & $2.85_{-0.07}^{+0.09}$ & $2.27_{-0.03}^{+0.11}$ & - & - & $9.67_{-0.09}^{+0.07}$ & 100 & $1.43/542$\ & & & $2.09_{-0.08}^{+0.07}$ & $2$ & $1.25\pm0.05$ & $11.74_{-1.62}^{+0.64}$ & $8.31_{-0.16}^{+0.07}$ & $43\pm10$ & $1.09/541$\ & & & $2.34\pm0.09$ & $2.4$ & $1.28\pm0.07$ & $10.65_{-0.99}^{+1.59}$ & $8.17_{-0.14}^{+0.08}$ & $54\pm10$ & $1.09/541$\ & & & $1.59_{-0.01}^{+0.09}$ & - & $1.42_{-0.04}^{+0.06}$ & $10.68_{-0.82}^{+0.79}$ & $7.66\pm0.07$ & 0 & $1.33/542$\ & & & $2.46\pm0.03$ & $2.39\pm0.02$ & - & - & $8.68_{-0.06}^{+0.07}$ & 100 & $1.34/472$\ & & & $2.01\pm0.03$ & $2$ & $1.16\pm0.05$ & $10.86_{-0.65}^{+0.75}$ & $8.48_{-0.10}^{+0.08}$ & $65\pm6$ & $1.06/471$\ & & & $2.22\pm0.05$ & $2.4$ & $1.48_{-0.04}^{+0.05}$ & $6.06_{-0.56}^{+0.53}$ & $8.36_{-0.12}^{+0.04}$ & $73\pm10$ & $1.08/471$\ & & & $1.57\pm0.02$ & - & $1.52\pm0.02$ & $9.31_{-0.22}^{+0.23}$ & $7.87_{-0.06}^{+0.05}$ & 0 & $1.81/472$\ & & & $2.29\pm0.03$ & $2.35\pm0.02$ & - & - & $6.74_{-0.05}^{+0.04}$ & 100 & $1.12/524$\ & & & $2.01\pm0.03$ & $2$ & $0.92\pm0.05$ & $12.82_{-1.13}^{+1.33}$ & $6.76_{-0.06}^{+0.04}$ & $78\pm4$ & $1.05/523$\ & & & $2.25\pm0.03$ & $2.4$ & $1.72_{-0.14}^{+0.24}$ & $2.28_{-0.78}^{+0.69}$ & $6.65_{-0.16}^{+0.03}$ & $92\pm8$ & $1.11/523$\ & & & $1.43$ & - & $1.52$ & $8.08$ & $6.05$ & 0 & $3.08/524$\ & & & $3.13_{-0.39}^{+0.15}$ & $2.29_{-0.03}^{+0.05}$ & - & - & $10.6\pm0.12$ & 100 & $1.77/504$\ & & & $2.0_{-0.11}^{+0.07}$ & $2$ & $1.25_{-0.04}^{+0.08}$ & $13.28_{-1.55}^{+1.06}$ & $8.69_{-0.26}^{+0.16}$ & $29\pm7$ & $1.11/503$\ & & & $2.19_{-0.12}^{+0.07}$ & $2.4$ & $1.26_{-0.05}^{+0.06}$ & $12.72_{-1.26}^{+1.28}$ & $8.61_{-0.13}^{+0.08}$ & $37\pm13$ & $1.12/503$\ & & & $1.66_{-0.02}^{+0.09}$ & - & $1.42_{-0.07}^{+0.04}$ & $11.31_{-0.67}^{+1.27}$ & $8.37_{-0.32}^{+0.04}$ & 0 & $1.23/504$\ & & & $2.0\pm0.03$ & $2$ & $1.29\pm0.03$ & $9.94_{-0.3}^{+0.32}$ & $6.7_{-0.04}^{+0.06}$ & $41\pm5$ & $1.07/519$\ &&& $2.15\pm0.05$ & $2.4$ & $1.4\pm0.02$ & $8.23\pm0.28$ & $6.64_{-0.06}^{+0.05}$ & $47\pm7$ & $1.12/519$\ &&& $1.72\pm0.02$ & - & $1.46\pm0.01$ & $9.22_{-0.17}^{+0.18}$ & $6.43_{-0.04}^{+0.03}$ & 0 & $1.43/520$\ & & & $2.65_{-0.11}^{+0.03}$ & $2.58_{-0.08}^{+0.07}$ & - & - & $6.76_{-0.07}^{+0.05}$ & 100 & $1.22/503$\ & & & $1.82_{-0.03}^{+0.07}$ & $2$ & $1.08_{-0.06}^{+0.04}$ & $11.3_{-0.95}^{+1.89}$ & $6.52_{-0.14}^{+0.06}$& $64\pm11$& $1.11/502$\ & & & $2.14_{-0.03}^{+0.1}$ & $2.4$ & $1.12\pm0.07$ & $9.49_{-1.49}^{+1.39}$ & $6.35_{-0.09}^{+0.07}$&$76\pm12$ & $1.09/502$\ & & & $1.11_{-0.06}^{+0.04}$ & - & $1.48_{-0.02}^{+0.07}$ & $8.13_{-0.74}^{+0.57}$ & $5.8_{-0.15}^{+0.05}$ & 0 & $1.75/503$\ & & & $2.95_{-0.39}^{+0.29}$ & $2.65\pm0.19$ & - & - & $7.69_{-0.16}^{+0.1}$ & 100 & $1.33/364$\ & & & $2.04_{-0.05}^{+0.1}$ & $2$ & $1.12_{-0.06}^{+0.08}$ & $14.23_{-0.96}^{+2.52}$ & $6.75_{-0.21}^{+0.12}$ & $35\pm10$ & $1.07/363$\ & & & $2.26_{-0.15}^{+0.11}$ & $2.4$ & $1.13_{-0.07}^{+0.09}$ & $13.1_{-2.17}^{+2.36}$ & $6.67_{-0.18}^{+0.16}$ & $46\pm14$ & $1.06/363$\ & & & $1.64\pm0.07$ & - & $1.31\pm0.07$ & $12.06_{-1.75}^{+1.32}$ & $6.23_{-0.13}^{+0.07}$ & 0 & $1.21/364$\ & & & $2.09_{-0.08}^{+0.07}$ & $2$ & $0.95\pm0.03$ & $21.91_{-1.86}^{+0.98}$ & $5.84_{-0.08}^{+0.07}$ & $23\pm4$ & $1.14/471$\ & & & $2.23_{-0.08}^{+0.04}$ & $2.4$ & $0.94_{-0.02}^{+0.03}$ & $21.92_{-1.91}^{+2.16}$ & $5.81_{-0.11}^{+0.04}$ & $32\pm6$ & $1.14/471$\ & & & $1.74_{-0.06}^{+0.1}$ & - & $1.1_{-0.05}^{+0.03}$ & $17.26_{-1.52}^{+1.58}$ & $5.49_{-0.14}^{+0.04}$ & 0 & $1.45/472$\ & & & $2.08\pm0.02$ & $2$ & $1.11\pm0.02$ & $17.17_{-0.42}^{+0.44}$ & $7.76_{-0.05}^{+0.04}$ & $26\pm3$ & $1.22/525$\ & & & $2.21\pm0.03$ & $2.4$ & $1.13\pm0.02$ & $15.8_{-0.34}^{+0.35}$ & $7.73\pm0.04$ & $37\pm4$ & $1.22/525$\ & & & $1.9\pm0.02$ & - & $1.23\pm0.01$ & $15.06\pm0.23$ & $7.48\pm0.03$ & 0 & $1.65/526$\ & & & $2.06\pm0.04$ & $2$ & $1.16\pm0.02$ & $16.16_{-0.50}^{+0.54}$ & $6.99\pm0.06$ & $12\pm5$ & $0.98/449$\ & & & $2.12_{-0.06}^{+0.05}$ & $2.4$ & $1.17\pm0.02$ & $15.68_{-0.39}^{+0.42}$ & $6.98\pm0.06$ & $16\pm7$ & $0.98/449$\ & & & $1.98\pm0.02$ & - & $1.21\pm0.01$ & $15.34_{-0.33}^{+0.34}$ & $6.88_{-0.04}^{+0.03}$ & 0 & $1.01/450$\ & & & $2.1\pm0.03$ & $2$ & $1.11\pm0.02$ & $18.36_{-0.51}^{+0.54}$ & $7.53_{-0.08}^{+0.04}$ & $14\pm3$ & $0.88/470$\ & & & $2.17\pm0.04$ & $2.4$ & $1.12\pm0.02$ & $17.71_{-0.42}^{+0.44}$ & $7.51_{-0.06}^{+0.05}$ & $21\pm5$ & $0.88/470$\ & & & $1.99\pm0.02$ & - & $1.17\pm0.01$ & $16.99_{-0.31}^{+0.32}$ & $7.37_{-0.04}^{+0.03}$ & 0 & $0.98/471$ [ccc|cc|cc|c|c|c]{} $ID^a$ & ------------ MJD- -56000$^b$ ------------ & -------------- Exp. XRT$^c$ s -------------- & -------------------- $N_H^d$, $10^{22}$cm$^{-2}$ -------------------- & $\Gamma^e$ & ----------- $T_{in}$, keV $^f$ ----------- & ------------------------- $R_{in} cos^{-1/2}(i)$, km $^g$ ------------------------- & Flux$^h$ & $f_{po}^j$ & $\chi_N^2/dof$\ & & & $1.88\pm0.03$ & $2$ & $1.17\pm0.02$ & $16.78_{-0.42}^{+0.45}$ & $7.46_{-0.05}^{+0.06}$ & $3\pm2$ & $1.02/471$\ & & & $1.89_{-0.04}^{+0.05}$ & $2.4$ & $1.17\pm0.02$ & $16.62_{-0.34}^{+0.35}$ & $7.45_{-0.05}^{+0.07}$ & $4\pm2$ & $1.02/471$\ & & & $1.85\pm0.02$ & - & $1.18\pm0.01$ & $16.53\pm0.31$ & $7.43\pm0.05$ & 0 & $1.02/472$\ & & & $2.05\pm0.03$ & $2$ & $1.12\pm0.02$ & $16.69_{-0.40}^{+0.41}$ & $6.54_{-0.04}^{+0.03}$ & $12\pm3$ & $1.07/499$\ & & & $2.1\pm0.04$ & $2.4$ & $1.13\pm0.01$ & $16.20_{-0.33}^{+0.34}$ & $6.53_{-0.04}^{+0.03}$ & $17\pm4$ & $1.07/499$\ & & & $1.96\pm0.02$ & - & $1.18\pm0.01$ & $15.67_{-0.26}^{+0.27}$ & $6.44_{-0.03}^{+0.04}$ & 0 & $1.15/500$\ & & & $1.98\pm0.04$ & $2$ & $1.1\pm0.02$ & $17.46_{-0.58}^{+0.62}$ & $6.24_{-0.05}^{+0.06}$ & $9\pm4$ & $1.12/417$\ && & $2.02\pm0.05$ & $2.4$ & $1.1\pm0.02$ & $17.08_{-0.49}^{+0.52}$ & $6.23_{-0.07}^{+0.06}$ & $13\pm6$ & $1.12/417$\ &&& $1.91\pm0.03$ & - & $1.14\pm0.01$ & $16.56_{-0.38}^{+0.39}$ & $6.15_{-0.06}^{+0.04}$ & 0 & $1.15/418$\ & & & $1.98\pm0.05$ & $2$ & $1.05\pm0.03$ & $17.87_{-0.79}^{+0.86}$ & $5.21_{-0.07}^{+0.05}$ & $7\pm5$ & $1.0/342$\ & & & $2.02\pm0.07$ & $2.4$ & $1.06\pm0.03$ & $17.59_{-0.67}^{+0.73}$ & $5.21_{-0.07}^{+0.06}$ & $10\pm6$ & $1.0/342$\ & & & $1.93\pm0.03$ & - & $1.09\pm0.01$ & $17.08_{-0.51}^{+0.53}$ & $5.15\pm0.04$ & 0 & $1.01/343$\ & & & $2.23_{-0.15}^{+0.13}$ & $2$ & $0.94_{-0.05}^{+0.06}$ & $22.90_{-3.34}^{+3.17}$ & $4.69_{-0.18}^{+0.00}$ & $6\pm3$ & $1.08/373$\ & & & $2.25_{-0.13}^{+0.15}$ & $2.4$ & $0.95_{-0.03}^{+0.05}$ & $22.19_{-2.54}^{+3.70}$ & $4.68_{-0.30}^{+0.05}$ & $10\pm5$ & $1.07/373$\ & & & $2.11_{-0.05}^{+0.13}$ & - & $0.99_{-0.05}^{+0.04}$ & $20.82_{-2.28}^{+3.25}$ & $4.62_{-0.09}^{+0.02}$ & 0 & $1.09/374$\ & & & $1.96\pm0.03$ & $2$ & $1.0\pm0.02$ & $17.13_{-0.53}^{+0.56}$ & $3.73\pm0.03$ & $8\pm3$ & $1.15/424$\ &&& $2.0\pm0.04$ & $2.4$ & $1.0\pm0.02$ & $16.92_{-0.48}^{+0.51}$ & $3.73_{-0.03}^{+0.02}$ & & $1.15/424$\ &&& $1.9\pm0.02$ & - & $1.04\pm0.01$ & $16.12_{-0.34}^{+0.35}$ & $3.68_{-0.03}^{+0.02}$ & 0 & $1.2/425$\ & & & $2.13\pm0.03$ & $2$ & $0.96\pm0.01$ & $19.70_{-0.61}^{+0.65}$ & $3.90_{-0.04}^{+0.02}$ & $9\pm2$ & $1.04/415$\ & & & $2.18\pm0.04$ & $2.4$ & $0.95\pm0.01$ & $19.48_{-0.58}^{+0.62}$ & $3.89\pm0.03$ & $15\pm4$ & $1.03/415$\ & & & $2.05\pm0.02$ & - & $1.01\pm0.01$ & $18.04_{-0.39}^{+0.40}$ & $3.82_{-0.03}^{+0.02}$ & 0 & $1.14/416$\ & & & $2.06\pm0.03$ & $2$ & $0.92\pm0.02$ & $19.03_{-0.67}^{+0.70}$ & $3.27_{-0.03}^{+0.02}$ & $14\pm2$ & $1.08/404$\ & & & $2.12\pm0.04$ & $2.4$ & $0.92\pm0.02$ & $18.60_{-0.62}^{+0.66}$ & $3.26\pm0.03$ & $21\pm4$ & $1.08/404$\ & & & $1.94\pm0.02$ & - & $0.99\pm0.01$ & $16.68_{-0.37}^{+0.39}$ & $3.17\pm0.02$ & 0 & $1.27/405$\ & & & $1.92\pm0.03$ & $2$ & $0.93\pm0.02$ & $18.66_{-0.63}^{+0.66}$ & $3.36_{-0.03}^{+0.02}$ & $11\pm3$ & $1.12/400$\ & & & $1.96\pm0.04$ & $2.4$ & $0.94\pm0.02$ & $18.28_{-0.57}^{+0.61}$ & $3.35_{-0.03}^{+0.02}$ & $17\pm4$ & $1.12/400$\ & & & $1.83\pm0.02$ & - & $0.99\pm0.01$ & $16.89_{-0.37}^{+0.38}$ & $3.28\pm0.02$ & 0 & $1.24/401$\ & & & $2.02\pm0.03$ & $2$ & $0.9\pm0.02$ & $19.51_{-0.68}^{+0.71}$ & $3.14_{-0.03}^{+0.02}$ & $15\pm2$ & $1.03/409$\ & & & $2.08\pm0.03$ & $2.4$ & $0.9\pm0.02$ & $19.05_{-0.64}^{+0.68}$ & $3.13_{-0.03}^{+0.02}$ & $24\pm4$ & $1.03/409$\ & & & $1.89\pm0.02$ & - & $0.98\pm0.01$ & $16.8_{-0.37}^{+0.38}$ & $3.03\pm0.02$ & 0 & $1.27/410$\ & & & $2.09\pm0.03$ & $2$ & $0.88\pm0.01$ & $20.73_{-0.73}^{+0.77}$ & $3.28_{-0.02}^{+0.03}$ & $18\pm2$ & $1.08/409$\ & & & $2.16\pm0.03$ & $2.4$ & $0.88\pm0.02$ & $20.26_{-0.71}^{+0.76}$ & $3.27_{-0.03}^{+0.02}$ & $26\pm3$ & $1.07/409$\ & & & $1.94\pm0.02$ & - & $0.98\pm0.01$ & $17.15_{-0.38}^{+0.39}$ & $3.15\pm0.02$ & 0 & $1.48/410$\ & & & $2.09\pm0.07$ & $2$ & $0.89\pm0.03$ & $19.23_{-1.43}^{+1.61}$ & $2.86_{-0.07}^{+0.04}$ & $16\pm5$ & $0.96/208$\ & & & $2.15\pm0.08$ & $2.4$ & $0.89\pm0.03$ & $18.79_{-1.34}^{+1.54}$ & $2.85_{-0.07}^{+0.04}$ & $24\pm8$ & $0.95/208$\ & & & $1.95\pm0.05$ & - & $0.97\pm0.02$ & $16.35_{-0.73}^{+0.77}$ & $2.75\pm0.03$ & 0 & $1.06/209$\ & & & $1.91\pm0.03$ & $2$ & $0.89\pm0.01$ & $18.36_{-0.64}^{+0.67}$ & $2.62\pm0.02$ & $13\pm2$ & $1.0/387$\ & & & $1.96\pm0.03$ & $2.4$ & $0.89\pm0.02$ & $18.01_{-0.61}^{+0.65}$ & $2.62_{-0.03}^{+0.02}$ & $20\pm3$ & $1.01/387$\ & & & $1.80\pm0.02$ & - & $0.96\pm0.01$ & $16.03_{-0.37}^{+0.38}$ & $2.54_{-0.02}^{+0.01}$ & 0 & $1.23/388$\ & & & $1.73\pm0.04$ & $2$ & $0.94\pm0.02$ & $16.36_{-0.77}^{+0.83}$ & $2.60_{-0.04}^{+0.02}$ & $7\pm4$ & $1.09/314$\ & & & $1.75_{-0.06}^{+0.05}$ & $2.4$ & $0.94\pm0.02$ & $16.1_{-0.69}^{+0.75}$ & $2.59_{-0.04}^{+0.03}$ & $10\pm6$ & $1.09/314$\ & & & $1.68\pm0.03$ & - & $0.97\pm0.01$ & $15.41_{-0.49}^{+0.50}$ & $2.56\pm0.02$ & 0 & $1.11/315$\ & & & $1.82\pm0.05$ & $2$ & $0.89\pm0.03$ & $17.91_{-1.06}^{+1.18}$ & $2.37_{-0.05}^{+0.03}$ & $8\pm5$ & $1.04/259$\ & & & $1.84_{-0.07}^{+0.06}$ & $2.4$ & $0.89\pm0.03$ & $17.6_{-0.96}^{+1.07}$ & $2.37_{-0.04}^{+0.03}$ & $12\pm8$ & $1.05/259$\ & & & $1.75\pm0.04$ & - & $0.93\pm0.01$ & $16.58_{-0.62}^{+0.65}$ & $2.32_{-0.03}^{+0.02}$ & 0 & $1.07/260$ ID Model Slope Offset, keV ---- ------------------------ ------------------------ ------------------------- po $1.02\pm0.01$ $0.03\pm0.02$ $(po+disk) \Gamma=1.5$ $0.97_{-0.0}^{+0.01}$ $0.1\pm0.02$ $(po+disk) \Gamma=1.8$ $1.0\pm0.01$ $0.05\pm0.02$ $(po+disk) \Gamma=2$ $1.02_{-0.0}^{+0.01}$ $0.02\pm0.02$ po $1.02\pm0.01$ $0.06_{-0.03}^{+0.02}$ $(po+disk) \Gamma=1.5$ $0.99\pm0.01$ $0.12\pm0.02$ $(po+disk) \Gamma=1.8$ $1.02_{-0.01}^{+0.0}$ $0.05_{-0.01}^{+0.02}$ $(po+disk) \Gamma=2$ $1.04_{-0.0}^{+0.01}$ $0.01\pm0.02$ po $1.03\pm0.02$ $0.04_{-0.03}^{+0.04}$ $(po+disk) \Gamma=1.5$ $0.99\pm0.01$ $0.12_{-0.02}^{+0.03}$ $(po+disk) \Gamma=1.8$ $1.0\pm0.01$ $0.08_{-0.03}^{+0.02}$ $(po+disk) \Gamma=2$ $1.02\pm0.01$ $0.04\pm0.02$ po $0.89\pm0.01$ $0.18_{-0.03}^{+0.02}$ $(po+disk) \Gamma=2$ $1.04_{-0.01}^{+0.02}$ $-0.01\pm0.03$ $(po+disk) \Gamma=2.4$ $1.05_{-0.02}^{+0.01}$ $-0.03_{-0.03}^{+0.04}$ disk $1.04_{-0.01}^{+0.02}$ $-0.0_{-0.04}^{+0.03}$ po $1.0_{-0.01}^{+0.02}$ $-0.04_{-0.02}^{+0.03}$ $(po+disk) \Gamma=2$ $1.07_{-0.01}^{+0.02}$ $-0.06\pm0.02$ $(po+disk) \Gamma=2.4$ $1.08_{-0.01}^{+0.02}$ $-0.09_{-0.03}^{+0.02}$ disk $1.08_{-0.02}^{+0.01}$ $-0.03_{-0.02}^{+0.03}$ po $0.95\pm0.01$ $-0.0_{-0.0}^{+0.04}$ $(po+disk) \Gamma=2$ $1.04_{-0.02}^{+0.01}$ $-0.05_{-0.02}^{+0.04}$ $(po+disk) \Gamma=2.4$ $1.04\pm0.01$ $-0.07_{-0.02}^{+0.03}$ disk $1.03_{-0.02}^{+0.01}$ $-0.01\pm0.03$ po $1.04_{-0.02}^{+0.01}$ $-0.08\pm0.03$ $(po+disk) \Gamma=2$ $1.02\pm0.01$ $0.02_{-0.03}^{+0.02}$ $(po+disk) \Gamma=2.4$ $1.05\pm0.01$ $-0.04_{-0.02}^{+0.01}$ disk $1.03_{-0.02}^{+0.01}$ $0.1_{-0.03}^{+0.04}$ po $1.01_{-0.01}^{+0.02}$ $-0.03\pm0.06$ $(po+disk) \Gamma=2$ $1.07_{-0.03}^{+0.02}$ $-0.06_{-0.04}^{+0.05}$ $(po+disk) \Gamma=2.4$ $1.08\pm0.02$ $-0.08\pm0.04$ disk $1.08_{-0.03}^{+0.01}$ $-0.03_{-0.02}^{+0.06}$ $(po+disk) \Gamma=2$ $1.06\pm0.02$ $-0.1\pm0.03$ $(po+disk) \Gamma=2.4$ $1.07\pm0.02$ $-0.12_{-0.02}^{+0.03}$ disk $1.07\pm0.02$ $-0.07\pm0.03$ $(po+disk) \Gamma=2$ $1.08\pm0.03$ $-0.03\pm0.01$ $(po+disk) \Gamma=2.4$ $1.07_{-0.01}^{+0.03}$ $-0.03\pm0.01$ disk $1.07_{-0.02}^{+0.03}$ $-0.02\pm0.1$ : Parameters of the gain fit command used in fitting the XRT spectra[]{data-label="tbl-gain-pars"} When fitting the broadband spectra, we added the cross-calibration constants between the three instruments to take into account the nonsimultaneity of the observations. A difference of the cross-calibration constants between the JEM-X and ISGRI/IBIS instruments is observed when working with the INTEGRAL data (see, e.g., Filippova et al. 2014a). [cc|ccc]{} $ID^a$ & MJD-56000$^b$ & ------- Exp. $XRT$ s ------- & ---------- Exp. $ ISGRI$ s ---------- & --------- Exp. $JEM-X$ s --------- \ $01$ & 736 & $79$ & 4735 & 11244\ $03$ & 742 & $1925$ & 44637 & 10080\ $05$ & 751 & $1832$ & 4219 & 5006\ $07$ & 761 & $1613$ & 25806 & 68580\ $09$ & 771 & $1872$ & 5791 & 10738 DATA ANALYSIS ============= Light Curve during the 2014 Outburst ------------------------------------ {width="\textwidth"} Figure 1 shows the source’s light curve from Swift/XRT and Swift/BAT data in the 0.5-10 (the data points were averaged over one observation) and 15-50 keV (the data points were averaged over one day) energy bands, respectively. To convert the 15-50 keV flux to units corresponding to the flux from the Crab Nebula, we used the relation 1 Crab =0.22 counts s$^{-1}$ cm$^{-2}$. It can be seen from the figure that in the 15-50 keV energy band the flux from the source reached a maximum $\sim0.3$ Crab $\sim15$ days after the outburst onset, it then dropped by a factor of 1.5 in 5 days and was approximately at a constant level for $\sim25$ days, after which it began to exhibit a peak-shaped variability that lasted for about 50 days and passed into the outburst decay phase. The emission from the source in this energy band ceased to be recorded $\sim140$ days later. In the 0.5-10 keV energy band the outburst reached its maximum ($\sim$1.1 Crab) with a delay relative to the maximum in the 15-50 keV energy band, 55 days after the onset, but the source exhibited a peak-shaped flux variability almost immediately from the outburst onset. On completion of this activity, 135 days after the outburst onset, the flux reached a constant level $\sim$150 counts s$^{-1}$ ($\sim$400 mCrab) and remained so for $\sim$30 days, after which it began to drop. The observations ceased on $\sim$240 day of the outburst, because the source fell into a region near the Sun inaccessible to observations. At this time the 0.5-10 keV flux was $\sim$50 counts s$^{-1}$ ($\sim$140 mCrab). In Fig. 2a the flux in the soft energy band (0.5- 10 keV) is plotted against the source’s hardness (the ratio of the 4-10 and 0.5-4 keV fluxes). It can be seen from the figure that the diagram has a shape that resembles the upper part of a typical “q”$\,$ diagram. Spectral Analysis during the 2014 Outburst ------------------------------------------ When fitting the spectra, we used typical models describing the source’s spectrum: (1) in the low/hard state-a power law with a high-energy cutoff and low-energy absorption, phabs\*cutoffpl (or in the case of fitting only the Swift/XRT data- a power law and low-energy absorption, phabs\*powerlaw); (2) in the intermediate states-a power law with a high-energy cutoff, a multitemperature accretion disk, and low-energy absorption, phabs\*(cutoffpl + diskbb) (or in the case of fitting only the Swift/XRT data-a power law, a multitemperature accretion disk, and low-energy absorption, phabs\*(powerlaw + diskbb)); (3) in the high/soft state- a multitemperature accretion disk and low-energy absorption, phabs\*diskbb, or the previous model, phabs\*(powerlaw + diskbb). The quality of the available data does not allow us to apply the more complex spectral model including the reflection of Comptonized radiation from the accretion disk that was used in fitting the NuSTAR data (Miller et al. 2015; Mereminskiy et al. 2019). The results of fitting the Swift/XRT spectra are presented in Table 1. The errors in the parameters are given for a 90% confidence interval. The table also provides the contribution of the unabsorbed power-law component to the total unabsorbed flux, the absorbed 0.8-10 keV flux, and the inner accretion disk radius estimated from the normalization in the diskbb model $N = (R_{in}/D_{10 kpc})^2 cos i$, where $R_{in}$ is the “apparent”$\,$ inner disk radius in km, $D_{10 kpc}$ is the distance to the source in units of 10 kpc, and $i$ is the inclination to the plane of the sky. The distance to the source was taken to be 8.5 kpc. When fitting the spectra obtained only from the Swift/XRT data, the available 0.8-10 keV energy band does not allow unambiguous constraints to be placed on the parameters of the power-law component in the case of using the multicomponent phabs\*(powerlaw + diskbb) model. Therefore, when fitting the data at the initial outburst phases (from observation 03 to 09), we fixed the photon index at 1.5 and 1.8, which were derived when fitting the broadband spectra, at 2, and, in the subsequent observations, at 2 and 2.4, typical for the intermediate state (Remillard and McClintock 2006; Belloni and Motta 2016). {width="75.00000%"} To fit the broadband spectra, we also used the phabs\*highecut\*simpl\*diskbb model (the parameter $E_c$ in the highecut model was frozen at a minimum value of 0.0001 keV, which allowed the cutoffpl model to be imitated) that takes into account the physical cutoff of the power-law component at low energies in a simplified way. When estimating the errors in the parameters of the phabs\*highecut\*simpl\*diskbb model, we fixed the absorption $N_H$ (and the parameter $f_{scat}$ for observation 03) at their values found. The results of fitting the broadband spectra are presented in Table 4. It can be seen from the table that this model gives a systematically larger inner accretion disk radius than does the model with the cutoffpl component, while the remaining parameters do not differ greatly. It can be seen from Tables 1 and 4 that in observations 01-03 the model with a power law describes well the spectra, while in observation 03 the system’s spectrum can also be described by the power-law model with a multitemperature disk (the $\chi^2$ value is almost the same for the phabs\*(powerlaw +diskbb) and phabs\*powerlaw models). Beginning from the fourth observation, fitting the data by the model of a multitemperature disk with a power law is more preferable than that by the model only with a power law. Up to observation 23, fitting the data by a power law or a multitemperature disk with low-energy absorption gives a $\chi^2$ value systematically poorer than does the multicomponent model. In this case, the low-energy absorption, the temperature, and the inner radius of the accretion disk depend on the chosen photon index, i.e., the available data allow only the range of values (given in Table 1) in which the model parameters can lie to be specified. [a-cross-calibration constants for JEM-X and XRT relative to ISGRI, respectively; b-interstellar absorption; c-photon index; d- cutoff energy of the cutoffpl model; e-disk fraction subject to Comptonization (simpl); f-accretion disk temperature; g-inner disk radius for a distance to the system of 8.5 kpc; h-contribution of the power-law component to the total 0.8-10 keV flux; i-absorbed 0.8-10 keV flux of the broadband model, in units of $10^{-9}$ erg cm$^{-2}$ s$^{-1}$. The errors are given for a 90% confidence interval.]{} In observations 24-42 the data are well fitted by the model of a multitemperature disk with low-energy absorption. Although adding the power-law component when fitting some of the spectra formally reduces the $\chi^2$ value, the contribution of the powerlaw component is so small ($˜30\%$) that a change in the photon index within the range 2-2.4 has virtually no effect on the remaining model parameters. Temporal Variability during the 2014 Outburst --------------------------------------------- To analyze the source’s variability, we constructed its power spectra in several energy bands: 0.5-10 (F), 0.5-3 (A), and 3-10 keV (B). QPOs were detected in the power spectra constructed for observations 02-09. As the best-fit model for the power spectra obtained in these observations we used a model consisting of two Lorentz profiles (one Lorentz profile described the broadband To analyze the source’s variability, we constructed its power spectra in several energy bands: 0.5-10 (F), 0.5-3 (A), and 3-10 keV (B). QPOs were detected in the power spectra constructed for observations 02-09. As the best-fit model for the power spectra obtained in these observations we used a model consisting of two Lorentz profiles (one Lorentz profile described the broadband $$\begin{array}{l} P(f)=\frac{N_{qpo}}{\pi} \frac{\delta f_{qpo}/2}{(f-f_{qpo})^2+(\delta f_{qpo}/2)^2}+\\\frac{N_{sub}}{\pi} \frac{\delta f_{sub}/2}{(f-f_{sub})^2+(\delta f_{sub}/2)^2}+P_{noise} \end{array}$$ where $f_{qpo}$ and $\delta f_{qpo}$ are the frequency and width of the Lorentzian responsible for the QPOs, $f_{sub}$ and $\delta f_{sub}$ are the frequency and width of the Lorentzian responsible for the broadband noise, $N_{qpo}$ and $N_{sub}$ are the normalizations of the QPO and broadband noise components, $P_{noise}$ is the constant responsible for the Poisson noise level. In the subsequent analysis we assumed that $f_{sub}$ = 0. This model describes well the power spectra of black hole candidates in the low/hard and low intermediate states (Belloni and Motta 2016). To determine the parameters of the best-fit model, we used the maximum likelihood method (see Leahy et al. 1983; Vikhlinin et al. 1994). As the likelihood function we used the product of the probability density functions for a $\chi^2$ distribution with 2n degrees of freedom: $$L=\prod{f_{\chi^2_{2n}}({{ P_{i,src} 2n}\over{P_{i,model}}})},$$ where $P_{i,src}$ is the measured rms power of the source in the i-th frequency bin, $P_{i,model}$ is the rms power of the source in the same bin obtained from the model, and n is the number of bins into which the light curve was partitioned. [cc|cccc|cc|c]{} $ID$ & Band$^a$ & $\delta f_{qpo}$, Hz $^b$ & $f_{qpo}$, Hz $^c$ & $\delta f_{zl}$, Hz $^d$ & $Q=\frac{f_{qpo}}{\delta f_{qpo}}$ $^e$ & $rms_{qpo}, \%$ $^f$ & $rms_{tot}, \%$ $^f$ &$2\Delta L$ $^g$( $log(p)$ $^h$)\ & --- F --- & $0.03\pm0.01$ & $0.106\pm0.004$ & $0.33\pm0.03$ & $3.4\pm1.3$ & $13.9\pm1.8$ & $30.0\pm1.0$ & $39$($-7.8$)\ & --- A --- & $0.04\pm0.02$ & $0.103\pm0.005$ & $0.32_{-0.05}^{+0.04}$ & $8.0\pm4.2$ & $13.5_{-2.5}^{+2.3}$ & $27.0\pm1.0$ & $28$($-5.4$)\ & --- B --- & $0.03_{-0.02}^{+0.01}$ & $0.113_{-0.012}^{+0.004}$ & $0.31_{-0.04}^{+0.03}$ & $4.2\pm2.5$ & $14.5_{-2.7}^{+2.4}$ & $32.0\pm1.0$ & $27$($-5.2$)\ & F & $0.03\pm0.01$ & $0.377_{-0.003}^{+0.004}$ & $0.62_{-0.04}^{+0.05}$ & $11.7\pm3.4$ & $9.9_{-1.0}^{+0.9}$ & $27.0\pm1.0$ & $81$($-16.7$)\ & A & $0.03\pm0.01$ & $0.379\pm0.005$ & $0.57\pm0.09$ & $11.6\pm5.1$ & $8.7_{-1.3}^{+1.2}$ & $23.0\pm1.0$ & $37$($-7.3$)\ & B & $0.04\pm0.01$ & $0.377_{-0.004}^{+0.005}$ & $0.61\pm0.05$ & $10.3\pm3.6$ & $11.2\pm1.3$ & $31.0\pm1.0$ & $58$( $-11.8$)\ & F & $0.29\pm0.03$ & $2.182\pm0.012$ & $1.59\pm0.28$ & $7.6\pm0.9$ & $10.3\pm0.4$ & $14.0\pm1.0$ & $338$($-72.2$)\ & A & $0.32_{-0.11}^{+0.1}$ & $2.208\pm0.039$ & $1.47_{-0.84}^{+0.64}$ & $6.8\pm2.3$ & $7.2_{-0.9}^{+0.8}$ & $10.0\pm1.0$ & $45$($-9.0$)\ & B & $0.21\pm0.03$ & $2.19\pm0.01$ & $5.09_{-0.67}^{+0.64}$ & $10.4\pm1.5$ & $13.5\pm0.6$ & $25.0\pm1.0$ & $295$($-62.9$)\ & F & $0.34\pm0.04$ & $1.69\pm0.01$ & $0.89_{-0.20}^{+0.19}$ & $5.0\pm0.6$ & $12.6\pm0.5$ & $15.0\pm0.0$ & $262$($-55.8$)\ & A & $0.31_{-0.13}^{+0.11}$ & $1.73\pm0.04$ & $1.08_{-0.66}^{+0.44}$ & $5.6\pm2.3$ & $7.7\pm1.1$ & $10.0\pm1.0$ & $31$($-6.1$)\ & B & $0.24_{-0.03}^{+0.04}$ & $1.7\pm0.01$ & $5.17_{-0.77}^{+0.83}$ & $7.2\pm1.1$ & $15.5\pm0.8$ & $27.0\pm1.0$ & $235$( $-49.7$)\ & F & $0.4_{-0.17}^{+0.15}$ & $5.07\pm0.06$ & $1.5_{-0.34}^{+0.32}$ & $12.7\pm5.3$ & $6.7\pm0.9$ & $12.0\pm1.0$ & $32$($-6.3$)\ & A & - & - & $0.75_{-0.52}^{+0.21}$ & - & - & $6.0_{-2.0}^{+1.0}$ & $0(-)$\ & B & $0.24\pm0.10$ & $5.06\pm0.03$ & $2.18_{-0.48}^{+0.49}$ & $21.1\pm9.1$ & $8.9\pm1.2$ & $19.0\pm1.0$ & $34$($-6.7$)\ & F & $0.48\pm0.08$ & $2.48\pm0.03$ & $0.93_{-0.19}^{+0.18}$ & $5.2\pm0.9$ & $10.6\pm0.6$ & $15.0\pm1.0$ & $120$($-25.1$)\ & A & $0.34_{-0.24}^{+0.20}$ & $2.41\pm0.09$ & $0.99_{-0.58}^{+0.35}$ & $7.1\pm4.9$ & $6.4_{-1.7}^{+1.5}$ & $10.0\pm1.0$ & $13$($-2.3$)\ & B & $0.49\pm0.09$ & $2.52\pm0.03$ & $1.12_{-0.22}^{+0.20}$ & $5.1\pm1.0$ & $15.2_{-1.0}^{+0.9}$ & $22.0\pm1.0$ & $90$($-18.7$)\ & F & $0.54_{-0.35}^{+0.33}$ & $5.09\pm0.14$ & $1.08_{-0.38}^{+0.37}$ & $9.4\pm6.2$ & $7.2\pm1.6$ & $12.0\pm1.0$ & $14$($-2.5$)\ & A & - & - & $2.63_{-1.44}^{+1.04}$ & - & - & $12.0\pm2.0$ &$0(-)$\ & B & - & - & $9.55_{-3.78}^{+3.34}$ & - & - & $21.0\pm3.0$ & $0(-)$\ & F & $0.31\pm0.06$ & $2.19\pm0.02$ & $0.28\pm0.08$ & $7.1\pm1.5$ & $11.7\pm0.8$ & $14.0\pm1.0$ & $91$($-18.9$)\ & A & $0.39_{-0.23}^{+0.21}$ & $2.25\pm0.09$ & $0.47_{-0.31}^{+0.10}$ & $5.8\pm3.5$ & $7.4_{-1.5}^{+1.6}$ & $10.0\pm1.0$ & $18$($-3.4$)\ & B & $0.18\pm0.05$ & $2.15\pm0.02$ & $2.7_{-0.90}^{+0.92}$ & $11.9\pm3.1$ & $15.0_{-1.3}^{+1.4}$ & $25.0\pm2.0$ & $80$($-16.5$) To find the error in the parameters of the best-fit model, we used the Monte Carlo method. The data were randomly selected 1000 times around the best-fit model $P_{i,model}$ according to the law $P_i=\chi^2_{2n}(P_{i,model}/2n)$. The randomly selected data were also fitted by the model using the maximum likelihood method. The sought-for error in a parameter was found as a difference between the mean value and the lower (upper) limit on the parameter corresponding to the 16% (84%) quantile of the distribution. {width="75.00000%"} {width="75.00000%"} To determine the QPO significance, we calculated the doubled difference of logarithmic likelihood functions $2 \log(L_{qpo}/L_{null})$, where $L_{qpo}$ and $L_{null}$ are the values of the likelihood function for the model with and without QPOs, respectively, and the probability that this difference is a random variable. The difference of the likelihood functions has a $\chi^2_k$ distribution (Cash 1979), where k is the difference of the numbers of free parameters in the models with and without QPOs, in our case, k = 3. The results of fitting the power spectra with QPOs are presented in Table 5. No QPOs were recorded in observation 06 in the A band and in observation 08 in both á and ÷. For these power spectra we calculated an upper limit on the QPO fractional rms at 90$\%$ confidence by assuming the QPO frequency and quality factor in the A and B band to coincide with those in the F band. For observation 06 in the A band the upper limit is $r_{qpo}~<~6\%$ ; in observation 08, $r_{qpo}~<~10$ and 15$\%$ for the A and B bands, respectively. It can be seen from Table 5 that the QPO frequency does not depend on the energy band. Note that, as follows from the literature, the systems with black hole candidates show both no correlation between the QPO frequency and energy and direct and inverse proportionality (Yan et al. 2012; Li et al. 2013a, 2013b). Figure 3 presents the power spectra with QPOs in the full energy band (0.5-10 keV) for several observations (02, 03, and 05). The QPO frequency is clearly seen to change from observation to observation. To determine the type of QPOs, it is necessary to measure the parameters of both the QPO peak itself and the broadband noise. It follows from Fig. 3 and Table 5 that broadband noise whose total fractional rms is greater than 10$\%$ is present in the power spectra under study at low frequencies, which, as was said in the Introduction, is characteristic for type- C QPOs. We constructed the dependence of the QPO frequency on the flux in the soft (0.5-10 keV) and hard (15-50 keV) energy bands (Fig. 4). Since the contribution of the disk component in the 15- 50 keV energy band is minor, this may be considered as the dependence of the QPO frequency on the flux in the power-law component. It can be seen from the figure that the dependence of the QPO frequency on the flux in the soft and hard energy bands is direct and inverse, respectively; such a behavior is also characteristic for type-C QPOs (Motta et al. 2011). Stiele et al. (2011) showed that type-B QPOs are observed only at certain photon indices of the spectral component describing the Comptonized radiation. At the transition stage from the hard to soft state the photon index must be greater than or of the order of 2.2. In our case, it can be seen from Table 4 that the photon index is less than or of the order of 2, which again provides evidence for type-C QPOs. For several systems, it was shown on the basis of Fourier spectroscopy that the corona makes a major contribution to the system’s variability (Churazov et al. 2001; Sobolewska and Zycki 2006), i.e., the fractional rms must decrease with decreasing contribution of the power-law component to the flux, which we observe. In observations 02 and 03, when the fractional rms in the A and B energy bands is determined by the power-law component, the total fractional rms in the soft energy band (A) is smaller than that in the hard (B) energy band by a factor of 1.2-1.3. At the same time, in observations 04-09, when the disk component is also present in the soft energy band, the fractional rms in the A band is smaller than that in the B band by a factor of 2-3. For observations 10-42, when no QPOs were recorded, we determined the total fractional rms in the 0.5-10 keV energy band. For this purpose, we fitted the power spectra either by a power law with a constant or only by a constant by the maximum likelihood method. If we failed to extract the powerlaw component, then the source’s limiting white noise power was estimated. For this purpose, we searched for $P_{source}$ at which the measured power in the frequency range 0.01-50 Hz was the 10% quantile of the normal distribution $\mathcal{N}(P_{noise}+P_{source}, (P_{noise}+P_{source})/Nn)$, where N is the number of frequency bins and $P_{source}$ is the measured Poisson noise level. A diagram of the derived dependence of the total fractional rms on hardness (the ratio of the 4-10 and 0.5-4 keV fluxes) is shown in Fig. 2b. It follows from the figure that the total fractional rms of the source in the time of observations decreased from $\sim$30 to $\sim$8% or less. Observed States during the 2014 Outburst ---------------------------------------- It follows from our spectral analysis (see Tables 1 and 4) that during observations 01-03 the photon index of the power-law component was $\sim 1.2 - 1.4$, while it follows from Table 5 that the total fractional rms during observations 02 and 03 was 30 and 27%, respectively. The values of these parameters suggest that the system was in the low/hard state in the period from observation 01 to 03. From observation 04 to 09 the total fractional rms decreased to 12-15%, type-C QPOs were observed in the variability power spectrum, and the accretion disk contribution, along with the power-law component, is recorded in the source’s energy spectrum, which is typical for the hard intermediate state. The total fractional rms was $10\pm2$ % during observation 10, $8\pm3\%$ in observation 11, and 14-16% in observations 12 and 13, which also provides evidence for the hard intermediate state. From observation 14 to 23 the source’s energy spectrum is still described by the model of an accretion disk with a power-law component, but the fractional rms of the source dropped below 10%. In many observations we managed to obtain only upper limits at a level &lt;10%. Thus, it can be concluded that the system passed to the soft intermediate state between observations 13 and 14. During observation 11 (when not only the total fractional rms, but also the contribution of the powerlaw component to the flux from the system decreased almost by a factor of 2 compared to the adjacent observations, see Table 1) the system may have passed to the soft intermediate state, but this cannot be asserted based on the available data. Despite the fact that fitting the data by the phabs\*(diskbb +powerlaw) and phabs\*diskbb models gives identical $\chi^2$ values, we think that the first model is most probable, because the source exhibits a significant flux (150 mCrab) in the 15-50 keV energy band. From observation 24 to 42 the source’s energy spectra are well described by the model of a multitemperature disk with low-energy absorption, i.e., it can be argued that the system passed to the high/soft state. Although adding the power-law component when fitting some of the spectra formally reduces the $\chi^2$ value, nevertheless, first, the contribution of the power-law component is minor ($˜30\%$) and, second, a weak power-law can also be observed in the high/soft state (Belloni and Motta 2016). It is worth noting that the powerlaw model has no low-energy cutoff and the estimate of the contribution from the power-law component to the total flux is an upper limit, i.e., the fraction of the nonthermal component is actually smaller. Furthermore, it can be seen from Fig. 1 that after observation 23 the source is barely recorded in the 15-50 keV energy band, which also provides evidence for the transition to the high/soft state. The source’s characteristic spectra corresponding to the low/hard, intermediate low/hard, and high/soft states (observations 01, 09, and 42, respectively) are presented in Fig. 5. {width="\textwidth"} Figure 6 shows the periods of time and the states in which the source was and presents the time dependences of the accretion disk temperature at the inner radius, the contribution of the power-law component to the total flux, and the total fractional rms. The accretion disk temperature at the inner radius and the contribution of the power-law component were taken from the model in which the photon index was fixed at 2. {width="\textwidth"} On the whole, our results on the transitions between states are consistent with those from Yan and Yu (2017) and Wang et al. (2018). Mini-Outbursts of the System ---------------------------- We performed an analysis of the light curve for GRS 1739-278 over the entire period of observations since its discovery aimed at searching for undetected outbursts. From 1996 to mid-2011 the source was regularly observed by the ASM/RXTE all-sky monitor in the 1.2-12 keV energy band (Levine et al. 1996). According to these data, after the bright 1996 outburst the source exhibited no outburst activity and was not detected. Since 2005 the source has been observed almost continuously by the Swift/BAT telescope in the 15-50 keV energy band. From 2005 to 2014 no outbursts were detected on the light curve and the mean flux was $1.0\pm0.4$ mCrab. After the end of the 2014 outburst the mean flux from the source rose to $9.7 \pm 0.2$ mCrab. {width="\textwidth"} {width="\textwidth"} {width="75.00000%"} {width="50.00000%"} A detailed analysis of the Swift/BAT light curve after the 2014 outburst showed that, apart from the mini-outbursts mentioned in the literature, the system exhibited several more similar events. To determine the statistical significance of the detected outbursts, we performed an analysis in which we partitioned the light curves into bins containing these outbursts (indicated by the gray rectangles in Fig. 7) and fitted the time dependence of the flux by two models: a constant and a constant with a Gaussian profile as the first approximation for the outburst profile. The outburst detection significance was defined as the probability of the difference of the $\chi^2$ values for both best-fit models (F-test). The results of our analysis are shown in Fig. 8; the outburst detection significance is given above each panel. It follows from the figure that the mini-outbursts mentioned in the literature had a significance of 7-8$\sigma$ (mini-outbursts 2, 3, and 8), while the detected four mini-outbursts have a significance of 4-5.5$\sigma$. Since outburst 7 has a low significance, 2.6$\sigma$, we did not include it in our final conclusions. After the failed 2016 outburst the system returned to a quiescent state with a mean flux of $5.0 \pm 0.3$ mCrab. Evolution of the Outbursts of the System in 1996, 2014, and 2015 ---------------------------------------------------------------- Using the RXTE/ASM archival data, we constructed the hardness-intensity diagram for the 1996 outburst and compared it with that for the 2014 outburst and the 2015 mini-outbursts, when the system passed to the high/soft state. To make a proper comparison of the diagrams, it is necessary to take into account the difference in the energy bands and characteristics of the instruments. For this purpose, we calculated the hardness using the best-fit spectral models for several states. In the 1996 outburst we chose the KVANT/TTM observations on February 6-7, 1996, and the RXTE/PCA observations on March 31, 1996, and May 29, 1996 (see Borozdin et al. 1998). For the 2014 outburst we used observations 01, 11, and 32. The first observations for each outburst correspond to the time of the largest recorded hardness, the second observations refer to the time of the source’s maximum soft X-ray intensity, and the third observations refer to the high/soft state, when the power-law contribution to the total flux is minor. Using this sample of observations, we calculated the hardness for the 5-10 and 1.5-5 keV energy bands. The hardness ratios were found to be 1.29, 0.57, and 0.30, respectively, in the 1996 outburst and 1.29, 0.64, and 0.23 in the 2014 outburst. Having calculated the hardness from the light curves in the above reference observations, we found the ratios of the “true” $\,$ (based on the models) and observed (based on the light curves) hardnesses. The scatter of these ratios relative to the mean value is $12-15\%$; therefore, we used the mean value as a coefficient to convert the observed hardness-intensity diagram to the true one. The same coefficient was used For the 2014 outburst and the 2015 mini-outbursts. For the convenience of comparing the diagram shapes, we also normalized the flux to its maximum value. The derived diagrams are presented in Fig. 9. It can be seen from the figure that during the bright outbursts the curves have a similar shape, with the behavior of the source on the hardness-intensity diagram during the bright outbursts differing significantly from its behavior during the mini-outbursts. We estimated the luminosity at which a minimum hardness was reached during the outbursts by assuming the distance to the system to be 8.5 kpc: $L_{1.2-12 \text{\,keV}}\sim 1.5\times 10^{37}$ erg s$^{-1}$ for the 1996 outburst, $L_{0.5-10 \text{\,keV}}\sim 2\times 10^{37}$ erg s$^{-1}$ for the 2014 outburst, and$L_{0.5-10 \text{\,keV}}\sim(5-6)\times 10^{36}$ erg s$^{-1}$ for the mini-outbursts. Yu et al. (2007) and Wu et al. (2010) constructed the dependence of the peak flux in the 20-160 keV energy band during the low/hard state on the time between the current and previous peak fluxes in the low/hard state for GX 339-4 (a low-mass binary system with a black hole candidate) and attempted to fit this dependence by a linear law. We constructed the same dependence of the peak flux in the 15-50 keV energy band in the low/hard state on the time to the previous peak in the low/hard state for GRS 1739-278 by taking into account the bright 1996 and 2014 outbursts and the 2015 mini-outbursts (Fig. 10). The time of the transition to the hard state at the end of the 2014 outburst was taken from Wang et al. (2018). The linear dependence that best fits the data for GRS 1739-278 looks as follows: $F_{hard}(\Delta T)=(0.043\pm 0.003) \frac{\text{mCrab}}{\text{day}}\Delta T+(27\pm 2 ) \text{mCrab}$. CONCLUSIONS =========== In this paper we performed a joint study of the spectral and temporal evolution of GRS 1739-278 during the 2014 outburst and made a comparative analysis of the system’s behavior during the remaining outbursts mentioned in the literature and in the periods between them. Our results can be briefly summarized as follows. - We showed that during the 2014 outburst the system passed to the hard intermediate state 22 days after the outburst onset, to the soft intermediate state 66 days later (possibly exhibiting this state on day 55 and returning to the hard intermediate state no later than 4 days after), and to the high/soft state 145 days later. - QPOs in the frequency range 0.1-5 Hz were detected during the outburst of GRS 1739-278 in 2014. All QPOs are type-C ones. No energy dependence of the QPO frequency was found. - We showed that after the 2014 outburst the system passed to the regime of mini-outburst activity and, apart from the three mini-outbursts mentioned in the literature (Yu and Yan 2017; Mereminskiy et al. 2017), we detected four more mini-outbursts with a comparable ($\sim$20 mCrab) flux in the hard energy band (15-50 keV). - We showed that the hardness-intensity diagram for the 2015 mini-outbursts, during which the system exhibited the transition to the high/soft state, differs from that for the bright 1996 and 2014 outbursts: the minimum hardness during the mini-outbursts was reached at fluxes of at least $60-80\%$ of the peak one, while in the bright outbursts the minimum hardness was reached at fluxes of $\sim10\%$ of the peak one. The 0.5-10 keV luminosity of the source corresponding to these times differed approximately by a factor of 3: $L_{0.5-10\text{\, keV}}\sim 2\times 10^{37}$ erg s$^{-1}$ for the bright outburst and $L_{0.5-10\text{\,keV}}\sim(5-6)\times 10^{36}$ erg s$^{-1}$ for the mini-outbursts. - We constructed the dependence of the peak flux in the hard energy band during the low/hard state on the time interval between outbursts. This dependence can be fitted by a linear law, which may point to the dependence of the system’s peak flux in the low/hard state on the mass of the accretion disk being accumulated. ACKNOWLEDGMENTS =============== This work was financially supported by RSF grant no. 14-12-01287. We used the data provided by the UK Swift Science Data Center at the University of Leicester and the INTEGRAL Science Data Centers at the University of Geneva and the Space Research Institute of the Russian Academy of Sciences. REFERENCES ========== - 1\. K. A. Arnaud, Astron. Data Anal. Software Syst. V **101**, 17 (1996). - 2\. T. M. Belloni, Lect. Notes Phys. **794**, 53 (2010). - 3\. T. M. Belloni and S. E. Motta, Astrophys. Space Sci. Lib. **440**, 61 (2016). - 4\. K. N. Borozdin and S. P. Trudolyubov, Astrophys. J. **533**, L131 (2000). - 5\. K. Borozdin, N. Alexandrovich, R. Sunyaev, et al., IAU Circ. **6350** (1996). - 6\. K.N. Borozdin, M. G. Revnivtsev, S. P. Trudolyubov, et al., Astron. Lett. **24**, 435 (1998). - 7\. D. N. Burrows, J. E. Hill, J. A. Nousek, et al., Space Sci. Rev. **120**, 165 (2005). - 8\. F. Capitanio, T. Belloni, M. del Santo, et al., Mon. Not. R. Astron. Soc. **398**, 1194 (2009). - 9\. W. Cash, Astrophys. J. **228**, 939 (1979). - 10\. P. Durouchoux, I. A. Smith, K. Hurley et al., IAU Circ. **6383**, 1 (1996). - 11\. P. A. Evans, A. P. Beardmore, K. L. Page, et al., Astron. Astrophys. **469**, 379 (2007). - 12\. C. Ferrigno, E. Bozzo, M. del Santo, et al., Astron. Astrophys. **537**, L7 (2012). - 13\. E. Filippova, E. Bozzo, and C. Ferrigno, Astron. Astrophys. **563**, A124 (2014a). - 14\. E. Filippova, E. Kuulkers, N.M. Skadt, et al., Astron. Telegram **5991**, 1 (2014b). - 15\. F. Furst, M. A. Nowak, J. A. Tomsick, et al., Astrophys. J. **808**, 122 (2015). - 16\. M. R. Gilfanov, Lect. Notes Phys. **794**, 17 (2010). - 17\. S. Grebenev, R. Sunyaev,M. Pavlinsky, et al., Astron. Astrophys. Suppl. Ser. **97**, 281 (1993). - 18\. S. Grebenev, R. Sunyaev, and M. Pavlinsky, Adv. Space Res. **19**, 15 (1997). - 19\. J. Homan, R. Wijnands,M. van der Klis, et al., Astrophys. J. **132**, 377 (2001). - 20\. A. Ingram, C. Done, and P. C. Fragile,Mon. Not. R. Astron. Soc. **397**, L101 (2009). - 21\. M. van der Klis, in Timing Neutron Stars, Ed. by H. Ogelman and E. P. J. van den Heuvel, NATO ASI Ser. C **262**, 27 (1988). - 22\. H. A. Krimm, S. T. Holland, R. H. D. Corbet, et al., Astrophys. J. **209**, 14 (2013). - 23\. H. A. Krimm, S. D. Barthelmy, W. Baumgartner, et al., Astron. Telegram **5986**, 1 (2014). - 24\. D. A. Leahy,W. Darbro,R. F. Elsner, et al., Astrophys. J. **266**, 160 (1983). - 25\. A.M. Levine, H. Bradt, W. Cui, et al., Astrophys. J. **469**, L33 (1996). - 26\. Z. B. Li, J. L. Qu, L.M. Song, et al., Mon. Not. R. Astron. Soc. **428**, 1704 (2013a). - 27\. Z. B. Li, S. Zhang, J. L. Qu, et al., Mon. Not. R. Astron. Soc. **433**, 412 (2013b). - 28\. I. Mereminskiy, R. Krivonos, S. Grebenev, et al., Astron. Telegram **9517**, 1 (2016). - 29\. I. A. Mereminskiy, E. V. Filippova, R. A. Krivonos, S. A. Grebenev, R. A. Burenin, and R. A. Sunyaev, Astron. Lett. **43**, 167 (2017). - 30\. I. A. Mereminskiy, A. N. Semena, S.D. Bykov, et al., Mon. Not. R. Astron. Soc. **482**, 1392 (2019). - 31\. J. M. Miller, J. A. Tomsick, M. Bachetti, et al., Astrophys. J. **799**, L6 (2015). - 32\. S.Motta, T.Munoz-Darias, and T.Belloni,Mon.Not. R. Astron. Soc. **408**, 1796 (2010). - 33\. S. Motta, T. Munoz-Darias, P. Casella, et al., Mon. Not. R. Astron. Soc. **418**, 2292 (2011). - 34\. J. Paul, L. Bouchet, E. Churazov, et al., IAU Circ. **6348**, 1 (1996). - 35\. R. A. Remillard and J. E. McClintock, Ann. Rev. Astron. Astrophys. **44**, 49 (2006). - 36\. J. Rodriguez, S. Corbel, E. Kalemci, et al., Astrophys. J. **612**, 1018 (2004). - 37\. M. del Santo, T.M. Belloni, J. A. Tomsicket al.,Mon. Not. R. Astron. Soc. **456**, 3585 (2016). - 38\. D. M. Smith, W. A. Heindl, C. B. Markwardt, et al., Astrophys. J. **554**, L41 (2001). - 39\. M. A. Sobolewska and P. T. Zycki, Mon. Not. R. Astron. Soc. **370**, 405 (2006). - 40\. J. F. Steiner, R. Narayan, J. E. McClintock, et al., Publ. Astron. Soc. Pacif. **121**, 1279 (2009). - 41\. H. Stiele, S. Motta, T. Munoz-Darias, et al., Mon. Not. R. Astron. Soc. **418**, 1746 (2011). - 42\. Y. Tanaka and N. Shibazaki, Ann. Rev. Astron. Astrophys. **34**, 607 (1996). - 43\. M. Vargas, A. Goldwurm, J. Paul, et al., Astron. Astrophys. **313**, 828 (1996). - 44\. A. Vikhlinin, E. Churazov, M. Gilfanov, et al., Astrophys. J. **424**, 395 (1994). - 45\. R. Walter, R. Rohlfs, M. T. Meharga, et al., in Proceedings of the 8th Integral Workshop on The Restless Gamma-ray Universe INTEGRAL 2010, PoS(INTEGRAL2010)**162**. - 46\. S.Wang, N. Kawai,M. Shidatsu, et al., Publ. Astron. Soc. Jpn. **70**, 67 (2018). - 47\. Y. X.Wu,W.Yu, Z. Yan, et al., Astron. Astrophys. **512**, A32 (2010). - 48\. S. P. Yan, J. L. Qu, G. Q. Ding, et al., Astron. Astrophys. Suppl. Ser. **337**, 137 (2012). - 49\. Z. Yan and W. Yu, Mon. Not. R. Astron. Soc. **470**, 4298 (2017). - 50\. W. Yu and Z. Yan, Astrophys. J. **701**, 1940 (2009). - 51\. W. Yu, M. van der Klis, and R. Fender, Astrophys. J. **611**, L121 (2004). - 52\. W.Yu, F. K.Lamb,R.Fender, et al., Astrophys. J. **663**, 1309 (2007). - 53\. A. A. Zdziarski,M.Gierlinski, J.Mikolajewska, et al., Mon. Not. R. Astron. Soc. **351**, 791 (2004). [^1]: <https://heasarc.gsfc.nasa.gov/docs/heasarc/caldb/swift/docs/xrt/SWIFT-XRT-CALDB-09_v19.pdf> [^2]: <https://heasarc.gsfc.nasa.gov/docs/heasarc/caldb/swift/docs/xrt/SWIFT-XRT-CALDB-09_v20.pdf> [^3]: <https://www.isdc.unige.ch/integral/download/osa/doc/10.2/osa_um_ibis/index.html> [^4]: <https://www.isdc.unige.ch/integral/download/osa/doc/10.2/osa_um_jemx/index.html>

ArXiv

--- abstract: 'In this paper we obtain Sobolev estimates for weak solutions of first order variational Mean Field Game systems with coupling terms that are local functions of the density variable. Under some coercivity conditions on the coupling, we obtain first order Sobolev estimates for the density variable, while under similar coercivity conditions on the Hamiltonian we obtain second order Sobolev estimates for the value function. These results are valid both for stationary and time-dependent problems. In the latter case the estimates are fully global in time, thus we resolve a question which was left open in [@ProSan]. Our methods apply to a large class of Hamiltonians and coupling functions.' address: - 'Department of Mathematics, Baylor University, One Bear Place, Waco, TX 97328, USA' - 'Department of Mathematics, UCLA, 520 Portola Plaza, Los Angeles, CA 90095, USA' author: - 'P. Jameson Graber' - 'Alpár R. Mészáros' bibliography: - 'jameson\_alpar.bib' title: Sobolev regularity for first order Mean Field Games --- Introduction ============ The theory of Mean Field Games (briefly MFG in the sequel) saw the light more than a decade ago, thanks to the works of J.-M. Lasry and P.-L. Lions on the one hand (see [@LasLio06i; @LasLio06ii; @LasLio07]) and M. Huang, R. Malhamé and P. Caines, on the other hand (see [@HuaMalCai]). Their main motivation was to study limits of Nash equilibria of (stochastic or deterministic) differential games when the number of players tends to infinity. Since then, it became a very lively and active branch of the theory nonlinear partial differential equations. In its most simple form an MFG can be described in an informal way as follows. As data, one considers the given quantities: $T>0$ (the time horizon), $H:{\mathbb{T}}^d\times{\mathbb{R}}^d\to{\mathbb{R}}$ (the Hamiltonian), $\phi_T:{\mathbb{T}}^d\to{\mathbb{R}}$ (the final cost of the agents), $f:{\mathbb{T}}^d\times{{\mathscr P}}({\mathbb{T}}^d)\to{\mathbb{R}}$ (the running cost of the agents) and $m_0\in{{\mathscr P}}({\mathbb{T}}^d)$ (the initial distribution of the agents), where ${\mathbb{T}}^d:={\mathbb{R}}^d/{\mathbb{Z}}^d$ denotes the $d$-dimensional flat torus and ${{\mathscr P}}({\mathbb{T}}^d)$ stands for the set of nonnegative Borel probability measures on ${\mathbb{T}}^d$. A typical agent predicts the evolution of the agents’ density $m:[0,T]\times{\mathbb{T}}^d\to[0,+\infty)$ and to find her/his optimal strategy, solves the control problem $$\label{prob:control} \inf_{\a}\left\{\int_t^T L(\gamma(s),\a(s))+f(\gamma(s),m(s,\gamma(s)))\dd s +\phi_T(\gamma(T))\right\}=:\phi(t,x)$$ subject to $$\left\{ \begin{array}{ll} \gamma'(s)=\alpha(s), & s\in(t,T]\\ \gamma(t)=x\in{\mathbb{T}}^d. \end{array} \right.$$ Here the Lagrangian $L:{\mathbb{T}}^d\times{\mathbb{R}}^d\to{\mathbb{R}}$ is the Legendre-Fenchel transform of $H$ w.r.t. the second variable. The optimal strategy is given in feedback form, hence for the agent it is optimal to play $-D_\xi H(\gamma(s),\nabla\phi(s,\gamma(s)))$. Having this velocity field as drift, the evolution of the agents’ density is given by the solution of a continuity equation. If the prediction coincides with this evolution, the game has a Nash equilibrium. Thus, searching for Nash equilibria for MFG is equivalent to solving the following system of PDEs: $$\label{eq:MFG_intro} \left\{ \begin{array}{ll} -\partial_t\phi + H(x,\nabla\phi) = f(x,m), & {\rm{in}}\ (0,T)\times{\mathbb{T}}^d,\\[5pt] \partial_t m -\nabla\cdot (D_\xi H(x,\nabla\phi)m)=0, & {\rm{in}}\ (0,T)\times{\mathbb{T}}^d,\\[5pt] m(0,\cdot)=m_0,\ \ \phi(T,\cdot)=\phi_T, & {\rm{in}}\ {\mathbb{T}}^d. \end{array} \right.$$ With a well-chosen time rescaling, one can introduce stationary MFG systems as long time average of time dependent ones. These take the form $$\label{eq:MFG_intro_stat} \left\{ \begin{array}{ll} \l+ H(x,\nabla\phi) = f(x,m), & {\rm{in}}\ {\mathbb{T}}^d,\\[5pt] -\nabla\cdot (D_\xi H(x,\nabla\phi)m)=0, & {\rm{in}}\ {\mathbb{T}}^d,\\[5pt] {\displaystyle}\int_{{\mathbb{T}}^d}m\dd x=1, \; {\displaystyle}\int_{{\mathbb{T}}^d}\phi\dd x=0, \; m\ge 0, & {\rm{a.e.\ in\ }}{\mathbb{T}}^d, \end{array} \right.$$ where an additional variable $\lambda\in{\mathbb{R}}$ appears in the first equation, which plays the role of the ergodic constant. We refer to [@CarLasLioPor; @CarLasLioPor2] and [@CarGra Section 4] for results on the limiting procedure and well-posedness of . The case when the running cost $f$ is monotone and regularizing (nonlocal) in the measure variable, is well understood in the literature. Already in the original works of J.-M. Lasry and P.-L. Lions (see also [@Lions-course]) has been shown the well-posedness of both systems and . By contrast, when $f$ is a local function of $m$, the question of well-posedness of these systems is more subtle. Notice also that in the lack of sufficient regularity for the density variable $m$ (when $f$ is a local function of $m$), the control problem is not meaningful, while the system may still have solutions in a suitable weak sense. In the case when $f$ is a local function of $m$, PDE techniques, which are used successfully for second order systems (see for instance [@GomPimSan15; @GomPimSan16; @GomPimVos; @Por]), can no longer be used to show the well-posedness of the system . Also, in general one cannot hope for the existence of classical solutions of . Nevertheless, as it was already mentioned in [@LasLio06ii] that when $f$ is non-decreasing in its second variable and $H$ convex in its second variable, at least formally, systems and can be seen as first order optimality conditions of two convex optimization problems in duality. In the first order case (or when a degenerate diffusion is present) these arguments were made rigorous in a series of papers (see [@Car15; @CarGra; @CarGraPorTon]), and the existence of weak solutions (understood in a suitable sense) was shown. An important and interesting question in this context is the regularity of solutions of and . A first result in this direction was obtained in [@CarPorTon], which, it must be emphasized, provides first and foremost a regularity result for viscosity solutions of first-order Hamilton-Jacobi equations. Using techniques involving the inverse Hölder inequality, the authors obtained $W^{1,1}_{{\rm{loc}}}((0,T)\times{\mathbb{T}}^d)$ estimates and a.e. differentiability of $\phi$, depending only on the summability of the right-hand side of the equation. In the context of mean field games, this amounts to requiring a sufficiently high growth condition on $f(x,m)$. Notice that their estimates are not owing to the variational structure of the mean field games system; in particular the density variable plays no role. A second direction in the search for Sobolev estimates for first order MFGs was initiated in [@ProSan] (see also [@San17]), where the authors obtained $H^1_{{\rm{loc}}}((0,T]\times{\mathbb{T}}^d)$ estimates for a well-chosen function of $m$, where $(\phi,m)$ is a weak solution of . Their technique is based on a so-called [*regularity by duality method*]{}, which relies on the fact that is the optimality condition for two convex optimization problems in duality. In fact, these techniques are inspired from [@CarMesSan] and more precisely they have their roots in [@Bre99] (see also [@AmbFig2]), where Y. Brenier introduced a similar approach to obtain regularity estimates for the pressure field arising from generalized solutions of the incompressible Euler equations. In a nutshell, the basic idea of this method is the following. One considers special competitors for the optimization problem involving the $m$ variable that are translations of the optimizers in time and space, then one compares the energy of these new competitors to the energy of the optimizers. Finally, assuming a sort of coercivity/monotonicity property of the running cost $f$ in the second variable (see the precise assumption in Sections \[sec:prelim\]-\[sec:stationary\]), one can obtain a differential quotient estimate for a function of $m$ via the difference of the energies and deduce the Sobolev estimate. A drawback of this method is that since the optimization problem is subject to the continuity equation with an initial condition, when one constructs the new competitors (via the time-space translations), one must preserve the initial condition, so it seems impossible to construct translations that allow to vary the initial time as well. We shall give more details on this method in Section \[sec:prelim\]. Let us mention that in the framework of first order models, P.-L. Lions in his lectures ([@Lions-course]) showed – at least formally – how to obtain a priori estimates in the case of the [*planning problem*]{} (where the initial and final density of the agents is prescribed). In particular, by rewriting the system as a very degenerate elliptic system in time and space, he obtains global $L^\infty$ estimates for the density variable $m$ and $W^{1,\infty}$ estimates for the value function $\phi$. However, his techniques require strong smoothness assumptions on the prescribed densities, Hamiltonian and coupling and positivity (and monotonicity) of the coupling function. Nevertheless, our objective in this paper is different: it is to obtain first order Sobolev estimates for the $m$ variable and second order Sobolev estimates for $\phi$. Thus, it seems that such an approach using degenerate elliptic equations to obtain these regularity estimates cannot be applied in our case. In the same context enters the recent paper [@LavSan], where the authors obtain local in time $L^\infty$ estimates for first order MFG models (with quadratic Hamiltonians), using variational techniques. Their approach, however, is very different from the one of Lions and from ours as well. Actually, the same techniques as in [@ProSan; @San17] can be used – more or less in a straight forward manner – for stationary problems (since there is no time involved there) as well, so as our preparatory results, we present how to gain Sobolev estimates for the solutions of . More precisely, we have \[thm:intro\_stat\] We assume the standard regularity and growth conditions on the data $f$ and $H$ (see Sections \[sec:prelim\] and \[sec:stationary\] for the precise assumptions). Let $(\phi,m,\l)$ be a solution of . Then, one has a second order Sobolev estimate for $\phi$, i.e. $$J_* (H(\cdot,\nabla\phi)+\l)\in H^1({\mathbb{T}}^d),$$ and a first order Sobolev estimate for $m$, i.e. $$J(m)\in H^1({\mathbb{T}}^d),$$ where $J,J_*:{\mathbb{R}}\to{\mathbb{R}}$ are functions depending on $f$ with a precise growth condition that depends on the growth condition of $f$ in its second variable. As our main result, we obtain global in time Sobolev estimates for weak solutions of first order MFG systems up to the initial (and final) time. While we keep the variational framework and rely on the convex duality, because of the obstruction described above, our approach is different from the one used in [@ProSan]. Our main results can be summarized in an informal way in the theorem below. \[thm:intro\_time\] We assume that $m_0,\phi_T \in W^{2,\infty}({{\mathbb T}}^d)$ and we assume some regularity and growth conditions on $H$ (which has a growth like $|\cdot|^r$ for some $r>1$ in its second variable) and $f$ (which has a growth like $|\cdot|^{q-1}$ for some $q>1$ in its second variable). We refer to Sections \[sec:prelim\] and \[sec:time\] for the precise assumptions. Then there exist a constant $C>0$ depending only on the data such that - $\|m^{\frac{q}{2} - 1}\nabla m\|_{L^2([0,T]\times{\mathbb{T}}^d)} \leq C;$ - $\|m^{1/2}D (j_1(\nabla \phi))\|_{L^2([0,T]\times{\mathbb{T}}^d)} \leq C;$ - if $r=2$, then $\|\partial_t (m^{q/2})\|_{L^1([0,T]\times{\mathbb{T}}^d)} \leq C.$ Here $j_1:{\mathbb{R}}^d\to{\mathbb{R}}^d$ is a function depending on $H$ and has a growth like $|\cdot|^{r/2}.$ The core idea of our technique is the following. The space translations of the optimizers will actually solve some auxiliary optimization problems, where the data are the space translations of the data of the original problems. This observation leads us to obtain fully global in time differential quotient estimates in the space variable. Then using the continuity equation – when the Hamiltonian has a quadratic growth in its second variable – the first order Sobolev estimates in time for $m$, globally in time, will be a consequence. Let us mention that while in [@ProSan] the authors considered only the case when $H(x,\xi)=\frac12|\xi|^2$, most of the computations seem to be adaptable to more general Hamiltonians to obtain [*local in time*]{} Sobolev estimates, but to gain estimates up to the final time $t=T$, it seems crucial for them that $D_\xi H$ is a Lipschitz function. In comparison with this, our techniques allow us to consider quite general Hamiltonians (with no restriction on the growth condition) to obtain the global in time estimates for the space variable. Nevertheless, for the global in time estimates in the time variable, it seems that we have a similar obstruction. Let us remark also that the estimate Theorem \[thm:intro\_time\]-(ii) seems to appear naturally in the context of Hamilton-Jacobi equations in connection with the nonlinear adjoint method introduced by L.C. Evans in [@Eva]. Nevertheless, we believe that it is interesting to see this kind of second order estimates in the framework of first order Hamilton-Jacobi equations arising from MFGs, especially if we take into consideration that solutions to these equations a priori had to be understood in a very weak sense. As a last remark, let us mention that it seems that our techniques to obtain global in time Sobolev estimates can be applied for other type of problems as well, in particular in the case of the first order [*planning problem*]{} (see [@Lions-course]). These estimates – together with other results on this kind of systems – is the subject of a subsequent work ([@GraMes2]). The structure of the paper is simple. In Section \[sec:prelim\] we collect the existing results on the well-posedness of first order variational MFG systems. The main assumptions, the convex duality results for the optimizations problems are also presented here. We also give a short summary on the main ideas of the technique “regularity by duality” presented in [@ProSan] in the framework of MFG. In Section \[sec:stationary\] we present the Sobolev estimates for first order stationary MFG systems and we prove Theorem \[thm:intro\_stat\]. Finally, Section \[sec:time\] is devoted to the proof of the global in time Sobolev estimates presented in Theorem \[thm:intro\_time\]. Preliminary results and the regularity by duality method {#sec:prelim} ======================================================== We list the standing assumptions which are sufficient to ensure the well-posedness of both and . [**Standing assumptions:** ]{} the following conditions on the data were used to get the result in [@CarGra]. 1. (Conditions on the Hamiltonian) $H:{{\mathbb T}}^d \times {{\mathbb R}}^d \to {{\mathbb R}}$ is continuous in both variables, convex and differentiable in the second variable, with $D_\xi H$ continuous in both variables. Moreover, $H$ has superlinear growth in the gradient variable: there exist $r >1 $ and $C >0$ such that $$\label{eq:hamiltonian_bounds} \frac{1}{rC}|\xi|^r-C \leq H(x,\xi) \leq \frac{C}{r}|\xi|^r + C,\ \ \forall\ (x,\xi)\in{\mathbb{T}}^d\times{\mathbb{R}}^d.$$ We denote by $H^*(x,\cdot)$ the Fenchel conjugate of $H(x,\cdot)$, which, due to the above assumptions, satisfies $$\label{eq:hamiltonian_conjugate_bounds} \frac{1}{r'C}|\zeta|^{r'}-C \leq H^*(x,\zeta) \leq \frac{C}{r'}|\zeta|^{r'} + C,\ \ \forall\ (x,\zeta)\in{\mathbb{T}}^d\times{\mathbb{R}}^d,$$ where $r' = r/(r-1)$ is the conjugate of $r$. 2. (Conditions on the coupling) Let $f$ be continuous on ${{\mathbb T}}^d \times (0,\infty)$, strictly increasing in the second variable, and there exists a constant $C>0$ satisfying $$\label{eq:coupling_growth} \frac{1}{C}|m|^{q-1} - C \leq f(x,m) \leq C|m|^{q-1} + C, ~~ \forall ~ m \geq 1, \forall\ x\in{\mathbb{T}}^d.$$ 3. (Conditions on the antiderivative of $f$) We define $F$ so that $F(x,\cdot)$ is an antiderivative of $f(x,\cdot)$ on $(0,\infty)$, that is, $$F(x,m) = \int_1^m f(x,s)\dd s, ~~ \forall ~ m > 0.$$ It follows that $F$ is continuous on ${{\mathbb T}}^d \times (0,\infty)$, is strictly convex and differentiable in the second variable, and satisfies the growth condition $$\label{eq:cost_growth} \frac{1}{qC}|m|^q - C \leq F(x,m) \leq \frac{C}{q}|m|^q + C,~~~ \forall ~ m \geq 1,\ \forall\ x\in{\mathbb{T}}^d.$$ For $m < 0$ we set $F(x,m) = +\infty$. We denote by $F(x,0)$ the limit $\lim_{m \to 0^+} F(x,m)$, which may be finite or $+\infty$. We will denote throughout the conjugate exponent of $q$ by $p = q'$. Define $F^*(x,\cdot)$ to be the Fenchel conjugate of $F(x,\cdot)$ for each $x\in{\mathbb{T}}^d$. Note that $$\label{eq:cost_growth_star} \frac{1}{pC}|a|^p - C \leq F^*(x,a) \leq \frac{C}{p}|a|^p + C, ~~ \forall a \geq 0,\ \forall\ x\in{\mathbb{T}}^d.$$ Preliminary results on the stationary system -------------------------------------------- Let us recall here the stationary MFG system. $$\label{MFGergo} \left\{\begin{array}{cll} {\rm{(i)}}& \lambda +H(x,D \phi) =f(x, m(x)), &{\rm{in}}\ {\mathbb{T}}^d,\\[5pt] {\rm{(ii)}} & -\nabla\cdot (mD_\xi H(x, \nabla \phi))=0, &{\rm{in}}\ {\mathbb{T}}^d, \\[5pt] {\rm{(iii)}}& {\displaystyle}\int_{{{\mathbb T}}^d} m\dd x= 1; \; {\displaystyle}\int_{{\mathbb{T}}^d}\phi\dd x=0;\; & m\geq 0,\ \ {\rm{a.e\ in\ }} {\mathbb{T}}^d. \end{array}\right.$$ \[def:weaksolergoMFG\] We say that a triple $(\lambda, \phi,m )\in {{\mathbb R}} \times W^{1, pr}({{\mathbb T}}^d)\times L^q({{\mathbb T}}^d)$ is a weak solution of if - $m\geq 0$ a.e. in ${\mathbb{T}}^d$, $\displaystyle \int_{{{\mathbb T}}^d}m\dd x=1$, ${\displaystyle}\int_{{\mathbb{T}}^d}\phi\dd x=0$ and $mD_\xi H(\cdot,\nabla \phi)\in L^{1}({{\mathbb T}}^d)$, - Equation -(i) holds in the following sense: $$\label{eq:aeergo} \displaystyle \quad \lambda +H(x,\nabla \phi(x))= f(x,m(x)) \quad \; \mbox{\rm a.e. in $\{m>0\}$}$$ and $$\label{eq:distribergo} \quad \lambda +H(x,\nabla \phi(x))\leq f(x,m) \quad \mbox{\rm a.e. in}\; {{\mathbb T}}^d,$$ - Equation -(ii) holds: $$\label{eqcontdefergo} \displaystyle \quad -\nabla\cdot( mD_\xi H(x,\nabla \phi))= 0\quad {\rm in }\; {{\mathbb T}}^d,$$ in the sense of distributions. In the sequel we summarize the results on the well-posedness of system . We define, on $\overline{ {{\mathcal K}}}_0:= {{\mathbb R}}\times W^{1,pr}({{\mathbb T}}^d)$, the functional $$\label{DefmathcalA} {\mathcal A}(\lambda,\phi)= \int_{{{\mathbb T}}^d} F^*\left(x,\lambda+H(x,\nabla \phi(x)) \right) \dd x - \lambda.$$ Our first optimization problem is $$\label{PB:dual2ergo} \inf_{(\lambda,\phi)\in \overline{ {{\mathcal K}}}_0} \mathcal A(\lambda,\phi)$$ To describe the second optimization problem, let us denote by $\overline{{{\mathcal K}}}_1$ the set of pairs $(m, w)\in L^1({{\mathbb T}}^d) \times L^1({{\mathbb T}}^d;{{\mathbb R}}^d)$ such that $m\geq 0$ a.e. in ${\mathbb{T}}^d$, $\displaystyle \int_{{{\mathbb T}}^d}m(x)\dd x=1$, and $w$ is divergence free, i.e. it satisfies in the sense of distributions $$\label{conteqergo} \nabla\cdot w=0,\ \ {\rm in}\ \ {{\mathbb T}}^d.$$ We define on $\overline{{{\mathcal K}}}_1$ the functional $${\mathcal B}(m,w)= \int_{{{\mathbb T}}^d} m(x) H^*\left(x, -\frac{w(x)}{m(x)}\right)+ F(x,m(x)) \dd x ,$$ where we use the usual convention to define $H^*(x,-w/m)$, i.e. $$\label{conventionH*} mH^*\left(x,-\frac{w}{m}\right) = \left\{ \begin{array}{ll} +\infty, & \mbox{\rm if $m=0$ and $w\neq 0$,} \\ 0, & \mbox{\rm if $m=0$ and $w = 0$.} \end{array} \right.$$ Since $H^*$ and $F$ are bounded below and $m\geq 0$ a.e., the integral in ${\mathcal B}(m,w)$ is well defined in ${{\mathbb R}}\cup\{+\infty\}$. The second optimal control problem is the following: $$\label{Pb:mw2ergo} \inf_{(m,w)\in \overline{{{\mathcal K}}}_1} \mathcal B(m,w)\;.$$ \[Lem:dualiteergo\] We have $$\label{minmax} \min_{(\lambda,\phi)\in \overline{ {{\mathcal K}}}_0}{\mathcal A}(\lambda,\phi) = - \min_{(m,w)\in \overline{{{\mathcal K}}}_1} {\mathcal B}(m,w).$$ Moreover, the minimum in the right-hand side is achieved by a unique pair $(m,w)\in \overline{{{\mathcal K}}}_1$ satisfying $(m,w)\in L^q( {{\mathbb T}}^d)\times L^{\frac{r'q}{r'+q-1}}( {{\mathbb T}}^d)$. \[theo:mainexergo\] There exists at least one solution $(\lambda,\phi,m)$ to the stationary MFG system . Moreover, the pair $(\lambda,m)$ is unique. If $(m,w)\in \overline{{{\mathcal K}}}_1$ is a minimizer of and $(\lambda, \phi)\in \overline{ {{\mathcal K}}}_0$ is a minimizer of , then $(\l,\phi,m)$ is a solution of the MFG system and $w= -mD_\xi H(\cdot,\nabla \phi)$ a.e.. Conversely, any solution $(\lambda,\phi, m)$ of is such that the pair $(m,-mD_\xi H(\cdot,\nabla \phi))$ is the minimizer of while $(\l,\phi)$ is a minimizer of . Preliminary results on the time dependent system ------------------------------------------------ Let us recall the time dependent MFG system that we will study. $$\label{eq:mfg}\left\{ \begin{array}{cll} {\rm{(i)}} & -\partial_t\phi + H(x,\nabla \phi) = f(x,m), & {\rm{in\ }} (0,T)\times{\mathbb{T}}^d, \\ (ii) & \partial_t m - \nabla\cdot \left(mD_\xi H(x,\nabla \phi)\right) = 0, & {\rm{in\ }} (0,T)\times{\mathbb{T}}^d, \\ (iii) & \phi(T,x) = \phi_T(x), m(0,x) = m_0(x), & {\rm{in\ }} {\mathbb{T}}^d. \end{array}\right.$$ First, we need to impose some additional standard assumption on the initial and final data. 1. (Conditions on the initial and final conditions) $m_0$ is a probability measure on ${{\mathbb T}}^d$ which is absolutely continuous with respect to Lebesgue measure, having density (which we also call $m_0$) in $C({{\mathbb T}}^d)$. We suppose moreover that $m_0>0$ on ${{\mathbb T}}^d$. We assume that $\phi_T:{{\mathbb T}}^d\to {{\mathbb R}}$ is a Lipschitz continuous function on ${{\mathbb T}}^d$. \[def:weak\] A pair $(\phi,m) \in BV((0,T) \times {{\mathbb T}}^d) \times L^q((0,T) \times {{\mathbb T}}^d)$ is called a weak solution to the system (\[eq:mfg\]) if it satisfies the following conditions. 1. $\nabla \phi\in L^r([0,T]\times{\mathbb{T}}^d)$ and the maps $mf(\cdot, m)$, $mH^*\left(\cdot,-D_\xi H(\cdot,\nabla \phi)\right) $ and $mD_\xi H(\cdot,\nabla \phi)$ are integrable, 2. $\phi$ satisfies a first-order Hamilton-Jacobi inequality $$\label{eq:hjb_weak} -\partial_t \phi + H(x,\nabla \phi) \leq f(x,m)$$ in the sense of distributions (tested against smooth non-negative functions), the boundary condition $\phi(T,\cdot)\leq \phi_T$ in the sense of trace and the following equality $$\begin{gathered} \int_0^T \int_{{{\mathbb T}}^d}m(t,x) \left[ H(x,\nabla \phi(t,x))-\nabla \phi(t,x)\cdot D_\xi H(x, \nabla \phi(t,x)) -f(x,m(t,x))\right]\dd x\dd t \\ = \int_{{{\mathbb T}}^d} (\phi_T(x)m(T,x))-\phi(0,x)m_0(x))\dd x \label{eq:ibp_weak} \end{gathered}$$ 3. $m$ satisfies the continuity equation $$\label{eq:continuity_weak} \partial_t m - \nabla\cdot(mD_\xi H(x,\nabla \phi)) = 0 ~~~\ {\rm{in}}~~(0,T) \times {{\mathbb T}}^d, ~~~ m(0,\cdot) = m_0,$$ in the sense of distributions. We consider two optimal control problems which are in duality, see [@Car15]. By the abuse of notion, we use the same notations ${{\mathcal A}}$ and ${{\mathcal B}}$ for the two functionals as in the stationary case. First, the control of the continuity equation reads as: minimize $${{\mathcal B}}(m,w) = \int_0^T \int_{{{\mathbb T}}^d} mH^*\left(x,-\frac{w}{m}\right) + F(x,m) \dd x \dd t + \int_{{{\mathbb T}}^d} \phi_T m(T) \dd x$$ over ${{\mathcal K}}_1 = \{(m,w) \in L^1({{\mathbb T}}^d) \times L^1({{\mathbb T}}^d;{{\mathbb R}}^d) : \partial_t m + \nabla \cdot w = 0, \ m(0) = m_0 \}$, where the continuity equation holds in the sense of distributions. (As usual, for $(m,w) \in {{\mathcal K}}_1$, the measure-valued function $[0,T]\ni t\mapsto m(t)\in{{\mathscr P}}({\mathbb{T}}^d)$ is continuous in time in the sense of weak (narrow) convergence of probability measures, cf. [@AmbCri]). Second, we control the Hamilton-Jacobi equation: minimize $${{\mathcal A}}(\phi) = \int_0^T \int_{{{\mathbb T}}^d} F^*(x,-\partial_t\phi + H(x,\nabla \phi))\dd x \dd t - \int_{{{\mathbb T}}^d} \phi(0)m_0 \dd x$$ over the set ${{\mathcal K}}_0$ consisting of maps $\phi \in C^1([0,T] \times {{\mathbb T}}^d)$ such that $\phi(T,\cdot) = \phi_T$. Set $E_1 = C^1([0,T] \times {{\mathbb T}}^d), E_0 = C([0,T] \times {{\mathbb T}}^d;{{\mathbb R}}) \times C([0,T] \times {{\mathbb T}}^d;{{\mathbb R}}^d)$. On $E_1$ we define $${{\mathcal F}}(\phi) = -\int_{{{\mathbb T}}^d} \phi(0)m_0 + \chi_{{{\mathcal K}}_0}(\phi).$$ On $E_0$ define $${{\mathcal G}}(a,b) = \int_0^T \int_{{{\mathbb T}}^d} F^*(x,-a(t,x) + H(x,b(t,x))) \dd x \dd t.$$ Then ${{\mathcal A}}$ can be written ${{\mathcal A}}(\phi) = {{\mathcal F}}(\phi) + {{\mathcal G}}(\Lambda(\phi))$ where $\Lambda : E_1 \to E_0$ is given by $\Lambda(\phi) = (\partial_t \phi,\nabla \phi)$. The two problems are in duality: \[thm:duality\] We have $$\label{eq:duality} \inf_{\phi \in {{\mathcal K}}_0} {{\mathcal A}}(\phi) = -\min_{(m,w) \in {{\mathcal K}}_1} {{\mathcal B}}(m,w).$$ Moreover, the minimum on the right-hand side is achieved by a unique pair $(m,w) \in {{\mathcal K}}_1$ which must satisfy $m \in L^q([0,T] \times {{\mathbb T}}^d)$ and $w \in L^{\frac{r'q}{r'+q-1}}([0,T] \times {{\mathbb T}}^d)$. Use the Fenchel-Rockafellar Theorem, which shows that $$\inf_\phi {{\mathcal F}}(\phi) + {{\mathcal G}}(\Lambda(\phi)) = \max_{(m,w)} -{{\mathcal G}}^*(-(m,w)) - {{\mathcal F}}^*(\Lambda^*(m,w)).$$ The right-hand side is just $-\min {{\mathcal B}}$. See [@Car15] for details. We can relax the second problem in the following way. ${{\mathcal K}}$ will be defined as the set of all pairs $(\phi,\alpha) \in BV([0,T]\times{\mathbb{T}}^d) \times L^1([0,T]\times{\mathbb{T}}^d)$ such that $\nabla \phi \in L^r([0,T] \times {{\mathbb T}}^d)$, $\phi(T,\cdot) \leq \phi_T$ in the sense of traces, $\alpha_+ \in L^p((0,T) \times {{\mathbb T}}^d)$, $\phi \in L^\infty([t,T]\times{{\mathbb T}}^d)$ for every $t \in (0,T)$, and $$\label{eq:inequality_in_distribution} -\partial_t \phi + H(x,\nabla \phi) \leq \alpha.$$ in the sense of distributions. Then on ${{\mathcal K}}$ we define the functional $${{\mathcal A}}(\phi,\alpha) = \int_0^T \int_{{{\mathbb T}}^d} F^*(x,\alpha(t,x))\dd x \dd t - \int_{{{\mathbb T}}^d} \phi(0,x)m_0(x)\dd x.$$ We have the following, due to [@CarGra]. ${\displaystyle}\inf_{\phi \in {{\mathcal K}}_0} {{\mathcal A}}(\phi) = \min_{(\phi,\alpha) \in {{\mathcal K}}} {{\mathcal A}}(\phi,\alpha)$, the latter attained by some $(\phi,\alpha) \in {{\mathcal K}}$ of which $\phi$ can be obtained as a limit of smooth functions $\phi_n$. Minimizers of ${{\mathcal A}}(\phi,\alpha)$ and ${{\mathcal B}}(m,w)$ precisely characterize weak solutions to the MFG system. \[thm:minimizers\_weak\] ${\rm{(i)}}$ If $(m,w) \in {{\mathcal K}}_1$ is a minimizer of ${{\mathcal B}}$ and $(\phi,\alpha) \in {{{\mathcal K}}}$ is a minimizer of ${{\mathcal A}}$, then $(\phi,m)$ is a weak solution of (\[eq:mfg\]) and $\alpha(t,x) = f(x,m(t,x))$ almost everywhere. ${\rm{(ii)}}$ Conversely, if $(\phi,m)$ is a weak solution of (\[eq:mfg\]), then there exist functions $w, \alpha$ such that $(\phi,\alpha) \in {{\mathcal K}}$ is a minimizer of ${{\mathcal A}}$ and $(m,w) \in {{\mathcal K}}_1$ is a minimizer of ${{\mathcal B}}$. ${\rm{(iii)}}$ If $(\phi,m)$ and $(\phi',m')$ are both weak solutions to (\[eq:mfg\]), then $m = m'$ almost everywhere while $\phi = \phi'$ almost everywhere in the set $\{m > 0\}$. Before we continue, let us remark that in [@CarGra], Theorem \[thm:minimizers\_weak\] holds only under the extra assumption $r > d(q-1)$. The reason for this assumption was to have a priori upper bounds on the solution $\phi$. However, in [@CarGraPorTon] it was shown that this is not necessary; one has in general a priori bounds on the positive part $\phi_+$ in $L^\eta$ for some $\eta > 1$. In particular, we have \[thm:a priori bounds for hj\] Suppose $\phi \in L^r((0,T);W^{1,r}({{\mathbb T}}^d))$ satisfies, in the sense of distributions, $$\left\{ \begin{array}{lrl} {\rm{(i)}} & -\partial_t \phi + c_0|\nabla \phi|^r \leq \alpha, & {\rm{in\ }}(0,T)\times{\mathbb{T}}^d,\\[5pt] {\rm{(ii)}} & \phi(T,x) \leq \phi_T(x), & {\rm{a.e.\ in\ }}{\mathbb{T}}^d, \end{array}\right.$$ for some $r > 1, c_0 > 0, \alpha \in L^p([0,T] \times {{\mathbb T}}^d)$ and $\phi_T \in L^\infty({{\mathbb T}}^d)$. Then there exists a constant $C = C(p,d,r,c_0,T,\|\alpha\|_{L^p([0,T] \times {{\mathbb T}}^d)},\|\phi_T\|_{L^\eta({{\mathbb T}}^d)})$ such that $$\|\phi_+\|_{L^\infty([0,T];L^\eta({{\mathbb T}}^d))} + \|\phi_+\|_{L^\gamma([0,T] \times {{\mathbb T}}^d)} \leq C$$ where $\eta = \frac{d(r(p-1)+1)}{d-r(p-1)}$ and $\gamma = \frac{rp(1+d)}{d-r(p-1)}$ if $p < 1 + \frac{d}{r}$ and $\eta = \gamma = +\infty$ if $p > 1 + \frac{d}{r}$. Using Theorem \[thm:a priori bounds for hj\], one can modify the proof of Theorem 2.9 in [@CarGra] to get existence of minimizers for the relaxed problem ${\displaystyle}\min_{(\phi,\alpha) \in {{\mathcal K}}} {{\mathcal A}}(\phi,\alpha)$. Indeed, let $\phi_n \in {{\mathcal K}}_0$ be a minimizing sequence, so that $$\int_0^T \int_{{{\mathbb T}}^d} F^*(x,\alpha_n)\dd x \dd t - \int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x \to \inf_{\phi \in {{\mathcal K}}_0} {{\mathcal A}}(\phi)$$ where $\alpha_n = -\partial_t\phi_n + H(x,\nabla\phi_n)$. For $\e\in{\mathbb{R}}$, set $K_\e = \sup_{x \in {{\mathbb T}}^d} F(x,\e)$ and note that $F^*(x,a) \geq a\e - K_\e$ for all $a\in{\mathbb{R}}$. By the bounds it follows that $$\label{eq:minimizer-bounds1} \int_0^T \int_{{{\mathbb T}}^d} F^*(x,\alpha_n)\dd x \dd t \geq \frac{1}{pC}\|(\alpha_n)_+\|_{L^p}^p - \e \int_0^T \int_{{{\mathbb T}}^d} (\alpha_n)_-\dd x \dd t - K_\e T - C.$$ Multiplying $(\alpha_n)_- = (\alpha_n)_+ + \partial_t \phi_n - H(x,\nabla \phi_n)$ by $m_0$ then integrating, using and the fact that $m_0$ is bounded above and below by positive constants, we deduce $$\begin{gathered} \label{eq:minimizers-bounds2} \int_0^T \int_{{{\mathbb T}}^d} (\alpha_n)_-\dd x \dd t \leq C\int_0^T \int_{{{\mathbb T}}^d}(\alpha_n)_+ m_0 \dd x \dd t + \int_{{{\mathbb T}}^d} \phi_T m_0 - \phi_n(0)m_0 \dd x + CT \\ \leq C\|(\alpha_n)_+\|_{L^p} - \int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x + C.\end{gathered}$$ Since $$\int_0^T \int_{{{\mathbb T}}^d} F^*(x,\alpha_n)\dd x \dd t - \int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x \leq C$$ we can combine and to get $$\frac{1}{pC}\|(\alpha_n)_+\|_{L^p}^p - \e C\|(\alpha_n)_+\|_{L^p} - (1 - \e) \int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x \leq C + K_\e T.$$ By fixing $\e > 0$ small enough and using once more, we obtain a priori upper bounds on both $\|(\alpha_n)_+\|_p$ and $-\int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x$. On the other hand, since $$\begin{gathered} -\int_{{{\mathbb T}}^d} \phi_n(t)m_0 \dd x \leq -\int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x + \int_0^t \int_{{{\mathbb T}}^d} (\alpha_n - H(x,\nabla \phi_n))m_0 \dd x \\ \leq -\int_{{{\mathbb T}}^d} \phi_n(0)m_0 \dd x + C\|(\alpha_n)_+\|_p + C,\end{gathered}$$ we also have an upper bound on $-\int_{{{\mathbb T}}^d} \phi_n(t)m_0 \dd x$ that holds uniformly in time. Combining this with Theorem \[thm:a priori bounds for hj\], we deduce that $\phi_n$ is bounded in $L^\infty([0,T];L^1({{\mathbb T}}^d))$. Moreover, implies that $(\alpha_n)_-$ is bounded in $L^1$. The rest of the proof now follows that of Theorem 2.9 in [@CarGra] and Proposition 5.4 in [@CarGraPorTon]. In particular, a modified version of $\alpha_n$ now has a weak limit $\alpha$ in $L^1([0,T]\times{\mathbb{T}}^d)$, and $\phi_n$ is seen to have a weak limit $\phi$ in $BV([0,T]\times{\mathbb{T}}^d)$, and one shows (cf. [@CarGraPorTon]) that $(\phi,\alpha) \in {{\mathcal K}}$ is a minimizer of ${{\mathcal A}}$. The above arguments imply in particular that both optimization problems, and hence the MFG system have a solution, without imposing the joint assumption on the growth condition of $f$ and $H$. The regularity by duality method -------------------------------- The idea to obtain Sobolev estimates by duality on the $m$ variable in the system used in [@ProSan] dates back to [@Bre99] (see also [@AmbFig2] and [@CarMesSan]) and it is as follows. Consider the two optimization problems involving the functionals ${{\mathcal A}}$ and ${{\mathcal B}}$ and suppose that these have minimizers $(\phi,\a)$ and $(m,w)$ respectively in the corresponding spaces as described above. The duality implies in particular that $${{\mathcal A}}(\phi,\a)+{{\mathcal B}}(m,w)=0.$$ Now, consider $\eta\in{\mathbb{R}}$ small and $\d\in{\mathbb{T}}^d$ such that $|\d|$ is small and construct some special competitors by translations for the problem involving ${{\mathcal B}}$, i.e. set $$(m^{\eta,\d},w^{\eta,\d})(t,x):=(m(t+\eta,x+\d),w(t+\eta,x+\d)).$$ Let us remark that in order to preserve in particular the initial condition $m(0,\cdot)=m_0$, one needs to use some special cutoff functions in time, when one considers time translations, but let us skip this detail in this informal description. Suppose that $F$ satisfies the strong coercivity condition . Then, the key part of the analysis is to show that $$\label{ineq:error} {{\mathcal B}}(m^{\eta,\d},w^{\eta,\d})\le {{\mathcal B}}(m,w)+ C(\eta^2+|\d|^2)$$ for some constant $C>0$ independent of $\eta$ and $\d$. Then, one can show some $H^1_{{\rm{loc}}}$ estimates for $m$. Indeed, first we will have $${{\mathcal A}}(\tilde\phi,\tilde\a)+{{\mathcal B}}(\tilde m,\tilde w)\ge \|J(\tilde m)-J_*(\tilde\a)\|^2_{L^2}$$ for any competitors $(\tilde\phi,\tilde\a)$ and $(\tilde m,\tilde w)$ respectively. This means in particular by the duality that for the optimizers $(\phi,\a)$ and $(m,w)$ one has $J(m)=J_*(\a).$ Second, by these last two observations if we add the quantity ${{\mathcal A}}(\phi,\a)$ to both sides of , one gets $$\|J(m)-J(m^{\eta,\d})\|^2_{L^2}\le C(\eta^2+|\d|^2),$$ thus the Sobolev estimate follows for $J(m)$. A warm-up: Sobolev estimates for the stationary problem {#sec:stationary} ======================================================= Our goal, in fact, is to show that the weak solutions $(\phi,m,\l)$ given by Theorem \[theo:mainexergo\] are such that for $\phi$ one can derive a second order Sobolev estimate, while in the same time one can obtain first order Sobolev estimates for $m$. These results are in the spirit of those in [@ProSan]. We will make some extra assumptions: $$\label{eq:H5} \begin{array}{ll} F^*(\cdot, \a)\in C^{1,1}({\mathbb{T}}^d),\ \forall\ \a\in {\mathbb{R}}, & D_xF^*(x,\cdot)\in{\rm{Lip_{loc}}}({\mathbb{R}}),\\ [5pt] D_\a F^*\in{\rm{Lip_{loc}}}({\mathbb{T}}^d\times{\mathbb{R}})\ \ {\rm{and}} & H(\cdot,\xi) \in C^{1,1}({\mathbb{T}}^d), \ \forall\ \xi\in{\mathbb{R}}^d. \end{array}\tag{H5}$$ For $h \in {{\mathbb T}}^d$, we define the translates $$\phi_h(x) := \phi(x-h).$$ Choose $(\lambda,\phi) \in \overline{{{\mathcal K}}_0}$ such that ${{\mathcal A}}(\lambda,\phi)$ is finite. Observe that $${{\mathcal A}}(\lambda,\phi_h) = \int_{{{\mathbb T}}^d} F^*(x,\lambda + H(x,\nabla \phi(x-h)))\dd x - \lambda = \int_{{{\mathbb T}}^d} F^*(x+h,\lambda + H(x+h,\nabla \phi(x)))\dd x - \lambda$$ is also finite. Suppose $(\lambda,\phi) \in \overline{{{\mathcal K}}_0}$ is such that ${{\mathcal A}}(\lambda,\phi)$ is finite. Let $A:{{\mathbb T}}^d \to {{\mathbb R}}$ be given by $A(h) := {{\mathcal A}}(\lambda,\phi_h)$. Then $A$ is $C^{1,1}({\mathbb{T}}^d)$. Observe that $$\begin{aligned} A(y+h)-A(y) &= \int_{{{\mathbb T}}^d} [F^*(x+h,\lambda + H(x+h,\nabla \phi_y(x))) - F^*(x,\lambda + H(x,\nabla \phi_y(x)))]\dd x \\ &= \int_{{{\mathbb T}}^d} D_xF^*(x,\lambda + H(x,\nabla \phi_y(x)))\cdot h\dd x\\ &+ \int_{{\mathbb{T}}^d} D_\a F^*(x,\lambda + H(x,\nabla \phi_y(x))) D_x H(x,\nabla \phi_y(x))\cdot h \dd x + o(|h|). \end{aligned}$$ It follows that $A \in C^1({\mathbb{T}}^d)$ with $$\begin{gathered} DA(h) = \int_{{{\mathbb T}}^d} [D_xF^*(x,\lambda + H(x,\nabla \phi_h(x))) + D_\a F^*(x,\lambda + H(x,\nabla \phi_h(x))) D_x H(x,\nabla \phi_h(x))] \dd x \\ = \int_{{{\mathbb T}}^d} [D_xF^*(x+h,\lambda + H(x+h,\nabla \phi(x))) + D_\a F^*(x+h,\lambda + H(x+h,\nabla \phi(x))) D_x H(x+h,\nabla \phi(x))] \dd x. \end{gathered}$$ By the assumption we see that $DA(h)$ is Lipschitz continuous, as desired. \[lem:h\^2\] Suppose $(\lambda,\phi) \in \overline{{{\mathcal K}}_0}$ is such that ${{\mathcal A}}(\lambda,\phi)$ is minimized. Then $A(h) = {{\mathcal A}}(\lambda,\phi_h) = {{\mathcal A}}(\lambda,\phi) + O(|h|^2)$. By the previous Lemma, we can expand $A(h) = A(0) + DA(0)h + O(|h|^2)$, and since $h = 0$ is a minimizer for $A$ we have $DA(0) = 0$. Now we are going to assume the following coercivity estimate on $F$: there exist $J,J_*:[0,+\infty)\to{\mathbb{R}}$ and $c_0>0$ such that $$\label{coercivity} F(x,m) + F^*(x,a) - ma \geq c_0|J(m) - J_*(a)|^2.\tag{H6}$$ Here the functions $J,J_*$ can be derived explicitly for many common examples of $F$ (see the examples in [@ProSan Section 3]). We deduce that $$\begin{aligned} {{\mathcal A}}(\lambda,\phi) + {{\mathcal B}}(m,w) &= \int_{{{\mathbb T}}^d} \left[F^*\left(x,\lambda+H(x,\nabla \phi(x)) \right) + m(x) H^*\left(x, -w(x)/m(x)\right)+ F(x,m(x))\right]\dd x - \lambda \\ &\geq \int_{{{\mathbb T}}^d} c_0|J(m)-J_*(\lambda + H(x,\nabla \phi(x)))|^2\dd x\\ &+ \int_{{\mathbb{T}}^d}m(x)\left[H(x,\nabla \phi(x)) + H^*\left(x, -w(x)/m(x)\right)\right]\dd x \\ &\geq \int_{{{\mathbb T}}^d} \left[c_0|J(m)-J_*(\lambda + H(x,\nabla \phi(x)))|^2 - \nabla \phi(x) \cdot w(x)\right]\dd x \\ &= c_0\int_{{{\mathbb T}}^d} |J(m)-J_*(\lambda + H(x,\nabla \phi(x)))|^2\dd x\end{aligned}$$ for any $(\lambda,\phi) \in \overline{ {{\mathcal K}}}_0$ and $(m,w) \in \overline{ {{\mathcal K}}}_1$ such that $w \in L^{(q'r)'}({{\mathbb T}}^d) = L^{r'q/(r'+q-1)}({{\mathbb T}}^d)$ (note that $\nabla\cdot w = 0$ in the sense of distributions, hence $\int_{{{\mathbb T}}^d} \nabla \phi(x) \cdot w(x)\dd x = 0$). In particular, suppose that $(\lambda,\phi) \in \overline{ {{\mathcal K}}}_0$ and $(m,w) \in \overline{ {{\mathcal K}}}_1$ are optimal, that is, ${{\mathcal A}}(\lambda,\phi)$ and ${{\mathcal B}}(m,w)$ are minimized. Consider the translates $\phi_h(x) := \phi(x-h)$. Then we have the estimates $$\begin{aligned} \|J_*(H(\cdot,\nabla \phi_h) + \lambda)& - J_*(H(\cdot,\nabla \phi) + \lambda) \|_{L^2({{\mathbb T}}^d)}^2 \\ &\leq 2\|J_*(H(\cdot,\nabla \phi_h) + \lambda) - J(m) \|_{L^2({{\mathbb T}}^d)}^2 + 2 \|J(m) - J_*(H(\cdot,\nabla \phi) + \lambda) \|_{L^2({{\mathbb T}}^d)}^2 \\ &\leq \frac{2}{c_0}\left({{\mathcal A}}(\lambda,\phi_h) + {{\mathcal B}}(m,w) \right) + \frac{2}{c_0} \left({{\mathcal A}}(\lambda,\phi) + {{\mathcal B}}(m,w) \right) \\ &= \frac{2}{c_0}\left({{\mathcal A}}(\lambda,\phi_h) - {{\mathcal A}}(\lambda,\phi) \right) = O(|h|^2),\end{aligned}$$ where in the last line we have used the duality result of Lemma \[Lem:dualiteergo\] and the estimate from Lemma \[lem:h\^2\]. This proves \[thm:stat\_main\] $J_*(H(\cdot,\nabla \phi) + \lambda)$ is in $H^1({{\mathbb T}}^d)$. Let us remark that the very same approach applies (even with simpler computations) to obtain Sobolev estimates for the $m$ variable. Indeed, let us assume that the following regularity conditions hold true. $$F(\cdot, m)\in C^{1,1}({\mathbb{T}}^d),\ \forall\ m\in[0,+\infty),\ \ {\rm{and}}\ \ H^*(\cdot,\zeta) \in C^{1,1}({\mathbb{T}}^d), \ \forall\ \zeta\in{\mathbb{R}}^d. \tag{H5'}$$ Then, considering any competitor $(m,w)$ such that ${{\mathcal B}}(m,w)$ is finite, for $h\in{\mathbb{T}}^d$ defining the translates $(m_h,w_h)$, it is easy to show that the map ${\mathbb{T}}^d\ni h\mapsto {{\mathcal B}}(m_h,w_h)$ is $C^{1,1}$. A similar reasoning as for the proof of Theorem \[thm:stat\_main\] yields the estimate for $J(m)$. More precisely, one has \[thm:stat\_main2\] $J(m)$ is in $H^1({\mathbb{T}}^d)$. Observe also that formally Theorem \[thm:stat\_main\] implies Theorem \[thm:stat\_main2\]. In fact, since $\l+H(x,\nabla\phi)=f(x,m)$, Theorem \[thm:stat\_main\] implies that $J_*(f(\cdot,m))$ is in $H^1({\mathbb{T}}^d)$. Since the coupling $f$ satisfying the growth condition , $J$ grows like $|\cdot|^{q/2}$ and $J_*$ has a growth like $|\cdot|^{q'/2}$, from where $J_*(f(\cdot,m))$ has exactly the same growth as $J(m)$. However, this reasoning stays formal, one of the main obstructions being that $\l+H(x,\nabla\phi)=f(x,m)$ is an equality a.e. only on $\{m>0\}$. Thus, in order to obtain the Sobolev estimate for $m$, one needs to use the regularity by duality machinery, as described above. Sobolev estimates for the time-dependent case {#sec:time} ============================================= Our goal in this section is to prove global in time Sobolev estimates for the time dependent system . We will see that the technique is quite different from the duality method used in the previous section. The main reason for this is that we need to perturb the data, and not just the solution. In this context, let us assume some extra monotonicity/coercivity conditions for $f$ and $H$. [**Additional assumptions**]{} 1. (Conditions on the coupling) Let $f$ be continuous on ${{\mathbb T}}^d \times (0,\infty)$, strictly increasing in the second variable, satisfying . Moreover, we will assume that $f(x,m)$ is Lipschitz with respect to $x$, specifically $$\label{f Lipschitz in x} |f(x,m) - f(y,m)| \leq Cm^{q-1}|x-y| \ \forall x,y \in {{\mathbb T}}^d, \ m \geq 0;$$ and also that $f(x,m)$ is strongly monotone in $m$, i.e. there exists $c_0 > 0$ such that $$\label{f strongly monotone} \left( f(x,\tilde m) - f(x,m)\right)(\tilde m - m) \geq c_0\min\{\tilde m^{q-2},m^{q-2}\}|\tilde m-m|^2 \ \forall \tilde m, m \geq 0, \ \tilde m \neq m.$$ (Note: if $q < 2$ one should interpret $0^{q-2}$ as $+\infty$ in . In this way, when $\tilde m = 0$, for instance, reduces to $f(x,m)m \geq c_0m^q$, as in the more regular case $q \geq 2$.) 2. (Coercivity assumptions.) We assume that there exist $j_1,j_2:{{\mathbb R}}^d \to {{\mathbb R}}^d$ and $c_0>0$ such that $$\label{eq:Hcoercivity} H(x,\xi) + H^*(x,\zeta) - \xi \cdot \zeta \geq c_0|j_1(\xi) - j_2(\zeta)|^2, \ \ \forall\ \xi,\zeta\in{\mathbb{R}}^d.$$ Notice that with this assumption we require that a sharper version of Young’s inequality (i.e. $H(x,\xi) + H^*(x,\zeta) - \xi \cdot \zeta \geq 0$) takes place. In particular, in light of , one can check that if $H(x,\xi)=\frac{1}{r}|\xi|^r$ ($r>1$), then holds true with the choices of $j_1(\xi) := |\xi|^{r/2-1}\xi$, $j_2(\zeta) := |\zeta|^{r'/2-1}\zeta$ and $c_0$ can be computed explicitly depending only on $r$. The following “canonical" examples satisfy all of the hypotheses on the data: $$f(x,m) = c_1(x)m^{q-1}, \ \ H(x,\xi) = c_2(x)|\xi|^r,$$ where $c_1,c_2$ are continuous functions bounded below by a positive constant, and $c_1$ is Lipschitz. In order to satisfy the hypotheses of Proposition \[prop:space-regularity\], we would also require $c_2$ to have $W^{2,\infty}$ regularity. Hypothesis (H7) is akin to (H6), in that the monotonicity estimates on $f(x,\cdot)$ are directly related to the convexity of $F(x,\cdot)$. In particular, if $F(x,m) = \frac{1}{q}c_1(x)m^q$ with $c_1$ as described in the previous remark, then (H7) is satisfied (because $f(x,m) = c_1(x)m^{q-1}$) while (H6) is satisfied with $J(m) = m^{q/2}$. Then the result of Proposition \[prop:space-regularity\] below can be interpreted as giving an $H^1$ estimate on $J(m)$. Regularity in space ------------------- Let $\delta \in {{\mathbb T}}^d$ be given. Here we will consider perturbations of the data which are translations in the space variable. To wit, we use the superscript $\delta$ to denote translation by $\delta$, i.e. $\alpha^\delta(t,x) = \alpha(t,x+\delta),F^\delta(x,m) = F(x+\delta,m),$ and so on for all other data. Accordingly, we will write $${{\mathcal A}}^\delta = {{\mathcal A}}_{F^\delta, m_0^\delta}, \ \ {{\mathcal B}}^\delta = {{\mathcal B}}_{ H^\delta, F^\delta, \phi_T^\d}, \ \ {{\mathcal K}}^\delta = {{\mathcal K}}( \phi_T^\delta, H^\delta), \ \ {{\mathcal K}}_1^\delta = {{\mathcal K}}_1( m_0^\delta).$$ Let $(\phi,\alpha) \in {{\mathcal K}}$ be a minimizer for ${{\mathcal A}}$ and $(m,w) \in {{\mathcal K}}_1$ be a minimizer for ${{\mathcal B}}$. To construct minimizers for ${{\mathcal A}}^\delta$ and ${{\mathcal B}}^\delta$, we can use translations in space of $(\phi,\alpha)$ and $(m,w)$. Indeed, we have that $(\phi^\delta,\alpha^\delta) \in {{\mathcal K}}^\delta$ and that ${{\mathcal A}}^\delta(\phi^\delta,\alpha^\delta) = {{\mathcal A}}(\phi,\alpha)$. One can then deduce that $(\phi^\delta,\alpha^\delta)$ is a minimizer for ${{\mathcal A}}^\delta$, since any competitor can be translated backwards to get a competitor for $(\phi,\alpha)$, which is a minimizer for ${{\mathcal A}}$. In the same way $(m^\delta,w^\delta)(t,x) = (m,w)(t,x+\delta)$ is a minimizer for ${{\mathcal B}}^\delta$. We will use these facts in the following proof. \[prop:space-regularity\] Let $m_0,\phi_T \in W^{2,\infty}({{\mathbb T}}^d)$, and assume that $H$ is twice continuously differentiable in $x$ with $$\label{eq:D_x^2 H} |D_x^2 H(x,\xi)| \leq C|\xi|^{r} + C.\tag{H9}$$ Then $\|m^{\frac{q}{2} - 1}\nabla m\|_{L^2([0,T]\times{\mathbb{T}}^d)} \leq C$ and $\|m^{1/2}D (j_1(\nabla \phi))\|_{L^2([0,T]\times{\mathbb{T}}^d)} \leq C$. *Step 1.* As above, let $(m,w) \in {{\mathcal K}}_1(m_0)$ be a minimizer for ${{\mathcal B}} = {{\mathcal B}}_{H,F,\phi_T}$. Initially we take a minimizing sequence of smooth functions $\phi_n \in {{\mathcal K}}_0(\phi_T)$ such that ${{\mathcal A}}(\phi_n) \to \inf {{\mathcal A}}$. Here ${{\mathcal A}} = {{\mathcal A}}_{F,m_0}$. Denoting the translates as above, we see that $\phi_n^\delta \in {{\mathcal K}}_0(\phi_T^\delta)$ is also a minimizing sequence for ${{\mathcal A}}^\delta = {{\mathcal A}}_{F^\delta,m_0^\delta}$. Following the argument in [@CarGra] we get that, up to a subsequence, $\phi_n \to \phi$ in $L^\gamma([0,T]\times{\mathbb{T}}^d)$ with $1\leq \gamma<d/(d-1)$, $\nabla \phi_n \rightharpoonup \nabla \phi$ weakly in $L^r([0,T]\times{\mathbb{T}}^d;{\mathbb{R}}^d)$ as $n\to +\infty$, and for $\alpha_n := -\partial_t \phi_n + H(x,\nabla \phi_n)$ we have $\tilde \alpha_n := \alpha_n \chi_{\{{\alpha_n} \geq -l_n\}} \to \alpha$ in $L^1([0,T]\times{\mathbb{T}}^d)$ as $n\to +\infty$, for any given sequence $(l_n)_{n\in\mathbb{N}}$ such that $l_n \to \infty$ as $n\to+\infty$; moreover, $(\phi,\alpha) \in {{\mathcal K}}$ is a minimizer of ${{\mathcal A}}$. In particular, for all $M > 0$ we can pass to a subsequence on which $\alpha_n \chi_{\{\alpha_n \geq -M\}} \to \alpha \chi_{\{\alpha \geq -M\}}$ in $L^p([0,T]\times{\mathbb{T}}^d)$. We now show that (up to a subsequence) $(\nabla \phi_n)$ converges weakly to $\nabla \phi$ in $L^r_m([0,T] \times {{\mathbb T}}^d;{{\mathbb R}}^d)$. To see this, first use $\phi_n$ as a test function in $\partial_t m + \nabla \cdot w = 0$ to get $$\label{eq:phi_n-test-m} \int_{{{\mathbb T}}^d} \phi_T m(T) - \phi_n(0)m_0 = \int_0^T \int_{{{\mathbb T}}^d} (H(x,\nabla \phi_n)-\alpha_n)m + \nabla \phi_n \cdot w \dd x\dd t.$$ Using we get $$\frac{1}{C}\int_0^T \int_{{{\mathbb T}}^d} |\nabla \phi_n|^r m \leq \|\phi_T\|_\infty + \int_{{{\mathbb T}}^d}|\phi_n(0)|m_0 + \int_0^T \int_{{{\mathbb T}}^d} (\alpha_n)_+ m + |\nabla \phi_n|m\left|\frac{w}{m}\right| \dd x\dd t,$$ then apply Young’s inequality to get $$\label{eq:bound-on-phi_n} \frac{1}{C}\int_0^T \int_{{{\mathbb T}}^d} |\nabla \phi_n|^r m \leq \|\phi_T\|_\infty + \int_{{{\mathbb T}}^d}|\phi_n(0)|m_0 + \int_0^T \int_{{{\mathbb T}}^d} (\alpha_n)_+ m + Cm\left|\frac{w}{m}\right|^{r'} \dd x\dd t$$ where, as usual, $C$ takes on a larger value. As argued in [@CarGraPorTon], $\limsup_{n \to \infty} \int_{{{\mathbb T}}^d}|\phi_n(0)|m_0 \leq C$. Since $(\alpha_n)_+$ is bounded in $L^p$ and $m\left|\frac{w}{m}\right|^{r'}$ is integrable, this means that the right-hand side of is bounded by a constant, as desired. Thus, we can assume that (up to a subsequence) $(\nabla \phi_n)$ converges weakly to some $\xi$ in $L^r_m([0,T] \times {{\mathbb T}}^d;{{\mathbb R}}^d)$. We now claim that $\xi = \nabla \phi$, $m$-a.e. in $[0,T]\times{\mathbb{T}}^d$. That is, we claim that $m\xi = m\nabla \phi$ a.e. It is enough to show that $$\int_0^T \int_{{{\mathbb T}}^d} \psi \cdot \nabla \phi_n \ m \to \int_0^T \int_{{{\mathbb T}}^d} \psi \cdot \nabla \phi \ m$$ for all $\psi \in C_c^\infty((0,T) \times {{\mathbb T}}^d; {{\mathbb R}}^d)$. Let $K > 0$ (large) and $\e > 0$ (small). Since $(\nabla\phi_n)$ converges weakly to $\nabla\phi$ in $L^r([0,T] \times {{\mathbb T}}^d;{{\mathbb R}}^d)$, we have $$\label{eq:nabla phi_n convergence} \iint_{\{m \leq K\}} \psi \cdot \nabla \phi_n \ m \dd x \dd t \to \iint_{\{m \leq K\}} \psi \cdot \nabla \phi \ m \dd x \dd t.$$ On the other hand, Young’s inequality yields $$\label{eq:zero} \left|\iint_{\{m {\geq} K\}} \psi \cdot \nabla \phi_n m \dd x \dd t\right| \leq \frac{\e^r}{r}\|\psi\|_\infty \int_0^T \int_{{{\mathbb T}}^d} |\nabla \phi_n|^r m \dd x \dd t + \frac{\e^{-r'}}{r'}\|\psi\|_\infty \iint_{\{m {\geq} K\}} m \dd x \dd t$$ so that, by , we have $$\limsup_{n \to \infty} \left|\iint_{\{m {\geq} K\}} \psi \cdot \nabla \phi_n m \dd x \dd t\right| \leq C\frac{\e^r}{r}\|\psi\|_\infty + \frac{\e^{-r'}}{r'}\|\psi\|_\infty \iint_{\{m {\geq} K\}} m \dd x \dd t.$$ Since $m$ is bounded in $L^q([0,T]\times{\mathbb{T}}^d)$, one has $$\iint_{\{m\ge K\}}m\dd x\dd t\le \left(\iint_{\{m\ge K\}}1\dd x\dd t\right)^{1/q'}\|m\|_{L^q},$$ thus Chebyshev’s inequality yields that $\iint_{\{m\ge K\}}m\dd x\dd t\to 0,$ as $K\to+\infty.$ Thus, we let $K \to \infty$ and then $\e \to 0$ to get in $$\limsup_{K \to \infty}\limsup_{n \to \infty} \left|\iint_{\{m {\geq} K\}} \psi \cdot \nabla \phi_n m \dd x \dd t\right| = 0,$$ In fact, by similar arguments, $$\limsup_{K \to \infty}\limsup_{n \to \infty} \left|\iint_{\{m {\geq} K\}} \psi \cdot \nabla \phi_n \ m \dd x \dd t-\iint_{\{m {\geq} K\}} \psi \cdot \nabla \phi \ m \dd x \dd t\right| = 0,$$ which, combined with , proves the claim. *Step 2.* Now use $\phi_n^\delta$ and $\phi_n$ as test functions in $\partial_t m + \nabla \cdot w = 0$ and $\partial_t m^\delta + \nabla \cdot w^\delta = 0$ respectively to get $$\label{eq:phi_n^delta-test-m} \int_{{{\mathbb T}}^d} \phi_T^\delta m(T) - \phi_n^\delta(0)m_0 = \int_0^T \int_{{{\mathbb T}}^d} (H(x+\delta,\nabla \phi_n^\delta)-\alpha_n^\delta)m + \nabla \phi_n^\delta \cdot w \dd x\dd t$$ and $$\label{eq:phi_n-test-m^delta} \int_{{{\mathbb T}}^d} \phi_T m^\delta(T) - \phi_n(0)m^\delta_0 = \int_0^T \int_{{{\mathbb T}}^d} (H(x,\nabla \phi_n)-\alpha_n)m^\delta + \nabla \phi_n \cdot w^\delta \dd x\dd t$$ Combine with and use the regularity of $H$ to get $$\begin{gathered} \label{eq:space-regularity1} \int_{{{\mathbb T}}^d} (\phi_T^\delta - \phi_T) m(T) - (\phi^\delta_n(0)-\phi(0))m_0 \\ = \int_0^T \int_{{{\mathbb T}}^d} (H(x+\delta,\nabla \phi^\delta_n) + H^*(x,-w/m) + \nabla \phi^\delta_n \cdot w/m - \alpha_n^\delta + f(m))m \dd x\dd t \\ =\int_0^T \int_{{{\mathbb T}}^d} (H(x,\nabla \phi^\delta_n) + H^*(x,-w/m) + \nabla \phi^\delta_n \cdot w/m - \alpha_n^\delta + f(m))m \dd x\dd t \\ + \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \langle D_x H(x + s\delta,\nabla \phi^\delta_n),\delta \rangle \dd s\dd x\dd t. \end{gathered}$$ Likewise, combine with (but applied to translates) to get $$\begin{gathered} \label{eq:space-regularity2} \int_{{{\mathbb T}}^d} (\phi_T - \phi_T^\delta) m^\delta(T) - (\phi_n(0)-\phi^\delta(0))m^\delta_0 \\ = \int_0^T \int_{{{\mathbb T}}^d} (H(x,\nabla \phi_n) + H^*(x+\delta,-w^\delta/m^\delta) + \nabla \phi_n \cdot w^\delta/m^\delta - \alpha_n + f^\delta(m^\delta))m^\delta \dd x\dd t \\ = \int_0^T \int_{{{\mathbb T}}^d} (H(x-\delta,\nabla \phi^{-\delta}_n) + H^*(x,-w/m) + \nabla \phi_n^{-\delta} \cdot w/m - \alpha_n^{-\delta} + f(m))m \dd x\dd t \\ = \int_0^T \int_{{{\mathbb T}}^d} (H(x,\nabla \phi^{-\delta}_n) + H^*(x,-w/m) + \nabla \phi_n^{-\delta} \cdot w/m - \alpha_n^{-\delta} + f(m))m \dd x\dd t \\ - \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \langle D_x H(x - s\delta,\nabla \phi^{-\delta}_n),\delta \rangle \dd s\dd x\dd t. \end{gathered}$$ Note that, by the changes of variables $x \mapsto x + \delta$ for the first integral and $x \mapsto x - \delta$ for the second, followed by the translation $s \mapsto 1-s$, we get $$\begin{gathered} \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \langle D_x H(x + s\delta,\nabla \phi^{\delta}_n),\delta \rangle \dd s\dd x\dd t - \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \langle D_x H(x - s\delta,\nabla \phi^{-\delta}_n),\delta \rangle \dd s\dd x\dd t \\ = \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \langle D_x H(x + s\delta,\nabla \phi_n) - D_x H(x - s\delta,\nabla \phi_n),\delta \rangle \dd s\dd x\dd t \\ = \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \int_{-s}^s \langle D_{xx}^2 H(x + r\delta,\nabla \phi_n)\delta,\delta \rangle \dd r\dd s\dd x\dd t. \end{gathered}$$ Then adding together and we get $$\begin{gathered} \label{eq:space-regularity3} \int_0^T \int_{{{\mathbb T}}^d} \left(H(x,\nabla \phi^\delta_n) +H(x,\nabla \phi^{-\delta}_n)+ 2H^*(x,-w/m) + \nabla \phi^\delta_n \cdot w/m+\nabla \phi^{-\delta}_n \cdot w/m\right)m \dd x\dd t \\ = \int_{{{\mathbb T}}^d} \left(\phi_T^\delta + \phi_T^{-\delta} - 2\phi_T\right)m(T)\dd x - \int_{{{\mathbb T}}^d} (\phi_n(0)(m_0^\delta + m_0^{-\delta}) - 2\phi(0)m_0)\dd x \\ + \int_0^T \int_{{{\mathbb T}}^d} \left(\alpha_n^\delta + \alpha_n^{-\delta} - 2f(m)\right)m \dd x \dd t + \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \int_{-s}^s \langle D_{xx}^2 H(x + r\delta,\nabla \phi_n)\delta,\delta \rangle \dd r\dd s\dd x\dd t. \end{gathered}$$ Now we pass to the limit as $n\to+\infty$ in each term above. First, we since $H$ is convex in the second variable and $(\nabla \phi_n^{\pm\delta})$ converges weakly to $\nabla \phi^{\pm\delta}$ in $L^r_m([0,T]\times{\mathbb{T}}^d;{\mathbb{R}}^d)$, by weak lower semicontinuity we get $$\label{ineq:liminf1} \int_0^T \int_{{{\mathbb T}}^d} \left(H(x,\nabla \phi^\delta) +H(x,\nabla \phi^{-\delta})\right)m \dd x\dd t\le\liminf_{n\to+\infty}\int_0^T \int_{{{\mathbb T}}^d} \left(H(x,\nabla \phi^\delta_n) +H(x,\nabla \phi^{-\delta}_n)\right)m \dd x\dd t.$$ Then, by the previous weak convergence and by the fact that $w/m\in L^{r'}_m([0,T]\times{\mathbb{T}}^d;{\mathbb{R}}^d)$, we have that $$\label{ineq:liminf2} \int_0^T \int_{{{\mathbb T}}^d} \left(\nabla \phi^\delta \cdot w/m+\nabla \phi^{-\delta} \cdot w/m\right)m \dd x\dd t=\lim_{n\to+\infty}\int_0^T \int_{{{\mathbb T}}^d} \left(\nabla \phi^\delta_n \cdot w/m+\nabla \phi^{-\delta}_n \cdot w/m\right)m \dd x\dd t$$ Second, let us compute $$\begin{gathered} \label{eq:space-regularity4} \liminf_{n \to \infty} \left\{\int_{{{\mathbb T}}^d} \left(\phi_T^\delta + \phi_T^{-\delta} - 2\phi_T\right)m(T)\dd x - \int_{{{\mathbb T}}^d} \left(\phi_n(0)(m_0^\delta + m_0^{-\delta}) - 2\phi(0)m_0\right)\dd x \right. \\ + \left. \int_0^T \int_{{{\mathbb T}}^d} \left(\alpha_n^\delta + \alpha_n^{-\delta} - 2f(m)\right)m \dd x \dd t + \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \int_{-s}^s \langle D_{xx}^2 H(x + r\delta,\nabla \phi_n)\delta,\delta \rangle \dd r\dd s\dd x\dd t \right\}.\\ \leq \limsup_{n \to \infty} \left\{\int_{{{\mathbb T}}^d} \left(\phi_T^\delta + \phi_T^{-\delta} - 2\phi_T\right)m(T)\dd x - \int_{{{\mathbb T}}^d} \left(\phi_n(0)(m_0^\delta + m_0^{-\delta}) - 2\phi(0)m_0\right)\dd x \right. \\ + \left. \int_0^T \int_{{{\mathbb T}}^d} \left(\alpha_n^\delta + \alpha_n^{-\delta} - 2f(m)\right)m \dd x \dd t + \int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \int_{-s}^s \langle D_{xx}^2 H(x + r\delta,\nabla \phi_n)\delta,\delta \rangle \dd r\dd s\dd x\dd t \right\}. \end{gathered}$$ Let us recall that $(\phi_n)_{n\ge 0}$ converges in $L^1([0,T]\times{\mathbb{T}}^d)$ to $\phi\in BV([0,T]\times{\mathbb{T}}^d)$ and $(\phi_n(0))_{n\ge 0}$ is bounded in $L^1({\mathbb{T}}^d)$. In particular $\partial_t\phi_n{\rightharpoonup}\partial_t\phi$, as $n\to+\infty$, weakly-$\star$ in ${{\mathscr M}}([0,T]\times{\mathbb{T}}^d).$ These imply furthermore that $$\label{ineq:liminf3} \limsup_{n\to+\infty}- \int_{{{\mathbb T}}^d} \phi_n(0)(m_0^\delta + m_0^{-\delta}) \dd x\le - \int_{{{\mathbb T}}^d} \phi(0)(m_0^\delta + m_0^{-\delta})\dd x$$ Indeed, on the one hand we have $$\begin{aligned} -\int_{{\mathbb{T}}^d}\phi_n(0,x)\psi(x)\dd x&=-\int_{{\mathbb{T}}^d}\phi_T(x)\psi(x)\dd x+\int_0^T\int_{{\mathbb{T}}^d}\partial_t\phi_n\psi\dd x\dd t\\ &\to-\int_{{\mathbb{T}}^d}\phi_T(x)\psi(x)\dd x+\int_0^T\int_{{\mathbb{T}}^d}\partial_t\phi\psi\dd x\dd t\\ &=\int_{{\mathbb{T}}^d}(\phi(T,x)-\phi_T(x))\psi(x)\dd x-\int_{{\mathbb{T}}^d}\phi(0,x)\psi(x)\dd x\end{aligned}$$ as $n\to+\infty$ for all $\psi\in C({\mathbb{T}}^d)$. On the other hand, since $\phi(T,x)\le\phi_T(x)$ for a.e. $x\in{\mathbb{T}}^d$, for $\psi\ge 0$ one can conclude that $$\limsup_{n\to+\infty}- \int_{{{\mathbb T}}^d} \phi_n(0)\psi \dd x\le - \int_{{{\mathbb T}}^d} \phi(0)\psi\dd x,$$ which implies as desired. As for the third term, we take an arbitrary $M > 0$ and get $$\begin{aligned} \limsup_{n\to \infty} \int_0^T \int_{{{\mathbb T}}^d} &\left(\alpha_n^\delta + \alpha_n^{-\delta} - 2f(m)\right)m \dd x \dd t\\ &\leq \limsup_{n\to \infty} \int_0^T \int_{{{\mathbb T}}^d} \left(\alpha_n^\delta \chi_{\{\alpha_n^\delta \geq -M\}} + \alpha_n^{-\delta}\chi_{\{\alpha_n^{-\delta} \geq -M\}} - 2f(m)\right)m \dd x \dd t \\ &= \int_0^T \int_{{{\mathbb T}}^d} \left(f^\delta(m^\delta)\chi_{\{f^\delta(m^\delta) \geq -M\}} + f^{-\delta}(m^{-\delta})\chi_{\{f^{-\delta}(m^{-\delta}) \geq -M\}} - 2f(m)\right)m \dd x \dd t, \end{aligned}$$ where we used the fact (see for instance [@CarGra Theorem 3.5-(i)]) that since $(\phi,\alpha)$ is a minimizer of the primal problem and $(m,w)$ is a minimizer of the dual problem, one has that $\alpha=f(\cdot,m)$ a.e. This, after letting $M \to \infty$ becomes $$\limsup_{n\to \infty} \int_0^T \int_{{{\mathbb T}}^d} \left(\alpha_n^\delta + \alpha_n^{-\delta} - 2f(m)\right)m \dd x \dd t \leq \int_0^T \int_{{{\mathbb T}}^d} \left(f^\delta(m^\delta) + f^{-\delta}(m^{-\delta}) - 2f(m)\right)m \dd x \dd t.$$ Lastly, by we can assert $$\label{eq:space-regularity8} \left|\int_0^T \int_{{{\mathbb T}}^d} \int_0^1 \int_{-s}^s \langle D_{xx}^2 H(x + r\delta,\nabla \phi_n)\delta,\delta \rangle \dd r\dd s\dd x\dd t \right| \leq C(\|\nabla \phi_n\|_{r}^{r} + 1)|\delta|^2 \leq C|\delta|^2.$$ Thus, when passing to the $\liminf$ in as $n\to+\infty$, the previous inequalities and estimations imply, $$\begin{gathered} \label{eq:space-regularity3-limit} \int_0^T \int_{{{\mathbb T}}^d} \left(H(x,\nabla \phi^\delta) + H^*(x,-w/m) + \nabla \phi^\delta \cdot w/m\right)m \dd x\dd t\\ +\int_0^T\int_{{\mathbb{T}}^d}\left(H(x,\nabla \phi^{-\delta})+ H^*(x,-w/m) +\nabla \phi^{-\delta} \cdot w/m\right)m \dd x\dd t \\ \le \int_{{{\mathbb T}}^d} \left(\phi_T^\delta + \phi_T^{-\delta} - 2\phi_T\right)m(T)\dd x - \int_{{{\mathbb T}}^d} \phi(0)(m_0^\delta + m_0^{-\delta} - 2m_0)\dd x \\ + \int_0^T \int_{{{\mathbb T}}^d} \left(f^\delta(m^\delta) + f^{-\delta}(m^{-\delta}) - 2f(m)\right)m \dd x \dd t + C|\delta|^2 \end{gathered}$$ Now, the assumption on $H$ together with on the one hand and the inequality $|a+b|^2 \leq 2(|a|^2+|b|^2),\ \forall a,b\in{\mathbb{R}}^d$ on the other hand yield $$\begin{gathered} \label{eq:space-regularity5} \frac{c_0}{2}\int_0^T \int_{{{\mathbb T}}^d} \left(|j_1(\nabla \phi^\delta) - j_1(\nabla \phi^{-\delta})|^2\right) m \dd x \dd t \\ \leq \int_{{{\mathbb T}}^d} \left(\phi_T^\delta + \phi_T^{-\delta} - 2\phi_T\right)m(T)\dd x - \int_{{{\mathbb T}}^d} \phi(0)\left(m_0^\delta + m_0^{-\delta} - 2m_0\right)\dd x \\ + \int_0^T \int_{{{\mathbb T}}^d} \left(f^\delta(m^\delta) + f^{-\delta}(m^{-\delta}) - 2f(m)\right)m \dd x \dd t + C|\delta|^2. \end{gathered}$$ *Step 3.* We estimate the terms in . First, we have $$\label{eq:space-regularity6} \int_{{{\mathbb T}}^d} \phi(0)\left(2m_0 - m_0^\delta-m_0^{-\delta}\right)\dd x \leq C|\delta|^2\int_{{{\mathbb T}}^d} |\phi(0)|\dd x,$$ where $C$ depends on $\|m_0\|_{W^{2,\infty}}$, and we know that $\int_{{{\mathbb T}}^d} |\phi(0)|\dd x$ is well-defined and finite (see [@CarGraPorTon Lemma 5.1]). Similarly, $$\label{eq:space-regularity7} \int_{{{\mathbb T}}^d} m(T)\left(2\phi_T - \phi_T^\delta-\phi_T^{-\delta}\right)\dd x \leq C|\delta|^2\int_{{{\mathbb T}}^d} m(T)\dd x=C|\delta|^2,$$ where $C$ depends on $\|\phi_T\|_{W^{2,\infty}}$. There is one more term to estimate. We rewrite it using change of variables and then apply Equations and to get $$\begin{gathered} \label{eq:space-regularity9} \int_0^T \int_{{{\mathbb T}}^d} (f^\delta(m^\delta) + f^{-\delta}(m^{-\delta}) - 2f(m))m \dd x \dd t = -\int_0^T \int_{{{\mathbb T}}^d} (f^\delta(m^\delta) - f(m))(m^{\delta} - m) \dd x \dd t \\ = - \iint_{\{m^\delta \leq m\}} (f^\delta(m^\delta) - f(m^\delta))(m^{\delta} - m) \dd x \dd t - \iint_{\{m^\delta \leq m\}} (f(m^\delta) - f(m))(m^{\delta} - m) \dd x \dd t \\ - \iint_{\{m < m^\delta\}} (f^\delta(m^\delta) - f^\delta(m))(m^{\delta} - m) \dd x \dd t - \iint_{\{m < m^\delta\}} (f^\delta(m) - f(m))(m^{\delta} - m) \dd x \dd t \\ \leq C\int_0^T \int_{{{\mathbb T}}^d} |\delta|\min\{(m^\delta)^{q-1},m^{q-1}\}|m^{\delta} - m| \dd x \dd t - c_0\int_0^T \int_{{{\mathbb T}}^d} \min\{(m^\delta)^{q-2},m^{q-2}\}|m^\delta - m|^2 \dd x \dd t \\ \leq C|\delta|^2 \int_0^T \int_{{{\mathbb T}}^d} \min\{m^\delta,m\}^{q} \dd x \dd t - \frac{c_0}{2}\int_0^T \int_{{{\mathbb T}}^d} \min\{(m^\delta)^{q-2},m^{q-2}\}|m^\delta - m|^2 \dd x \dd t, \end{gathered}$$ where, we used Young’s inequality in the last inequality, and the expression $\min\{(m^\delta)^{q-2},m^{q-2}\}|m^\delta - m|^2$ is treated as zero whenever $m^\delta = m$ (even in the case $q < 2$). Since $$\int_0^T \int_{{{\mathbb T}}^d} \min\{m^\delta,m\}^{q} \dd x \dd t \leq \int_0^T \int_{{{\mathbb T}}^d} m^q \dd x \dd t \leq C,$$ we see from plugging , , and into that $$\begin{gathered} \frac{c_0}{2}\int_0^T \int_{{{\mathbb T}}^d} \left(|j_1(\nabla \phi^\delta) - j_1(\nabla \phi^{-\delta})|^2\right) m \dd x \dd t \\ + \frac{c_0}{2}\int_0^T \int_{{{\mathbb T}}^d} \min\{(m^\delta)^{q-2},m^{q-2}\}|m^\delta - m|^2 \dd x \dd t \leq C|\delta|^2, \end{gathered}$$ where $C$ depends only on the data. The result now follows from dividing both sides by $|\delta|^2$ and letting $\delta \to 0$. As a potential application of Proposition \[prop:space-regularity\], one might hope to derive second order in space regularity for $\phi$. For this, one would like some summability estimate on $m^{-1}$. Since, thanks to [@CarGra], the growth of $f(x,m)$ is unrestricted for small values of $m$, all of our results are still valid for data satisfying, for example, $F(x,m) \geq \frac{1}{C}m^{-s}- C$ for some $s \geq 1$, thus providing the desired estimate. Such a case would correspond to extreme congestion penalization, in which players experience a great benefit by moving toward unoccupied spaces. In this context, the following corollary is meaningful. Suppose $m^{-1} \in L^s([0,T]\times{\mathbb{T}}^d)$ for some $s \geq 1$. Then $$\label{eq:space-regularity-j_1(nabla phi_n)} \|\nabla (j_1(\nabla \phi))\|_{\frac{2s}{s+1}}^2 \leq C\|m^{-1}\|_{s},$$ where, for simplicity of writing $\|\cdot\|_{\eta}$ denotes the norm $\|\cdot\|_{L^\eta([0,T]\times{\mathbb{T}}^d)}$. Observe that by Hölder’s inequality and Proposition \[prop:space-regularity\], we have $$\|j_1(\nabla \phi^\delta) - j_1(\nabla \phi^{-\delta})\|_{\frac{2s}{s+1}}^2 \leq \|m^{-1}\|_{s}\int_0^T \int_{{{\mathbb T}}^d} (|j_1(\nabla \phi^\delta) - j_1(\nabla \phi^{-\delta})|^2) m \dd x \dd t\le C\|m^{-1}\|_{s}|\d|^2,$$ which implies the thesis of this corollary dividing both sides by $|\d|^2$ and taking $\d\to 0$. Some time regularity -------------------- Assume that $r = 2$, so that the Hamiltonian is of quadratic growth. In line with this assumption, we will assume that $$\label{eq:D^2 H bounded} |D_{pp}^2 H(x,\xi)| \leq C, \ |D_{xp}^2 H(x,\xi)| \leq C(|\xi| + 1) \ \ \forall x \in {{\mathbb T}}^d, \forall \xi \in {{\mathbb R}}^d.\tag{H10}$$ Then the result of Proposition \[prop:space-regularity\] reads $$\label{eq:space-regularity-r=2} \|m^{q/2 -1}\nabla m\|_{L^2([0,T]\times{\mathbb{T}}^d)}, \|m^{1/2}D^2 \phi\|_{L^2([0,T]\times{\mathbb{T}}^d)} \leq C.$$ Recalling that $w = -mD_\xi H(x,\nabla\phi)$, we compute $$-\nabla \cdot w = \nabla m \cdot D_\xi H(x,\nabla\phi) + m \ \mathrm{tr} (D_{xp}^2 H(x,\nabla \phi)) + m \ \mathrm{tr}(D_{pp}^2 H(x,\nabla \phi)D^2 \phi)$$ and deduce from that $$m^{q/2 -1}|\nabla \cdot w| \leq Cm^{q/2 -1}|\nabla m|(|\nabla \phi| + 1) + Cm^{q/2}(|\nabla \phi| + 1) + Cm^{q/2 -1/2}m^{1/2}|D^2 \phi|.$$ Now since $\|H(x,\nabla \phi)\|_{L^1} \leq C$, we have that $|\nabla \phi| \in L^2([0,T]\times{\mathbb{T}}^d)$ by . Moreover, since $q > 1$ and $\|m\|_{L^q} \leq C$ it follows (by Hölder’s inequality) that $\|m^{q/2 -1/2}\|_{L^2} \leq C$. Therefore implies, by Hölder’s inequality, that $$\label{eq:div w regularity} \|m^{q/2 -1}\nabla \cdot w\|_{L^1([0,T]\times{\mathbb{T}}^d)} \leq C.$$ We deduce the following: \[Global $L^1$ bounds on $\partial_t m$\]\[cor:time-reg\] Assume $r = 2$ and that holds. Suppose $(m,w)$ is the minimizer of ${{\mathcal B}}$. Then there exist $C>0$ depending only on the data such that $$\|\partial_t (m^{q/2})\|_{L^1([0,T]\times{\mathbb{T}}^d)} \leq C.$$ Apply to the equation $\partial_t m = -\nabla \cdot w$. The argument breaks down if $r \neq 2$. For simplicity, set $H(x,\xi) = \frac{1}{r}|\xi|^r$. The result of Proposition \[prop:space-regularity\] reads $$\label{eq:space-regularity-r neq 2} \|m^{q/2 -1}\nabla m\|_{L^2}, \|m^{1/2}{(|\nabla\phi|^{r/2-1}D^2 \phi+(r/2-1)|\nabla\phi|^{r/2-2}D^2\phi\nabla\phi\otimes\nabla\phi)}\|_{L^2} \leq C.$$ We compute $$-\nabla \cdot w = \nabla m \cdot |\nabla \phi|^{r-2}\nabla \phi + m |\nabla \phi|^{r-2}\Delta \phi+{m(r-2)|\nabla\phi|^{r-4}(D^2\phi\nabla\phi)\cdot\nabla\phi}.$$ If $r < 2$, then the last two terms of the r.h.s. are degenerate: it is possible that $m|\nabla \phi|^{r-2}\Delta \phi$ is not integrable, since even though one would have $L^2$ regularity for $|\nabla \phi|^{r/2-1}D^2 \phi$, there are no estimates on the remaining factor of $m|\nabla \phi|^{r/2-1}$. On the other hand, if $r > 2$, then the first term on the r.h.s. is problematic: $|\nabla \phi|^{r-2}\nabla \phi \in L^{r'}([0,T]\times{\mathbb{T}}^d)$, but $r' < 2$, and thus $L^2$-regularity for $\nabla (m^{q/2})$ is not enough to ensure that the product $\nabla (m^{q/2}) \cdot |\nabla \phi|^{r-2}\nabla \phi$ is integrable. [Acknowledgements]{} We thank Filippo Santambrogio for the fruitful discussions during this project. In particular, some discussions with him and his paper [@San17] inspired us in writing this paper. We thank also Pierre Cardaliaguet for his remarks and valuable suggestions on the manuscript. The first author was supported by the National Science Foundation through Grant DMS-1612880. The second author was partially supported by the Gaspar Monge Program for Optimization and Operation Research (PGMO) via the project [*VarMFGPDE*]{}. He would like to thank the hospitality of Tohoku University and Tohoku Forum for Creativity, Sendai, Japan in the framework of the Thematic Program on [*Nonlinear Partial Differential Equations for Future Applications*]{} in Summer 2017.

ArXiv

--- address: - 'Biostatistics Group, Department of Health Sciences, University of Leicester, Leicester, UK' - 'Genetic Epidemiology Group, Department of Health Sciences, University of Leicester, Leicester, UK' - 'Roche Innovation Centre, F. Hoffmann-La Roche Ltd, Basel, Switzerland' - 'Institute of Heath Research, University of Exeter Medical School, Exeter, UK' - 'Centre for Research on Health and Social Care Management, Bocconi University, Milan, Italy' author: - Tasos Papanikos - John Thompson - Keith Abrams - 'Nicolas St[ä]{}dler' - Oriana Ciani - Rod Taylor - Sylwia Bujkiewicz bibliography: - 'bibli.bib' title: 'A Bayesian hierarchical meta-analytic method for modelling surrogate relationships that vary across treatment classes using aggregate data' --- Introduction {#Introduction} ============ New advances in science have led to discovering of promising therapies which often are targeted to specific patient populations, for example defined by a genetic biomarker. This leads to clinical trials of smaller size, whilst the increased effectiveness of these therapies reduces the number of events or deaths and consequently lead to measurement of treatment effect on overall survival (OS) with large uncertainty. Therefore, surrogate endpoints allowing the measurement of treatment effect with higher precision have been investigated to accelerate the availability of these treatments to the patients. These alternative endpoints often can be considered a cost effective replacement of final clinical outcome, as they are particularly useful when they can be measured earlier, easier, more frequently compared to the final clinical endpoint or if they require smaller sample size and shorter follow up times [@burzykowski2006evaluation]. Potential surrogate endpoints have been investigated as candidate endpoints in clinical trials in a number of disease areas. However, before these candidate endpoints are used, either as primary endpoints in trial design or in regulatory decision making, they need to be validated [@fleming2012biomarkers]. In practice, the most common approach to validate a candidate outcome is to examine whether it satisfies three levels of association, proposed by the International Conference on Harmonisation Guidelines on Statistical Principles for Clinical Trials (ICH) [@phillips2003e9]. First, the biological plausibility of the association of the surrogate and final outcomes is investigated which, involves biological rather than statistical considerations. Furthermore, the individual level association is evaluated to establish whether the candidate surrogate endpoint can be used to predict the course of the disease in an individual patient. Last but not least, the study level association is investigated to ensure that the treatment effects on the final outcome can be predicted from the effect on the surrogate endpoint. Study level association requires data from a number of randomised controlled trials (RCTs) and can be investigated carrying out a bivariate meta-analysis [@daniels1997meta; @buyse2000validation; @bujkiewicz2015uncertainty; @lassere2008biomarker]. In this paper we focus on the third level of association only. A bivariate meta-analytical method that was developed by Daniels & Hughes [@daniels1997meta] can be used to validate a candidate surrogate endpoint, by evaluating the association pattern between the treatment effects on the surrogate and the final outcomes, and to predict treatment effects on the final clinical outcome from the effects on surrogate endpoint. This method, implemented in a Bayesian framework, can be used to evaluate a surrogate endpoint in a disease area overall, or in each treatment class separately through a subgroup analysis. Traditionally, surrogate relationships between treatment effects on a surrogate endpoint and treatment effects on a final outcome investigated in a disease area using data from all trials regardless of treatment classes or trials of the same class of treatments. For instance, in advance colorectal cancer (aCRC) progression free survival (PFS), tumor response (TR) or time to progression (TTP) have been investigated as potential surrogate endpoints for OS [@buyse2007progression; @giessen2013progression; @ciani2015meta; @chirila2012meta]. In previous work, Buyse et al. [@buyse2007progression] found a strong association between treatment effects on PFS and OS in this disease area, by including in their meta-analysis studies on one treatment class only (modern chemotherapy). More recently, Ciani et al. [@ciani2015meta] investigated the surrogate relationship in aCRC across all modern treatments, including a range of targeted therapies, which led to suboptimal surrogate relationship in this disease area. They concluded that in aCRC the association patterns could vary across treatment classes and a surrogate relationship observed in a specific treatment class may not directly apply across other treatment classes or lines of treatment. This may be particularly important for targeted treatments used only in a subset of population. For example anti-EGFR treatments are recommended for patients without a KRAS/panRAS mutation as these mutations are associated with resistance to the anti-EGFR therapies [@KRASmut2017; @heterogeneityofCrC2013] and the association pattern might be different for this particular treatment class in this subset of population with this unique characteristic. Furthermore, Giessen et al. [@giessen2013progression] who investigated the surrogate relationships in aCRC including all available treatments and subgroups of therapies, inferred that for validation of surrogacy in targeted treatments such as anti-EGFR therapies or anti-VEGF treatments further research is required once more data become available. Consequently, the assumption that a surrogate relationship remains the same across different treatment classes or lines of treatment does not seem reasonable in aCRC, which may be the case in other disease areas. Therefore potential differences in surrogate relationships across classes should be investigated. This can be achieved by performing subgroup analysis using a standard model (e.g. Daniels and Hughes model [@daniels1997meta]) or extending the standard model by adding another level to the hierarchical structure of the model for a surrogate relationship accounting for differences between treatment classes. In this paper, we propose two new methods which allow different degrees of borrowing of information for surrogate relationships across treatment classes aiming to obtain estimates of surrogate relationships with higher precision [@efron1975; @louis1984; @louis2000]. The first approach assumes full exchangeability of surrogate relationships exploiting the similarity of surrogate relationships and borrowing information across treatment classes. The second method is able to relax this assumption, by allowing for partial exchangeability [@neuenschwander2016robust] of surrogate relationships across treatment classes to avoid excessive borrowing of information from distinctly different treatment classes. In this model, the parameters describing surrogate relationships can be either exchangeable or non-exchangeable giving more flexibility when the assumption of exchangeability is not reasonable. The modelling techniques were demonstrated using an example in advanced colorectal cancer where the surrogate relationships may vary across treatment classes [@ciani2015meta]. To assess models’ performance and compare them with subgroup analysis we carried out a simulation study. In the remainder of this paper, we present the standard model in Section \[Standard\] , the two proposed models are introduced in Section \[New\], the results of the simulation study are demonstrated in Section \[SimStdy\] and the illustrative example as well as the results from its analysis are presented in Section \[Application\]. The paper concludes with a discussion in Section 7. Standard surrogacy model {#Standard} ======================== To investigate surrogate relationships within treatment classes using aggregate data, we performed subgroup analysis adopting a standard surrogacy model that was introduced by Daniels and Hughes [@daniels1997meta] for the study level evaluation of potential surrogate markers. Equation (\[eq:1\]) corresponds to the within-study model where $Y_{1i}$, $Y_{2i}$ are the estimates of treatment effects on surrogate endpoint and on the final outcome (for example log odds ratios for TR and log hazard ratio for OS). These effects follow a bivariate normal distribution with $\mu_{1i}$ and $\mu_{2i}$ corresponding to the true treatment effects on the surrogate and the final clinical outcome respectively while, $\sigma_{1i}$, $\sigma_{2i}$ and $\rho_{wi}$ are the within-study standard deviations for both outcomes and the within-study correlations between the treatment effects on the two outcomes for each study $i$. $$\begin{gathered} \begin{pmatrix}Y_{1i}\\ Y_{2i} \end{pmatrix} \sim N \begin{pmatrix} \begin{pmatrix} \mu_{1i}\\ \mu_{2i} \end{pmatrix}\!\!&,& \begin{pmatrix} \sigma_{1i}^{2} & \sigma_{1i}\sigma_{2i}\rho_{wi} \\ \sigma_{1i}\sigma_{2i}\rho_{wi} & \sigma_{2i}^{2} \end{pmatrix} \end{pmatrix}{\addtocounter{equation}{1}\tag{\theequation}}\label{eq:1}\\ \mu_{2i}|\mu_{1i} \sim N(\lambda_{0}+\lambda_{1}\mu_{1i},\psi^{2}){\addtocounter{equation}{1}\tag{\theequation}}\label{eq:2} \\ \end{gathered}$$ At the between-studies level (\[eq:2\]), the true effects on the surrogate endpoint $\mu_{1i}$ are modelled as fixed effects, and the true effects on the final outcome $\mu_{2i}$ have linear relationship with the true effects on the surrogate $\mu_{1i}$. This relationship plays a very important role as it can be used to predict $\mu_{2i}$ from known $\mu_{1i}$ in a new study $i$. The parameters $\lambda_{0}$, $\lambda_{1}$, $\psi^{2}$ correspond to the intercept, the slope and the conditional variance of the linear model and measure the shape of the relationship and the strength of association between the treatment effects on the surrogate endpoint and the effects on the final outcome. In the Bayesian framework, the Daniels & Hughes model was implemented by assuming no prior knowledge about surrogate relationship by using vague prior distributions. This allows the data to dominate the posterior distribution even if the dataset is relatively small. The following prior distributions can be used: $\mu_{1i} \sim N(0,a)$, $\lambda_{0} \sim N(0,a)$, $\lambda_{1} \sim N(0,a)$, $\psi \sim N(0,b)I(0,)$, where $N(0,b)I(0,)$ denotes a normal distribution truncated [@bland1999measuring] at the mean $\mu=0$ with standard deviation $s=b$. The parameters $a,b$ are chosen to be sufficiently large and depend on the scale of data. By adapting this method in our research, we applied this standard model to subsets of data that consist of only one class of treatment examining the surrogate relationship of each subgroup separately, taking motivation from similar analyses in clinical trials [@berry1990SubgroupAnal; @grouin2005Subgroups]. This kind of analysis is very practical when association patterns in a given disease area are different and the treatment classes consist of many studies. By performing subgroup analysis using the standard model, we explored potential differences in the association patterns across treatment classes and use them as a reference for results obtained with the newly developed methods. Criteria for surrogacy {#Criteria} ---------------------- As we mentioned previously, the parameters $\lambda_{0}$, $\lambda_{1}$, $\psi^{2}$ play a very important role, as they are used to evaluate surrogacy. A good surrogate relationship should imply that $\lambda_{1}\ne0$ as slope establishes the association between treatment effects on the surrogate and the final outcome. Subsequently, having $\psi^{2}=0$ implies that $\mu_{2i}$ could be perfectly predicted given $\mu_{1i}$. The parameter $\lambda_{0}$ corresponds to the intercept and is expected to be zero for a good surrogate relationship. This ensures that no treatment effect on the surrogate endpoint will imply no effect on the final outcome. These three criteria proposed by Daniels & Hughes [@daniels1997meta], will be referred to as surrogacy criteria in the remainder of this paper. A simple way to examine these surrogacy criteria is to check whether or not zero is included in the 95% credible intervals (CrIs) of $\lambda_{0}$, $\lambda_{1}$ and to compute the Bayes factor for the hypothesis $H_{1}$: $\psi^{2}=0$. The model with $\psi^{2}=0$ is a nested model within the standard model [@kass1995reference], so in order to compare these models, Bayes factors can be computed using the Savage Dickey density ratio [@verdinelli1995computing]. To implement the Savage Dickey density ratio, proper prior distributions for $\psi$ are needed. In our research a moderately informative half normal prior distribution $N(0,2)I(0,)$ was used for the conditional standard deviation. A strong association pattern requires zero to be included in the CrI of $\lambda_{0}$, zero not to be included in the CrI of $\lambda_{1}$ and the Bayes factor of $\psi^{2}$ to be greater than 3.3 [@jeffreys1998theory]. In this paper we used the evaluation framework proposed by Daniels and Hughes. However, there are other definitions and criteria for surrogacy in the literature. Detail review of other evaluation frameworks can be found in Lassere et al.[@lassere2008biomarker]. Cross-validation {#CrossVal} ---------------- One of the main aims of this paper was to explore whether the two hierarchical methods, that we propose in the next section, improve the predictions of treatment effect on the final outcome (by reducing bias and/or uncertainty) compared to subgroup analysis using the standard model. To evaluate this, a cross-validation procedure was carried out. It is a similar to the ’leave-one-study-out’ procedure that was described by Daniels & Hughes [@daniels1997meta] and it is repeated as many times as the number of studies in the data set. In a simulated data scenario, this can be used to draw inferences about predicting the true effect on the final endpoint $\mu_{2i}$ in a ’new’ study $i$, however, in a real data scenario true effects are unknown and therefore, we can only compare the observed values $Y_{2i}$ with their predicted intervals. For each study $i (i=1,..,N)$, treatment effect on the final endpoint $Y_{2i}$ is omitted and assumed unknown. This effect is then predicted from the observed effect on the surrogate endpoint $Y_{1i}$ and by taking into account the treatment effects on both outcomes from the remaining studies. In a Bayesian framework it can be achieved by performing Markov chain Monte Carlo (MCMC) simulation. The mean predicted effect is equal to the true effect $\hat{\mu}_{2i}$ predicted by MCMC simulation and the variance of the predicted effect is equal to $\sigma_{2i}^{2}+var(\hat{\mu}_{2i}|Y_{1i},\sigma_{1i},Y_{1(-i)},Y_{2(-i)})$ where $Y_{1,2(-i)}$ denote the observed treatment effects from the remaining studies without the study that is omitted in $i^{th}$ iteration [@daniels1997meta]. We then checked whether the 95% predictive interval (constructed using the variance) included the observed value of the treatment difference on the final outcome. Methods for surrogate endpoint evaluation incorporating aggregate data from different treatment classes {#New} ======================================================================================================= When subgroup analysis is used to investigate the study level surrogate relationships within treatment classes the validation process may fail due to limited data resulting in estimates of the parameters describing surrogate relationships obtained with considerable uncertainty [@berry1990SubgroupAnal]. We propose two hierarchical models to investigate surrogate relationships within treatment classes allowing different degrees of borrowing of information about the parameters of interest, as alternative approaches to subgroup analysis with the standard model. These models were developed to investigate the study level association and therefore they can only be applied to aggregate data (e.g logHR or logOR). They allow for the association patterns to vary across classes taking advantage of the attractive statistical properties of exchangeability [@efron1975; @louis1984; @louis2000]. Hierarchical model with full exchangeability (F-EX) {#F-EX} --------------------------------------------------- Our first approach extends the standard model accounting for differences in study level surrogacy across different treatment classes [@berry2013Hierarchical; @thall2003; @chugh2009phase]. Similarly as in the standard model, at the within-study level we assume that correlated and normally distributed observed treatment effects $Y_{1ij}$ and $Y_{2ij}$ (e.g logHR or logOR) in each study $i$ estimate the true treatment effects $\mu_{1ij}$ and $\mu_{2ij}$ on the surrogate and final outcomes respectively. In addition, by introducing index $j$ we account for the differences between the classes. $$\begin{gathered} \begin{pmatrix}Y_{1ij}\\ Y_{2ij} \end{pmatrix} \sim N \begin{pmatrix} \begin{pmatrix} \mu_{1ij}\\ \mu_{2ij} \end{pmatrix}\!\!,& \begin{pmatrix} \sigma_{1ij}^{2} & \sigma_{1ij}\sigma_{2ij}\rho_{wij} \\ \sigma_{1ij}\sigma_{2ij}\rho_{wij} & \sigma_{2ij}^{2} \end{pmatrix} \end{pmatrix}\\ \mu_{2ij}|\mu_{1ij} \sim N(\lambda_{0j}+\lambda_{1j}\mu_{1ij},\psi_{j}^{2}){\addtocounter{equation}{1}\tag{\theequation}}\label{eq:3}\\ \lambda_{0j} \sim N(\beta_{0},\xi_{0}^{2}), \lambda_{1j} \sim N(\beta_{1}, \xi_{1}^{2})\\ \newline \end{gathered}$$ The parameters $\sigma_{1ij}$, $\sigma_{2ij}$, $\rho_{wij}$ correspond to the within-study variances and within-study correlations for each study $i$ in treatment class $j$. The observed estimates $Y_{1ij}$, $Y_{2ij}$, $\sigma_{1ij}$, $\sigma_{2ij}$ are aggregate data extracted from systematic review RCTs whilst, the within-study correlations $\rho_{wij}$ can be calculated using a bootstrapping method from individual patient data (IPD). Similarly as in the standard model, the true effects $\mu_{1ij}$ on the surrogate endpoint are modelled as fixed effects. In contrast to the standard surrogacy model, this method assumes unique surrogate relationships between true treatment effects on the surrogate endpoint and the final outcome across treatment classes in a single model, allowing for borrowing of information across them. Each relationship between the true effects on the surrogate endpoint $\mu_{1ij}$ and the final outcome $\mu_{2ij}$ is described by a linear model where, $\lambda_{0j}$ denotes the intercept of the $j^{th}$ treatment class and $\lambda_{1j}$ establishes the relationship between treatment effects on surrogate and final outcomes within the treatment class $j$. To evaluate whether a candidate endpoint is considered a valid surrogate endpoint in a given treatment class, all three surrogacy criteria need to be met for this particular class. Implementing this model in the Bayesian framework, we place non-informative prior distributions on the model parameters such as: $\beta_{0}, \beta_{1} \sim N(0,a)$ and $\xi_{0},\xi_{1} \sim N(0,b)I(0,)$, $\mu_{1ij} \sim N(0,a)$ and $\psi_{j}\sim N(0,b)I(0,)$. Similarly as in the standard model $a,b$ are chosen to be sufficiently large and depend on the scale of data. F-EX model extends the standard model (described in section \[Standard\]) by including an additional layer of hierarchy to the linear relationship between true effects on the surrogate and the final outcome, assuming that slopes and intercepts are fully exchangeable across treatment classes. This can be implemented by placing common normal distributions on $\lambda_{0j}$ and $\lambda_{1j}$ with means and variances $\beta_{0}$, $\xi_{0}^{2}$ and $\beta_{1}$, $\xi_{1}^{2}$ leading to borrowing of information across treatment classes. Hierarchical models have desirable statistical properties that allow us to improve our inferences taking advantage of borrowing of information from other treatment classes. The exchangeable estimates, however, are shrunk towards the means $\beta_{0}$, $\beta_{1}$ and the amount of shrinkage depends on the number of studies within each class, the between treatment class heterogeneity [@neuenschwander2016robust] and the number of treatment classes. Although these statistical properties are very attractive in terms of potential reduction of uncertainty around the parameters of interest, they are advantageous only when the assumption of exchangeability is reasonable, otherwise there is a danger of excessive shrinkage. Hierarchical model with partial exchangeability (P-EX) {#P-EX} ------------------------------------------------------ F-EX method can be extended into a method with partial exchangeability similarly as the method proposed by Neuenschwander et al. [@neuenschwander2016robust]. This model is able to relax the assumption of exchangeability allowing the parameters of interest for each class to be either exchangeable with all or some of the parameters from other treatment classes or non-exchangeable. The proposed method is more flexible compared to F-EX model, in particular in data scenarios where the assumption of exchangeability is not reasonable for some of the treatment classes. The within and the between study level of this model is exactly the same as in the method with full exchangeability where, $Y_{1ij}$, $Y_{2ij}$ are the treatment effects on the surrogate and final clinical outcomes and they follow a bivariate normal distribution with mean values corresponding to the true treatment effects $\mu_{1ij}$ and $\mu_{2ij}$ on the two outcomes. $$\begin{gathered} \begin{pmatrix}Y_{1ij}\\ Y_{2ij} \end{pmatrix} \sim N \begin{pmatrix} \begin{pmatrix} \mu_{1ij}\\ \mu_{2ij} \end{pmatrix}\!\!,& \begin{pmatrix} \sigma_{1ij}^{2} & \sigma_{1ij}\sigma_{2ij}\rho_{wij} \\ \sigma_{1ij}\sigma_{2ij}\rho_{wij} & \sigma_{2ij}^{2} \end{pmatrix} \end{pmatrix}\\ \mu_{2ij}|\mu_{1ij} \sim N(\lambda_{0j}+\lambda_{1j}\mu_{1ij},\psi^{2}_{j})\\ \lambda_{0j}\sim N(\beta_{0},\xi_{0}^{2}){\addtocounter{equation}{1}\tag{\theequation}}\label{eq:4}\\ \lambda_{1j}=\left\{\fontsize{10}{25} \begin{array}{cc} \lambda_{1j} \sim N(\beta_{1},\xi_{1}^{2}) & \quad\text{if } p_{j} = 1 \\ \lambda_{1j} \sim N(0,b) & \quad\text{if } p_{j} = 0 \end{array} \right. \newline \end{gathered}$$ However, the parameters of slopes are modelled in a different way compared to F-EX model. In this approach two possibilities arise for these parameters for each treatment class $j$. When $p_{j}=1$ the parameter $\lambda_{1j}$ can be exchangeable with some or all the parameters of the slopes from the other treatment classes via an exchangeable component. It follows a common normal distribution with other slopes as in F-EX model. On the other hand, when $p_{j}=0$ the slope can be non-exchangeable with any slopes from the other treatment classes. In this case a vague prior distribution can be placed on the parameter, as in the standard model. The method evaluates the degree of borrowing of information for each parameter $\lambda_{1j}$ by using these two components with respective mixture weights. The main advantage of this method is that it allows the mixture weights to be inferred from the data. In each MCMC iteration, the sampler chooses between the two components by using a Bernoulli distribution $p_{j}\sim Bernoulli(\pi_{j})$. By calculating the posterior mean of this Bernoulli distribution we derive the mixture weights of each treatment class. The hyper-parameters $\pi_{j}$ of the Bernoulli prior distribution can be either fixed or, in a fully Bayesian framework, they can follow a prior distribution for example, a Beta distribution $\pi_{j}\sim Beta(1,1)$. We have used fixed $\pi_{j}$, since placing a prior distribution required longer chains to converge and provided almost the same results. In a special case where $p_{j}=1$ for all treatment classes, P-EX model reduces to full exchangeability model as it uses only the exchangeable component. Having $p_{j}=0$ for all treatment classes makes the P-EX model equivalent to subgroup analysis using the standard model as only the non-exchangeable component is used to estimate $\lambda_{1j}$ in this case. In a Bayesian framework vague prior distributions can be placed on the parameters $\beta_{0},\beta_{1}$, $\xi_{0},\xi_{1}$, $\mu_{1ij}$ as in F-EX model. Software Implementation and computing ===================================== All models were implemented in OpenBUGS [@R2w] where posterior estimates were obtained using MCMC simulations performing 50000 iterations (after discarding 20000 iterations as burn-in period). The OpenBUGS code of F-EX and P-EX models can be found in the supplementary material (sections D.3, D.4). Convergence was assessed visually by checking the history, chains and autocorrelation plots using graphical tools in OpenBUGS and R. All estimates are presented as means with corresponding 95% CrIs. The median was used only for the estimates of the conditional variances as a measure of central tendency since their posterior distributions were very skewed. The cross-validation procedure was performed in R using R2OpenBUGS [@R2w] package to execute OpenBUGS code multiple times. Simulation study {#SimStdy} ================ The proposed hierarchical methods allow different levels of borrowing of information for the parameters of interest. F-EX model assumes exchangeability for slopes whilst, the P-EX model allows for partial exchangeability for these parameters. We carried out a simulation study to assess the performance of the hierarchical methods and to compare them with subgroup analysis conducted using the standard model. We evaluated the performance of the methods in distinct data scenarios generated assuming different strengths of association within classes, different levels of similarity of the association patterns across classes and different number of studies per class. We evaluated the models’ ability to identify treatment classes with strong association patterns and to make predictions of the treatment effect on the final outcome in a new study from a treatment effect measured on the surrogate endpoint. Data generation process and scenarios {#GenerationPr} ------------------------------------- We simulated data under nine different scenarios generating 1000 replications for each scenario. Each replication included average treatment effects on the surrogate and the final outcome (and corresponding standard errors (SEs) and within-study correlations) from a number of studies of treatments belonging to five treatment clases. We assumed that the data in each treatment class had a different heterogeneity pattern. Therefore, to have a control over such heterogeneity patterns when simulating the data we needed to make an assumption about the distribution of the true effects both on the surrogate and the final endpoints. The standard model by Daniels & Hughes assumes fixed effect for the true effects on the surrogate endpoint (no common distribution) making difficult to control the heterogeneity patterns when simulating the data. To avoid this issue, we simulated data using a product normal formulation of bivariate random effect meta-analysis (BRMA) (equation set (\[eq:5\])), assuming normal random effects on the surrogate endpoint. Apart from this assumption, this method is the same as Daniels & Hughes model using a bivariate normal distribution to describe the within-study variability and a linear relationship to model the association between the surrogate and the final outcome. $$\begin{gathered} \begin{pmatrix}Y_{1ij}\\ Y_{2ij} \end{pmatrix} \sim N \begin{pmatrix} \begin{pmatrix} \mu_{1ij}\\ \mu_{2ij} \end{pmatrix}\!\!,& \begin{pmatrix} \sigma_{1ij}^{2} & \sigma_{1ij}\sigma_{2ij}\rho_{wij} \\ \sigma_{1ij}\sigma_{2ij}\rho_{wij} & \sigma_{2ij}^{2} \end{pmatrix} \end{pmatrix}\\ \mu_{1ij}\sim N(\eta_{1j},\psi_{1j}^{2}){\addtocounter{equation}{1}\tag{\theequation}}\label{eq:5}\\ \mu_{2ij}|\mu_{1ij} \sim N(\eta_{2ij},\psi_{2j}^{2})\\ \eta_{2ij} = \lambda_{0j}+\lambda_{1j}\mu_{1ij}\\ \psi_{1j} = \frac{\psi_{2j}}{|\lambda_{1j}|\sqrt{(1/\rho_{bj}^2)-1}} \end{gathered}$$ Simulating data from this model, however, can lead to results obtained with increased uncertainty, as the models used to analyse the data make fewer distributional assumptions. To generate the data, we pursued the following steps: 1. Set the number of classes $N=5$ 2. Simulate the data for each class separately using BRMA model (eq. \[eq:5\]) under 3 main designs. 3. Create three sets of scenarios: two with fixed number of studies ($n_{j}$=16 and $n_{j}$=8, j=1,...,5) per treatment class and one with unbalanced classes ($n_{1}$ = 4, $n_{2}$ = 8, $n_{3}$ = 6, $n_{4}$ = 10, $n_{5}$ = 7 ). We applied the three sets of scenarios to each design. In total, we have 9 scenarios (3 designs $\times$ 3 sets = 9 scenarios). 4. Simulate the true effects using model (eq. (5)) The values of the parameters are listed in Table 1 and a short description of each design can be found below: **Design 1:** In the first design, our aim was to illustrate the properties of exchangeability. We simulated data in five treatment classes assuming high degree of similarity for their slopes and intercepts. The data in each treatment class were simulated assuming strong association (see surrogacy criteria in Section \[Criteria\] ) for each individual class but weak overall. **Design 2:** The second design illustrates the case where there is a treatment class with very different association pattern (slope) compared to the other classes. This implies that the assumption of exchangeability is in doubt for this parameter in this particular class. Similarly as in the first scenario, we assumed strong association for each individual class. **Design 3:** The last design focuses on the association patterns of strengths that vary across treatment classes, investigating whether the proposed methods can estimate a strong association pattern better compared to subgroup analysis with the standard model and whether they can distinguish between the different association patterns despite borrowing of information across treatment classes. To achieve this, we generated three out of five treatment classes with strong association and the remaining two classes with a weak association. \[Design\] ------------------------------------------- ------------------------------------------- ------------------------------------------- $\lambda_{11} = 0.40$, $\rho_{b1} = 0.89$ $\lambda_{11} = 0.60$, $\rho_{b1} = 0.93$ $\lambda_{11} = 0.40$, $\rho_{b1} = 0.90$ $\lambda_{12} = 0.45$, $\rho_{b2} = 0.90$ $\lambda_{12} = 1.55$, $\rho_{b2} = 0.99$ $\lambda_{12} = 0.50$, $\rho_{b2} = 0.70$ $\lambda_{13} = 0.50$, $\rho_{b3} = 0.91$ $\lambda_{13} = 1.60$, $\rho_{b3} = 0.99$ $\lambda_{13} = 0.60$, $\rho_{b3} = 0.93$ $\lambda_{14} = 0.55$, $\rho_{b4} = 0.92$ $\lambda_{14} = 1.65$, $\rho_{b4} = 0.99$ $\lambda_{14} = 0.70$, $\rho_{b4} = 0.75$ $\lambda_{15} = 0.60$, $\rho_{b5} = 0.93$ $\lambda_{15} = 1.70$, $\rho_{b5} = 0.99$ $\lambda_{15} = 0.80$, $\rho_{b5} = 0.95$ $\lambda_{0j} = 0$ $\lambda_{0j} = 0$ $\lambda_{0j} = 0$ $\sigma_{1ij,2ij}=0.1$ $\sigma_{1ij,2ij}=0.1$ $\sigma_{1ij,2ij}=0.1$ $\rho_{wij} = 0.4$ $\rho_{wij} = 0.4$ $\rho_{wij} = 0.4$ $\psi_{2j} = 0.08$ $\psi_{2j} = 0.08$ $\psi_{21,23,25} = 0.08$ $\psi_{22,24} = 0.30$ $\eta_{1j} = 0.3$ $\eta_{1j} = 0.3$ $\eta_{1j} = 0.3$ ------------------------------------------- ------------------------------------------- ------------------------------------------- : Simulation designs Performance measures {#Perfmeasures} -------------------- To evaluate the goodness of fit of the models, we calculated the coverage probability of the 95% CrIs of $\lambda_{1j}$ and the 95% predictive intervals of $\mu_{2ij}$. The absolute bias and the root mean square error (RMSE) of $\hat{\lambda}_{1j}$ and $\hat{\mu}_{2ij}$ were also monitored and reported in the tables. In order to investigate potential decrease in the degree of uncertainty of the estimates as a result of borrowing of information across treatment classes, we calculated ratios of the width of the 95% CrIs. The width ratio $w_{\lambda_{1j}^{FEX,(PEX)}}$/$w_{\lambda_{1j}^{subgr}}$ was defined as the ratio of the widths of the CrIs of $\lambda_{1j}$ from F-EX or P-EX to the width of the CrIs of $\lambda_{1j}$ from subgroup analysis using the standard model. Similarly, the width ratio $w_{\mu_{2ij}^{FEX,(PEX)}}$/$w_{\mu_{2ij}^{subgr}}$ was the ratio of the 95% predictive intervals of the true effects $\mu_{2ij}$ from F-EX or P-EX to the width of the predictive intervals of $\mu_{2ij}$ from subgroup analysis using the standard model. We also monitored the largest Monte Carlo error (MCE) of the simulations as an index of accuracy of the Monte Carlo samples. Furthermore, a cross-validation procedure was applied to each method across the simulated data scenarios. In the simulation study, the true effect on the final endpoint $\mu_{2ij}$ was known, since it had been simulated ,therefore the cross-validation procedure was applied on the true effects (in real data scenarios we compare the predicted effect with the observed effect) by checking whether the simulated value of the true effect $\hat\mu_{2ij}$ was included in the predictive interval of $\mu_{2ij}$. Results ------- All the tables in the results section list the performance of the posterior means of $\hat{\lambda}_{1j}$, the performance of the posterior means of $\hat{\mu}_{2ij}$ as well as the probabilities of estimating a strong association pattern (see definition in Section \[Criteria\]) for each class across methods. The following section presents the results of the analysis by reporting the coverage probabilities of the CrIs of $\lambda_{1j}$ and $\mu_{2ij}$ for each scenario (by taking the mean of coverage probabilities across classes), the overall absolute bias and RMSE of $\hat{\lambda}_{1j}$ and $\hat{\mu}_{2ij}$, the width ratios of $\lambda_{1j}$ and $\mu_{2ij}$ for each scenario (by calculating the mean of the width ratios of $\lambda_{1j}$ across classes and the mean of the width ratios of $\mu_{2ij}$ across studies and classes), the MCE and the probability to estimate a strong association pattern by fitting each model. Detailed results for the performance of $\hat{\lambda}_{1j}$ and $\hat{\mu}_{2ij}$ for each class separately and across methods are listed in the Supplementary material (see section B and section C). ### Performance of the estimates $\hat{\lambda}_{1j}$ Table \[Performancelambda1\] presents the results across the nine scenarios reporting averages of the measures we monitored for $\hat{\lambda}_{1j}$ over the five classes of treatment. The performance of the models varied in terms of the coverage probability of the 95% CrIs of $\lambda_{1j}$ across scenarios. In the scenarios 1, 4 and 7 where the number of studies per class was relatively high, the models achieved 95% coverage probabilities. However, in the scenarios where the number of studies was smaller the coverage probability was higher due to increased uncertainty and likely to the fact that the model we used in the generation process was slightly different from models used to fit the data. Monte Carlo errors were small across most of the scenarios implying good accuracy of the Monte Carlo samples and that convergence was achieved in those scenarios across all the methods. However, in scenarios where the data were limited (scenarios 3, 6, 9) subgroup analysis with the standard model yielded larger MCEs. This implies that subgroup analysis requires longer chains to achieve the same level of convergence as the other two models. In the first three scenarios (1st design), where the treatment classes were very similar in terms of patterns (similar slopes), F-EX and P-EX were superior compared to subgroup analysis as they gave posterior means of slopes with lower absolute bias, RMSE and reduced uncertainty (narrower 95% CrIs) due to borrowing of information across classes. P-EX model achieved almost the same level of borrowing of information as F-Ex model, with mixtures weights were very close to 1 across treatment classes (see details in the supplementary material D.1 where the mixture weights are listed). Overall, the proposed hierarchical models performed better compared to subgroup analysis but the difference was more pronounced in the scenarios with small number of studies. In the second design (scenarios 4, 5, 6), where the exchangeability assumption was not reasonable for one of the classes, P-EX model yielded the most robust results. The model resulted in the posterior means with the smallest absolue bias and RMSE, reducing the degree of borrowing of information for the class with the distinctly different (the mixture weights in this class were $p_{1}=0.56$, $p_{1}=0.31$ and $p_{1}=0.80$ respectively) whilst still borrowing information across the remaining classes ($p_{2},p_{3},p_{4},p_{5}\approx 0.97$). On the other hand, F-EX performed poorer compared to the other methods in scenario 6 with unbalanced and relatively small number of studies per class, leading to more biased results. This indicates that F-EX model is not appropriate when the assumption of exchangeability is not reasonable. Subgroup analysis using the standard model achieved decent performance only in the forth scenario where there were sufficient data. In the third design (scenarios 7, 8, 9) the proposed models achieved superior performance compared to subgroup analysis for the estimates of $\lambda_{1j}$, similarly as in the first three scenarios. The last column of Table \[Performancelambda1\] shows the probabilities of estimating a strong association pattern across the data scenarios and models. F-EX and P-EX methods estimated the surrogacy (based on the three surrogacy criteria) better compared to subgroup analysis across all scenarios. In the first design (scenarios 1, 2, 3) where the association was designed to be strong for all the classes, F-EX and P-EX models predicted a strong association pattern in more than 85% of the simulations. Subgroup analysis predicted the 81% of them in the 1st scenario but its performance reduced noticeably in the 2nd and 3rd scenario where the data were more sparse. In the second design (scenarios 4, 5, 6) with strong association patterns across all classes, P-EX and F-EX estimated more than 87% of the association patterns across these three scenarios. Subgroup analysis performed well only in the 4th scenario predicting the 89% of the association patterns but its performance gradually reduced as the number of studies was decreased in scenario 5 and 6. Scenario Methods RMSE MCE ---------- --------------------------------- ------------------- ------ ------ ------ ------ ------- ------ $\textbf{1st Design}$ 1st Fixed ($n_j=16$) Subgroup Analysis 0.95 0.08 0.10 0.003 0.81 F-EX model 0.95 0.06 0.07 0.72 0.002 0.85 P-EX model 0.96 0.06 0.07 0.72 0.002 0.85 2nd Fixed ($n_j=8$) Subgroup Analysis 0.98 0.11 0.15 0.005 0.71 F-EX model 0.97 0.07 0.09 0.60 0.003 0.89 P-EX model 0.97 0.07 0.09 0.61 0.003 0.90 3rd Unbalanced Subgroup Analysis 0.99 0.13 0.18 0.017 0.56 ($n_{1}=4$,$n_{2}=8$,$n_{3}=6$, F-EX model 0.99 0.07 0.09 0.52 0.003 0.89 $n_{4}=10$,$n_{5}=7$) P-EX model 0.99 0.07 0.09 0.53 0.004 0.88 $\textbf{2nd Design}$ 4th Fixed ($n_j=16$) Subgroup Analysis 0.95 0.09 0.11 0.007 0.89 F-EX model 0.94 0.08 0.10 0.90 0.005 0.91 P-EX model 0.94 0.07 0.09 0.86 0.004 0.91 5th Fixed ($n_j=8$) Subgroup Analysis 0.97 0.14 0.17 0.007 0.88 F-EX model 0.96 0.12 0.15 0.86 0.005 0.92 P-EX model 0.97 0.10 0.12 0.78 0.005 0.92 6th Unbalanced Subgroup Analysis 0.98 0.15 0.20 0.025 0.72 ($n_{1}=4$,$n_{2}=8$,$n_{3}=6$, F-EX model 0.96 0.17 0.21 0.70 0.011 0.88 $n_{4}=10$,$n_{5}=7$) P-EX model 0.97 0.14 0.18 0.70 0.011 0.87 $\textbf{3rd Design}$ 7th Fixed ($n_j=16$) Subgroup Analysis 0.95 0.11 0.14 0.003 F-EX model 0.95 0.09 0.11 0.79 0.002 P-EX model 0.95 0.09 0.11 0.79 0.003 8th Fixed ($n_j=8$) Subgroup Analysis 0.97 0.17 0.22 0.006 F-EX model 0.96 0.11 0.14 0.67 0.004 P-EX model 0.96 0.11 0.14 0.67 0.004 9th Unbalanced Subgroup Analysis 0.98 0.19 0.25 0.021 ($n_{1}=4$,$n_{2}=8$,$n_{3}=6$, F-EX model 0.97 0.12 0.15 0.56 0.005 $n_{4}=10$,$n_{5}=7$) P-EX model 0.97 0.12 0.15 0.57 0.005 : Performance of $\hat{\lambda}_{1j}$[]{data-label="Performancelambda1"} Table \[surrogacy\] presents the results from the last 3 scenarios (3rd design), where the surrogate relationships varied across classes. F-EX and P-EX methods were able to estimate a strong association pattern with higher probability compared to subgroup analysis in the classes where the association was designed to be strong. At the same time, the methods successfully identified classes with strong association patterns from a mixture of classes with weak and strong association patterns, even for the scenarios with relatively few studies per class where subgroup analysis failed almost completely to identify. The probabilities of estimating a strong association per class in designs 1 and 2 are presented in the Supplementary material (see section B.1, B.2, B.3, B.4, B.5, B.6). \[key\] [lllccc]{} Scenario&&$\textbf{Treatment classes}$ &Subgroup Analysis&F-EX model&P-EX model\ 7th&Fixed ($n_{j}=16$)&1$^{st}$ class & 0.82 & 0.84 &0.84\ &&2$^{nd}$ class$^*$ & 0.00 & 0.00 &0.00\ &&3$^{rd}$ class & 0.83 & 0.85 &0.85\ &&4$^{th}$ class$^*$ & 0.00 & 0.00 &0.00\ &&5$^{th}$ class & 0.80 & 0.80 &0.80\ 8th&Fixed ($n_{j}=8$)&1$^{st}$ class & 0.78 & 0.89 &0.89\ &&2$^{nd}$ class$^*$ & 0.04 & 0.05 &0.05\ &&3$^{rd}$ class & 0.80 & 0.90 &0.90\ &&4$^{th}$ class$^*$ & 0.06 & 0.06 &0.06\ &&5$^{th}$ class & 0.85 & 0.87 &0.86\ 9th&Unbalanced &1$^{st}$ class & 0.06 & 0.82 &0.80\ &($n_{1}=4$,$n_{2}=8$,&2$^{nd}$ class$^*$ & 0.06 & 0.07 &0.07\ &$n_{3}=6$,$n_{4}=10$,&3$^{rd}$ class & 0.65 & 0.91 &0.91\ &$n_{5}=7$) &4$^{th}$ class$^*$ & 0.03 & 0.03 &0.03\ & &5$^{th}$ class & 0.82 & 0.89 &0.89\ ### Performance of predictions $\hat{\mu}_{2ij}$ Table \[PerformancePred\] shows the results from cross-validation procedure which resulted in the posterior means ($\hat{\mu}_{2ij}$) and 95% predictive intervals of the true effects $\mu_{2ij}$. It presents the same measures as Table \[Performancelambda1\] averaged over the five classes. In scenarios 1, 4 and 7, the models achieved 95% coverage due to the large amount of data, however, in the remaining scenarios where the number of studies was smaller the models yielded higher coverages probabilities. F-EX and P-EX had small Monte Carlo errors across all scenarios, however, subgroup analysis gave on average significantly larger MCEs compared to the proposed methods in scenarios 3, 6, 9. More specifically, in the treatment classes where the surrogate relationship was designed to be weak or the data were limited the MCEs were higher (see details in the supplementary material C.3, C.6, C.9). This indicates that subgroup analysis with the standard model requires longer chains for its posteriors to achieve the same level of convergence as the other two methods. In the first three scenarios, F-EX and P-EX models outperformed subgroup analysis in terms of the absolute bias, RMSE and the uncertainty of $\hat{\mu}_{2ij}$. However, there was no winner between them as both methods had almost the same degree of borrowing of information resulting in 7%, 20%, and 33% narrower predictive intervals compared to subgroup analysis across these three scenarios respectively. In the scenarios 4 ,5 ,6 (2nd desing), P-EX yielded posterior means with the smallest absolute bias, RMSE and CrIs with the smallest width ratio across classes. Furthermore, P-EX method gave the most robust results for the ’extreme’ treatment class relaxing 44%, 70% and 20% the borrowing of information in this class across the scenarios (see the mixture weights in the supplementary material D.1). In the 6th scenario F-EX performed poorer compared to P-EX model leading to biased results especially for the treatment class where the surrogacy was different and the exchangeability assumption unreasonable (1st class in the section B.6 of the supplementary material). Subgroup analysis performed almost equally well as the P-EX model in the 4th scenario where the number of studies per class was relatively large. The last three scenarios (3rd design) gave similar results as the first three in terms of the uncertainty, the absolute bias and the RMSE of $\hat{\mu}_{2ij}$. F-EX P-EX models performed equally well, whilst subgroup analysis with the standard model was the worst approach resulting in inflated predictive intervals, larger RMSE and worse MCE in all cases. Scenario Methods RMSE MCE ---------- --------------------------------- ------------------- ------ ------ ------ ------ ------- $\textbf{1st Design}$ 1st Fixed ($n_j=16$) Subgroup Analysis 0.95 0.09 0.11 0.003 F-EX model 0.95 0.08 0.10 0.93 0.002 P-EX model 0.95 0.08 0.10 0.93 0.002 2nd Fixed ($n_j=8$) Subgroup Analysis 0.98 0.11 0.13 0.010 F-EX model 0.98 0.08 0.10 0.80 0.004 P-EX model 0.98 0.08 0.10 0.80 0.004 3rd Unbalanced Subgroup Analysis 0.99 0.12 0.18 0.023 ($n_{1}=4$,$n_{2}=8$,$n_{3}=6$, F-EX model 0.99 0.08 0.11 0.67 0.005 $n_{4}=10$,$n_{5}=7$) P-EX model 0.99 0.09 0.11 0.68 0.008 $\textbf{2nd Design}$ 4th Fixed ($n_j=16$) Subgroup Analysis 0.95 0.13 0.18 0.009 F-EX model 0.95 0.13 0.18 0.97 0.008 P-EX model 0.96 0.12 0.17 0.96 0.008 5th Fixed ($n_j=8$) Subgroup Analysis 0.99 0.16 0.20 0.015 F-EX model 0.98 0.15 0.19 0.92 0.009 P-EX model 0.98 0.14 0.18 0.87 0.008 6th Unbalanced Subgroup Analysis 0.99 0.18 0.23 0.021 ($n_{1}=4$,$n_{2}=8$,$n_{3}=6$, F-EX model 0.99 0.18 0.22 0.80 0.009 $n_{4}=10$,$n_{5}=7$) P-EX model 0.99 0.15 0.19 0.77 0.010 $\textbf{3rd Design}$ 7th Fixed ($n_j=16$) Subgroup Analysis 0.95 0.16 0.23 0.006 F-EX model 0.95 0.16 0.22 0.96 0.004 P-EX model 0.95 0.16 0.22 0.96 0.004 8th Fixed ($n_j=8$) Subgroup Analysis 0.98 0.18 0.26 0.017 F-EX model 0.97 0.16 0.22 0.85 0.006 P-EX model 0.97 0.16 0.22 0.85 0.006 9th Unbalanced Subgroup Analysis 0.98 0.20 0.28 0.027 ($n_{1}=4$,$n_{2}=8$,$n_{3}=6$, F-EX model 0.97 0.17 0.21 0.72 0.008 $n_{4}=10$,$n_{5}=7$) P-EX model 0.97 0.17 0.21 0.72 0.009 : Performance of $\hat{\mu}_{2ij}$[]{data-label="PerformancePred"} Discussion of the results ------------------------- The aim of the simulation study was to illustrate and assess the performance of the methods under different scenarios. The models gave 95% coverage probabilities in the scenarios 1, 4, and 7 where the number of studies was sufficiently large (16 for each class). However, in the remaining scenarios the coverage probabilities were higher than 95%, which means that the methods derived more conservative CrIs of parameters than expected. This is largely due to the sparsity of the data in these scenarios but may also be partly due to different models being used to simulate and analyse the data as explained in section \[GenerationPr\]. In the first design (scenarios 1, 2, 3) where the assumption of exchangeability was reasonable, F-EX and P-EX models performed better than the subgroup analysis giving on average narrower 95% CrIs of $\lambda_{1j}$ and 95% predictive intervals of $\mu_{2ij}$. This indicates that P-EX model successfully identified the correct level of borrowing of information inferring that the mixture weights should be very close to 1. P-EX model was the best choice in all the scenarios of the second design (scenarios 4, 5, 6) where there was a treatment class with distinctly different slope. It relaxed the degree of borrowing of information for the ’extreme’ treatment class, giving the most accurate posterior means of the slopes. Moreover, P-EX model was the best choice in terms of predictions of the true effect on the final endpoint, reducing the width of predictive intervals by 4%, 13% and 23% compared to subgroup analysis in each scenario respectively. Last but not least, the proposed methods estimated the strong association patterns better compared to subgroup analysis across all data scenarios. In particular, in scenarios 3, 6 and 9, where the data were sparse, the proposed hierarchical methods were able to estimate surrogacy significantly better compared to the subgroup analysis. This illustrates well the benefits of using hierarchical methods when data are limited. Furthermore, as illustrated by scenarios 7,8 and 9, F-EX and P-EX could easily distinguish between the different association patterns as they identified treatment classes with strong association patterns and at the same time did not overestimate the strength of the association in the classes where the association was designed to be weak. Application: Advanced colorectal cancer {#Application} ======================================= Data ---- We illustrate the proposed methodology in an example in aCRC. The data were obtained from a systematic review conducted by Ciani et al. [@ciani2015meta] which included 101 RCTs published between 2003 and 2013, evaluating multiple interventions in aCRC. The review consist of trials that report treatment effects on OS or/and on alternative endpoints such as progression-free-survival or tumor response (PFS, TR). OS was defined as the time from randomisation to time of death, PFS was set as the time from randomisation to tumor progression or death from any cause. Tumor response was estimated using objective tumor measurements which are measured using imaging methods and determined according to the Response Evaluation Criteria in Solid Tumors guidelines [@therasse2000Tumors] or the World Health Organization recommendations [@world1979handbook]. The RCTs in the systematic review contain five treatment classes: the class of chemotherapies, the anti-epidermal growth factor receptor (Anti-EGFR) monoclonal antibodies class, angiogenesis inhibitors, other molecular-targeted agents (MTA) and intrahepatic arterial (IHA) chemotherapies . Ciani et al. [@ciani2015meta] investigated surrogate relationships between treatment effects on potential surrogate endpoints (TR and PFS) and on the final clinical outcome (OS). They found that the surrogate relationships between treatment effects on these endpoints were suboptimal. Furthermore, they stated that PFS was an acceptable surrogate endpoint for OS, whereas TR should not be used as a surrogate endpoint for this final outcome. They concluded that good surrogacy observed in previous studies, that included traditional chemotherapy trials in aCRC may not apply directly across other classes of treatments. More details about the studies and how the systematic review was designed can be found in Ciani et al. [@ciani2015meta]. We refer these data as ’Ciani data’ in the remainder of this paper. In our example, we focused on a subset of these data examining the surrogacy between treatment effects on TR and PFS and treatment effects on PFS and OS including data from three treatment classes. We obtained data from 35 studies reporting treatment effect on PFS and OS where, 15 of them belonged to the chemotherapy treatment class, 9 of them investigated anti-EGFR therapies and 11 anti-angiogenic treatments. To investigate surrogate relationships between treatment effects on TR and PFS we used data from 35 studies reporting treatment effects on these endpoints; 17 of them investigated chemotherapies, 8 and 10 studies anti-EGFR and anti-angiogenic treatments respectively. TR can be evaluated as a surrogate endpoint to treatment effect on PFS, as treatment effects on TR is typically measured earlier compared to treatment effects on PFS. Figure \[fig:1\] provides a graphical representation of the data set we used. It illustrates the association patterns between the treatment effects across classes on each pair of outcomes. ![Scatterplots of treatment effects on PFS-OS and TR-PFS[]{data-label="fig:1"}](DoublePair.pdf){height="7cm" width="15cm"} Individual patient data (IPD) were available from four RCTs[@bennouna2013continuation; @rothenberg2008capecitabine; @cassidy2011xelox; @hurwitz2004bevacizumab], which were used to estimate the within-study correlations. By applying a bootstrap method (see section A in the supplementary material) we estimated two sets of within-study correlations: for each of the two pairs of outcomes one correlation corresponding to each treatment class. We assumed that within treatment classes the within-study correlations are the same across studies. Scale of the outcomes --------------------- The treatment effects on OS and PFS were modelled on the log hazard ratio scale $logHR(OS)$, $logHR(PFS)$, whereas the treatment effects on TR were modelled on log odds ratio $logOR(TR)$ scale. We retrieved the corresponding standard errors of $logHR(PFS)$ and $logHR(OS)$ on PFS and OS from the 95% confidence intervals and by using the standard formulae for the standard errors of $logOR(TR)$ . Results of data analysis {#ApplicResults} ------------------------ The first aim of our analysis was to explore potential differences in association patterns across treatment classes. To investigate this, we applied the two proposed models and subgroup analysis using standard model to the data and derived posterior distributions for the parameters of the surrogate relationships for each treatment class. We obtained the posteriors mean of the intercepts $\hat{\lambda}_{0j}$, the slopes $\hat{\lambda}_{1j}$ and posterior median of conditional variances $\hat{\psi}^{2}_{j}$ with corresponding 95% CrIs across treatment classes. By checking the surrogacy criteria (described in section \[Criteria\]) we were able to infer whether or not a candidate endpoint is a valid surrogate in each treatment class. We carried out a cross-validation procedure (section \[CrossVal\]) to investigate how well the models predict the true treatment effect on the final clinical outcome. The measures we monitored were the absolute error of the predictions, the ratios of the width of the 95% predictive intervals from P-EX or F-EX to the width of the 95% predicted interval obtained from subgroup analysis and the largest MCE. ### Results across models and treatment classes   **Subgroup analysis with the standard model** The results of subgroup analysis presented in the first two columns of Table \[Estimates\] showed strong association between the treatment effects on PFS and the effects on OS in the class of chemotherapies and the anti-angiogenic treatment class with all three criteria for surrogacy satisfied (the 95% CrIs of $\lambda_{01}$ and $\lambda_{03}$ included zero, the 95% CrIs of $\lambda_{11}$ and $\lambda_{13}$ did not contain zero and there was substantial evidence using Bayes factors in favour of the hypotheses $H_{1}:\psi_{1}^{2}=0$, and $H_{1}:\psi_{3}^{2}=0$ (see details about Bayes factors in the supplementary material D.2)). In contrast, we can infer that the surrogate relationship between treatment effects on PFS and the effects on OS in the anti-EGFR treatment class was weak, as the 95% CrI of the posterior distribution of the slope included zero. Investigating the surrogacy on TR-PFS pair we found a similar pattern, thus we can infer that there was an acceptable surrogate relationship between treatment effects on TR and PFS in the chemotherapy and the anti-angiogenic classes. The relationship was negative overall, since the slopes were negative across classes. On the other hand, the surrogacy criteria indicated poor surrogacy between the treatment effects on TR and the treatment effects on PFS for anti-EGFR class, since the 95% CrI of the slope $\lambda_{12}$ included zero.  **F-EX model** The results of F-EX model are presented in columns 3 and 4 of Table \[Estimates\]. For the PFS-OS pair of outcomes, the association patterns were very similar in the anti-angiogenic and chemotherapy treatment classes as both classes satisfied the surrogacy criteria and the slopes were of similar magnitude. The 95% CrIs of the intercepts $\lambda_{01}$ and $\lambda_{03}$ included zero indicating that zero treatment effect on the surrogate implies zero treatment effect on the final outcome for these two classes. The intervals of the slopes $\lambda_{11}$ and $\lambda_{13}$ did not contain zero indicating positive association as the two slopes were positive. The conditional variances in these two classes were small indicating strong association which was supported by the analysis using Bayes factors (see details about the Bayes factors in the supplementary material D.2). On the other hand, the association was weak in the anti-EGFR treatment class failing to meet one of the criteria, as the 95% CrI of the slope $\lambda_{12}$ included zero. On the contrary, for TR-PFS pair of outcomes all three surrogacy criteria were satisfied across all the treatment classes taking advantage of the assumption of exchangeability for the parameters $\lambda_{0j}$ and $\lambda_{1j}$. This implies that TR was an acceptable surrogate endpoint for PFS across treatment classes in this data set.   **P-EX model** P-EX model allows the parameters of slope of each treatment class to be either exchangeable or non-exchangeable with parameters of slopes from other classes yielding parameters with partial exchangeability. For both pairs of outcomes, fixed values for the hyper-parameters $\pi_{j}=(0.5,0.5,0.5)$ were chosen assuming that exchangeability and non-exchangeability were $\textit{apriori}$ equally likely. As in the case of F-EX model, the surrogacy criteria were estimated for each class separately and then a cross-validation procedure followed, however, for this model we also monitored the mixture weights by calculating the posterior means of $p_{j}$ in order to measure the degree of borrowing of information across classes (Table \[Estimates\] columns 5, 6). For the PFS-OS pair, the weights increased from their prior values ($\pi_{j}=0.5$) to 0.968 in the class of chemotherapy, to 0.965 in the anti-EGFR class and to 0.966 in the anti-angiogenic treatment class indicating that borrowing of information was reduced approximately 3.5% for each class compared to F-EX model. Looking at the results from P-EX model we drew the same inferences as from F-EX model, inferring that the association patterns were strong in the anti-angiogenic and the chemotherapy classes, but weak in the anti-EGFR treatment class where the 95% CrI of the slope $\lambda_{12}$ included zero. In contrast to this, for TR-PFS pair the mixture weights were smaller than on PFS-OS pair due to the slightly larger between treatment class heterogeneity. There was 7.1% reduction in borrowing of information in anti-angiogenic class compared to F-EX models, whilst the weights for the chemotherapies and anti-EGFR agents were 0.944 and 0.95 respectively. All three surrogacy criteria were fulfilled across treatment classes despite the decrease in levels of borrowing of information, indicating that TR was an acceptable surrogate for PFS across treatment classes in the Ciani data. ### Results of the cross-validation procedure After estimating the surrogacy criteria across treatment classes, we carried out cross-validation procedure to predict the treatment effects $\mu_{2i}$ on the final outcome. The results in Table \[CV\] showed that the cross-validation procedure of subgroup analysis with the standard model gave predictive intervals of the effects on the final outcome containing the corresponding observed estimates $Y_{2i}$ in the 97% of the studies for both pairs of outcomes confirming good fit of the model. The cross-validation procedure yielded the most accurate posterior means of the true effects on the final endpoint (small absolute error) in the treatment class of chemotherapies, where the number of the available studies was large and performed poorly in terms of accuracy of predictions in the anti-EGFR class (large absolute error) where the surrogacy was weak and the number of studies small. Similarly, subgroup analysis with the standard model was less accurate in targeted treatment classes for the TR-PFS pair of outcomes where the number of studies was smaller. The results from the cross-validation procedure of F-EX model showed that the method fitted the data well. All of the predicted intervals of $\mu_{2ij}$ contained the observed values of the treatment effects on the final outcome on PFS-OS pair and all but one on TR-PFS pair. The cross-validation procedure yielded the posterior means of $\mu_{2ij}$ with the smallest absolute error in chemotherapy treatment class on PFS-OS pair and performed equally well in terms of its accuracy in the other two classes. In contrast to this, higher absolute error were observed in the anti-angiogenic class on TR-PFS pair indicating that the assumption of exchangeability for the parameters describing the surrogate relationships was fairly strong and it was likely to cause ’overshrinkage’ in this particular class. The results obtained for the width ratios imply that F-EX method gave intervals of the true effect on the final endpoint with smaller degree of uncertainty compared to subgroup analysis. There was a small decrease in the uncertainty of the predictions of $\mu_{2ij}$ on PFS-OS pair for the chemotherapy treatment class, as the cross-validation procedure of F-EX model yielded 1.2% narrower intervals compared to subgroup analysis. Furthermore, significantly reduced uncertainty was observed in the other two treatment classes for PFS-OS pair, 13.8% in the anti-EGFR treatment class and 7% in the anti-angiogenic, where the number of studies was smaller. On the contrary, very limited decrease in the degree of uncertainty was observed for the TR-PFS pair of outcomes across all classes. Overall on this pair, the predictive intervals were only 1.3% narrower compared to subgroup analysis. The benefit was small (3.2% reduction of the width of the predictive interval) even for the anti-EGFR treatment class where there were only 8 studies for this pair. Focusing on the results from the cross-validation procedure using P-EX model, all the intervals of the predicted treatment effects on the final outcome contained the observed treatment effects on PFS-OS pair and all but one on the TR-PFS pair. The absolute error was smaller in chemotherapy treatment class for the PFS-OS pair where the number of studies was large and significantly higher in the other two classes. In contrast to this, the cross-validation procedure with P-EX gave almost equally accurate estimates in the anti-EGFR and the chemotherapy treatment classes on TR-PFS pair. However, the absolute error was higher in the anti-angiogenic treatment class where the association was much stronger compared to the other two classes indicating potential excessive borrowing of information from the other classes. This is likely due to the assumption of full exchangeability for the intercepts. The method predicted the effects on the final outcome with reduced uncertainty giving more precise estimates ($\hat{\mu}_{2ij}$) compared to subgroup analysis in the anti-EGFR class on PFS-OS pair reducing the uncertainty by 13.6%. On the other hand, the predicted effects $\hat\mu_{2ij}$ had almost the same degree of uncertainty as those from subgroup analysis for TR-PFS pair. The intervals were only 1% narrower on average across all classes compared to the subgroup analysis. -------- ------------------------- -------- -------- -------- -------- -------- -------- -- -------- -------- -- Models Measures PFS-OS TR-PFS PFS-OS TR-PFS PFS-OS TR-PFS PFS-OS TR-PFS 1.000 0.941 0.888 1.000 1.000 1.000 0.971 0.971 Absolute error (median) 0.047 0.108 0.140 0.132 0.099 0.145 0.090 0.123 MCE (max) 0.002 0.004 0.006 0.004 0.003 0.004 0.003 0.004 1.000 0.941 1.000 1.000 1.000 1.000 1.000 0.971 Absolute error (median) 0.041 0.104 0.102 0.112 0.123 0.206 0.089 0.128 Width ratio (median) 0.988 0.985 0.862 0.968 0.930 0.997 0.950 0.987 MCE (max) 0.002 0.004 0.005 0.005 0.003 0.003 0.003 0.004 1.000 0.941 1.000 1.000 1.000 1.000 1.000 0.971 Absolute error (median) 0.041 0.104 0.126 0.114 0.109 0.206 0.092 0.128 Width ratio (median) 0.989 0.989 0.864 0.975 0.931 0.999 0.957 0.990 MCE (max) 0.002 0.004 0.005 0.005 0.003 0.003 0.003 0.004 -------- ------------------------- -------- -------- -------- -------- -------- -------- -- -------- -------- -- : Predictions of $\mu_{2ij}$ across treatments and models[]{data-label="CV"} Comparison of the results from F-EX, P-EX and those from subgroup analysis -------------------------------------------------------------------------- Figure 2 presents 95% CrIs of the slopes $\lambda_{1j}$ and intercepts $\lambda_{0j}$ across the treatment classes and methods of estimation. Comparing the aforementioned methods in regards to the surrogacy criteria on the PFS-OS pair, we can conclude that F-EX model estimated the parameters of the surrogate relationships with reduced uncertainty compared to the subgroup analysis and P-EX model taking advantage of borrowing of information across classes. P-EX relaxes the assumption of exchangeability reducing the effect of borrowing of information on average by 3.6%. It gave narrower CrIs of the parameters of interest compared to subgroup analysis but slightly larger than those obtained form F-EX model. Furthermore, both F-EX and P-EX methods can distinguish between the different association patterns avoiding to give over-shrunk estimates of the slopes and the intercepts, although they allow different degrees of borrowing of information for the slopes. In particular, this pair of outcomes (PFS-OS) illustrates well the impact of number of studies per class on the degree of borrowing of information. In general, borrowing of information is determined by the number of studies within treatment classes, between treatment classes heterogeneity, as well as the number of treatment classes. In this case, the fewer studies we have within a treatment class, the bigger is the impact of borrowing of information resulting in higher reduction in uncertainty of the estimates of surrogate relationships. This effect was particularly strong for the anti-EGFR treatment class. ![95% Credible intervals of $\lambda_{1j}$ and $\lambda_{0j}$ for the PFS-OS pair of outcomes[]{data-label="fig:3"}](PFSOSintervals.pdf){height="8.5cm" width="16.57cm"} On the other hand, TR-PFS pair is a good example to illustrate the performance of the hierarchical methods when between treatment class heterogeneity is relatively large. In this case, subgroup analysis performed equally well as the proposed methods in terms of uncertainty of the CrIs of the paramaters describing the surrogate relationships. For instance by fitting F-EX and P-EX models, we did not observe any decrease in uncertainty around $\lambda_{1j}$ and $\lambda_{0j}$ across classes. This is because the between treatment classes heterogeneity was relatively large for TR-PFS pair and hence there was not much shrinkage. Furthermore using subgroup analysis, the surrogacy criteria failed in the anti-EGFR class (zero was included in the 95% CrI of the slope) where only 8 studies available). However, the 95% CrI in the anti-EGFR class just contains zero and overlaps substantially with the 95% CrI of the slope for chemotherapy treatment class. By applying P-EX and F-EX models, we were able to draw different inferences for the surrogacy in the anti-EGFR class as these methods allow for borrowing of information for the parameters describing the surrogate relationships from the other classes. As illustrated in Figure 3, both hierarchical models moved the 95% CrI of the slope in the direction of the CrIs of the other two classes resulting in the surrogacy criteria being satisfied across all treatment classes. ![95% Credible intervals of $\lambda_{1j}$ and $\lambda_{0j}$ for the TR-PFS pair of outcomes[]{data-label="fig:4"}](TRPFSintervals.pdf){height="8.5cm" width="16.57cm"} When carrying out cross-validation procedure, we wish to ensure that not only predictive intervals contain the observed values but also that they are sufficiently narrow. In general, adding a hierarchical structure to slopes and intercepts reduces the uncertainty and leads to more precise predictions compared to those obtained from subgroup analysis. For the PFS-OS pair of outcomes, the accuracy of the predictions was very similar across all methods (similar absolute error) but the uncertainty varied depending of the level of borrowing of information. F-EX model gave on average the most precise estimates ($\hat{\mu}_{2ij}$) having the narrowest 95% predictive intervals of the effect on the final outcome (smallest width ratio seen in Tables 3, 5, 7) reducing the overall uncertainly by 5%. The benefit was smaller in the chemotherapy class where the number of studies was much larger compared to the anti-EGFR treatment class where we had only 8 studies available. Overall, P-EX performed better than subgroup analysis and equally well with F-EX regarding the uncertainty of the predictions. This indicates that the assumption of exchangeability seems to be plausible for this pair of outcomes and P-EX model was able to identify this. For the TR-PFS pair, subgroup analysis with the standard model was a robust approach in terms of the accuracy of its predictions. Although the overall absolute error was very similar across models, F-EX and P-EX yielded higher absolute error compared to subgroup analysis in the anti-angiogenic class. This implies that the posterior means of the true effects were to some extent ’overshrunk’ due to excessive borrowing of information from the other classes. P-EX model was implemented allowing for partial exchangeability only for the slopes, this decision is likely to affect the performance of the model in terms of its predictions on TR-PFS pair of outcomes. However, the model can be extended allowing for partial exchangeability also of the intercepts or the conditional variances and different combinations of these assumptions can be explored and models compared using DIC. Similarly, there was no significant decrease in the degree of uncertainty of the estimates $\hat{\mu}_{2ij}$ of F-EX and P-EX models. The results indicate that the hierarchical methods performed slightly better compared to subgroup analysis in terms of uncertainty only in the class of chemotherapy and the anti-EGFR treatment class giving 1.5% and 3% narrower predictive intervals respectively. This kind of behaviour might be caused by the relatively large between treatment class heterogeneity and the assumption of full exchangeability of the intercepts. Discussion ========== We developed two hierarchical models allowing to account for distinct treatment classes when examining the surrogate relationships. The proposed models may be particularly useful in surrogate endpoint evaluation in complex diseases where different treatment classes of different mechanism of action and potential different association patterns within those classes exist. These models investigate potential differences in study level surrogacy across treatment classes in a particular disease area and can help to identify treatment classes with strong association patters, even when data are relatively sparse. F-EX model is somewhat restrictive, assuming full exchangeability for the parameters describing the surrogate relationships across treatment classes. In many situations the assumption of exchangeability may be too strong given the heterogeneity between treatment classes. In such circumstances, a more flexible model such as P-EX may be a better choice. P-EX model can infer an appropriate level of borrowing of information from the data, reducing the degree of borrowing of information through the mixture weights, thus relaxing the assumption of exchangeability when it is not fully reasonable. It evaluates whether the association pattern between treatment effects (logHR or logOR) on the surrogate and the final endpoint in a specific treatment class differs from the other patterns in other classes. F-EX model is appropriate only when the degree of similarity of surrogate relationships is relatively high. It can offer substantial gains in precision, reduced RMSE of the posterior means of the parameters describing surrogate relationships and it can improve the predictions of the true effects on the final endpoint. For example, F-EX model gave posterior means of the slopes and predicted effects with reduced uncertainty (smaller credible intervals) compared to subgroup analysis for the first simulated data scenario and for the illustrative example on PFS-OS pair where the parameters describing the surrogate relationship were similar and the assumption of full exchangeability was reasonable. These findings are consistent with the results from other hierarchical Bayesian methods which assume full exchangeability and were developed in other research areas [@berry2013Hierarchical; @thall2003]. However, P-EX model achieves the same degree of borrowing of information in such data scenarios making less assumptions compared to F-EX model. P-EX model regulates the degree of borrowing of information using its exchangeable and non-exchangeable components with respective mixture weights. For instance, when between treatment class heterogeneity is relatively large or there is a treatment class with distinctly different pattern, P-EX model has the advantage of avoiding the excessive borrowing of information, as illustrated in the second design of the simulation study. All the above illustrate the benefits of partial exchangeability, as described by Neuenschwander et al. [@neuenschwander2016robust] in their work. Subgroup analysis using the standard model is a simple method which performs well when there are sufficient data available for each treatment class, but it produces estimates with higher bias and uncertainty when data within a treatment class are limited. In any other situation P-EX should be preferred as it regulates the degree of borrowing of the slopes. Although the proposed methods provide additional robustness to the CrIs and the posterior means of the parameters describing the surrogate relationships compared to subgroup analysis, potential limitations should always be kept in mind. First, in real data scenarios it can be challenging to find data sets with sufficient number of treatment classes. The small number of treatment classes can affect the performance of hierarchical methods substantially [@mcneish2016modeling] reducing the impact of borrowing of information. For instance, fitting P-EX model to the illustrative example (in aCRC with three treatment classes) led to a situation where in some of the MCMC iterations only one class was deemed exchangeable by the model which is not possible since there were no other classes to exchange information with. However, in our example it did not affect the performance of the model as it occurred only in the 0.5% of the MCMC iterations. On the other hand, there is no upper limit to the number of classes we can have. In general, the more classes the better it is for the models to borrow information across them. Another limitation of the illustrative example is that treatment switching was applied in a subset of trials in this data set. Patients were allowed to switch from the treatment that was initially assigned to them to the other treatment arm in the trial. Most commonly patients switched after progression from control to experimental arm in particular, if there was sufficient evidence during the trial that the experimental treatment was better than control [@latimer2016treatment]. Treatment switching has diminishing effect on the difference in treatment effects on OS when applying intention-to-treat analysis, and the effect is often obtained with larger uncertainty. This makes the estimation of surrogacy between treatment effects on the surrogate and treatment effects on the final outcome very challenging. Many adjustment methods have been proposed, however, their validity is often questionable[@latimer2016treatment]. Additionally, the evaluation of PFS as a surrogate endpoint is distinctive compared to other surrogate endpoints as PFS can be considered as nested outcome within OS outcome. These factors may explain the different findings for the two pairs of outcomes (PFS-OS and TR-OS). Furthermore, as it was mentioned in section \[Application\], each treatment class consist of studies with multiple treatment comparisons. According to Daniels and Hughes [@daniels1997meta] and Shanafelt et. al [@shanafelt2004chemotherapy] different treatment comparisons and the use of active or inactive control interventions may influence the surrogate relationship. This could potentially be resolved by classifying treatment according the treatment class comparison (for example anti-angiogenic therapies versus chemotherapy) which potentially would lead to more treatment classes, but with reduced number of studies per class. To continue with the same issue, in this paper the treatment classes were defined according to the class of the experimental treatment regardless the control. Alternatively, we could classify them according to the treatment contrasts taking into account the class of the control group, however, this could result in fewer studies per class. A network meta-analysis model was developed for this problem by Bujkiewicz et al. [@bujkiewicz2019bivariate]. Additionally the evaluation framework proposed by Daniels and Hughes (see section \[Criteria\]) examine whether zero is contained in the CrIs of $\lambda_{1}$ and $\lambda_{0}$. However, the sparsity of data may lead to increased uncertainty around the intercept and slope. This increased uncertainty is also likely to manifest itself in increased conditional variance, thus invalidating the third criterion. Unsurprisingly, for sparse data it is unlikely that all the surrogacy criteria hold and this problem is more likely to occur in subgroup analyses. Our proposed methods alleviate this problem as shown in some of the scenarios of the simulation study. However, we used the criteria mainly for the purpose of model comparison. In real life scenarios, when evaluating a potential surrogate endpoint for use in clinical trials or regulatory decision making, the decision of whether the surrogate endpoint may be used to make the prediction of the clinical benefit should be based on the balance between the strength of the surrogate relationship and the need for the decision to be made about the effectiveness of the new treatment [@alonso2016applied]. Moreover, the strength (or weakness) of the surrogate relationship will manifest itself in the width of the predicted interval of the treatment effect on the final outcome. A larger interval around the intercept and slope will result in a larger interval around the predicted effect and hence increased uncertainty about the regulatory or clinical decision made based on such prediction. The implication of this is that perhaps we do not need precise surrogacy criteria and instead we need only look at the predictions [@bujkiewicz2019bivariate]. The quality of predictions can be evaluated through a cross-validation procedure (see section \[CrossVal\]). A possible extension of these methods is to add another layer of hierarchy accounting for the different treatments within a treatment class. However, a relatively large number of studies for each treatment and number of treatments per class would be required to fit such model. As we mentioned in section 6.4, P-EX model could also be extended by making additional partial-exchangeability assumptions about the intercepts and the conditional variances, however, this may lead to over-parameterising the model. Furthermore, taking advantage of the setting proposed by Bujkiewicz et al. [@bujkiewicz2015bayesian], both hierarchical models can be extended to allow for modelling multiple surrogate endpoints (or the same surrogate endpoint but reported at multiple time points) as joint predictors of treatment effect on the final outcome. Further research is also needed to extend the proposed methodology to Binomial data or to time to event data where the assumption of normality is not plausible. Moreover, to overcome the convergence issues caused by vague prior distributions on the hyper-parameter of the mixture weights ($\pi_{j}$), alternative prior distributions should be developed by extending the P-EX in a similar way as proposed by Kaizer et al. [@kaizer2017bayesian]. In summary, we developed hierarchical Bayesian methods for evaluating surrogate relationships within treatment classes whilst borrowing of information for surrogate relationships across treatment classes. We believe that the proposed methods have a lot of potential for improving the validation of surrogate endpoints in the era of personalized medicine, where the surrogacy may depend on the mechanism of action of specific targeted therapies. Acknowledgements ================ This research used the ALICE/SPECTRE High Performance Computing Facility at the University of Leicester and was partly funded by the Medical Research Council, grant no. MR/L009854/1 awarded to Sylwia Bujkiewicz. Keith Abrams is partially supported as a UK National Institute for Health Research (NIHR) Senior Investigator Emeritus (NI-SI-0512-10159). We also thank the anonymous reviewers for their comments, which helped to improve the quality of the manuscript.

ArXiv

--- abstract: 'When data productions and consumptions are heavily unbalanced and when the origins of data queries are spatially and temporally distributed, the so called *in-network data storage paradigm* supersedes the conventional *data collection paradigm* in *wireless sensor networks* (WSNs). In this paper, we first introduce *geometric quorum systems* (along with their metrics) to incarnate the idea of in-network data storage. These quorum systems are “geometric” because curves (rather than discrete node sets) are used to form quorums. We then propose GeoQuorum as a new quorum system, for which the quorum forming curves are parameterized. Though our proposal stems from the existing work on using curves to guide data replication and retrieval in dense WSNs, we significantly expand this design methodology, by endowing GeoQuorum with a great flexibility to fine-tune itself towards different application requirements. In particular, the tunability allows GeoQuorum to substantially improve the load balancing performance and to remain competitive in energy efficiency. Both our analysis and simulations confirm the performance enhancement brought by GeoQuorum.' author: - bibliography: - 'GeoQuorum.bib' title: 'GeoQuorum: Load Balancing and Energy Efficient Data Access in Wireless Sensor Networks' --- Introduction {#sec:intro} ============ Since their inception, *wireless sensor networks* (WSNs) bear the task of intensive data collection through their large scales and dense deployments, which represents a significant improvement over traditional sensing systems [@AkyildizSSC-CM02]. However, the low-cost devices (those tiny sensor nodes) involved in a WSN are also constraining factors to their missions: the limited energy storage of a node heavily confines its ability of intensively transmitting the acquired data. As a result, we have been witnessing a great volume of research developments, aiming at tackling the conflict between the need for low power operations and the requirements of large scale data gathering. The related research proposals have been mainly focusing on two issues: namely load balancing and energy efficiency [@AhnK-MobiHoc06]. Whereas the former is concerning not imposing a too heavily communication load upon a small set of nodes, the latter is managing to reduce the total communication load taken by the whole WSN. From a conventional point of view, a WSN needs to collect data from a large set of nodes to a particular (often small) set of nodes. The resulting *convergecast* type of data transmission pattern under this assumption makes the above two objectives contradict each other. For example, shortest path routing, as the most energy efficient communication protocol, may lead to very unbalanced load distribution, hence “kill” those heavily loaded nodes long before other nodes running out of their battery [@ChangT-Infocom00; @LuoH-ToN10]. Fortunately, the convergecast data collection paradigm is not the only way of acquiring data from a WSN. In particular, when data productions and consumptions are heavily unbalanced (e.g. queries from human users to the sensory data are less frequent than generation of those data), or when the queries may originate in a spatially and temporally distributed manner, other data access schemes involving *in-network data storage* (e.g., [@RatnasamyKYYEGS-WSNA02; @ShengLM-MobiHoc06]) are more meaningful. More importantly, we may strike a better tradeoff between load balancing and energy efficiency under such a data access paradigm, as our paper will demonstrate. Here we simply provide an illustration of a data access paradigm using in-network data storage in Fig. \[fig:qillustrate\]. ![Comparison of two data access paradigms.[]{data-label="fig:qillustrate"}](quorumillustrate){width="\columnwidth"} It is clear that, whereas the convergecast collects the data at a single point, the in-network storage replicates data at various nodes, to which a later data query is directed. The latter paradigm endows a greatly flexibility to the data access: it can be performed whenever and wherever needed. Last but not not least, the in-network data storage is technically enabled by the new developments on flash memory storage [@MathurDGS-IPSN06], which is ultra energy-efficient and has huge capacity. In this paper, we are focusing on a particular design methodology, *quorum systems*, under the in-network data storage paradigm. Based on this methodology, data produced by sensor nodes and queries generated by human users are both directed to certain *quorums* (subsets of nodes). As the intersection between quorums are guaranteed in the design phase, users may access the sensory data without directly communicating with the sources that generate those data. Although quorum systems exist in distributed systems [@MalkhiR-DC98] and have been applied to wired and wireless networking (e.g., [@ZhouSR-TCS02; @HaasL-ToN99; @LuoEH-TMC04]), we are reviving them in the sensor networking scenarios. Moreover, our design method, namely, *geometric quorum systems* (GQS), leverages on the recent developments in using geometric principles to guide the protocol implementations in WSNs, e.g., [@SarkarZG-MobiCom06; @ZengSLGG-Infocom10]. In particular, we propose GeoQuorum where the quorums are formed by parameterized curves. Tuning the parameters that determine the quorums allows us to flexibly identify desired tradeoffs between load balancing and energy efficiency. Through both analysis and simulations, we further demonstrate that our design outperforms the existing ones in terms of both load balancing and energy efficiency. In summary, our main contributions are: - A formal definition of GQS and the related metrics. - A thorough analysis of the existing quorum system designs for WSNs against the defined metrics. - A general conformal geometry based quorum design methodology that applies to WSNs with any shape of the network areas. - A specific quorum system, GeoQuorum, formed by parameterized curves, allowing a flexible tradeoff to be made between load balancing and energy efficiency. The remaining of this paper goes as follows. In Sec. \[sec:qs\], we define quorum systems (in the traditional sense) and their metrics, and we also briefly review the application of quorum systems in networked settings, in particular a recent geometry-based quorum system design. We focus on GQS in Sec. \[sec:gqs\]. Starting with the conformal geometry basics and network model in Sec. \[sec:bkgccg\], we formally define GQS in Sec. \[sec:gqsmd\], we then analyze the performance of existing designs in Sec. \[sec:gqsals\] and propose GeoQuorum in Sec. \[sec:gqsgq\]. We report the simulation results in Sec. \[sec:sim\] and conclude our paper in Sec. \[sec:con\]. Fundamental of Quorum Systems {#sec:qs} ============================= Basic Definitions {#sec:qsbs} ----------------- Quorum systems represent a fundamental abstraction for coordination among the nodes of a distributed system (e.g., a set of networked nodes). In its traditional sense, a quorum system is defined upon a finite set (also termed *universe*) $\mathcal{U} = \{u_1, u_2, \cdots, u_n\}$ of nodes. In particular, the following definition characterizes a quorum system [@MalkhiR-DC98]. A quorum system $\mathcal{Q} \subset 2^\mathcal{U}$ is a set of subsets of $\mathcal{U}$ such that every two subsets intersect. Each $Q \in \mathcal{Q}$ is called a *quorum*. Given a quorum system $\mathcal{Q}$, networked nodes may make use of it to perform coordinations, such as sharing information. A node may choose to *access* a quorum by either *writing* to or *reading* from it. Thanks to the intersection property, a read access will find the desired data from some quorum that stores the data written by another node. Note that the goal of a read access is to search for the data, whereas data delivery is carried out by a certain routing scheme that is independent of the quorum system. Taking into account the inherent asymmetry between read and write accesses, we may redefine the quorum system in a asymmetric fashion as follows [@LuoEH-TMC04], the earlier definition hence specifies *symmetric* quorum systems. An asymmetric quorum system $\mathcal{Q} \subset 2^\mathcal{U}$ consists of two disjoint sets, $\mathcal{Q}^R$ and $\mathcal{Q}^W$, of subsets of $\mathcal{U}$, such that each subset in $\mathcal{Q}^R$ intersects every subset in $\mathcal{Q}^W$. Each subset in $\mathcal{Q}^R$ (resp. $\mathcal{Q}^W$) is called a *read* (resp. *write*) *quorum*. Metrics on Quorum Systems ------------------------- We introduce two metrics to measure the performance of quorum systems, namely, *load* and *robustness*. ### Load This metric measures the computational load taken by individual nodes due to their participation in various quorums. Obviously, it depends not only on how a quorum system is constructed, but also on what strategy individual nodes adopt to access the system. An access strategy $S$ consists of an *access rate* $\lambda_S$ and a probability measure $P_S$ on $\mathcal{Q}$, i.e., $\sum_{Q \in \mathcal{Q}} P_S(Q) = 1$. The strategy is *pure* if $P_S(Q) = 1$ for some $Q \in \mathcal{Q}$; otherwise it is *mixed*. \[def:as\] For asymmetric quorum systems, we replace $\mathcal{Q}$ by $\mathcal{Q}^R$ or $\mathcal{Q}^W$, depending on which access operation is under consideration. The load induced by $S$ on a node $u_i$ is $$\ell_S(i) = \sum_{Q \in \mathcal{Q}:u_i \in Q} \lambda_S P_S(Q).$$ The *system load* induced by $S$ on a quorum system $\mathcal{Q}$ is the maximal load induced by $S$ on any node in $\mathcal{U}$, i.e., $$I\!\!L_S(\mathcal{Q}) = \max_{u_i \in \mathcal{U}} \ell_S(i).$$ \[def:ld\] Intuitively, this metric measures the evenness of load distribution within the whole system: the lower the system load, the more balanced the load is distributed. ### Robustness As another important metric, *robustness* indicates the ability of a quorum system to cope with node failures (viz. its fault tolerance). Many measures have been proposed for this metric, we choose the most straightforward one: the size of the intersection between two quorums. The robustness of a quorum system $\mathcal{Q}$ is the size of the minimum intersection between an arbitrary pair of quorums $$I\!\!R(\mathcal{Q}) = \min_{Q_i, Q_j \in \mathcal{Q}} |Q_i \cap Q_j|.$$ For asymmetric quorum systems, $Q_i, Q_j \in \mathcal{Q}$ is hence replaced by $Q_i \in \mathcal{Q}^W$ and $Q_j \in \mathcal{Q}^R$. \[def:rob\] It is straightforward to see that, if the system robustness is $k$, then any node failures involving less than $k$ nodes will not affect the intersection property of the system. Related Work on Conventional Quorum Systems ------------------------------------------- Traditional quorum systems are confined in 2D space, and hence only allow for limited designs, such as the grid shown in Figure \[fig:grid\](a), ![Quorums system based on 2D linear curves.[]{data-label="fig:grid"}](Grid){width=".8\columnwidth"} or the B-Grid [@NaorW-SIAMJC98], shown in Figure \[fig:grid\](b), for improving the robustness. Similar ideas were re-introduced into *mobile ad hoc networks* (MANETs) and WSNs [@HaasL-ToN99; @YeLCLZ-MobiCom02; @LiuHZ-Sensys04], though sometimes under different names. These designs are often so rigid that they allow very little tunability that adapts a system to various application requirements. To improve the flexibility of the quorum systems, *probabilistic quorum systems* [@MalkhiRW-IC01] were introduced to relax the intersection rule (making it a random variable) and to leave more freedom in trading load for robustness; they were later applied to MANETs to cope with node mobility [@HaasL-ICC99]. Interested readers are referred to [@LuoHE-MobiHoc03; @LuoEH-TMC04; @FriedmanKA-DSN08] for more recent developments in probabilistic quorum systems. In general, probabilistic quorum systems are designed to cope with system dynamics (e.g., node mobility), hence they are trading system efficiency for higher robustness. As we explained in Sec. \[sec:intro\], energy efficiency is a crucial issue in WSNs, whereas nodes in WSNs are often static. Consequently, we advocate a deterministic design for quorum systems, while relying on other techniques (rather than pure randomization) to improve its flexibility. Quorum Systems in A Projective Space {#sec:qsps} ------------------------------------ Recently, a new design methodology for (deterministic) quorum systems was proposed in [@SarkarZG-MobiCom06]. This method suggests using projective map to first “lift” the 2D network area onto a 3D surface, a sphere, then design quorum systems on the 3D surface, and finally project the designed system back to the 2D area. As the system design is done in the 3D space, it allows much more diversity in “shaping” the quorums, and hence has a potential to deliver more flexible system designs. The practicality of this design approach is backed by the *trajectory based forwarding* [@NiculescuN-MobiCom03], which allows a routing path to be defined by a continuous curve. Given a certain data type, two designs are proposed in [@SarkarZG-MobiCom06].[^1] - Symmetric quorum systems with each quorum represented by a great circle. The access strategy for a write access is pure as the corresponding great circle is fixed by two points: the node accesses a quorum and the geographical hash $h$ of the data type. - Asymmetric quorum systems with write quorums represented by great circles and read quorums by latitude circles. While the access strategy for a write access is the same as the first design, that of a read access also becomes pure, as the circle of each read quorum is also defined by the node access the quorum and the geographical hash $h$ of the data type. We illustrate the two designs in Fig. \[fig:drulings\]. ![Quorum systems designed in [@SarkarZG-MobiCom06]. We use red (resp. blue) color to indicate quorums accessed by a write (resp. read) access. For quorums in red, the corresponding geographical hash location $h$ and its antipodal point $\bar{h}$ are shown. We also use pentagrams to represent the intersection between quorums, and triangles to represent the nodes that access a quorum.[]{data-label="fig:drulings"}](DRulings){width=".8\columnwidth"} The projective map used in [@SarkarZG-MobiCom06] is stereographic projection, which is easy to use due to its closed-form expression. However, using stereographic projection, the north pole is mapped to infinity on the plane. Consequently, as the sphere cannot be fully covered by the (mapped) network area, conclusions drawn through geometric analysis on the sphere surface may not be valid for the original 2D network area. For example, two curves intersecting on the sphere may not retain the property within the network area, as the intersection may be out of the boundary: only one of the two intersections between two great circles is guaranteed in [@SarkarZG-MobiCom06]. In terms of quorum system design, apart from presenting heuristics, no rigorous definitions and metrics are provided for the quorum systems, thus no formal analysis is given in [@SarkarZG-MobiCom06] to evaluation the performance of the designed system. Also, only planar curves are used to represent quorums on the 3D surface, which significantly confines the design flexibility. In addition, as we will show later, any two planar curves can intersect at up to two points on a sphere, the system robustness is hence fixed and cannot be tuned. Geometric Quorum Systems for Data Access {#sec:gqs} ======================================== In this section, we first introduce the geometry background and define our network model, along with the properties and metrics of geometric quorum systems (GQS). Then we analyze the performance of existing designs, based on the defined metrics. Finally, we present our asymmetric quorum systems, GeoQuorum, that makes use of spatial curves to substantially improve the flexibility in fine-tuning system performance. Background on Computational Conformal Geometry {#sec:bkgccg} ---------------------------------------------- *Computational conformal geometry* (CCG) is an emerging research field spanning computer science and pure mathematics. It focuses on developing the computational methodologies on discrete surfaces to discover conformal geometric invariants. Intuitively speaking, a conformal map is a function that preserves the angles. Due to its shape preserving properties, computational conformal geometry has broad applications in both pure theoretic research, such as mathematics, theoretical physics, and engineering applications, such as computer graphics, computer-aided design, computer vision, etc. See [@Gu09] for a recent survey of CCG. Here we briefly discuss the CCG tools that we will use in this paper, which is also illustrated in Fig. \[fig:mapping\]. Given a simply connected shape $D\in\mathbb{R}^2$ with boundary $\partial D$ and a node set inside it, we first construct a Delaunay triangulation using [@Shewchuk96], then we construct a closed surface by double covering: make a copy of $D$, denoted by $D'$, reverse its orientation and glue $D$ and $D'$ along the common boundary. The resulted surface $\bar D$ is a closed surface of genus 0, i.e., topologically equivalent to a sphere. We then compute a harmonic function $\phi$ mapping $\bar D$ to the unit sphere, i.e., $\phi:\bar{D}\rightarrow\mathbb{S}^2$ such that $\bigtriangleup\phi=0$ where $\bigtriangleup$ is the Laplace-Beltrami operator. This map has the following promising properties: - $\phi$ is conformal, thus, there is no angle distortion [@DBLP:journals/tmi/GuWCTY04]; - Due to symmetry, $D$ and $D'$ are mapped to the north and south hemispheres respectively, i.e., $\phi(D)$ and $\phi(D')$ are mirror reflected with respect to the equator. - The boundary $\partial D$ is mapped to the equator. As $\partial D$ could be arbitrary 2D simple curve, this algorithm applies for arbitrary 2D simply connected region $D$. Here we should emphasize that there are other ways to compute the conformal mappings on the sphere, such as the stereographic projection used in [@SarkarZG-MobiCom06]. However, the harmonic map based method as mentioned above allows us to map arbitrary simply connected region to cover the whole sphere, thus avoiding various issues involved in stereographic projection. Network Model and Geometric Design Basics {#sec:gqsmd} ----------------------------------------- We represent a WSN by $\mathcal{U}$, with $u_i \in \mathcal{U}$ being a sensor node. $\mathcal{U}$ also serves as the universe upon which a quorum system can be defined. ### Geometric Model of WSNs We apply the tool discussed in Sec. \[sec:bkgccg\] to map the network area to a sphere of **unit radius**. For the reverse projection, any curve that passes across the equator has its upper and lower sections projected separately to the two network areas. Then two (projected) sections are combined to get the projection on the original network area, as shown in Fig. \[fig:mapping\](e). This improved map allows us to perform geometric analysis on the whole sphere surface. In order to facilitate the analysis of geometric quorum systems in Sec. \[sec:gqsals\], we assume that the tessellation on the sphere surface is regular and use the vertices to represent sensor nodes. Such a model makes sense in dense WSNs where nodes are uniformly distributed; the tessellation vertices can be considered as representatives of the nodes in a close-by region. We use this model only to simplify analysis, our numerical simulations still take arbitrarily deployed WSNs as input, and the tessellation on a sphere simply results from the triangulation of the network nodes, as shown in Fig. \[fig:mapping\](c). ### Geometric Quorum Systems We extend the conventional definitions for quorum systems (presented in Sec. \[sec:qsbs\]) to geometric quorums system (GQS).   A GQS $\mathcal{Q}$ is a set of curves in space $\mathcal{A}$ ($\mathcal{U} \subset \mathcal{A}$), such that every two curves intersect. Each curve in $\mathcal{Q}$ defines a quorum. The definition for asymmetric quorum systems is omitted; one simply splits $\mathcal{Q}$ into $\mathcal{Q}^W$ and $\mathcal{Q}^R$, and intersection is only required between the two sets. We keep using the same definition for access strategy (*Definition \[def:as\]*), load (*Definition \[def:ld\]*), and robustness (*Definition \[def:rob\]*), but the interpretations are slightly different. In particular, a sensor node $u_i \in \mathcal{U}$ is also a tessellation vertex of $\mathcal{A}$, and $u_i \in Q$ means $u_i$ is a vertex of a triangle passed by the curve $Q$. The system load defined for a WSN is the maximum energy consumption for a certain tessellation vertex to **transmit** the data (for write) or queries (for read) to the quorum at which they aim, representing the load balancing effect of the quorum system. Unlike traditional distributed systems, the energy efficiency (or total energy consumption of the whole WSN) is also a major concern of WSNs. Let $M(Q)=|\{u \in \mathcal{A}|u \in Q\}|$ be a measure of the total energy consumption of a quorum $Q$, we further define a metric to measure this performance aspect. The total load induced by $S$ on a certain quorum $\mathcal{Q}$ is $$I\!\!L_T(\mathcal{Q}) = \sum_{Q \in \mathcal{Q}} \lambda_S P_S(Q) M(Q).$$ \[def:tld\] In general, each node may take a different access strategy. To simplify the analysis, we only distinguish between two types of strategies, namely $S_R$ and $S_W$ for read and write respectively. Existing Quorum System Designs {#sec:gqsals} ------------------------------ In this section, we analyze the performance of the two quorum systems designed in [@SarkarZG-MobiCom06], namely $\mathcal{Q}_G$ and $\mathcal{Q}_L$ described in Sec. \[sec:qsps\]. We will point out that, if we design quorum systems based only on planar curves, the system robustness is very limited and the system load can be very high. More importantly, such a design may be lack of flexibility to cope with a high asymmetry between $S_R$ and $S_W$, which can lead to an unnecessarily high total load. Based on our analysis, we also propose certain remedies to improve these two designs. ### Symmetric Quorum Based on Great Circles ($\mathcal{Q}_G$) We first show that **the system robustness is limited**. Instead of considering only great circles, the following result encompasses all possible planar curves on a sphere. Any two distinct planar curves intersect at most at two points on a sphere. Any planar curve on the sphere surface is the intersection between the sphere surface and a certain plane (*cutting plane* hereafter), and two planar curves intersect each other iff their corresponding cutting planes intersect each other. In addition, the intersection of two planar curves is a subset of that of the two cutting planes. Now, as the intersection of two cutting planes (of two distinct planar curves) is bounded to be a line, and as a line may intersect the sphere surface at (at most) two points, the proposition follows. The following corollary is immediate. The robustness of $\mathcal{Q}_G$ is no more than 2. The implication of this result is very clear: if two nodes of a WSN fail, the intersection property of the quorum system may be violated. If we consider a specific pair of quorums, the probability of the two intersecting nodes fail simultaneously is not negligible. Moreover, the robustness is fixed no matter what planar curves are used to define a pair of quorums. Secondly, we demonstrate that the access strategy defined by geographical hash leads to **very unbalanced load distribution**. It is straightforward to see that the write strategy is pure. In particular, the quorum that can be chosen by a writing node is fixed by that node and the hash location $h$. \[prp:load\] Let $\mathcal{D} = \{d_1, d_2, \cdots, \}$ be a set of data types in a WSN, and let $N(d)$ be the number of nodes contributing to a data type $d \in \mathcal{D}$. We have $$I\!\!L_S(\mathcal{Q}_G) = \lambda_{S_W} \max_{d \in \mathcal{D}} N(d) + \lambda_{S_R} \mathcal{O}(1)$$ The proof is based on *Definition \[def:ld\]*. The first term is incurred by the write access. As the write strategy is pure, we have $\ell_{S_W}(i) = \sum_{Q \in \mathcal{Q}:u_i \in Q} \lambda_{S_W} \mathbf{1}_Q$. Therefore, the load is computed by counting how many quorums include a node $u_i$. Because every write quorum for a given data type includes the corresponding hash location, the maximum loaded node is at the hash location whose corresponding data type has the largest number of contributors. The second term is introduced by the read access. As the read strategy can be mixed, the optimal strategy is apparently a uniform distribution over all possible quorums, given the assumption of a regular tessellation on the sphere surface. Consequently, the load introduced by a read access is identical for all nodes and can be bounded by a constant (which depends on the density of the tessellation vertices). Note that the first term is fully determined by the nature of the sensory data produced by a WSN. In the case that many nodes are contributing to the same data type, the system load can be extremely high. In practice, this result suggests that, for every node contributing to a certain data type, a certain amount of communication load is imposed upon the nodes around the hash location at a rate $\lambda_{S_W}$. Therefore, the more nodes contributing to this data type, the sooner the nodes around the hash location will run out of their energy storage. We show the two sets of quorums in Fig. \[fig:existing\](a); we plot the 2D project in the network area to avoid confusion in 3D representations. One extreme case is shown in the top figure, where the hash location coincides the network center, ![Geometric quorum systems designed using planar curves in 3D. We mark write (reps. read) quorums by red (resp. blue) color, and we also show the corresponding geographical hash location $h$ and its antipodal point $\bar{h}$.[]{data-label="fig:existing"}](Existing){width="\columnwidth"} imposing very high load there.[^2] One possible **remedy** $\mathcal{Q}_G^m$ is to adopt a mixed strategy for write access. For example, it is straightforward to see that, if the nodes contributing to a certain data type are uniformly distributed, a mixed strategy with uniform distribution among all possible quorums (great circles) may balance the load. However, as great circle is the longest planar curve on a sphere, always using it as quorum can lead to unnecessarily high total load, in particular if, for example, the read and write strategies have very different access rates.[^3] ### Asymmetric Quorum Based on Great Circles and Latitude Curves ($\mathcal{Q}_L$) According to Sec. \[sec:qsps\], $\mathcal{Q}_L$ shares the same write quorums with $\mathcal{Q}_G$. Therefore, the same drawbacks, namely low robustness and high system load, persist. Actually, what $\mathcal{Q}_L$ improves (comparing with $\mathcal{Q}_G$) is total load, as shown by the following proposition. For a given data type, if the nodes that access (by read or write) a quorum system are uniformly distributed on the sphere, the part of the total load contributed by the read access of $\mathcal{Q}_L$, $I\!\!L_T^R(\mathcal{Q}_L)$, is $\frac{\pi}{4}$ of that of $\mathcal{Q}_G$. Let $R_G$ and $R_L$ be the radius of a great circle and a latitude circle, respectively. Based on the assumption of a regular tessellation on the sphere surface, we use the length of a quorum $Q$ (a curve) to represent $M(Q)$, the measure of its total energy consumption. As shown in Fig. \[fig:latitude\], ![The relation between $R_G$ and $R_L$.[]{data-label="fig:latitude"}](Latitude){width=".4\columnwidth"} $R_L$ is determined by $R_G$ and the angle $\phi$. Hence we have $$\begin{aligned} \mathbf{E}_\phi\left(M(Q_L)\right) &=& 2 \pi \int_{0}^{\frac{\pi}{2}} \mathrm{Pr}\{R_L = R_G\sin(\phi)\} R_L \nonumber \\ &=& 2 \pi R_G \int_{0}^{\frac{\pi}{2}} \sin^2(\phi) d\phi ~~=~~ \frac{\pi^2}{2} R_G \nonumber \end{aligned}$$ Since we have the total load contributed by a write quorum of $\mathcal{Q}_G$ as $2 \pi R_G$, the ratio $\frac{\pi}{4}$ follows. As the part of the total load contributed by the write access is the same for $\mathcal{Q}_G$ and $\mathcal{Q}_L$, the above result shows that load reduction brought by $\mathcal{Q}_L$ may become marginal if a lot of nodes are contributing to a certain data type. One immediate **remedy** to cut down both the system load and total load is $\mathcal{Q}_L^d$, a dual design of $\mathcal{Q}_L$. In other words, we swap the quorums for read and write, with read quorums given by great circles and write quorums given by latitude circles, as illustrated in Fig. \[fig:existing\](b). This design can reduce the system and total load if only a few write quorums are contributing to a data type, because, though all the read quorums for a certain data type pass through the hash location and its antipodal point, the read access may terminate (i.e., reaches an intersection with some write quorum) before reaching those locations. Unfortunately, the load reduction effect diminishes with an increasing number of write quorums. GeoQuorum: System with Spatial Quorums {#sec:gqsgq} -------------------------------------- Besides the various drawbacks we have pointed out, a main disadvantage of $\mathcal{Q}_G$ and $\mathcal{Q}_L$ is the lack of flexibility to be fine-tuned, hence they cannot adapt to different access rates. In this section, we present GeoQuorum as a new design. GeoQuorum makes use of spatial curves to form quorums, hence allows a great deal of freedom in fine-tuning the system performance. GeoQuorum is an asymmetric quorum system, with write and read quorums formed by different type of curves. Specifically, we have - **write quorums** are formed by circles with adjustable radius $R_W$, where $R_W$ can be tuned according to the access rate. - **read quorums** are formed by a special spherical spiral, defined as follows: $$\begin{aligned} % x &=& \cos \left(\theta + \theta_0 \right) \cos \phi \\ y &=& \sin \left(\theta + \theta_0 \right) \cos \phi \\ z &=& \sin \phi % \end{aligned}$$ where $\phi = a\theta$ and $\phi \in \left[-\frac{\pi}{2},\frac{\pi}{2}\right]$. Let $\alpha$ be the angle (with respect to the sphere center) between two consecutive loops ($\Delta \theta = 2\pi$), we have $\alpha = 2a\pi$. The parameter $a$ is determined by $R_W$ and the required robustness. - **access strategy** is mixed: - a **write quorum** is randomly chosen among all circles passing through the node that executes a write access, and - a **read quorum** starts from the node that executes a read access and ends at its antipodal point, with a randomly chosen $\theta_0$. We illustrate such a quorum system in Fig. \[fig:spiral\](a). ![Geometric quorum system designed using spatial curves in 3D.[]{data-label="fig:spiral"}](Spiral){width=".9\columnwidth"} Note that the current design is based on the assumption that $\lambda_W > \lambda_R$; otherwise we adopt a dual design where we swap the write and read quorums. We first show the relation between $R_W$ and $a$ by the following proposition. If $R_W \ge ka\pi, a \in (0,0.5)$, then the robustness of GeoQuorum is at least $2k$. \[prp:rbst\] As shown in Fig. \[fig:spiral\](b), if $2R_W \ge 2a\pi$ (i.e., $k=1$, the arc length delimited by the write quorum (a circle) is bounded to be larger than $2a\pi$, which is indeed the distance between two consecutive loops of the spherical spiral on the sphere surface. Consequently, the read quorum formed by the spherical curve has a least one loop passing through the write quorum, resulting in 2 intersections. Every time $k$ is increased by one (i.e., $2R_W \ge 2ka\pi$), an extra loop of the spiral passes the circle, given 2 more intersections, hence $2k$ in total. Therefore, given a certain robustness requirement, we have a one-to-one correspondence between $R_W$ and $a$: $R_W = ka\pi$, as choosing the smallest circle minimizes the incurred system and total load. Under the assumption that $\lambda_W > \lambda_R$, we may choose to tune $R_W$ according to $\lambda_W$ (the higher the rate the smaller $R_W$ is), then we match $a$ to $R_W$ based on the required robustness. Due to the use of mixed access strategy and the parameterized design, GeoQuorum can be tailored to meet the application requirements, such that both system load and total load can be reduced; which we will show in Sec. \[sec:comp\]. Interestingly, our design includes $\mathcal{Q}_G$ and $\mathcal{Q}_L$ as special cases. The following proposition shows that, by making $a$ sufficiently large, the spherical spiral becomes a great circle. If $a \rightarrow \infty$, the spiral curves becomes part of a great circle, whose orientation is determined by $\theta_0$. Let $\mathbf{C}(\theta)=(x(\theta,\phi), y(\theta,\phi), z(\theta,\phi))^{T}$ denote the spherical spiral curve, where $\phi=a\theta$. A simple computation shows that $\mathbf{C}''=-a^2(\mathbf{C}+\mathbf{D})$, where $\mathbf{D}=(d_x, d_y, 0)^T$ and $d_x=\frac{-1}{a^2}(\cos(\theta+\theta_0)\cos(a\theta)-2a\sin(\theta+\theta_0)\sin(a\theta))$, $d_y=\frac{-1}{a^2}(\sin(\theta+\theta_0)\cos(a\theta)+2a\cos(\theta+\theta_0)\sin(a\theta))$. As $\lim_{a\rightarrow\infty}\mathbf{D}=\mathbf{0}$, the curvature vector of $C$ is parallel to the normal vector of sphere when $a\rightarrow\infty$, i.e., the geodesic curvature of $C$ is zero. Observe that the geodesics on the sphere are great circles, so $\bf C$ is part of a great circle. Furthermore, the orientation of the great circle is determined by its binormal vector $\mathbf{C}'\times\mathbf{C}''$, which is parallel to $\mathbf{C}'\times\mathbf{C}$ when $a\rightarrow\infty$, but we also have $\lim_{a\rightarrow\infty}\frac{\mathbf{C}'\times\mathbf{C}}{\|\mathbf{C}'\times\mathbf{C}\|}=(\sin\theta_{0},\cos\theta_{0},0)^T$, hence the proposition follows. For $a \ge 0.5$, $R_W$ cannot be futher increased. Consequently, the range of $\phi$ has to be extended to $(-\frac{\pi}{2},\frac{3\pi}{2})$ to maintain the robustness. Therefore, if $a \rightarrow \infty$, GeoQuorum shares the same write quorum with $\mathcal{Q}_G$ and $\mathcal{Q}_L$: the great circles. Simulations {#sec:sim} =========== We hereby use simulation results to confirm the advantages of GeoQuorum over the existing designs, and also to demonstrate GeoQuroum’s ability of fine-tuning load and robustness. Simulation Settings ------------------- We randomly put nodes in an area (to be specified for each set of simulations). Then we use Delaunay triangulation to generate the connectivity graph, assuming that the sensor nodes are power controlled such that two nodes are connected by a wireless link iff there exists an edge (of the Delaunay triangulation) between them. We assume that the trajectory based forwarding [@NiculescuN-MobiCom03] is used to guide both write and read accesses, based on the corresponding curves that form the write and read quorums. As far as a curve passes through a triangle, all the three vertices are charged with a unit of communication load. This stems from the broadcast nature of wireless communication and the need for local coordination in the trajectory based forwarding. The data found by a read access are delivered to the node that initiates the access, through a routing scheme (e.g., shortest path routing) independent of the quorum system. As a result, we do not consider the load introduced by data delivery in the simulation, as it is a constant for all quorum systems. Comparing GeoQuorum with Existing Designs {#sec:comp} ----------------------------------------- We compare GeoQuorum with $\mathcal{Q}_G$ and $\mathcal{Q}_L$ introduced in [@SarkarZG-MobiCom06], as well as $\mathcal{Q}^m_G$, our remedy to $\mathcal{Q}_G$. Actually, a direct comparison is not possible, because though we faithfully take the quorum systems from [@SarkarZG-MobiCom06], the underlying geometric design tools are different, as we explained in Sec. \[sec:qsps\] and \[sec:bkgccg\]. Therefore, the following comparisons are based on our CCG design space. We assume WSNs with 5000 nodes. There is one data type, 500 nodes are contributing to it and 100 nodes may query it. We normalize the data query rate to 1 and vary the data production rate $r$ to test the system performance. Note that the actual write and read access rates (to a quorum system) are $500r$ and 100, respectively. Such an asymmetry between data production and consumption is reasonable, as otherwise multiple convergecasts may lead to better performance. For GeoQuorum, we set $R_W = 0.2\pi$ and $a = 0.2$. For each value of $r$, we obtain simulation results for 10 WSNs and we show the mean value and the standard deviation. We first compare the system load of the four quorum systems in Fig. \[fig:maxldcomp\], then their total load in Fig. \[fig:ttldcomp\]. ![System load comparisons.[]{data-label="fig:maxldcomp"}](MaxLoadComp){width=".85\columnwidth"} ![Total load comparisons.[]{data-label="fig:ttldcomp"}](TotalLoadComp){width=".85\columnwidth"} We illustrate the actual load distribution in Fig. \[fig:distrib\]. The following observations are immediate from these figures. - The load distribution of $\mathcal{Q}_G$ and $\mathcal{Q}_L$ are very unbalanced, exactly due to the existence of a hash location $h$ and its antipodal point $\bar{h}$, as pointed out by *Proposition \[prp:load\]*. - Both GeoQuorum and $\mathcal{Q}^m_G$ significantly reduce the system load, as they benefit from using mixed access strategy. - GeoQuorum incurs a much lower total load compared with all other three systems, due to its adaptivity to the asymmetry in data production and consumption. Although the performance of $\mathcal{Q}_G$ and $\mathcal{Q}_L$ appear to be very similar in the figures, they actually differ by about $1\%$ to $2\%$. As explained in Sec. \[sec:gqsals\], $\mathcal{Q}_L$ differs from $\mathcal{Q}_G$ only in read quorums, but as we assume that read access rate is far lower than the write access rate, this difference is “diluted”. Tuning the Load of GeoQuorum ---------------------------- We show the performance of our GeoQuorum under parameter fine-tuning in this section. In particular, we use the same 5000-node WSNs in Sec. \[sec:comp\] and the same four values of $r$. We tune the spiral parameter $a$ from 0.025 to 0.3 while increasing $R_W$ proportionally to maintain the same robustness. The results on system and total load are plotted in Fig. \[fig:maxldtn\] and \[fig:ttldtn\], respectively. We only show mean values, as the standard deviations are very small (partially due to the load balancing effect brought by GeoQuorum). ![Tuning the system load.[]{data-label="fig:maxldtn"}](MaxLoadTuning){width=".85\columnwidth"} ![Tuning the total load.[]{data-label="fig:ttldtn"}](TotalLoadTuning){width=".85\columnwidth"} In general, one always has to make a tradeoff between load balancing and energy efficiency. The tunability of GeoQuorum allows us to make different tradeoffs upon different application requirements. For example, when the data production rate is low ($r = 4$), $a \in (0.75, 1.5)$ appears to achieves a balanced performance in both system and total load. This region shifts towards smaller values with an increasing $r$. For $r = 10$, $a$ is better to be around 0.05. The flexibility of freely tuning the system performance is one of the major advantages of GeoQuorum over the existing designs. Tuning the Robustness of GeoQuorum ---------------------------------- If we just tune $a$ but keep $R_W$ constant, we will change the robustness of GeoQuorum. As shown by *Proposition \[prp:rbst\]*, the robustness can be tuned at a granularity of 2. Of course, increasing robustness comes at a cost of an increased total load. Again using the 5000-node WSNs, we show the relation between robustness and total load by Fig. \[fig:rbsttn\]. ![Tuning the robustness.[]{data-label="fig:rbsttn"}](RobustnessTuning){width=".85\columnwidth"} We consider two cases where $R_W = 0.6\pi$ and $R_W = 0.3\pi$. When we tune $a$ to linearly increase the robustness from 2 to 10, the total load increases by (roughly) following a power law. GeoQuorum in Irregular Network Area ----------------------------------- As all the previous simulations are based on 5000-node WSNs on square area, we simply demonstrate the ability of our design tools to cope with irregular areas in this section. Fig. \[fig:idistrib\] shows two WSNs deployed on irregular areas, with each consisting of 20000 nodes, one data type, 2000 contributors, 500 queriers. The load balancing effect of GeoQuorum is evident, though the load close to the network center is slightly higher than that close to the boundary. This slight unbalance is the cost one has to pay to maintain an acceptable total load: if one aims at fully balancing the load, larger quorums (longer curves) that make detour close to the boundary have to be used, leading to an unnecessarily high total load. Conclusion {#sec:con} ========== We have investigated the issue of data access in WSNs, aiming at balancing (communication) load distribution while maintaining energy efficiency. Specifically, we have revived the application of quorum systems in WSNs, and proposed the concept of geometric quorum systems based on a new development in combining computational conformal geometry with sensor networking. In particular, we have proposed GeoQuorum that makes use of parameterized spatial curves to form quorums, such that the system performance can be fine-tuned to meet different application requirements. Through both analysis and simulations, we have confirmed the advantages of GeoQuorum over existing proposals. Our future work will focus on the implementation aspect of geometric quorum systems in general. We are planning to deploy these curve formed quorum systems in a WSN testbed, in order to obtain better insights on the performance and practicality of such systems. [^1]: In the original paper, a quorum system design is termed a *double ruling scheme*. The two designs we discuss here are named *double rulings retrieval* and *distance-sensitive retrieval*. [^2]: This actually brings us back to the distributed hash tables (GHTs) based scheme [@RatnasamyKYYEGS-WSNA02], totally annihilating the benefit of quorum systems. [^3]: In reality, the popularity of different data types may vary a lot, this may further differentiate the access rates. Therefore, in order to cope with the variety of the access rates, we need a better quorum system design to fully utilize the freedom in the 3D space.

ArXiv

--- abstract: 'Let $p$ be an odd prime number. We study the problem of determining the module structure over the mod $p$ Steenrod algebra $\mathcal A(p)$ of the Dickson algebra $D_n$ consisting of all modular invariants of general linear group $GL(n,\mathbb F_p)$. Here $\mathbb F_p$ denotes the prime field of $p$ elements. In this paper, we give an explicit answer for $n=2$. More precisely, we explicitly compute the action of the Steenrod-Milnor operations $St^{S,R}$ on the generators of $D_n$ for $n=2$ and for either $S=\emptyset, R=(i)$ or $S=(s), R=(i)$ with $s,i$ arbitrary nonnegative integers.' author: - Nguyen Sum title: On the module structure over the Steenrod algebra of the Dickson algebra --- Introduction ============ Let $p$ be an odd prime number and let $GL_n = GL(n,\mathbb F_p)$ be the general linear group over the prime field $\mathbb F_p$ of $p$ elements. This group acts naturally on the algebra $P_n := E(x_1,x_2,\ldots, x_n)\otimes P(y_1,y_2,\ldots , y_n)$. Here and in what follows, $E(.,.,\ldots)$ and $P(.,., \ldots)$ are the exterior and polynomial algebras over $\mathbb F_p$ generated by the indicated variables. We grade $P_n$ by assigning $\deg x_i=1$ and $\deg y_i = 2.$ Dickson showed in [@di] that the invariant algebra $P(y_1,y_2,\ldots , y_n)^{GL_n}$ is a polynomial algebra over $\mathbb F_p$ generated by the Dickson invariants $Q_{n,s},\ 0\leqslant s<n$. In [@m1], Huynh Mui proved that the Dickson algebra $D_n=P_n^{GL_n}$ of invariants is generated by the Dickson invariants $Q_{n, s},\ 0 \leqslant s < n,$ and Mui invariants $R_{n; s_1, \ldots, s_k},\ 0 \leqslant s_1 < \ldots < s_k < n.$ It is well known that $P_n$ is a module over the Steenrod algebra $\mathcal A(p)$. The action of $\mathcal{A}(p)$ on $P_n$ is determined by the formulas $$\begin{aligned} \beta x_j &= y_j,\ \beta y_j = 0,\\ P^i(x_j) &= \begin{cases} x_j, &i=0,\\ 0, &i>0,\end{cases} \ \ P^i(y_j) = \begin{cases} y_j, & i=0,\\ y_j^p,&i=1,\\ 0, &i>1,\end{cases}\end{aligned}$$ and subject to the Cartan formulas $$\begin{aligned} \beta(xy) &= \beta(x)y +(-1)^{\deg x} x\beta(y),\\ P^r(xy) &= \sum_{i=0}^rP^i(x)P^{r-i}(y),\end{aligned}$$ for $x,y \in P_n$ and $\beta$ is the Bockstein homomorphism (see Steenrod [@ss]). Since this action commutes with the one of $GL_n$, it induces an action of $\mathcal{A}(p)$ on Dickson algebra $D_n$. So $D_n$ is a submodule of $P_n$. Note that the polynomial algebra $P(y_1,y_2,\ldots,y_n)$ is a submodule of $P_n$ and $P(y_1,y_2,\ldots,y_n)^{GL_n}$ is a submodule of the algebra $D_n$. Let $\tau_s, \xi_i$ be the Milnor elements of degrees $2p^s-1,\ 2p^i-2$ respectively in the dual algebra $\mathcal{A}(p)^*$ of $\mathcal{A}(p)$. In [@mi], Milnor showed that as an algebra, $$\mathcal{A}(p)^* = E(\tau_0,\tau_1,\ldots ) \otimes P(\xi_1,\xi_2,\ldots ).$$ Then $\mathcal{A}(p)^*$ has a basis consisting of all monomials $$\tau_S\xi^R \ =\ \tau_{s_1}\ldots \tau_{s_k}\xi_1^{r_1}\ldots \xi_m^{r_m},$$ with $S = (s_1,\ldots ,s_k),\ 0 \leqslant s_1 <\ldots <s_k , R = (r_1,\ldots ,r_m),\ r_i \geqslant 0 $. Let $St^{S,R} \in \mathcal{A}(p)$ denote the dual of $\tau_S\xi^R$ with respect to that basis. Then $\mathcal{A}(p)$ has a basis consisting of all operations $St^{S,R}$. For $S=\emptyset, R=(r)$, $St^{\emptyset, (r)}$ is nothing but the Steenrod operation $P^r$. So, we call $St^{S,R}$ the Steenrod-Milnor operation of type $(S,R)$. The operations $St^{S,R}$ have the following fundamental properties: – $St^{\emptyset,(0)}=1, \ St^{(0),(0)} = \beta.$ – $St^{S,R}(z) = 0$ if $z\in P_n$ and $\deg z < k + 2(r_1+r_2+\ldots+r_m)$. – The Cartan formula $$St^{S,R}(zt) =\sum_{\overset{\scriptstyle{S_1\cup S_2=S}}{R_1+R_2=R}}(-1)^{(\deg z+\ell(S_1))\ell(S_2)}(S:S_1,S_2)St^{S_1,R_1}(z)St^{S_2,R_2}(t),$$ where $ R_1=(r_{1i}),\ R_2=(r_{2i}),\ R_1+R_2=(r_{1i}+r_{2i}), S_1\cap S_2=\emptyset, z,t\in P_n$, $\ell(S_j)$ means the length of $S_j$ and $$(S:S_1,S_2)= \text{sign}\begin{pmatrix} s_1&\ldots &s_h& s_{h+1}&\ldots &s_k\\ s_{1,1}&\ldots &s_{1,h}&s_{2,1}&\ldots &s_{2,k-h}\end{pmatrix} ,$$ with $S_1=(s_{1,1},\ldots,s_{1,h}), s_{1,1}<\ldots <s_{1,h}$, $S_2=(s_{2,1},\ldots,s_{2,k-h}), s_{2,1} < \ldots < s_{2,k-h}$ (see Mui [@m2]). The action of $St^{S,R}$ on Dickson invariants $Q_{n,s}$ has partially been studied by many authors. This action for $S=\emptyset,\ \! R=(i)$ was explicitly determined by Madsen-Milgram [@mm], Smith-Switzer [@ss], Hung-Minh [@hm], Kechagias [@ke], Sum [@s2], Wilkerson [@wi]. This action for either $S=(s), R=(0)$ or $S=\emptyset, R=(0,\ldots,0,1)$ with 1 at the $i$-th place, was studied by Wilkerson [@wi], Neusel [@ne], Sum [@s3]. In this paper, we explicitly determine the action of the Steenrod-Milnor operations $St^{S,R}$ on Dickson invariants $Q_{2,0}, Q_{2,1}$ and Mui invariants $R_{2;0}, R_{2;1}, R_{2;0,1}$ for either $S=\emptyset, R=(i)$ or $S=(s), R=(i)$. In Section 2 we recall some results on the modular invariants of the general linear group $GL_2$ and the action of the Steenrod-Milnor operations on the generators of $P_2$. In Section 3, we compute the action of the Steenrod operations on Dickson-Mui invariants. Finally, in Section 4, we explicitly determine the action of the Steenrod-Milnor operations $St^{(s),(i)}$ on $Q_{2,0}, Q_{2,1}, R_{2;0}, R_{2;1}$ and $R_{2;0,1}$. Preliminaries ============= Let $u,v$ be nonnegative integers. Following Dickson [@di], Mui [@m1], we define $$[u;v] = \vmatrix y_1^{p^u}&y_2^{p^u}\\ y_1^{p^v} & y_2^{p^v} \endvmatrix,\ [1;u] = \vmatrix x_1&x_2\\ y_1^{p^u} & y_2^{p^u} \endvmatrix.$$ In particular, we set $$\begin{aligned} L_2 &= [0,1],\ L_{2,0}=[1,2],\ L_{2,1} = [0,2],\\ M_{2;0} &= [1;1],\ M_{2;1} = [1;0],\ M_{2;0,1} = x_1x_2.\end{aligned}$$ The polynomial $[u,v]$ is divisible by $L_2$. Then, Dickson invariants $Q_{2,0}, Q_{2,1}$ and Mui invariants $ R_{2;0}, R_{2;1}, R_{2;0,1}$ are defined by $$\begin{aligned} Q_{2,0}&=L_{2,0}/L_2,\ Q_{2,1} = L_{2,1}/L_2,\\ R_{2;0} &= M_{2;0}L_2^{p-2},\ R_{2;1} = M_{2;1}L_2^{p-2},\ R_{2;0,1} = M_{2;0,1}L_2^{p-2}.\end{aligned}$$ Now we prepare some data in order to prove our main results. First, we recall the following which will be needed in the next sections. Let $\alpha_i(a)$ denote the $i$-th coefficient in $p$-adic expansion of a non-negative integer $a$. That means $$a= \alpha_0(a)p^0+\alpha_1(a)p^1+\alpha_2(a)p^2+ \ldots ,$$ for $0 \leqslant \alpha_i(a) <p, i\geqslant 0.$ Denote by $I(u,v)$ the set of all integers $a$ satisfying $$\begin{aligned} & \alpha_i(a)+\alpha_{i+1}(a) \leqslant 1,\text{ for any } i,\\ & \alpha_i(a)=0, \text{ for either } i<u \text{ or } i\geqslant v-2.\end{aligned}$$ \[md2.9\]Under the above notations, we have $$[u,v] = \sum_{a\in I(u,v)}(-1)^aL_2^{p^u+p(p-1)a}Q_{2,1}^{\frac{p^{v-1}-p^u}{p-1}-(p+1)a}.$$ \[bd2.2\] Let $b$ be a nonnegative integer and $\varepsilon = 0, 1$. We have $$St^{S,R}(x_k^\varepsilon y_\ell^b) = \begin{cases} \binom bR x_k^\varepsilon y_\ell^{b + \vert R\vert}\ , &S = \emptyset\ ,\\ \varepsilon \binom bRy_k^{p^s}y_\ell^{b + \vert R\vert}\ , &S = (s), \ s \geqslant 0 ,\\ 0\ , &\text{otherwise.} \end{cases}$$ Here $\binom{b}{R} = \frac{b!}{(b-r_1-r_2-\ldots-r_m)!r_1!\ldots r_m!}$ for $r_1+r_2+\ldots+r_m\leqslant b$ and $\binom{b}{R}=0$ for $r_1+r_2+\ldots+r_m > b$ and $|R|= (p-1)r_1+(p^2-1)r_2+\ldots + (p^m-1)r_m$. Note that for $R=(i), \binom bR= \binom bi$ is the binomial coefficient. By convention, we set $\binom bi=0$ for $i<0$. Applying Lemma \[bd2.2\] to $P^i=St^{\emptyset,(i)}$, we get Let $b, i$ be nonnegative integers. Then we have $$P^iy_k^b = \binom niy_k^{b+(p-1)i}.$$ Since $\binom {p^e}i = 0$ in $\mathbb F_p$ for $1 <i<p^e$, we get \[hq2.4\] For any nonnegative integers $e, i$, $$P^iy_k^{p^e} = \begin{cases} y_k^{p^e}, & i=0,\\ y_k^{p^{e+1}}, & i=p^e,\\ 0, &\text{otherwise}. \end{cases}$$ Applying Corollary \[hq2.4\] and the Cartan formula to $[u,v]=y_1^{p^u}y_2^{p^v} - y_1^{p^v}y_2^{p^u}$, we obtain \[bd2.5\] Let $u,v, i$ be nonnegative integers. Then we have $$P^i{[u,v]} = \begin{cases} {[u,v]}, & i=0,\\ {[u+1,v]}, & i=p^u,\\ {[u,v+1]}, & i=p^v,\\ {[u+1,v+1]}, & i=p^u+p^v,\\ 0, &\text{otherwise}. \end{cases}$$ Since $L_2={[0,1]}, L_{2,0} = {[1,2]}, L_{2,1} ={[0,2]}$, from Lemma \[bd2.5\], we get \[hq2.6\] For any nonnegative integers $i$, $$\begin{aligned} P^iL_2 &= \begin{cases} L_2, &i=0,\\ L_2Q_{2,1}, &i= p,\\ L_2Q_{2,0}, &i= p+1,\\ 0, &\text{otherwise,}\end{cases}\\ P^iL_{2,0} &= \begin{cases} {[0,1]}=L_2Q_{2,0}, &i=0,\\ {[1,3]}= L_2Q_{2,0}Q_{2,1}^p, &i= p^2,\\ {[2,3]} = L_2Q_{2,0}^{p+1}, &i= p^2+p,\\ 0, &\text{otherwise,}\end{cases}\\ P^iL_{2,1} &= \begin{cases} {[0,2]}=L_2Q_{2,1}, &i=0,\\ {[1,2]} = L_2Q_{2,0}, &i= 1,\\ {[0,3]} = L_2(Q_{2,1}^{p+1}-Q_{2,0}^p), &i=p^2,\\ {[1,3]} = L_2Q_{2,0}Q_{2,1}^{p}, &i= p^2+1,\\ 0, &\text{otherwise.}\end{cases}\end{aligned}$$ Combining Lemma \[bd2.2\], Corollary \[hq2.4\] and the Cartan formula gives \[bd2.7\] Let $s, i$ be nonnegative integers. Then we have $$St^{(s),(i)}{[1;u]} = \begin{cases} {[s,u]}, &i=0,\\ {[s,u+1]}, &i=p^u,\\ 0, &\text{otherwise}. \end{cases}$$ Applying Lemma \[bd2.2\] and Corollary \[hq2.4\] to $M_{2;0}={[1;1]},\ M_{2;1} = {[1;0]}$, we obtain \[hq2.8\] For any nonnegative integer $i$, $$\begin{aligned} P^i(M_{2;0}) &= \begin{cases} {[1;1]}=M_{2;0}, &i=0,\\ {[1;2]}= M_{2;0}Q_{2,1}-M_{2;1}Q_{2,0}, &i = p,\\ 0, &\text{otherwise,}\end{cases}\\ P^i(M_{2;1}) &= \begin{cases} {[1;0]}=M_{2;1}, &i=0,\\ {[1;1]}= M_{2;0}, &i = 1,\\ 0, &\text{otherwise.}\end{cases}\end{aligned}$$ The action of the Steenrod operations on Dickson-Mui invariants =============================================================== First of all, we prove the following which was proved in Hung-Minh [@hm], by another method. \[dl3.1\] For any nonnegative integer $i$ and $s=0,1$, we have $$P^iQ_{2,s} = \begin{cases} (-1)^k\binom {k+s}rQ_{2,0}^{r+1-s}Q_{2,1}^{k+s-r},& i=kp+r, 0\leqslant r-s \leqslant k < p,\\ 0, &\text{otherwise.}\end{cases}$$ Recall that $\deg Q_{2,s} = 2(p^2-p^s) < 2p^2$. Hence, $P^iQ_{2,s} =0$ for $i \geqslant p^2$. Suppose that $i < p^2$. Then, using the $p$-adic expansion of $i$, we have $$i=kp+r\ \text{ for}\ 0\leqslant k,\ r<p .$$ We prove the theorem by induction on $k$. We have $P^0Q_{2,s} = Q_{2,s}$. According to Corollary \[hq2.6\], $$\begin{aligned} 0&=P^1L_{2,0} = P^1(L_2Q_{2,0}) = L_2P^1Q_{2,0},\\ L_2Q_{2,0}&=P^1L_{2,1} = P^1(L_2Q_{2,1}) = L_2P^1Q_{2,1}.\end{aligned}$$ These equalities imply $P^1Q_{2,0} = 0, P^1Q_{2,1} = Q_{2,0}$. For $1< r<p$, $P^rL_{2,s} = 0$ and $P^rL_2 =0$. Using the Cartan formula and Corollary \[hq2.6\], we have $$0=P^rL_{2,s} = P^r(L_2Q_{2,s}) = L_2P^rQ_{2,s}.$$ This implies $P^rQ_{2,s} = 0$. So the theorem holds for $k=0$ and any $0\leqslant r <p$. Suppose $0<k<p$ and the theorem is true for $k-1$ and any $0\leqslant r <p$. Using the Cartan formula, Corollary \[hq2.6\] and the inductive hypothesis, we get $$\begin{aligned} 0 &= P^iL_{2,s} = P^i(L_2Q_{2,s})\\ &= L_2P^iQ_{2,s} + L_2Q_{2,1}P^{i-p}Q_{2,s}+L_2Q_{2,0}P^{i-p-1}Q_{2,0}\\ &=L_2P^iQ_{2,s} + L_2Q_{2,1}(-1)^{k-1}\binom{k-1+s}rQ_{2,0}^{r+1-s}Q_{2,1}^{k+s-r-1}\\ &\hskip4cm+L_2Q_{2,0}(-1)^{k-1}\binom{k-1+s}{r-1}Q_{2,0}^{r-s}Q_{2,1}^{k+s-r}\\ &=L_2P^iQ_{2,s}+ (-1)^{k-1}\Big(\binom{k-1+s}r +\binom{k-1+s}{r-1}\Big)L_2Q_{2,0}^{r+1-s}Q_{2,1}^{k+s-r}.\end{aligned}$$ From this equality and the relation $\binom{k-1+s}r +\binom{k-1+s}{r-1}=\binom {k+s}r$, we see that the theorem holds for $k$. The proof is completed. To compute the action of $P^i$ on $R_{2;0}, R_{2;1}, R_{2;0,1}$ we need the following \[bd3.2\] Let $i$ be a nonnegative integer. Then we have $$P^iL_2^{p-2} = \begin{cases} (-1)^k(k+1)\binom kr L_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r}, &i=kp+r, 0\leqslant r \leqslant k < p,\\ 0, &\text{otherwise}. \end{cases}$$ Note that $\deg L_2^{p-2} = 2(p-2)(p+1) <2p^2$. So $P^iL_2^{p-2} = 0$ for $i \geqslant p^2$. Hence, it suffices to prove the theorem for $i=kp+r$ with $0\leqslant k,r < p$. Since $P^0 = 1$, we have $P^0L_2^{p-2} = L_2^{p-2}$. If $0< r <p$ then from Theorem \[dl3.1\], the Cartan formula and the relation $Q_{2,0} = L_2^{p-1}= L_2L_2^{p-2}$, we get $$0= P^rQ_{2,0} = L_2P^rL_2^{p-2}.$$ This implies $P^rL_2^{p-2} =0$. The lemma is true for $k=0$ and $0\leqslant r <p$. Suppose that $0<k<p$ and the lemma holds for $k-1$ and any $0\leqslant r <p$. Using the Cartan formula, Theorem \[dl3.1\], Corollary \[hq2.6\] and the inductive hypothesis, we have $$\begin{aligned} (-1)^k\binom kr Q_{2,0}^{r+1}Q_{2,1}^{k-r} &= P^iQ_{2,0} = P^i(L_2L_2^{p-2})\\ &= L_2P^iL_2^{p-2} + L_2Q_{2,1}P^{i-p}L_2^{p-2}+ L_2Q_{2,0}P^{i-p-1}L_2^{p-2}\\ &=L_2P^iL_2^{p-2} + L_2Q_{2,1}(-1)^{k-1}k\binom{k-1}rL_2^{p-2}Q_{2,0}^{r}Q_{2,1}^{k-r-1}\\ &\hskip2cm+ L_2Q_{2,0}(-1)^{k-1}k\binom{k-1}{r-1}L_2^{p-2}Q_{2,0}^{r-1}Q_{2,1}^{k-r}\\ &= L_2P^iL_2^{p-2} + (-1)^{k-1}k\Big(\binom{k-1}r +\binom{k-1}{r-1}\Big)Q_{2,0}^{r+1}Q_{2,1}^{k-r}.\end{aligned}$$ This equality and the relation $\binom{k-1}r+\binom{k-1}{r-1}=\binom kr$ imply the lemma for $k$ and any $0\leqslant r <p$. \[dl3.3\] Let $i$ be a nonnegative integer. We have $$P^iR_{2;0} =\begin{cases} (-1)^k\big((r+1)\binom krR_{2;0}Q_{2,0}^{r}Q_{2,1}^{k-r} +k\binom{k-1}rR_{2;1}Q_{2,0}^{r+1}Q_{2,1}^{k-r-1}\big),\\ \hskip5cm i=kp+r, 0\leqslant r \leqslant k < p,\\ 0, \hskip4.6cm\text{otherwise}. \end{cases}$$ Note that $\deg R_{2;0} = 2p^2-3 < 2p^2$. So $P^iR_{2;0} = 0$ for $i \geqslant p^2$. We prove the theorem for $i=kp+r$ with $0 \leqslant k,r <p$. For $k=r=0$, $P^0R_{2;0} = R_{2;0}$. For $k= 0, 0<r<p$, applying the Cartan formula and Corollary \[hq2.8\], we get $$P^rR_{2;0} = P^r(M_{2;0}L_2^{p-2}) = M_{2;0}P^rL_2^{p-2}=0.$$ The theorem holds for $k=0$ and $0\leqslant r <p$. Suppose that $0<k < p$. Using the Cartan formula, Corollary \[hq2.8\] and Lemma \[bd3.2\], we obtain $$\begin{aligned} P^iR_{2;0} &= P^i(M_{2;0}L_2^{p-2}) = P^0M_{2;0}P^iL_2^{p-2} + P^pM_{2;0}P^{i-p}L_2^{p-2}\\ &= M_{2;0}P^iL_2^{p-2} + (M_{2;0}Q_{2,1} - M_{2;1}Q_{2,0})P^{i-p}L_2^{p-2}\\ &= M_{2;0}(-1)^k(k+1)\binom kr L_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r}\\ &\quad + (M_{2;0}Q_{2,1} - M_{2;1}Q_{2,0})(-1)^{k-1}k\binom {k-1}rL_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r-1}\\ &= (-1)^k\Big(\big((k+1)\binom kr - k\binom{k-1}r\big)R_{2;0}Q_{2,0}^rQ_{2,1}^{k-r}\\ &\hskip5cm + k\binom{k-1}rR_{2;1}Q_{2,0}^{r+1}Q_{2,1}^{k-r-1}\Big).\end{aligned}$$ This equality and the relation $(k+1)\binom kr - k\binom{k-1}r = (r+1)\binom kr$ imply the theorem for $k$ and $0\leqslant r <p$. By an analogous argument as given in the proof of Theorem \[dl3.3\], we can easily obtain the following For any nonnegative integer $i$, we have $$\begin{aligned} P^iR_{2;1} &=\begin{cases} (-1)^k(k+1)\big(\binom krR_{2;1}Q_{2,0}^{r}Q_{2,1}^{k-r} +\binom{k}{r-1}R_{2;0}Q_{2,0}^{r-1}Q_{2,1}^{k-r+1}\big),\\ \hskip5cm i=kp+r, 0\leqslant r \leqslant k < p,\\ 0, \hskip4.6cm\text{otherwise}, \end{cases}\\ P^iR_{2;0,1} &= \begin{cases} (-1)^k(k+1)\binom krR_{2;0,1}Q_{2,0}^rQ_{2,1}^{k-r}, &i=kp+r, 0\leqslant r \leqslant k <p,\\ 0, &\text{otherwise}. \end{cases}\end{aligned}$$ On the action of the Steenrod-Milnor operations\ on Dickson-Mui invariants ================================================ In this section, we compute the action of $St^{(s),(i)}$ on Dickson-Mui invariants. It is easy to see that $St^{(s),(i)}Q_{2,s} = 0$. So we need only to compute the action of $St^{(s),(i)}$ on $R_{2;0}, R_{2;1}$ and $R_{2;0,1}$. First, we recall the following \[bd4.1\] For any nonnegative integers $s, i$, $$\begin{aligned} St^{(s),(i)}(M_{2;0}) &= \begin{cases} {[s,1]}, &i=0,\\ {[s,2]}, &i = p,\\ 0, &\text{otherwise,}\end{cases}\ \ St^{(s),(i)}(M_{2;1}) = \begin{cases} {[s,0]}, &i=0,\\ {[s,1]}, &i = 1,\\ 0, &\text{otherwise.}\end{cases}\end{aligned}$$ This lemma can easily be proved by using the Cartan formula and Lemma \[bd2.2\]. \[dl4.2\] Let $s, i$ be nonnegative integers. Then we have $$St^{(s),(i)}R_{2;0} =\begin{cases} (-1)^k(r+1)\binom krQ_{2,0}^{k+1}Q_{2,1}^{k-r}, \ \ \ \ s=0, i= kp+r, 0\leqslant r \leqslant k < p,\\ (-1)^{k+1}k\binom{k-1}{r}Q_{2,0}^{r+2}Q_{2,1}^{k-1-r}, \ \ \ s=1, i=kp+r, 0\leqslant r < k < p,\\ (-1)^{k+1}(k+1)\binom{k}{r}Q_{2,0}^{r+2}Q_{2,1}^{k-r},\ s=2, i=kp+r, 0\leqslant r \leqslant k < p,\\ \displaystyle{(-1)^{k+1}\binom kr\Big((k+1)\sum_{a\in I(1,s)}(-1)^aQ_{2,0}^{pa+r+2}Q_{2,1}^{\frac{p^{s-1}-p}{p-1}-(p+1)a+k-r}}\\ \qquad-\displaystyle{(k-r)\sum_{a\in I(2,s)}(-1)^aQ_{2,0}^{p(a+1)+r+2}Q_{2,1}^{\frac{p^{s-1}-p^2}{p-1}-(p+1)a+k-r-1}}\Big),\\ \hskip5.4cm s >2, i =kp+r, 0\leqslant r \leqslant k < p,\\ 0, \hskip5cm \text{otherwise}. \end{cases}$$ Since $\deg R_{2;0} = 2p^2- 3$, $St^{(s),(i)}R_{2;0} = 0$ for $i\geqslant p^2$. Suppose $i<p^2$, then using the $p$-adic expansion of $i$, we have $i=kp+r\ \text{ for}\ 0\leqslant k,\ r<p .$ We have $St^{(s),(0)}R_{2;0} = St^{(s),(0)}(M_{2;0}L_2^{p-2}) = St^{(s),(0)}M_{2;0}P^0L_2^{p-2}=[s,1]L_2^{p-2}$. For $0<r<p$, using the Cartan formula, Lemma \[bd3.2\] and Lemma \[bd4.1\], we get $$St^{(s),(r)}R_{2;0} = St^{(s),(r)}(M_{2;0}L_2^{p-2}) =St^{(s),(0)}(M_{2;0})P^rL_2^{p-2}= 0.$$ The above equalities and Proposition \[md2.9\] imply the theorem for $k=0$. For $0<k<p$, using the Cartan formula, Lemma \[bd3.2\] and Lemma \[bd4.1\] we obtain $$\begin{aligned} St^{(s),(i)}R_{2;0} &= St^{(s),(i)}(M_{2;0}L_2^{p-2})\\ &= St^{(s),(0)}M_{2;0}P^iL_2^{p-2} + St^{(s),(p)}M_{2;0}P^{i-p}L_2^{p-2}\\ &= [s,1](-1)^k(k+1)\binom krL_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r}\\ &\quad + [s,2](-1)^{k-1}k\binom{k-1}rL_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r-1}.\end{aligned}$$ Combining this equality and Proposition \[md2.9\], we obtain the theorem. For any nonnegative integers $s, i$, $$St^{(s),(i)}R_{2;1} =\begin{cases} (-1)^k(k+1)\binom k{r-1}Q_{2,0}^{r}Q_{2,1}^{k-r+1}, \ \ s=0, i= kp+r, 0\leqslant r \leqslant k < p,\\ (-1)^{k+1}(k+1)\binom{k}{r}Q_{2,0}^{r+1}Q_{2,1}^{k-r}, \ \ \ s=1, i=kp+r, 0\leqslant r \leqslant k < p,\\ \displaystyle{(-1)^{k+1}(k+1)\Big(\binom kr\sum_{a\in I(0,s)}(-1)^aQ_{2,0}^{pa+r+1}Q_{2,1}^{\frac{p^{s-1}-1}{p-1}-(p+1)a+k-r}}\\ \qquad+\displaystyle{\binom{k}{r-1}\sum_{a\in I(1,s)}(-1)^aQ_{2,0}^{pa+r+1}Q_{2,1}^{\frac{p^{s-1}-p}{p-1}-(p+1)a+k-r+1}}\Big),\\ \hskip5.6cm s >1, i =kp+r, 0\leqslant r \leqslant k < p,\\ 0, \hskip5.2cm \text{otherwise}. \end{cases}$$ Since $\deg R_{2;1} = 2(p^2-p) - 1$, $St^{(s),(i)}R_{2;1}=0$ for $i\geqslant p^2$. Suppose $ i < p^2$ and $i=kp+r$ with $0\leqslant k,r <p$. Using the Cartan formula and Lemma \[bd4.1\], we have $$St^{(s), (0)}R_{2;1} = St^{(s), (0)}(M_{2;1})L_2^{p-2}={[s,0]}L_2^{p-2}.$$ From this and Proposition \[md2.9\], we see that the theorem is true for $i=0$. For $i>0$, using the Cartan formula, Lemma \[bd4.1\] and Lemma \[bd3.2\], we obtain $$\begin{aligned} St^{(s), (i)}R_{2;1} &= St^{(s), (0)}M_{2;1}P^iL_2^{p-2} + St^{(s),(1)}M_{2;1}P^{i-1}L_2^{p-2}\\ &= {[s,0]}(-1)^{k}(k+1)\binom krL_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r}\\ &\qquad + {[s,1]}(-1)^{k}(k+1)\binom k{r-1}L_2^{p-2}Q_{2,0}^{r-1}Q_{2,1}^{k-r+1}\\ &= (-1)^{k}(k+1)\Big(\binom kr{[s,0]}L_2^{p-2}Q_{2,0}^rQ_{2,1}^{k-r}\\ &\hskip3cm + \binom k{r-1}{[s,1]}L_2^{p-2}Q_{2,0}^{r-1}Q_{2,1}^{k-r+1}\Big).\end{aligned}$$ Now the theorem follows from this equality and Proposition \[md2.9\]. Suppose $s,i$ are nonnegative integers. We have $$St^{(s),(i)}R_{2;0,1} = \begin{cases} (-1)^{k+1}(k+1)\binom kr R_{2;1}Q_{2,0}^rQ_{2,1}^{k-r},\\ \hskip5cm s=0, i=kp+r, 0\leqslant r\leqslant k<p,\\ (-1)^{k+1}(k+1)\binom krR_{2;0}Q_{2,0}^rQ_{2,1}^{k-r},\\ \hskip5cm s=1, i=kp+r, 0\leqslant r\leqslant k<p,\\ \displaystyle{(-1)^k(k+1)\binom kr\Big(R_{2;1}\sum_{a\in I(1,s)}(-1)^aQ_{2,0}^{pa+1+r}Q_{2,1}^{\frac{p^{s-1}-p}{p-1}-(p+1)a+k-r}}\\ \hskip3cm \displaystyle{- R_{2;0}\sum_{a\in I(0,s)}(-1)^aQ_{2,0}^{pa+r}Q_{2,1}^{\frac{p^{s-1}-1}{p-1}-(p+1)a+k-r}\Big)},\\ \hskip5cm s>1, i=kp+r, 0\leqslant r\leqslant k<p,\\ 0, \hskip4.6cm \text{otherwise}. \end{cases}$$ Since $\deg R_{2;0,1} = (p-2)(p+1)+2$, $St^{(s),(i)}R_{2;0,1} =0$ for $i \geqslant p^2$. Suppose $i<p^2$ and $i= kp+r$ with $0\leqslant k,r <p$. Using the Cartan formula and Lemma \[bd2.2\], we have $$St^{(s),(0)}(x_1x_2) = y_1^{p^s}x_2-x_1y_2^{p^s} = -{[1;s]} = (M_{2;1}{[1,s]} - M_{2;0}{[0,s]})/L_2.$$ Since $R_{2;0,1} = x_1x_2L_2^{p-2}$, using the Cartan formula, the above equality and Lemma \[bd3.2\], we get $$\begin{aligned} St^{(s),(i)}R_{2;0.1} &= St^{(s),(0)}(x_1x_2)P^iL_2^{p-2}\\ &= (M_{2;1}{[1,s]} - M_{2;0}{[0,s]})(-1)^k(k+1)\binom kr L_2^{p-3}Q_{2,0}^rQ_{2,1}^{k-r}.\end{aligned}$$ Combining this equality and Proposition \[md2.9\], we get the theorem. [99]{} L. E. Dickson, [*A fundamental system of invariants of the general modular linear group with a solution of the form problem*]{}, Trans. Amer. Math. Soc. [**12**]{} (1911), 75-98. N. H. V. Hung and P. A. Minh, [*The action of the mod $p$ Steenrod operations on the modular invariants of linear groups*]{}, Vietnam J. Math. [**23**]{} (1995), 39-56. N. E. Kechagias, [*The Steenrod algebra action on generators of rings of invariants of subgroups of ${GL_n(\mathbb Z/p\mathbb Z)}$*]{}, Proc. Amer. Math. Soc. [**118**]{} (1993), 943-952. I. Madsen and R. J. Milgram, [*The classifying spaces for surgery and cobordism of manifolds*]{}, Annals of Mathematics Studies No. 92, Princeton University Press, Princeton N.J (1979). J. Milnor, [*Steenrod algebra and its dual*]{}, Ann. of Math. [**67**]{} (1958), 150-171. H. Mui, [*Modular invariant theory and the cohomology algebras of symmetric groups*]{}, J. Fac. Sci. Univ. Tokyo Sec. IA Math. [**22**]{} (1975), 319-369. H. Mui, [*Cohomology operations derived from modular invariants*]{}, Math. Z. [**193**]{} (1986), 151-163. M. D. Neusel, [*Inverse invariant theory and Steenrod operations*]{}, Mem. Amer. Math. Soc. [**146**]{} (2000), x+158. L. Smith and R. Switzer, [*Realizability and non-realizability of Dickson algebras as cohomology rings*]{}, Proc. Amer. Math. Soc. [**89**]{} (1983), 303-313. N. E. Stennrod, [*Cohomology operations*]{}, Lectures by N. E. Steenrod written and revised by D. B. A. Epstein, Annals of Mathematics No. 50, Princeton University Press, Princeton N.J (1962). N. Sum, [*On the action of the Steenrod algebra on the modular invariants of special linear group*]{}, Acta Math. Vietnam. [**18**]{} (1993), 203-213. N. Sum, [*Steenrod operations on the modular invariants*]{}, Kodai Math. J. [**17**]{} (1994), 585-595. N. Sum, [*The action of the primitive Steenrod-Milnor operations on the modular invariants*]{}, Geometry and Topology Monographs, [**11**]{} (2007), 349-367. C. Wilkerson, [*A primer on the Dickson invariants*]{}, Contemporary Mathematics, Amer. Math. Soc. [**19**]{} (1983), 421-434. Department of Mathematics, University of Quynhon 170 An Duong Vuong, Quynhon, Vietnam E-mail address: [email protected]

ArXiv

--- abstract: 'In 1975, Ashtekar and Magnon [@AM] showed that an energy condition selects a unique quantization procedure for certain observers in general, curved spacetimes. We generalize this result in two important ways, by eliminating the need to assume a particular form for the (quantum) Hamiltonian, and by considering the surprisingly nontrivial extension to nonminimal coupling.' address: 'Department of Mathematics, Oregon State University, Corvallis, OR 97331, USA' author: - 'Jim Fischer & Tevian Dray' title: | [gr-qc/9807028]{}\ [**A NEW LOOK AT THE ASHTEKAR-MAGNON ENERGY CONDITION**]{} --- \#1 PACS: \#1 H 2[CE\_[H]{}]{} Introduction ============ Ashtekar and Magnon [@AM] were among the first to consider quantum field theory as seen by observers who were [*not*]{} static or stationary. Remarkably, they were able to give a quantization procedure for the scalar field for [*any*]{} family of hypersurface orthogonal observers in a curved spacetime. Their procedure is based on a single, natural condition: The classical and quantum energies should agree. However, for non-Killing observers, it is not obvious how to define either of these energies. Ashtekar and Magnon choose to use the stress-energy tensor of the scalar field for the classical energy, and to define the quantum energy in terms of a particular choice of quantum Hamiltonian operator. We extend their work in two important ways. First of all, we show that the construction itself fully determines the Hamiltonian operator, which therefore does not need to be specified in advance. Second, we show that the basic result holds for [*any*]{} choice of the classical energy satisfying certain simple properties. Not surprisingly, for the case of minimal coupling ($\xi=0$) considered by Ashtekar and Magnon, if we use the stress-energy tensor to define the classical energy, then we recover not only their complex structure but also their quantum Hamiltonian. However, when $\xi\ne0$, this approach runs into a serious problem: The resulting Hamiltonian and complex structure do not reduce to the known answers for static observers. We show how this problem can be resolved by using the classical Hamiltonian to define the classical energy rather than the stress-energy tensor. After setting up our formalism in Section \[Math\], we summarize the work of Ashtekar and Magnon in Section \[Energy\] and present our generalization in Section \[Result\]. Section \[Appl\] shows how to recover Ashtekar and Magnon’s result for $\xi=0$, as well as considering the case $\xi\ne0$. Finally, in Section \[Disc\] we discuss our results. Mathematical Preliminaries {#Math} ========================== Let $(M,g_{ab})$ be a globally hyperbolic spacetime with associated Levi-Civita connection $\nabla$. The action ${\cal S}$ for a scalar field $\Phi$ on $M$ is given by $$\begin{aligned} {\cal S} &=& \int_M {\cal{L}} \,\sqrt{-g}\,d^nx \label{action}\\ {\cal L} &=& -\frac{1}{2}\left( g^{ab}\nabla_a\Phi\nabla_b\Phi + (m^2 + \xi R)\Phi^2 \right) \label{lagrange}\end{aligned}$$ The Klein-Gordon equation, obtained by varying the action $\cal S$ with respect to $\Phi$, is $$g^{ab}\nabla_a\nabla_b\Phi - (m^2 + \xi R)\Phi = 0 \label{KG}$$ Let $V$ be the space of smooth, real-valued solutions of (\[KG\]) which have compact support on any (and hence every) Cauchy surface. Ashtekar and Magnon suggested that, as a real vector space, the one-particle Hilbert space ${\cal{H}}$ should be a copy of $V$. Introduce coordinates $(t=x_0,\dots,x_{n-1})$ on $M$ so that the hypersurfaces $\{t={\rm const}\}$ are Cauchy surfaces. We assume throughout that the vector field $t^a\nabla_a=\partial_t$ is hypersurface orthogonal. The standard $3+1$ formalism leads to a decomposition of the metric $g_{ab}$ in terms of its pullback $h_{ij}$ to $\Sigma$ and the lapse function $N=t^at_a$; the shift is zero. We denote the Levi-Civita connection on $(\Sigma,h_{ij})$ by $D_i$. On $V$ we have the (nondegenerate) symplectic structure $$\Omega(\Phi,\Psi) = \int_\Sigma \left(\Psi \nabla_a\Phi - \Phi\nabla_a\Psi \right)\, n^a d\Sigma \label{symstruc}$$ where $\Sigma$ is any Cauchy hypersurface, $n^a$ is the future pointing, unit normal vector field to $\Sigma$ and $d\Sigma= \sqrt{h}\, d^{n-1}x$ is the volume element on $\Sigma$ induced by the inclusion map. Let $J$ be any complex structure on $V$, that is a linear map from $V$ to itself satisfying $$J^2 = -1 \label{struct}$$ which allows us to view $V$ as a complex vector space. We will also assume that $J$ is compatible in the sense that $$\begin{aligned} \Omega(\Phi,J\Psi) &\geq& 0 \label{com1}\\ \Omega(J\Phi,J\Psi) &=& \Omega(\Phi, \Psi) \label{com2}\end{aligned}$$ for any $\Phi,\Psi\in V$. As discussed in [@AM], the $*$-algebra approach leads naturally to the inner product $$\left<\Phi,\Psi\right> = \frac{1}{2}\Omega(\Phi, J\Psi) + \frac{i}{2}\Omega(\Phi, \Psi) \label{innerproduct}$$ which is Hermitian under the above assumptions. A candidate for the one-particle Hilbert space ${\cal H}$ is then the Cauchy completion of $(V,J,\left<\_,\_\right>)$, so that the problem of identifying the one-particle Hilbert space of states is reduced to that of choosing a suitable complex structure $J$ on $V$. Solutions $\Phi\in V$ are completely determined by their initial data, [^1] so that $V$ is isomorphic to the vector space $\V$ of pairs of smooth, real-valued functions on $\Sigma$ which have compact support. The isomorphic image of $\Phi$ is then $$\hat\Phi = \pmatrix{\varphi\cr \pi} = \pmatrix{\Phi\vert_\Sigma\cr \sqrt{h}\,n^a\nabla_a\Phi\vert_\Sigma} \label{data}$$ We write $\tau = C^\infty_0(\Sigma,\R)$, with inner product $$(f,g) = \int_{\Sigma} fg \,d^{n-1}x\\$$ for $f,g\in\V$; note that $\V=\tau\oplus\tau$. We conclude this section with some results about symmetric operators. Any linear operator $Q$ on $V$ can be represented as a $2\times2$ matrix $\hat Q$ whose elements are linear operators on $\tau$. In particular, we write $$\begin{aligned} \J &=& \pmatrix{A& B\cr C& D}\\ \noalign{\medskip} \H &=& \pmatrix{W& X\cr Y& Z}\end{aligned}$$ We define $Q$ to be [*symmetric*]{} on $V$ if $$\left<\Phi,Q\Psi\right> = \left<Q\Phi,\Psi\right> \label{symmetry}$$ for all $\Phi,\Psi\in V$; $Q$ is [*antisymmetric*]{} if a relative factor of $-1$ is inserted in (\[symmetry\]). \[symtheo2\] Suppose that the linear operator $Q$ on $V$ satisfies $$\Omega(\Phi,Q\Phi) = 0 \label{Asymm}$$ Then $Q$ is symmetric. [**Proof:**]{} This follows immediately since (\[Asymm\]) implies that the expectation value of $Q$ is always real. \[JHHJ0\] Let $Q$ be a symmetric operator defined on ${\cal H}$. Then $[Q,J] = 0$. [**Proof:**]{} Splitting (\[symmetry\]) into its real and imaginary parts, we obtain $$\begin{aligned} \Omega(\Phi, JQ\Psi) &=& \Omega(Q\Phi, J\Psi) \label{realpart}\\ \Omega(\Phi, Q\Psi) &=& \Omega(Q\Phi, \Psi) \label{imajpart}\end{aligned}$$ Using (\[realpart\]) and (\[imajpart\]) we obtain: $$\begin{aligned} \Omega(\Phi,(JQ - QJ)\Psi) &=& \Omega(\Phi,JQ\Psi) - \Omega(\Phi,QJ\Psi) \\ &=& \Omega(Q\Phi, J\Psi) - \Omega(Q\Phi, J\Psi) \\ &=& 0.\end{aligned}$$ Since $\Phi$ and $\Psi$ are arbitrary (and $\Omega$ is non-degenerate) we conclude that $JQ -QJ = 0. $ The significance of Lemma \[JHHJ0\] comes from the fact that the Hamiltonian operator $H$ should be self-adjoint and hence symmetric. The total derivative of $J$ is given by $$\partial_t J + J [H,J] = 0 \label{totalderiv}$$ If $[H,J] = 0$, (\[totalderiv\]) reduces to $\partial_t J = 0$. Thus, if $H$ is self-adjoint, the time derivative of $J$ measures the amount of particle creation. Setting $Q=H$, the conditions (15) and (16) for the symmetry of $H$ become $$\begin{aligned} X &=& -X^{\dag}\label{firsteq}\\ Y &=& -Y^{\dag}\label{seceq}\\ W &=& Z^{\dag}\label{thirdeq}\end{aligned}$$ and $$\begin{aligned} (AX + BZ) &=& (AX + BZ)^{\dag}\label{fourtheq}\\ (CW + DY) &=& (CW + DY)^{\dag}\label{fiftheq}\\ (AW + BY) &=& -(CX + DZ)^{\dag}\label{sixeq}\end{aligned}$$ respectively. But an immediate consequence of Lemma 2 is that (\[realpart\]) and (\[imajpart\]) are equivalent. Thus, $H$ is symmetric if and only if (\[firsteq\])–(\[thirdeq\]) are satisfied, and this is equivalent to (\[fourtheq\])–(\[sixeq\]) being satisfied. Furthermore, these latter equations are precisely the condition for $JH$ to be antisymmetric, so that we have the further result $Q$ is symmetric if and only if $JQ$ is antisymmetric. The Ashtekar-Magnon Energy Condition {#Energy} ==================================== The essential ingredient in the result of Ashtekar and Magnon [@AM] is the [*energy condition*]{}. Given a Cauchy hypersurface $\Sigma$, one can define the classical energy and the quantum energy of a scalar field with respect to those observers orthogonal to $\Sigma$. The energy condition says that these energies should be equal. Ashtekar and Magnon showed that there is a unique complex structure $J$ on $\Sigma$ such that the energy condition is satisfied. Using the results of Section \[Math\], they have thus shown that the energy condition selects a unique quantization procedure. Ashtekar and Magnon define the classical energy of the scalar field with respect to $\Sigma$ (and the choice of scale implicit in $t^a$) to be $$CE_T = \int_\Sigma T_{ab} t^a n^b d\Sigma \label{CEAM}$$ where $$T_{ab} = \nabla_a\Phi \nabla_b \Phi - \frac{1}{2}g_{ab}(\nabla^c\Phi\nabla_c\Phi + m^2\Phi^2) \label{stresstensor1}$$ is the stress-energy tensor associated with the scalar field. Ashtekar and Magnon define the quantum energy of the scalar field with respect to $\Sigma$ (and $t^a$) to be the expectation value of the Hamiltonian operator $H$, i.e. $$QE_H = \left<\Phi,H\Phi\right> \label{QEAM}$$ But what is the Hamiltonian operator $H$? If the vector field $t^a$ is Killing, so that the spacetime is stationary, the usual definition for the Hamiltonian operator $H$ is $H\Phi=-J(\pounds_{t^a}\Phi)$, where $\pounds$ represents Lie differentiation. But in the present case the vector field $t^a$ is not necessarily Killing, and so the function $\pounds_{t^a}\Phi$ is not necessarily a solution of the Klein-Gordon equation. Therefore $H$, as defined above, is not necessarily a map into ${\cal H}$. To overcome this problem, Ashtekar and Magnon used initial data to define $H$. Let $\Phi\in V$ be a solution of the Klein-Gordon equation with initial data as in (\[data\]). Consider the data to be a function of $t$ and take its derivative; the result is in $\V$ and hence defines a solution $\dot\Phi\in V$. Explicitly, $\dot\Phi$ is the solution with initial data [^2] $$\pmatrix{\dot\varphi\cr \dot\pi} = \partial_t \pmatrix{\Phi\cr \sqrt{h}\,n^a\nabla_a\Phi} \bigg\vert_\Sigma \label{AMdata2}$$ It is straightforward but messy to use (\[KG\]) to rewrite the time derivatives in terms of spatial derivatives, resulting in $$\begin{aligned} \dot\varphi &=& {N\over\sqrt{h}} \,\pi \label{doteq1}\\ \dot\pi &=& \sqrt{h}\, (Nh^{ij}D_iD_j + h^{ij}D_iND_j - m^2N) \,\varphi \label{doteq2}\end{aligned}$$ Ashtekar and Magnon proceed to define the Hamiltonian operator $H$ by requiring $$\begin{aligned} H\Phi &=& -J\dot\Phi \label{schrodee}\end{aligned}$$ Using (\[CEAM\]) and (\[QEAM\]), the energy condition takes the form $$\left<\Phi,H\Phi\right> = \int_\Sigma T_{ab} t^a n^b d\Sigma \label{encond}$$ It is again straightforward but messy to verify that $$2\, \Omega(\Phi,\dot\Phi) = {\rm Re}\left<\Phi,H\Phi\right> = CE_T$$ so that the true content of the energy condition is $$2\, \Omega(\Phi,H\Phi) = {\rm Im} \left<\Phi,H\Phi\right> = 0 \label{true}$$ We now state without proof Ashtekar and Magnon’s main result. \[AMtheo\] [(Ashtekar and Magnon [@AM])]{} Let $(M,g_{ab})$ be a globally hyperbolic spacetime with Cauchy surface $\Sigma$, and let $V$ be as above. Then there exists a unique compatible complex structure $J$ on $V$ such that the energy condition is satisfied. In other words, there is a unique complex structure $J$ such that $$\Omega(\Phi,H\Phi) = 0 \label{encond2}$$ for every $\Phi\in\V$, where $H$ is defined in terms of $J$ via (\[schrodee\]). It is important to note that Ashtekar and Magnon assume a particular form of the Hamiltonian operator $H$, namely (\[schrodee\]). Extending Ashtekar and Magnon’s Result {#Result} ====================================== The main result of this section is Theorem \[theo1\] which, is a generalization of Theorem \[AMtheo\]. There are two main differences. First, we replace the energy condition with a more general condition, which allows some flexibility in defining the classical energy of the scalar field. Second, we eliminate the need for specifying the Hamiltonian operator $H$. Theorem \[theo1\] uniquely determines both the complex structure and the operator $\H$. \[theo1\] Let $(M,g_{ab})$ be a globally hyperbolic spacetime with Cauchy surface $\Sigma $. Let $F$ be a real, nonzero smooth function on $\Sigma$, let $K$ be the Cauchy completion of the inner product space $\tau$ with $\ds\left<f,g\right>_\tau = \int_\Sigma fgF^{-1} \,d^{n-1}x$ and let $G$ be a real, semi-bounded, positive-definite symmetric operator on $K$. Suppose we have a linear operator $H$ and a compatible complex structure $J$ defined on $V$ such that $$\left<\Phi,H\Phi\right> = \frac{1}{2}\int_\Sigma (F\pi^2 + \varphi G\varphi)\,d^{n-1}x \label{energycond}$$ for all $\Phi\in V$ with data $\hat\Phi=\pmatrix{\varphi\cr \pi} \in \V$. Then the operators $J$ and $H$ are unique and are given in terms of their action on $\V$ by $$\begin{aligned} \J &=& \pmatrix{0& (FG)^{-\frac{1}{2}}F\cr -F^{-1}(FG)^{\frac{1}{2}}& 0}\\ \noalign{\medskip} \H &=& \pmatrix{(FG)^{\frac{1}{2}}& 0\cr 0& F^{-1}(FG)^{\frac{1}{2}}F}\end{aligned}$$ [**Proof 1:**]{} The right-hand-side of (\[energycond\]) can be written as $$\frac{1}{2} \Omega(\Phi,E\Phi)$$ where $$\hat{E} = \pmatrix{~0& F\cr -G& 0}$$ Comparing real and imaginary parts of (\[energycond\]) yields for all $\Phi\in V$: $$\begin{aligned} \Omega(\Phi,JH\phi) &=& \Omega(\Phi,E\Phi) \label{one}\\ \Omega(\Phi,H\Phi) &=& 0 \label{two}\end{aligned}$$ and Lemma \[symtheo2\] now shows that both $H$ and $JH-E$ are symmetric. As discussed previously, $JH$ is antisymmetric if $H$ is symmetric, and (\[firsteq\])–(\[thirdeq\]) (with appropriate sign changes) show that $E$ is antisymmetric. Thus, $JH-E$ is both symmetric and antisymmetric, and we conclude that $$JH = E$$ or equivalently $$H = -JE \label{hje}$$ Using equation (\[hje\]) we see that equation (\[two\]) can be written as $$\Omega(\Phi,JE\Phi) = 0 \label{jeiszero}$$ Careful examination of the proof given by Ashtekar and Magnon shows that it relies only on $J$ being a compatible complex structure and on (\[jeiszero\]). We can thus use their proof to uniquely determine the complex structure $\J$ in terms of the operators $F$ and $G$, the only subtlety being the conditions on $G$ which allow square roots to be taken. Finally, writing out the multiplication in (\[hje\]) and using the identity $$BG = -FC$$ (which follows naturally from (\[jeiszero\]) and the symmetry properties) results in the given form for $\H$. This completes the proof, full details of which are given in [@Jim]. It turns out there is another proof of Theorem \[theo1\]. We provide this alternate proof here: [**Proof 2:**]{} Writing out the symmetry of $JH-E$ we obtain $$\begin{aligned} (AW + BY) & = & -(CX + DZ)^{\dag}\label{AXBZ}\\ ((AX + BZ - F) & = & (AX + BZ - F)^{\dag}\label{CWBY}\\ ((CW + DY + G) & = & (CW + DY + G)^{\dag}\label{CWDY}\\ X & = & -X^{\dag}\label{X}\\ Y & = & -Y^{\dag}\label{Y}\\ W & = & Z^{\dag}\label{ZW}\end{aligned}$$ and the compatibility of $J$ yields $$\begin{aligned} B & = & B^{\dag}\label{B}\\ C & = & C^{\dag}\label{C}\\ A & = & -D^{\dag}\label{A=-D}\\\end{aligned}$$ By using the symmetry and antisymmetry properties of the operators $A$ through $Z$, we can rewrite (\[AXBZ\])–(\[CWDY\]) as $$\begin{aligned} AW + BY & = & XC + WA \label{XCWA}\\ XD + WB - F & = & -(AX + BZ - F)\label{XDWB}\\ ZC + YA + G & = & -(CW + DY + G)\label{ZCYA}\end{aligned}$$ Taking the adjoint of (\[XCWA\]) yields $$ZD + YB = CX + DZ \label{ZDYB}$$ Solving for $F$ in (\[XDWB\]) yields $$F = \frac{1}{2}\left(AX + XD + WB + BZ\right)\label{F}$$ Multiplying (\[F\]) on the left by $A$ and on the right by $D$ and subtracting gives $$AF - FD = \frac{1}{2}\left(A^2X + AWB + ABZ - XD^2 - WBD -BZD\right) \label{AFFD1}$$ Multiplying (\[ZDYB\]) on the left by $B$ and (\[XCWA\]) on the right by $B$ and solving for $BZD$ and $AWB$ yields $$\begin{aligned} BZD & = & BCX + BDZ - BYB\label{BZD}\\ AWB & = & XCB + WAB - BYB\label{AWB}\end{aligned}$$ Substituting (\[BZD\]) and (\[AWB\]) into (\[AFFD1\]) yields $$AF-FD = \frac{1}{2}\left(A^2X + BCX - XD^2 + ABZ - WBD + BDZ - WAB - XCB\right)\label{AFFD2}$$ Finally, using (\[struct\]) we see that the right hand side of (\[AFFD2\]) is identically zero. Therefore we now know that $$D = F^{-1}AF\label{AFFD3}$$ Using the argument given by Ashtekar and Magnon in proving Theorem \[AMtheo\], we can conclude that $$A = 0 = D$$ It is then straightforward to determine $B$ and $C$, thus obtaining the complex structure $\J$, and to then solve for $\H$. Applications {#Appl} ============ Minimal Coupling ($ \xi = 0$) {#application1} ----------------------------- We first recover Ashtekar and Magnon’s result. Comparing (\[doteq1\])–(\[doteq2\]) and (\[schrodee\]) with (\[energycond\]) shows that we should set $$\begin{aligned} G &=& \frac{\sqrt h}{~N} \,\Theta \\ F &=& \frac{~N}{\sqrt h}\end{aligned}$$ where the operator $\Theta$ is defined by $$\Theta = -\left( N^2 h^{ij} D_i D_j + h^{ij} N D_i N D_j - m^2 N^2 \right)$$ Theorem \[theo1\] selects for us operators $\Jzero$ and $\Hzero$ (the zero in the subscripts is being used to emphasize that we are considering the minimally coupled case): $$\begin{aligned} \Jzero &=& \pmatrix{0& \Theta^{-\frac{1}{2}}\frac{~N}{\sqrt h}\cr \noalign{\smallskip} -\frac{\sqrt{h}}{~N}\Theta^{\frac{1}{2}}& 0} \label{jay}\\ \noalign{\medskip} \Hzero &=& \pmatrix{\Theta^{\frac{1}{2}}& 0\cr \noalign{\smallskip} 0& \frac{\sqrt{h}}{~N}\Theta^{\frac{1}{2}}\frac{~N}{\sqrt{h}}} \label{ache}\end{aligned}$$ which agree with [@AM]. Furthermore, we have $$\Jzero\Hzero\pmatrix{\varphi \cr \pi} = \pmatrix{\frac{~N}{\sqrt h}\pi\cr \noalign{\smallskip} -\frac{\sqrt h}{~N}\Theta\varphi}$$ and comparing with (\[doteq1\])–(\[doteq2\]) shows that $$\begin{aligned} \hat\Hzero\hat\Phi &=& -\Jzero\hat{\dot\Phi}\end{aligned}$$ as desired. Therefore the Hamiltonian operator $H$ and complex structure $J$ obtained via Theorem \[theo1\] satisfy an equation which mimics the Schrödinger equation. It is of course not always true that $\dot\Phi = t^a\nabla_a\Phi$. However, if the vector field $t^a$ is Killing, then the operators $H$ and $J$ determined by Theorem \[theo1\] would indeed satisfy [**the**]{} Schrödinger equation $$\begin{aligned} H\Phi & = & -J\pounds_t\Phi \label{shrodenew}\end{aligned}$$ and would therefore correctly reduce to the well-established theory for static spacetimes. Non-Minimal Coupling ($\xi \neq 0$) {#application2} ----------------------------------- For the second application of Theorem \[theo1\] we will allow non-zero values for the coupling constant $\xi$. As in the previous application we will need to define what is meant by the classical energy of the scalar field. If we stick with the definition for the classical energy which is given by (\[CEAM\]) we will find that Theorem \[theo1\] selects for us a Hamiltonian operator and complex structure. However, it turns out that these operators do not satisfy (\[schrodee\]) and hence, in the static limit, the operators would not satisfy (\[shrodenew\]). However, by choosing the classical energy of the field to be the surface integral of the classical Hamiltonian, we can still apply Theorem \[theo1\], and in this case we obtain a Hamiltonian operator and complex structure which do reduce to the usual Hamiltonian operator and complex structure when considering static spacetimes. A primary candidate for our definition of the classical energy of the field associated with the Cauchy surface $\Sigma$ and timelike vector field $t^a$ is the one given by Ashtekar and Magnon (\[CEAM\]), involving the stress-energy tensor $T_{ab}$. The stress-energy tensor is is obtained by varying the action (\[action\]) with respect to the metric $g_{ab}$ (for more details see [@BD], Chapter 3). [^3] $$\begin{aligned} T_{ab}= \frac{2}{\sqrt{-g}}\frac{\delta S}{\delta g^{ab}} & = & (1 - 2\xi )\nabla _a \Phi \nabla _b \Phi + (2\xi - \frac{ 1}{ 2})g_{ab}g^{cd}\nabla _c\Phi \nabla _d \Phi \nonumber\\ & & \mbox{} - 2\xi \Phi \nabla _a \nabla _b \Phi + 2\xi g_{ab}\Phi g^{cd}\nabla_c\nabla_d\Phi \Phi \nonumber\\ & & \mbox{} + \frac{1}{2}( g_{ab}m^2 + g_{ab}R\xi - 2\xi R_{ab} ) \phi ^2 \label{Stress}\end{aligned}$$ In the minimally coupled case, that is when $\xi = 0$, the stress-energy tensor reduces to (\[stresstensor1\]). Straightforward computations allow us to put the integral $\int_\Sigma T_{ab}n^a t^b d\Sigma$ in the following form: $$\int _\Sigma T_{ab}n^a n^b Nd\Sigma = \frac{1}{2} \int _\Sigma \left( \frac{~N}{\sqrt h}\pi ^2 + \frac{\sqrt h}{~N} \varphi \Gamma \varphi \right)\, d^{n-1}x$$ where $$\begin{aligned} -\Gamma & = & (1 - 4\xi)h^{ij} N D_i N D_j + N^2 (1 - 4\xi) h^{ij} D_i D_j \\ & & \mbox{ } - \xi R N^2 - 2\xi n^an^bR_{ab} - m^2 N^2\end{aligned}$$ Since $\Gamma$ is semi-bounded, symmetric and positive-definite on $K$ and $\frac {~N}{\sqrt h}$ is nonzero, we can apply Theorem \[theo1\] and obtain the following operators: $$\begin{aligned} \J_{\xi} & = & \pmatrix {0 & \Gamma^{-\frac{1}{2}}\frac{~N}{\sqrt h} \cr -\frac{{\sqrt h}}{~N}\Gamma ^{\frac{1}{2}} & 0}\label{jay2}\\ & &\nonumber \\ \H_{\xi} & = & \pmatrix{\Gamma^{\frac{1}{2}} & 0 \cr 0 & \frac{{\sqrt h}}{~N}\Gamma^{\frac{1}{2}}\frac{~N}{\sqrt h}} \label{ache2}\end{aligned}$$ We can proceed as we did in the minimally coupled case and find that the operators (\[jay2\]) and (\[ache2\]) satisfy $$\Hxi\pmatrix{\varphi \cr \pi} = -\Jxi\pmatrix{{\frac{~N}{\sqrt h}\pi}\cr \noalign{\smallskip} -\frac{\sqrt h}{~N}\Gamma\varphi} \label{shrode3}$$ As in the minimally coupled case we have that $$\frac{~N}{\sqrt h}\pi = \partial_t\Phi\vert_\Sigma\nonumber$$ However, the function $-\frac{\sqrt h}{~N}\Gamma\varphi$ does not equal the restricted time derivative of $\Pi=\sqrt{h}n^a\nabla_a\Phi$: $$\begin{aligned} -\frac{\sqrt h}{~N}\Gamma\varphi & = & \sqrt h(1 - 4\xi)h^{ij}D_iND_j\varphi + N(1 - 4\xi)h^{ij}D_i\varphi D_j\varphi \nonumber \\ & & \mbox{ } - (\xi RN + N^{-1}2\xi n^an^bR_{ab} + m^2N)\varphi \label{Psi}\\ \partial_t\Pi\vert_\Sigma & = & \sqrt hh^{ij}D_iND_j\varphi + Nh^{ij}D_i\varphi D_j\varphi \nonumber \\ & & \mbox{ } - (\xi RN + m^2N)\varphi \label{Pidot}\end{aligned}$$ There are several differences between (\[Psi\]) and (\[Pidot\]), so that (\[shrode3\]) does not mimic the Schrödinger equation. We conclude that using the stress-energy tensor to define the classical energy of the field when $\xi \neq 0$ will produce an undesirable choice for the Hamiltonian operator and complex structure. If $\xi \neq 0$, one is either forced to reconsider the definition of the classical energy of the scalar field or abandon the use of Theorem \[theo1\]. Fortunately there does exist at least one other natural method for defining the classical energy of the field; this alternate definition involves the classical Hamiltonian. With $\Phi$ and $\Pi$ as above, the classical Hamiltonian [^4] is defined to be $$\begin{aligned} \Ham & = & \Pi t^a\nabla_a\Phi - {\cal L}\\ & = & \frac{~N}{\sqrt h}\Pi^2 - {\cal L}\\ & = & \frac{1}{2}\left(\frac{~N}{\sqrt h}\Pi^2 + N{\sqrt h} h^{ij}\partial_i\Phi\partial_j\Phi + N\sqrt h(m^2 + \xi R)\Phi^2\right) \label{Ham}\end{aligned}$$ The alternate definition for the classical energy of the scalar field associated to the hypersurface $\Sigma$ and the vector field $t^a$ is the surface integral of this Hamiltonian: $$\CE2 = \int_\Sigma \Ham \,d^{n-1}x\label{Hamdensity}\\ = \frac{1}{2}\int_\Sigma \left(\frac{~N}{\sqrt h}\pi^2 + N{\sqrt h}h^{ij}D_i\varphi D_j\varphi + N{\sqrt h}(m^2 + \xi R)\varphi^2 \right)\,d^{n-1}x \label{Hamdensity2}$$ In the case of minimal coupling, this definition for the classical energy of the scalar field coincides with Ashtekar and Magnon’s definition which involves the stress-energy tensor. We now show that by using the surface integral of the classical Hamiltonian to represent the classical energy of the field, we are still able to apply Theorem \[theo1\]. Moreover, Theorem \[theo1\] determines a unique Hamiltonian operator and unique complex structure that reduce to the appropriate operators when considering static spacetimes. By using the definition of the Hamiltonian and applying integration by parts we obtain $$\CE2 = \frac{1}{2}\int_\Sigma \left( \frac{~N}{\sqrt h}\pi^2 -\varphi ({\sqrt h}(Nh^{ij}D_iD_j + h^{ij}D_iND_j - N (m^2 + R\xi))\varphi \right)\, d^{n-1}x$$ The operator $$-\Upsilon = N^2 h^{ij} D_i D_j + h^{ij} N D_i N D_j - N^2 (m^2 + R\xi)$$ is positive-definite, semi-bounded and symmetric on $K$. We can therefore apply Theorem \[theo1\] by letting $G = \frac{\sqrt h}{~N}\Upsilon$ and $F = \frac{~N}{\sqrt h}$. We obtain for $\Jxi$ and $\Hxi$ the following operator-valued matrices: $$\begin{aligned} \J_{\xi} & = & \pmatrix {0 & \Upsilon^{-\frac{1}{2}}\frac{~N}{\sqrt h} \cr -\frac{{\sqrt h}}{~N}\Upsilon ^{\frac{1}{2}} & 0} \label{jay3}\\ \noalign{\medskip} \H_{\xi} & = & \pmatrix{\Upsilon^{\frac{1}{2}}& 0\cr 0& \frac{{\sqrt h}}{~N}\Upsilon^{\frac{1}{2}}\frac{~N}{\sqrt h}} \label{ache3}\end{aligned}$$ Proceeding as in the previous cases, we find that the operators (\[jay3\]) and (\[ache3\]) satisfy $$\Hxi\pmatrix{\varphi \cr \pi} = -\Jxi\pmatrix{{\frac{~N}{\sqrt h}\pi} \cr -\frac{\sqrt h}{~N}\Upsilon\varphi} \label{shrode4}$$ However, this time we have $$\pmatrix{{\frac{~N}{\sqrt h}\pi} \cr -\frac{\sqrt h}{~N}\Upsilon\varphi} = \partial_t \pmatrix{\Phi\cr \sqrt{h}\,n^a\nabla_a\Phi} \bigg\vert_\Sigma$$ That is, (\[shrode4\]) does indeed reduce to the Schrödinger equation when $t^a$ is a Killing vector field. Discussion {#Disc} ========== In Section \[Energy\] we summarized Ashtekar and Magnon’s uniqueness result (Theorem \[AMtheo\]). By requiring the quantum energy of the scalar field to be equal to the classical energy of the field at each instant of time, they were able to uniquely specify a complex structure $\J$ at each instant of time. However, their result depended on the need to define the Hamiltonian operator. Our main result (Theorem \[theo1\]) was discussed in Section \[Result\]. We showed that the Ashtekar and Magnon energy condition uniquely determines not only the complex structure, but also the Hamiltonian operator. As shown in Section \[application1\], Ashtekar and Magnon’s result is thus a special case of Theorem \[theo1\]. An important consequence of Theorem \[theo1\] concerns the case of non-trivial coupling ($\xi\neq0$). We saw in Section \[application2\] that the usual definition for the classical energy produces a complex structure and Hamiltonian operator that do not reduce to the appropriate operators if the spacetime is assumed to be static. However, also in Section \[application2\], we showed that using the classical Hamiltonian to define the classical energy produces via Theorem \[theo1\] a complex structure and Hamiltonian operator which do have the correct limits in the static case. It is somewhat disturbing that the two obvious formulations of the classical energy, namely the stress-energy tensor and the classical Hamiltonian, fail to agree when $\xi\ne0$. The results of Section \[application2\] suggest that the latter is to be preferred. It would therefore be worthwhile to further examine the properties of the generalized classical energy (\[Ham\]). For instance, under what circumstances is it conserved? The stress-energy tensor (\[Stress\]) is obtained by varying the action (\[action\]) with respect to the metric $g_{ab}$. By suitably modifying the action prior to carrying out the variation, is it possible to obtain the same classical energy using the stress-energy tensor that is obtained using the classical Hamiltonian? We have shown that if $\xi\ne0$ then the complex structure and Hamiltonian operator obtained using the stress-energy tensor are different from those obtained using the classical Hamiltonian. What is the relationship between the two different Fock spaces which are associated with these two different choices for the classical energy? Finally, we emphasize that both Ashtekar and Magnon’s work and ours consider only hypersurface orthogonal observers. While this description lends itself naturally to globally hyperbolic spacetimes, in which one can always [*choose*]{} such observers, it does not address stationary but non-static observers, let alone more general rotating observers. Some preliminary ideas on how to deal with these cases can be found in [@Kulkarni; @Boersma]. Acknowledgments {#acknowledgments .unnumbered} =============== It is a pleasure to acknowledge discussions with Corinne Manogue. This work is based on a dissertation [@Jim] submitted to Oregon State University by JF in partial fulfillment of the requirements for the Ph.D. degree in mathematics. A. Ashtekar and Anne Magnon, [*Quantum Field Theory in Curved Space-Times*]{}, Proc. Roy. Soc. [**A346**]{}, 375 (1975). James P. Fischer, [**A New Look at the Ashtekar-Magnon Energy Condition**]{}, PhD Dissertation, Oregon State University, 1998. N. D. Birrell and P. C. W. Davies, [**Quantum Fields in Curved Space**]{}, Cambridge University Press, Cambridge, 1982. Stephen A. Fulling, [**Aspects of Quantum Field Theory in Curved Space-Time**]{}, Cambridge University Press, Cambridge, 1989. Jürgen Renn, [**Teilchenstrukturen in Expandierenden Universen**]{}, Diplomarbeit, Freie Universität Berlin, 1982. Robert M. Wald, [**Quantum Field Theory in Curved Spacetimes and Black Hole Thermodynamics**]{}, University of Chicago Press, Chicago, 1994. Tevian Dray, Ravi Kulkarni, and Corinne A. Manogue, [*Scalar Field Quantization in Stationary, Non-Static Spacetimes*]{}, Gen. Rel. Grav. [**24**]{}, 1255–1266 (1992). Stuart. F. Boersma, [**Parametric Manifolds**]{} PhD Dissertation, Oregon State University, 1994; Stuart Boersma and Tevian Dray, [*Parametric Manifolds I: Extrinsic Approach*]{}, J. Math. Phys. [**36**]{}, 1378–1393 (1995); Stuart Boersma and Tevian Dray, [*Parametric Manifolds II: Intrinsic Approach*]{}, J. Math. Phys. [**36**]{}, 1394–1403 (1995); Stuart Boersma and Tevian Dray, [*Slicing, Threading and Parametric Manifolds*]{}, Gen. Rel. Grav. [**27**]{}, 319–339 (1995). [^1]: Ashtekar and Magnon omit the factor of $\sqrt{h}$ from $\pi$. [^2]: The dot does not refer to a time derivative! (This construction is less intuitive with Ashtekar and Magnon’s choice of data.) [^3]: There is a sign error in the last term of the third equation of (3.196) on p. 88 in [@BD]. [^4]: The usual definition for the classical Hamiltonian assumes that $t^a$ is a Killing vector field [@SF; @JR]. Therefore it may be more appropriate to call this a generalized classical Hamiltonian.

ArXiv

--- abstract: 'We summarize Run I results relevant to an analysis of the CP asymmetry in $B\to J/\psi K_s$, the CDF upgrade plans for Run II, and some of the main $B$ physics goals related to the exploration of the origin of CP violation.' address: 'Fermilab M.S. 318, P.O. Box 500, Batavia IL 60510' author: - Fritz DeJongh title: '$B$ Physics with the CDF Run II Upgrade FERMILAB-Conf-95/408-E ' --- = 0.5cm \#1[| \#1|]{} INTRODUCTION ============ During the Run I data taking period, from 1992 through 1995, CDF has acquired 110  of $p \bar{p}$ collisions at a center of mass energy of 1800 GeV. This data has provided many results on $B$ physics [@Meschi], and provides a basis for extrapolating to Run II, which is scheduled to start in 1999 after major upgrades to both the accelerator and detector. We present herein a summary of Run I results relevant to an analysis of the CP asymmetry in $B\to J/\psi K_s$, the CDF upgrade plans for Run II, and some of the main $B$ physics goals related to the exploration of the origin of CP violation. TAGGED $B\to J/\psi K_s$ IN RUN I {#section_runi} ================================= In the first 60 pb$^{-1}$ of Run I, as shown in Figure \[fig:b01\_jpsi\_kshort\], we have observed 140 $B^0 \to J/\psi K_s$ events with signal-to-noise better than 1:1. We obtained this sample with a dimuon trigger that required both muons to have transverse momentum ($P_T$) greater than 2.0 GeV. To obtain the CP asymmetry we must tag the flavor of the $B$ meson at the time at which it was produced. Work is under way to use a combination of Run I data and Monte Carlo to establish the “effective tagging efficiency” $\epsilon (1-2w)^2$, where $\epsilon$ is the tagging efficiency and $w$ is the mistag fraction, for a variety of algorithms. We currently have results [@Manfred] for two methods, Jet Charge and Muon tagging, for a total effective tagging efficiency of $\approx 2\%$. These results indicate that an order of magnitude improvement in the statstical uncertainty on the CP asymmetry will lead to a competitive measurement in Run II. =3.0in ACCELERATOR IMPROVEMENTS FOR RUN II =================================== A project called Fermi III is underway [@fermi3] to upgrade the Fermilab accelerator complex to produce an order of magnitude higher luminosity in the Tevatron. The luminosity in Run I was limited by the antiproton current. The largest component of Fermi III is to replace the Main Ring, which is housed in the same tunnel as the Tevatron and provides the acceleration stage just prior to the Tevatron, with the Main Injector, which will be housed in a separate and new tunnel. The Main Injector will provide for higher proton intensity onto the antiproton production target, and larger aperture for antiproton transfer into the Tevatron. Combined with improvements to the antiproton cooling system, the antiproton stacking rate will increase by over a factor of three to $17 \times 10^{10}$ per hour. The Tevatron schedule and some basic parameters are shown in Figure \[fig:schedule\]. Our physics projections for Run II assume 2  of integrated luminosity. DETECTOR IMPROVEMENTS FOR RUN II ================================ The CDF detector is being upgraded to handle an order of magnitude higher luminosity, and 132 ns bunch spacing [@PAC]. The main goal is to maintain detector occupancies at Run I levels, although many of these upgrades also provide for qualitatively improved detector capabilities. Tracking Upgrades ----------------- The efficiency of the current tracking system would be significantly degraded at luminosities of $\tentt$: Primary track efficiency would drop by 10%, and $K_s$ efficiency would drop by 60%. A three part tracking upgrade is being planned to recover this efficiency: - A Central Straw Tracker (CST) will consist of 4 axial and 4 stereo superlayers of 8 to 12 straws each at radii of 50 to 140 cm. - An Intermediate Scintillating Fiber Tracker (IFT) will consist of six stereo and six axial layers of 500 micron diameter scintillating fibers read out by VLPCs. - A new Silicon Vertex Detector (SVX II). The SVX II will consist of 5 layers of double sided silicon from radii of 2.9 to 10 cm, arranged in 5 axial layers, 2 small angle ($1.2^\circ$) stereo layers, and 3 $90^\circ$ stereo layers. Since each of these detectors can be read out in less than 132 ns, occupancies will be held to Run I levels up to luminosities of $3 \times \tentt$ at 132 ns bunch spacing. Furthermore, each detector is potentially a stand-alone 3D tracker, providing for greater redundancy in pattern recognition. In addition to maintaining efficiency, these upgrades provide for new tracking capabilities: Precision vertexing in 3 dimensions, tracking to $\abs{\eta} <2$ and tracking down to $p_T > 100 \mevcc$. Time of Flight -------------- We are planning for a Time of Flight system consisting of 3 m long $4 \times 4$ cm scintillator blocks placed at a radius of 1.4 m (inside the solenoid) and read out with mesh dynode photomultiplier tubes. The 4 cm width results in less than 20% confusion from multiple hits and other noise sources. We expect 100 to 125 ps time resolution, for $>2\sigma$ $K/\pi$ separation in the momentum range from 0.3 to 1.6 . This momentum range includes 55% of kaons potentially useful for flavor tagging. Trigger and DAQ system ---------------------- CDF is planning for a trigger and DAQ upgrade to allow for higher data rates while increasing the sophistication of the trigger decision, as summarized in Figure \[fig:trigger\]. Data is stored in a 42 cell pipeline while awaiting the Level 1 trigger decision, and can be transferred to Level 2 without halting Level 1. Information available for the Level 1 decision will include calorimetry clusters, CST tracks with $p_T > 1.5\ \gevcc$, and electron and muon identification. At Level 2, SVX II information will also be available, and DEC Alpha based processors allow for programmable algorithms. A commercial switch will be used to assemble events into the Level 3 processors where a full event reconstruction will be performed for the final trigger decision. $B$ PHYSICS EXPECTATIONS FOR RUN II =================================== The challenge for $B$ physics in Run II is to develop efficient trigger algorithms for key final states, and efficient flavor tagging algorithms. In this section, we discuss possibilities for flavor tagging, measurements of $\sin(2\beta)$, $\sin(2\alpha)$, $B_s$ mixing, and the observation of rare decay modes. Topics not discussed here include [@pacp]: The CP angle $\gamma$ [@Snowmass_gamma], study of exclusive $b\to u$ semileptonic decays, measurement of $V_{cb}$ in semileptonic $\Lambda_b$ decays, and $B_c$ spectroscopy. Flavor tagging -------------- As discussed in Section \[section\_runi\] we currently have results for two flavor tagging methods, jet charge and muon tagging, for a total effective tagging efficiency of almost 2%. Our goal for Run II is to attain the following effective tagging efficiencies from various algorithms: - 2% from lepton tagging, using electrons as well as muons, and additional coverage planned for lepton identification in Run II. - 2% from same-side tagging, which exploits the charge correlation of the pions produced in the fragmentation process along with the $B$ meson [@Rosner]. For the $B_s$ case, the charge correlation of kaons identified in the time-of-flight system may result in a tagging efficiency of 5%. - 3% from kaon tagging, using tracks with high impact parameter identified as kaons in the time of flight system. - 4% from a jet charge algorithm expoiting 3D vertexing information from the SVX II, and stand-alone tracking information from the SVX II and Intermediate Fiber Tracker. While work is in progress to evaluate all these algorithms using current data, for now we assume an 8% flavor tagging efficiency for $B_d$ mesons, and 11% for $B_s$ mesons. CP Asymmetry in $B_d \to J/\psi K_s$: $\sin(2 \beta)$ {#bsection_sin2b} ----------------------------------------------------- With a large branching ratio and distinctive trigger signature, the decay mode $B_d \to J/\psi K_s$ is the leading candidate for the initial observation of CP violation in the $B$ system. Furthermore, the extraction of $\sin(2 \beta)$ from this asymmetry is essentially free of hadronic uncertainties. The current Standard Model predictions for $\sin(2\beta)$ are $\sin(2\beta) > 0.17$ [@Ali-London] and $\sin(2\beta) = 0.65 \pm 0.12$ [@ciuchini]. As discussed in Section \[section\_runi\], CDF in Run I has reconstructed $\approx2$ $J/\psi K_s$ events per , using a dimuon trigger with a $p_T$ threshold of 2.0  on each muon. Improvements for Run II include lowering the trigger $p_T$ threshold to 1.5  per muon, improving the muon coverage, and using the channel $J/\psi \to e^+ e^-$. Our goal for Run II is to reconstruct 10 $B^0 \to J/\psi K_s$ events per , for a yield of 20,000 events assuming 2 . We also expect to achieve much improved signal-to-noise by using the improved capability and coverage of the SVX II, but have conservatively assumed $S/N=$2:1. Assuming an effective tagging efficiency of 8%, we find an uncertainty on $\sin(2\beta)$ of $0.07$. In addition to the above expectation of 20,000 $B^0 \to J/\psi K_s$ events in $2\ {\rm fb}^{-1}$, the $B^0 \to J/\psi K_s$ yield can increase by employing (a) the increased tracking coverage and (b) new ways of triggering, such as requiring one lepton and one additional track with large impact parameter. While speculative, an additional factor of four or more in the number of reconstructed $B^0 \to J/\psi K_s$ events may be possible. CP Asymmetry in $B_s \to J/\psi\phi$ {#section_psiphi} ------------------------------------ Within the Standard Model, the the CP asymmetry in $B_d \to J/\psi K_s$ measures the weak phase of the CKM matrix element $V_{td}$, while the CP asymmetry in $B_s \to J/\psi\phi$ measures the weak phase of the CKM matrix element $V_{ts}$, which is expected to be very small. As emphasized by Y. Nir [@YNir], and Helen Quinn at this workshop, the channel $B_s \to J/\psi\phi$ may therefore provide a signature for a source of CP violation beyond the Standard Model. With the same trigger improvements as for $B_d \to J/\psi K_s$, we expect 12000 $B_s \to J/\psi\phi$ events in Run II. =3.0in The magnitude of a CP asymmetry in $B_s \to J/\psi\phi$ decays would be modulated by the frequency of $B_s$ oscillations. Thus, for a meaningful limit, we must be able to resolve $B_s$ oscillations. If we neglect resolution effects, we can expect a precision on the asymmetry of $\pm 0.09$. However, resolution effects smear the oscillations and produce an additional dilution. Our experience in Run I shows that if we determine the primary vertex event-by-event, the proper lifetime resolution for fully reconstructed $B$ decays is $\approx 30 \mu$m. Figure \[acp\_psiphi\] shows our expected precision on the asymmetry as a function of $x_s$. There will be an additional dilution if the $J/\psi \phi$ final state is not a pure CP eigenstate. CP Asymmetry in $B^0 \to \pi^+\pi^-$: $\sin(2 \alpha)$ ------------------------------------------------------ A measurement of $\sin(2 \alpha)$ in conjunction with $\sin(2 \beta)$ provides powerful constraints on the unitarity triangle [@Lauten]. The greatest challenge in this measurement is the trigger requirement at a luminosity of $1 \times 10^{32} {\rm cm}^{-2}{\rm sec}^{-1}$. Our plan (described in detail in [@Snowmass_pipi]) consists of 1. At Level-1: Require two tracks with $P_T>2$ , imposing $\Delta\phi$ cuts on opposite sign track pairs. 2. At Level-2: Require an impact parameter $> 100 ~\mu$m for each track. 3. At Level-3: Use the full event information for the final decision. After additional analysis requirements we expect $\approx 5$ $B^0 \to \pi^+\pi^-$ events per pb$^{-1}$. Due to the impact parameter cuts, the proper lifetime distribution starts at $\approx 1.5$ lifetimes, and therefore the dilution of the CP asymmetry due to mixing of the signal $B$ before it decays will be 0.82, rather than 0.47 as we assumed for $\sin(2\beta)$. To measure the CP asymmetry in $B^0 \to \pi^+\pi^-$ events one needs to determine the fraction of the signal from $B_d \to K^+ \pi^-$, $B_s \to K^- \pi^+$ and $B_s \to K^- K^+$ decays. This can be done using invariant mass and $dE/dx$ distributions in the high statistics untagged sample. Figure \[fig:b04\_pipimass\] displays the expected mass distribution for the combination of the above four signals, assuming a pion mass assignment for all tracks [@Snowmass_pipi]. The $B_d \to \pi^+ \pi^-$ and $B_d \to K^+ \pi^-$ peaks are separated by 40 , while we expect a mass resolution of $\approx 28 \mevcc$. As in Run I, we also expect $K/\pi$ separation from $dE/dx$ in the CST of better than $1\sigma$. Any CP asymmetry in the $K\pi$ background component can be determined from the ratio of numbers of $K^-\pi^+$ and $K^+\pi^-$ events in the untagged sample. Any CP asymmetry in the $K^+ K^-$ background component would be modulated by the $B_s$ mixing frequency rather than the $B_d$ mixing frequency. Therefore, the CP asymmetry in the tagged sample in conjunction with a fit to the untagged sample can yield the CP asymmetry for $B^0 \to \pi^+\pi^-$. =3.0in Another issue for this analysis is the combinatorial background under the $B$ peak. We can estimate this background level detector using a sample of high $E_T$ electron triggers from Run I. In the case that the electron results from the semileptonic decay of a $B$ hadron, we can search for the other $B$ in the event to decay to $\pi^+ \pi^-$. Using standard cuts on the decay vertex and the isolation of the two-track combination, we obtain an observed background, $N$, comparable to the expected signal, $S$ (less than one event), for $P_T > 4$  on each track: $S/N \approx 1:1$. Lowering the $P_T$ threshold to 2  will allow us to double our efficiency. We expect to do this with the Run II detector while maintaining $S/N$ better than 1:1 by exploiting the 3D information from the SVX II and optimizing cuts. The final issue related to the extraction of the angle $\alpha$ from the measured CP asymmetry in $B_d \to \pi^+\pi^-$ is the extraction of possible penguin contributions in addition to the tree diagram which is expected to dominate this decay mode. We can estimate this penguin contamination, and thus extract $\alpha$, from a combination of experimental measurements and theoretical inputs. In particular, a time-dependent analysis yields a measurement of the amplitude as well as the phase of the CP asymmetry, which oscillates with the mixing frequency. This latter phase would be zero in the absence of a penguin contribution. In addition, we use the average branching ratio $\left(Br(B^0 \to \pi^+\pi^-)+Br(\bar B^0 \to \pi^+\pi^-)\right)/2$. This quantity can be extracted from untagged $B_d \to \pi^+\pi^-$ decays and will therefore have a very small error. Other ingredients are the value of $Br(B \to \pi \ell \nu)$ and some theoretical input such as the magnitude of the tree diagram given $Br(B_d \to \pi^- \ell \nu)$. As an example, if the penguin amplitude, $A_p$, is small compared to the tree amplitude, $A_t$, (say, $A_p/A_t = 0.07$ as predicted by Deshpande [*et. al.*]{} [@DH1]) the extraction of $\alpha$ is relatively easy, and the theoretical constraints can be relatively crude. If $A_p/A_t \approx 0.2$, this becomes more challenging, but feasible. A detailed analysis can be found in reference [@Fritz_Paris]. In conclusion, assuming a flavor-tagging efficiency of $8\%$ as for the $J/\psi K_s$ case, and a conservative $S/N = 1/4$, we expect an overall uncertainty on $\sin(2 \alpha)$ of $\pm 0.10$. Measurement of --------------- The CDF $B$ physics goals for Run II include observation of the time dependence of $B_s$ and $B_d$ mixing, obervation of exclusive radiative penguin decays, and the observation of a lifetime difference $\delta \Gamma$ between the CP eigenstates of the $B_s$ meson. The ratio of $B_d$ to $B_s$ mixing parameters, $x_d/x_s$, is proportional to $\Vtdts^2$. The matrix elements for $B_d$ and $B_s$ mixing are related by SU(3), allowing the cancellation of many theoretical uncertainties in the ratio. Similarly, in the absence of long distance effects, the ratio of decay rates $B(B\to\rho\gamma)/B(B\to K^*\gamma)$ is proportional to $\Vtdts^2$. Again, since these final states are related by SU(3), many theoretical uncertainties cancel in the ratio. A smaller value of  implies a larger value of $x_s$, and a smaller rate for $B\to\rho\gamma$, and therefore both of these measurements become more difficult. However, the lifetime difference $\Delta \Gamma_s$ is proportional to $x_s$, and therefore this measurement becomes easier. Although the theoretical uncertainties on $\Delta \Gamma_s$ are larger, the combination of the three types of measurements discussed in this section should allow CDF to measure over the full range permitted by the standard model. ### $B_s$ Mixing ALEPH has shown that $x_s > 8.8$ [@aleph], implying that $x_s$ must be measured by fitting the time-dependent oscillation of the $B_s$. For the Run Ia $B_s$ lifetime measurement [@Bs-ctau], CDF reconstructed 70 $B_s\rightarrow D_s \ell \nu$ events. For Run II, triggering and reconstruction of this channel with very high statistics is straightforward, and we expect $\approx 10^5$ events. Due to the unreconstructed neutrino, knowledge of the $B$ momentum limits the measurement to values of $x_s < \sim 11$. However, improvements in the analysis technique may result in an improved momentum resolution. For example, 3D vertexing allows a determination of the four-momentum of the missing neutrino, although with a quadratic ambiguity. For fully reconstructed decays, the $x_s$ reach is limited by vertexing resolution, as discussed in section \[section\_psiphi\]. In order to determine the flavor of the $B_s$ at the time of the decay this measurement requires events of the type $B_s \rightarrow D_s n\pi$. The challenge for CDF is to trigger on, and isolate from background, signals of this type. We note that the presence of a time-of-flight system in CDF should significantly improve the reconstruction efficiency by allowing efficient selection of kaons and rejection of pions at low $P_T$, where the backgrounds are largest. One strategy is to trigger on a single lepton ($e$ or $\mu$), which will serve as the flavor tag, and then reconstruct $B_s$ decays in this sample [@Snowmass]. For a 6 GeV lepton threshold in Run II, there will be $\sim 2 \times 10^3$ $B_s$ mesons that have decayed within the CDF fiducial volume to the modes $B_s \rightarrow D_s + \pi$ and $B_s \rightarrow D_s + 3\pi$ with $D_s\rightarrow \phi \pi$ or $D_s \rightarrow K^{*\pm} K^\mp$. It is not yet known how many of these may be reconstructed with good signal-to-noise. It is likely that the lepton trigger threshold will be lower, and also that some of the decay products of the $B_s$ will be included in the trigger requirement as well. Another strategy is to use a fully hadronic trigger, as for $B \to \pi^+\pi^-$, in which case all tagging techniques may be applied. The final states of the $B_s$ we are trying to reconstruct are produced an order of magnitude more frequently than $B \to \pi^+\pi^-$. More work is needed to design such a trigger; one possibility is a two-track trigger optimized for $\phi \to K^+ K^-$. Although we do not have a solid projection of how many events we will reconstruct, we show in Figure \[acp\_psiphi\] our precision on $x_s$ if we succeed in reconstructing 2000 events from fully hadronic triggers, with an effective tagging efficiency of 11%, or equivalently, 800 events recontructed in events with a lepton trigger. ### Radiative $B$ Decays Measurements of radiative $B$ decays at CLEO sets an upper bound on of 0.76 [@CLEO-rare]. CDF has already installed a trigger to collect radiative penguin decays. The limited bandwidth available in the current trigger and data acquisition system require the trigger to have quite high thresholds (10 GeV photon plus two 2 GeV tracks). The expected yield with this trigger is $\approx 20$ $\gamma K^*$ events per 100 pb$^{-1}$. In Run II, we expect to lower the photon $E_t$ threshold to 5 GeV and the track $P_t$ threshold to 1.5 GeV, with a resulting yield of $\sim 135$ events per 100 pb$^{-1}$ or $\sim 2700$ for 2 fb$^{-1}$. The mass resolution of the reconstructed $B$ is dominated by the resolution on the photon energy and is $\sim 140$ MeV. We have studied our ability to reject combinatorial background using Run 1A photon data and have studied with Monte Carlo the discrimination against $B\rightarrow K^* \pi^0$ and $\rho \pi^0$ and from higher multiplicity penguin decays. These backgrounds are manageable. The mass resolution is not adequate to separate $\gamma \rho$ from $\gamma K^*$ on an event-by-event basis (see Figure \[fig-peng\]); however, a statistical separation is possible. In addition, the CTC dE/dx system should provide 1$\sigma$ $K$-$\pi$ separation in the momentum range of interest. These radiative $B$ decays can also be observed using converted photons. The probability for a photon to convert ($\sim 5\%$) will be offset by a lower photon $E_t$ threshold. Also, the mass resolution is $\sim 5$ times better than for the signals with unconverted photons, allowing a cleaner separation between $B\to\gamma K^*$ and $B\to\gamma\rho$. =3.0in At the Tevatron it is possible to study $B_s$ penguin decays as well. Information on  can be obtained in the same manner as above from studying the ratio of $B(B_s\rightarrow \gamma K^*)/B(B_s \rightarrow \gamma \phi)$. The size of the $B_s$ penguin sample is expected to be 1/2 to 1/3 the size of the $B_d$ sample. Comparison of the two results would help constrain the size of the long distance contributions to the decays. ### $\Delta \Gamma_s/\Gamma_s$ Browder [*et al.*]{} [@Browder] show that if $x_s=15$, a 7% difference in lifetime is expected. Several techniques can be used to determine $\Delta \Gamma_{B_s}$ [@Dunietz]. The statistical uncertainty on the $B_s$ lifetime from semileptonic $B$ decays in Run II will be well below $1\%$. With this constraint, the decay mode $B_s \rightarrow J/\psi \phi$ can be decomposed into its two CP components (via a helicity analysis) fitting a separate lifetime for each component (if this final state is a pure CP eigenstate, its lifetime can simply be compared to the average $B_s$ lifetime). Using Run Ia data, CDF has measured the helicity structure of the decays $B\rightarrow J/\psi K^*$ and $B\rightarrow J/\psi \phi$ [@randy]. In Run II, the $B_s \rightarrow J/\psi \phi$ helicity structure should be known to about 1%, and the lifetime difference should be determined to 2-3%. Rare $B$ decays --------------- In Run I, CDF has performed a search for the decay modes $B^\pm \to \mu^+\mu^- K^\pm, B^0 \to \mu^+\mu^- K^{*o}$ and $B_{d,s} \to \mu^+\mu^-$ [@Meschi]. Assuming the Standard Model Branching ratios [@Ali-ECFA] for $B^+ \to \mu^+\mu^- K^+$ and $B^0 \to \mu^+\mu^- K^{*o}$, we expect in Run II $\approx 400$ $B^+ \to \mu^+\mu^- K^+$ and $\approx 650$ $B^0 \to \mu^+\mu^- K^{*o}$ events. This will enable us to study both (a) the invariant mass distribution of the dimuon pair and (b) the forward-backward charge asymmetry in the decay. Both of these distributions are sensitive to physics beyond the Standard Model, e.g. the presence of a charged Higgs or charginos [@wyler], [@ali-model-indep]. We also expect to oberve the decays $B^\pm\to e^+ e^- K^\pm$ and $B^0 \to e^+ e^- K^{*0}$. The decays $B^\pm \to\ell^+\ell^+\pi^\pm$ and $B^0\to\ell^+\ell^-\rho^0$ are suppressed by an order of magnitude, but will be observable if we can achieve a high enough level of signal-to-noise. An observation of these decay modes would provide another determination of $|V_{td}/V_{ts}|$. CONCLUSIONS =========== The CDF Run I has provided much experience in doing $B$ physics at a hadron collider, including the reconstruction of $B \to J/\psi K_s$, and flavor tagging. This experience indicates that with an order of magnitude improvement in statistical sensitivity, we can obtain a competitive measurement of $\sin(2\beta)$ in Run II. We expect to obtain this factor from accelerator upgrades, which will provide an order of magnitude more integrated luminosity, and detector upgrades, which will improve trigger and flavor tagging efficiency. Other $B$ physics goals include the measurement of the CP asymmetry in $B\to \pi^+\pi^-$, the observation of $B_s$ mixing, and high statistics observations of certain rare decay modes. Many fundamental measurements involve the $B_s$, $B_c$, or $\Lambda_b$, and are thus complementary to $B$ physics programs at the $\Upsilon(4S)$. [99]{} E. Meschi, this workshop. M. Paulini, Proc. 6th Int. Symp. on Heavy Flavour Physics, Pisa, Italy (1995). see http://www-fermi3.fnal.gov/ The CDF Upgrade,\ CDF/DOC/CDF/PUBLIC/3171. Physics with CDF in Run II,\ CDF/DOC/CDF/PUBLIC/3172. R. Aleksan, B.Kayser and P.Sphicas, Proc. of the Workshop on $B$ physics at Hadron Accelerators, Snowmass, Colorado (1993). M. Gronau and J.L. Rosner, Phys. Rev. [**D49**]{}, 254 (1994). A. Ali, D. London, Z. Phys. [**C65**]{}, 431 (1995). M. Ciuchini [*et. al.*]{}, Z. Phys. [**C68**]{}, 239 (1995). Y. Nir, Proc. Workshop on $B$ physics at Hadron Accelerators, Snowmass, Colorado (1993). M.E. Lautenbacher, Proc. of the 27th Int. Conf. on High Energy Physics, Glasgow, Scotland (1994). J.D.Lewis, J.Muller, J.Spalding, P.Wilson, Proc. Workshop on $B$ physics at Hadron Accelerators, Snowmass, Colorado (1993). N.G. Deshpande and X.G. He, Phys. Rev. Lett. [**74**]{}, 26 (1995). F. DeJongh and P. Sphicas, FERMILAB-PUB-95/179-E. D. Buskulic [*et. al.*]{}, Phys. Lett. [**B356**]{}, 409 (1995). F. Abe [*et al*]{}., Phys. Rev. Lett. [**74**]{}, 4988 (1995). J. Skarha and A. B. Wicklund, Proc. of the Workshop on $B$ physics at Hadron Accelerators, Snowmass, Colorado (1993). F. Abe [*et al*]{}., Phys. Rev. Lett. [**75**]{}, 3068 (1995). The CLEO Collaboration, [*Phys. Rev.*]{} [**D49**]{} 5701, (1994). T.E. Browder, S. Pakvasa, UH 511-814-95. I. Dunietz, Phys. Rev. [**D52**]{}, 3048 (1995). A.Ali, C.Greub, T.Mannel, Proc. of the ECFA Workshop on the Physics of the European B Meson Factory (1993). A.Ali, T.Mannel and T.Morozumi, Phys. Lett. [**B273**]{}, 473 (1991);\ B.Grinstein, M.J.Savage and M.B.Wise, Nucl. Phys. [**B319**]{}, 271 (1989); W.Jaus and D.Wyler, Phys. Rev. [**D41**]{}, 3405 (1991). A.Ali, G.Giudice and T.Mannel, Proc. of the 27th Int. Conf. on High Energy Physics, Glasgow, Scotland (1994).

ArXiv

--- bibliography: - '../Article\_new\_rmf/RojasS\_submission\_ref.bib' --- [**Sergio Rojas**]{}\ [email protected]\ [Physics Department, Universidad Simón Bolívar, Ofic. 220, Apdo. 89000, Caracas 1080A, Venezuela.]{} [**Abstract**]{}. The amount of published research in Physics Education Research (PER) shows, on one hand, an increasing interest in the design and development of high performance physics teaching strategies, and, on the other hand, it tries to understand plausible ways on which the brain processes scientific information so that scientific thinking skills could be taught more effectively. As physics is a subject in which mathematical and conceptual reasoning can not be separated, instructors of physics face the problem of finding suitable advise on the most effective methods of teaching physics (i.e. how much time should be spent on intuitive conceptual reasoning and how much time in developing quantitative reasoning, and how to teach both in a unitary way). In spite the important efforts made by the PER community, the published results are overwhelming and confusing for the physics instructors in the sense that the conclusions that have arisen in those articles are in some instances controversial and far from being conclusive in pointing out a particular strategy to overcome the afore mentioned problem. Accordingly, based on the analysis of published PER work, we’ll be arguing that one of the major difficulties to overcome in the teaching of physics could be associated to the lack of a consistent and coherent methodological framework for teaching which integrates both aspects, conceptual and mathematical reasoning, in a systemic way of thinking. We will be presenting a set of plausible steps that could be applied to tackle the aforementioned difficulty. –More–(15[**Keywords**]{}: Physics Education Research; Students Performance. Introduction ============ It is not difficult to find instructors of general physics courses anxiously perusing articles published by the Physics Education Research (PER) community, searching essentially for advice on how to approach the teaching of physics effectively. Despite the demand for actually useful pedagogy of physics, PER has not produced so far any ultimately theory in this regards, and the great amount of published research on the subject, in addition of being controversial[@Lea:1994; @Hammer:1996; @Ehrlich:2002; @Meltzer:2002; @HoellwarthEtAll:2005; @Redish:2005; @Ates:2007; @Coletta:2008; @Ates:2008] might be overwhelming and even confusing to the physics instructors in the sense that being physics intrinsically a quantitative based subject, much of the recent research favors an overemphasis on qualitative (conceptual) physical aspects [@MualemEylon:2007; @HoellwarthEtAll:2005; @SabellaRedish:2007; @Walsh:2007], while standard mathematical abilities, which are crucial for understanding physical processes, are not stressed, or even taught, because, rephrasing a passage from a recent editorial[@Klein:2007], they interfere with the students’ emerging sense of physical insight. Consequently, physics instructors face the problem of finding suitable advice on how to approach the teaching of physics in the most efficient way and an answer to the question of how much time should be spent on intuitive conceptual reasoning and how much time in developing quantitative reasoning. Let’s mention, in passing, that the aforementioned editorial[@Klein:2007] has risen an interesting debate[@KleinDebate1:2007; @KleinDebate2:2007; @KleinDebate3:2007] regarding the benefits and shortcomings of science education reform in the United States in relation to its influence on the development of reasoning skills on the students, expressed in the way students use or apply the learned materials in their courses. The questions raised in this debate are of interest to be taken into consideration by the PER community and/or curriculum developers, particularly by those involved on science education reform in developing countries. On the other hand, physics instructors need also be mindful of the importance of selecting the most appropriately functional textbook, basically because innovative active-learning teaching schemes requires students to acquire basic and fundamental knowledge through reading textbooks. Correspondingly, innovative teaching strategies should be designed to help students in processing their ever thicker and heavier textbooks, which are laden with physical and mathematical insights[@YapWong:2007; @HoellwarthEtAll:2005; @CrouchMazur:2001; @James:2006; @CerbinKopp:2006]. Thus, the panorama regarding the learning of physics is even more dramatic on the side of the students. For one reason, in their struggle to fully participate in the process of learning, at the moment of trying to find suitable learning materials that could help them to go beyond classroom instruction (i.e. aiming to develop self-confidence on their own through exercising their role of active-learners), students face the dilemma of deciding which textbook could be helpful: perhaps a conceptual physics textbook (i.e. [@Hewitt:1993]), the student could wonder; or may be a calculus based physics textbook (i.e. [@HallidayResnickWalker:2000]); or why not an algebra based physics textbook (i.e. [@Giancoli:2004]); or what about a combination of all of them? could finally the student ponder scrabbling in his/her pocket/handbag to verify if the money in there could be enough to bring some extra weight at home (for a good account of the drama of choosing a textbook see for instance [@Dake:2007] and references there in). For another, in a typical course work for students majoring in science and/or engineering they usually must take more than one physics class. It could happen that in one term his/her physics instructor may emphasize quantitative reasoning over conceptual analysis, and in another term the respective instructor could rather accentuate conceptual learning over quantitative analysis, likely causing confusion for students, leading them to wonder which emphasis is correct. Finally, it is not difficult to find published results by the PER community on which it is shown directly or indirectly the inability of students to express, interpret, and manipulate physical results in mathematical terms. That is, students shows a clear deficiency in their training to exploit the mathematical solution of a problem (which sometimes could be obtained mechanically or by rote procedures) to enhance their knowledge regarding conceptual physics [@Hammer:1996; @Ehrlich:2002; @Loverude:2003; @SabellaRedish:2007; @RimoldiniSingh:2005; @Meredith:2008; @Rojas:2008]. More important, the analysis of published excerpts of student’s responses to interviews conducted by some researchers to further understand students’ way of reasoning while solving physics problems, shows that students lack of a structured methodology for solving physics problems [@Hammer:1996; @RimoldiniSingh:2005; @SabellaRedish:2007; @Walsh:2007]. These findings can not be surprising at all. In fact, none of the most commonly recommended physics textbooks (i.e. [@HallidayResnickWalker:2000; @TiplerMosca:2003; @SerwayJewett:2003]) make use of a consistently and clear problem-solving methodology when presenting the solution of the textbook worked out illustrative examples [@tipler:note1]. Moreover, the lack of a coherent problem-solving strategy can also be found in both the student and instructor manual solutions that usually accompany textbooks. Generally, standard textbooks problem-solving strategies encourage the use of a formula based scheme as compiled by the formulae summary found at the end of each chapter of the text, and this strategy seems to be spread out even in classroom teaching [@Hamed:2008]. Consequently, students merely imitates the way in which problems are handled in the textbooks, which is also perhaps the same way in which problems are solved by the instructor in class. For support of the previous assertion, one needs only browse the Internet for introductory physics courses and skim the solutions of problems posted by the course instructor. Furthermore, from the aforementioned students interviews excerpts one can also appreciate the lack of reasoning skills trying to associate or connect a way to solving a problem with the solution of other previously similar problem from another context (i.e. by using analogies). Again, the absence of this skill can not be surprising at all because students are just mirroring the unrelatedness way on which commonly used physics textbooks present the themes (i. e. the use of analogies are not fostered) [@Schneider:2004; @Donohoe:2008; @tipler:note2]. On the importance of a structured, systemic methodology to solve physics problems ================================================================================= To further motivate the subsequent discussion, let us summarize our introductory commentaries. We are essentially pointing out three major problems in the learning and teaching of physics: 1) the demand of the physics instructors for effective teaching strategies that explain how much time should be spent on teaching intuitive conceptual reasoning and how much time on developing students’ quantitative reasoning, and how to teach both aspects holistically 2) the students’ need for suitable textbooks that will help them develop mathematical abilities reasoning, which are essential for enhancing their knowledge of conceptual physics, and 3) a deficiency in the teaching of physics leading to students not being taught a coherent physics problem-solving strategy that enables them to engage in both mathematical and conceptual reasoning. A moment of though about the above summarized difficulties leads us to postulate the necessity of a systemic [@Bunge:2000; @Bunge:2004] approach which, from an operational point of view, could help instructors and students to achieve a better performance in the process of teaching and learning physics. In a broad sense, a systemic approach in the learning and teaching of physics could be represented as a framework involving the ordered triple composition-environment-structure together with a mechanism or modus operandi which integrates the teaching and learning process according to an approach allowing us to tackle the aforementioned difficulties from an efficient and unifying point of view (for more details refers to Bunge [@Bunge:2000; @Bunge:2004]). On the instructor side the need of a systemics approach in the teaching of physics could be justified by the advantage of using a methodology which would help them to incorporate both conceptual and mathematical reasoning systematically in their teaching. In this way, students will obtain the necessary training in their computational skills while learning how to use mathematical formulae to obtain the physics in the equations, even when they can obtain the mathematical solutions of a problem by rote procedures. In other words, students could apply “higher order thinking skills."[@Rigden:1987] via the mathematical understanding of a physics problem, which in turns involves meaningful learning which goes beyond the merely application of rote procedures. Moreover, using properly designed quantitative problems that require students to illustrate their conceptual learning and understanding will reveal much to instructors about their students’ learning and will provide invaluable feedback[@Reif:1981; @ReifScott:1999; @YapWong:2007; @DunnBarbanel:2000], and such problems can also be a powerful way to help students to understand the concepts of physics[@Rigden:1987; @DunnBarbanel:2000], a point emphasized by the Nobel prize-winning, great physicist Lev Davidovich Landau on the importance of first mastering the techniques of working in the field of interest because “fine points will come by itself." In Landau’s words, “You must start with mathematics which, you know, is the foundation of our science. \[...\] Bear in mind that by ’knowledge of mathematics’ we mean not just all kinds of theorems, but a practical ability to integrate and to solve in quadratures ordinary differential equations, etc."[@Lifshitz:1977] To further enhance their reasoning skills, the students would have the opportunity to increase their intuitive conceptual skills in the physics laboratory, where conceptual learning is reinforced by experience[@AufschnaiterAufschnaiter:2007; @Hanif:2009]. On the student side, the need of a systemics approach in the learning of physics could be justified by the usefulness of applying a working methodology which could help them to approach the learning of physics from a interrelate point of view. That is, that his/her knowledge of mathematics is useful to master ideas from physics, and that the use of analogies are important to approach the solution of physical, mathematical and engineering problems. In short, this kind of practical, unified problem-solving strategy will help students to pose and approach any kind of problem, and will teach students that physics is the primary subject to start developing these kind of analytical skills. In other words, with such an approach, students would internalize that it is in physics classes where they can start to apply what they have learned in their math classes and to find new non-formal approaches to performing computations[@Yeatts:1992]. One could resort to the anecdote of the cathedral building[@anecdote:2007; @anecdote:note] to further illustrate the need of keeping in mind the interconnectedness of reasoning skill, mathematics and physics. To paraphrase Heron and Meltzer, learning to approach problems in a systematic way starts from teaching and learning the interrelationships among conceptual knowledge, mathematical skills and logical reasoning[@HeronMeltzer:2005]. An example of this assertion is illustrated by an explanation of what happened to the Millennium Bridge disaster, . which stated “Existing theories of what happened on the bridge’s opening day focus on the wobbling of the bridge but have not addressed the crowd-synchronization dynamics. In our approach, wobbling and synchrony are inseparable. They emerge together, as dual aspects of a single instability mechanism, once the crowd reaches a critical size.”[@StrogatzEtAl:2005] The mistakes made in the construction of the Millennium Bridge help us to understand the need for teaching and learning based on a systemic approach which recognizes the interrelatedness of every aspect of the physical process (physics, mathematics, and engineering design). A systemic structured methodology to solve physics problems =========================================================== Earlier work on the importance and necessity of a problem-solving strategy can be found in the work of the great mathematician George Pólya [@Polya:1945; @Polya:1963], who made emphasis on the relevance of the systematicity of a problem-solving strategy for productive thinking, discovery and invention. Some of his views, either provocative or encouraging, about teaching and learning can be found spread out in some PER publications, like for instance that [*teaching is not a science*]{} (i. e. [@Hammer:1996]); on [*the aim of teaching*]{} (i. e. [@Rigden:1987; @Redish:2005]); that [*teaching is an art*]{} (i. e. [@Milner:2007]); on the importance of [*problem-solving skills*]{} (i. e. [@Reif:1981; @ReifScott:1999; @Ehrlich:2002; @Heller:1992a]), etc. For a further detailed account of Pólya’s work let’s refers to [@Polya:1945; @Polya:1963; @Schoenfeld:1985; @Schoenfeld:1987; @Schoenfeld:1992; @Lederman:2009]. In [*How to Solve It*]{}, Pólya set four general steps to be followed as a problem-solving strategy: 1. \[P1a\] [**Understanding the problem**]{}: some considerations to develop at this step involves drawing a figure and asking questions like What is the unknown? What is the condition? Is it possible to satisfy the condition? Is the condition sufficient to determine the unknown? Or is it insufficient? Or redundant? Or contradictory? Draw a figure. Introduce suitable notation. 2. \[P2b\] [**Devising a plan**]{}: some considerations to have in mind in order to develop this step involves looking at the unknown and trying to think of a familiar problem having the same or a similar unknown. Some questions to be ask are like Have you seen this before? Or have you seen the same problem in a slightly different form? 3. \[P3c\] [**Carrying out the plan**]{}: Be sure to check each step and make sure that the steps are correct. 4. \[P4d\] [**Looking back**]{}: some considerations to develop at this step involves asking questions like Can the results be checked? Can the results be derived differently? Can the result or the method be applied to solve or fully understand other problems? Surprisingly, these steps encompass “the mental processes and unconscious questions experts explore as they themselves approach problem solving.”[@Lederman:2009] These four step are also the base of some computational models devised to “model and explore scientific discovery processes.”[@Lederman:2009] Now, even though the aforementioned four steps seems very simply, their generality seems to be hard for novices to follow them. Thus, in order to have a problem-solving strategy more affordable to students we extended the four steps problem-solving strategy into a six steps strategy. We made our choice based on empirical observations after experimenting with a five step strategy reported in [@Heller:1992a]. Justification for a more detailed problem-solving strategy can be found in the words of Schoenfeld “First, the strategies are more complex than their simple descriptions would seem to indicate. If we want students to use them, we must describe them in detail and teach them with the same seriousness that we would teach any other mathematics.”[@Schoenfeld:1980] Nevertheless, we will further rationalize below the need for explicitly including a new step in our proposed methodology (see item \[M5\]). Thus, our six steps proposed strategy is as follows: 1. \[M1\] [**Understand the problem**]{}: in this stage students needs to actually be sure to what the problem is. In addition to making drawings to get a grasp of the problem formulation, eventually they might need to reformulate the problem in their own words, making that that they are obtaining all the giving information for solving the problem. This is a crucial step in the sense that [*if one does not know where are we going, any route will take us there*]{}. 2. \[M2\] [**Provide a qualitative description of the problem**]{}: in this stage students needs to think and write down the laws, principles or possible formulations that could help them to solve the problem. For instance students needs to consider any possible frame work of analysis that could help them to represent or describe the problem in terms of the principles of physics (i.e. Newtons law, energy conservation, momentum conservation, theorem of parallel axis for computing inertia moment, non-inertial reference system, etc.) If necessary, the drawings of the previous step could be complemented by the corresponding free-body and/or vector diagram. 3. \[M3\] [**Plan a solution**]{}: once the student have as many possibilities to approach the problem, he/she only needs to pick one strategy of solution and write down the corresponding mathematical formulation of the problem, avoiding as much as possible to plug numbers in the respective equations. Also, they need to think whether the problem at hand is similar to a previously solved one, and find out whether the information at hand would be enough to get a solution (i.e. if a set of algebraic equations is under or over determined or the number of provided boundary conditions are enough to solve a differential equation). Eventually, one might need to go back to the previous step in order to get the a well posed problem, perhaps by choosing another strategy. 4. \[M4\] [**Carrying out the plan**]{}: at this stage the student will try to find a solution to the mathematical formulation of the problem and perhaps they will need to go back in order to find a easier mathematical formulations of the problem. This is facilitated is the students have writing down alternatives of solution as they were suppose to do on item \[M2\]. 5. \[M5\] [**Verify the internal consistency and coherence of the used equations**]{}: at the moment of finding a solution of the involved mathematical equations, students need to verify whether the equations are consistent with what they represent and that the equations are dimensionally correct. Though this seems to be an unnecessary step, experience shows that the students too often does not verify the consistency and coherence of the equations they solve. And this mistake is also found to be performed by textbook writers, as discussed in a recent editorial [@Bohren:2009]. After verifying no mistakes or inconsistencies are found in the mathematical solution of the problem, students could then plug numbers in the obtained results to find, whether required or not, a numerical solution which in turn could be used in the next step to further evaluate the obtained result. In the provided illustrative example will show how a right answer could be obtained, though the internal consistency of a used equation is not right[@note:2eq1]. 6. \[M6\] [**Check and evaluate the obtained solution**]{}: once a solution have been obtained, its plausibility needs to be evaluated. Some questions could be asked in this regards can the results be derived differently? Can the solution be used to write down the solution of a less general problem? Can the solution be used to further understand the qualitative behavior of the problem? Is it possible to have a division by zero by changing a given parameter? Does it makes sense?, and so on. A first comment on our six steps problem-solving strategy is that it provides a unified, systemic way of approaching the solution of a physical problem encompassing both qualitative (steps \[M1\]-\[M3\]) and quantitative (steps \[M4\]-\[M6\]) reasoning, and instructors could make as much emphasis as they prefer on any of the set of steps, providing the students with a structured recipe on how to approach in detail the other side of the problem’s solution. Second, comparing our six steps problem-solving strategy with Pólya’s four steps ones, one could see that we have explicitly divided Pólya’s step one (P1) into two steps (\[M1\]-\[M2\]), and Pólya’s step three (P3) into two steps (\[M4\]-\[M5\]). A further comment on this problem-solving strategy is that we prefer to call the the second step (\[M2\]) [*Provide a qualitative description of the problem*]{} rather than [*Physics description*]{} as in [@Heller:1992a], because one share the idea that students tend to think that by providing a qualitative analysis of a problem they are also providing the solution required by a physicist, and that the mathematical solution of the problem is just uninteresting mathematics. Instead, we make emphasis in that a physical solution of a problem is a combination of both qualitative and quantitative reasoning. As stressed by the great physicist Lord Kelvin “I often say that when you can measure something and express it in numbers, you know something about it. When you can not measure it, when you can not express it in numbers, your knowledge is of a meager and unsatisfactory kind. It may be the beginning of knowledge, but you have scarcely in your thoughts advanced to the state of science, whatever it may be." [@Hewitt:1993] Freeman Dyson was more eloquent “...mathematics is not just a tool by means of which phenomena can be calculated; it is the main source of concepts and principles by means of which new theories can be created."[@Dyson:1964] Illustrative Example ==================== In the following example we will present an approach on how to introduce students in the use of our proposed six steps problem-solving strategy. Though each one of the steps has its importance, we will provide further evidence of why step five needs to be taught explicitly. It is pertinent to point out that, paraphrasing Pólya’s words, by proper training students could absorb the steps of our problem-solving strategy in such a way that they could perform the corresponding operations mentally, naturally, and vigorously. The dynamics of teaching is left to the instructor. In this article we are not pretending to show how the teaching should be carried out. Innovative teaching strategies can be found elsewere[@YapWong:2007; @HoellwarthEtAll:2005; @CrouchMazur:2001; @James:2006; @CerbinKopp:2006; @Heller:1992a; @Heller:1992b].  The problem statement --------------------- [**Problem**]{}: About its central axis, find the moment of inertia of a thin hollow right circular cone with radius $R$, lateral length $L$, and mass $M$ uniformly distributed on its surface with density $\sigma$. 1. [**Understand the problem**]{}: “It is foolish to answer a question that you do not understand. $\cdots$ But he should not only understand it, he should also desire its solution.”[@Polya:1945] Following Pólya’s commentary, before attempting to solve this problem, students need to have been exposed to a basic theory on computing moment of inertia ($I$). Particularly, students need to be familiar with the computation of $I$ for a thin circular ring about its main symmetric axis. To further understands the geometry of the present problem, students could, for example, be talked about the shape of an empty ice cream cone. After some discussion, a drawing better than the one shown in figure \[conicalfig\] could be presented on the board. Let’s mention that additional ways of presenting each step in meaningful ways can be found in [@Polya:1945; @Heller:1992a]. 2. [**Provide a qualitative description of the problem**]{}: In this step one could further motivate the discussion by associating the computation of $I$ with rotational motion quantities (i.e. kinetic energy, angular momentum, torque, etc.). One can even motivate the qualitative discussion by considering the hollow cone as a first crude approximation of a symmetric top or of a cone concrete mixer. The drawing of figure \[conicalfig\] could even be made more explicative. 3. [**Plan a solution**]{}: “We have a plan when we know, or know at least in outline, which calculations, computations, or constructions we have to perform in order to obtain the unknown. $\cdots$ We know, of course, that it is hard to have a good idea if we have little knowledge of the subject, and impossible to have it if we have no knowledge. $\cdots$ Mere remembering is not enough for a good idea, but we can not have any good idea without recollecting some pertinent facts.”[@Polya:1945] Accordingly, at this stage instructors could point out the superposition principle to solve the problem by slicing the hollow cone in a set of small, infinitesimal, rings distributed along the symmetrical axis of the cone. Thus, each infinitesimal ring will have in common the same rotational axis about which the moment of inertia of them is already known $dI = r^2 dm = r^2 \sigma dS$, where $r$ is the radius of each ring, while $dS$ represents the respective infinitesimal surface of each ring. 4. [**Carrying out the plan**]{}: To carried out the plan, it won’t be a surprise to choose the wrong $dS$. In fact, it is not difficult, at first sight, to choose wrongly (see figure \[conicalfig\]): $dS = 2\pi r dz = 2\pi (R/H) z dz$, which lets to $S=\pi R H$, as the hollow cone surface (this result is of course wrong). Using this surface element, the momemnt of inertia for the small ring takes the form $dI = 2\pi \sigma (H/R) r^3 dr$, which lets to $I = 2\pi \sigma (H/R) (R^4/4) = (1/2) (\sigma S) R^2 = M R^2/ 2$, as the required moment of inertia of the hollow cone (which is the right answer). It is not difficult to get students performing this sort of computations and they become uneasy when trying to convince them that in spite of having found a right result, it is specious because it was obtained via a wrong choice for $dS$. Eventually students might agree on the incorrectness of their procedure if asked to compute explicitly the cone’s mass. 5. [**Verify the internal consistency and coherence of the used equations**]{}: “Check each step. Can you see clearly that the step is correct? Can you prove that it is correct? $\cdots$ Many mistakes can be avoided if, carrying out his plan, the student check each step.”[@Polya:1945] Steeping on our teaching experience, it is too easy for students to perform without hesitation the just aforementioned wrong computations, as presented in the previous step. And it is not easy to get students to realize their mistake. For god shake, they have computed the right answer!!!: for a hollow thin cone, rotating about its symmetric axis, $I=M R^2/2$ !!!. In this situation, to make aware students of their mistake, the easy way is the experiment. Instructors could unfold several hollow cones to actually show the students that the respective surface is $S=\pi R L$, instead of the wrongly obtained $S= \pi R H$. Accordingly, we hope to have provided enough evidence for the need to, explicitly and repeatedly, mention to students on the need to check each computational step, including checking for dimensionality correctness. In this case, the right approach is to consider $dS = 2 \pi r dl = 2 \pi l (R/L) dl$, which yield $S= \pi R L$, the right answer for $S$. This choice for $dS$ lets to $dI = 2 \pi \sigma (L/R) r^3 dr$, which yields $I = 2 \pi \sigma (L/R) (R^4/4) = (1/2)(\sigma S) R^2 = M R^2/2$, the right answer. Considering that it is not hard to find stories on reported wrong results due to wrong or incomplete computations [@StrogatzEtAl:2005; @Veysey:883], this problem could also be used as an example of how computations of a physical quantity (the surface of a cone shell) can be used to judge a mathematical result (the wrong value for $S$) that is used in additional computations yielding a right answer.. 6. [**Check and evaluate the obtained solution**]{}: “Some of the best effects may be lost if the student fails to reexamine and to reconsider the completed solution.”[@Polya:1945] After gaining confidence on the obtained solution of the problem, it is necessary to spend sometime in evaluating its plausibility. Examining the solution of our problem one could ask: it is not striking that the rotational inertia for a hollow cone about its symmetric axis is the same as for a solid disk having the same uniformly distributed mass $M$ and radius equal to the cone’s base? Does not it a counter example for the statement that rotational inertia only depends on how the mass is distributed around the axis of rotation? Furthermore, if for some reason the wrong choice for the $dS$ was not caught in the previous step, it could be detected if analyzing the case of having a non constant $\sigma$. A further interpretation of the result can be found at [@Bolam:1961]. Concluding remarks ================== This article presents a six step problem-solving strategy, aiming to approach three major problems in the learning and teaching of physics: 1) the demand of the physics instructors for effective teaching strategies that could help in the teaching of intuitive conceptual and quantitative reasoning, and how to teach both aspects holistically 2) the students’ need for suitable methodology that could help students to fill the textbooks’ gap on enhancing their mathematical reasoning abilities, which are essential for reinforcing students’ knowledge of conceptual physics, and 3) a deficiency in the teaching of physics leading to students not being taught a coherent physics problem-solving strategy that enables them to engage in both mathematical and conceptual reasoning. Let’s finish by recalling a particular point of view which the great mathematician Pólya stressed very much in his writings about the art of teaching and learning, which, in some sense, can be considered as an “axiomatic thought” about the art of teaching an learning. He was emphatic on the fact that “[**[*for efficient learning, the learner should be interested in the material to be learnt and find pleasure in the activity of learning*]{}**]{}.” In order to reinforce the content of this expression one might recall the story of the Cathedral’s construction workers [@anecdote:note]. In other words, inspiration to learn is without doubt a necessary condition in order to have an efficient and effective teaching and learning environment. This, of course, is by no means a new discovery, and, paraphrasing Schoenfeld [@Schoenfeld:1977], some ideas to circumventing few of the barriers between the dedicated instructor and his/her students’ attitudes in “learning” the subject that is being taught has been set forward in [@Schoenfeld:1977; @Ehrlich:2007; @Duda:2008]. Nevertheless, one should keep in mind that “we know from painful experience that a perfectly unambiguous and correct exposition can be far from satisfactory and may appear uninspiring, tiresome or disappointing, even if the subject-matter presented is interesting in itself. The most conspicuous blemish of an otherwise acceptable presentation is the ’deus ex machina’.”[@Polya:1949] Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to Dr. Cheryl Pahaham, who kindly provided useful comments on improving this article. \[LastPage\]

ArXiv

--- abstract: '[*We calculate the susceptibility amplitude ratio near a generic higher character Lifshitz point up to one-loop order. We employ a renormalization group treatment with $L$ independent scaling transformations associated to the various inequivalent subspaces in the anisotropic case in order to compute the ratio above and below the critical temperature and demonstrate its universality. Furthermore, the isotropic results with only one type of competition axes have also been shown to be universal. We describe how the simpler situations of $m$-axial Lifshitz points as well as ordinary (noncompeting) systems can be retrieved from the present framework.*]{}' author: - 'C. F. Farias[^1]' - 'Marcelo M. Leite[^2]' title: '**Susceptibility amplitude ratio for generic competing systems**' --- Introduction ============ The occurrence of $m$-axial Lifshitz phase transitions [@Ho-Lu-Sh; @Ho] in various real physical systems (e.g., magnetic modulated materials [@Be; @Yo1; @Yo2; @Be1; @Hana; @We], high-$T_{c}$ superconductors [@H; @K; @S], liquid crystals [@Ra; @Za; @Ska], etc.) has increased the interest in the field-theoretic description of this subject in the last few years [@MC; @AL; @DS]. According to modern renormalization group arguments, critical phenomena of $m$-axial Lifshitz competing systems have their universality classes characterized by $(N,d,m)$, namely, the number of components of the (field) order parameter $N$ dwelling in $d$ space dimensions with $m(\equiv m_{2})$ space directions presenting alternate (repulsion-attraction) among the fields [@L; @Leite1; @CL1]. Another type of “competing axes” can be defined: if the alternate couplings are of the type attractive-repulsive-attractive and take place along $m_{3}$ spatial dimensions, one refers to $m_{3}$-axial third-character Lifshitz critical behavior (for the realization of the $m_{3}=1$ case, see [@Se1]). More generally, a $m_{L}$-fold $L$-th character Lifshitz behavior appears whenever short-ranged alternate interactions with L couplings of the type repulsion-attraction-repulsion-attraction-... are allowed [@Se2; @Ni1; @Ni2]. If all sorts of the aforementioned competing axes are present simultaneously in the critical system under consideration, its phase transitions are governed by the generic higher character Lifshitz critical behavior. The universality classes of these arbitrary competing systems are defined by the set $(N,d,m_{2},...,m_{L})$ [@Leite2; @LeiteII; @CL2]. The language of magnetic systems is particularly suitable to describe these systems. It is convenient to make the connection of these complex critical behaviors with the prototype of second order phase transitions in noncompeting systems: the Ising model ($N=1$) [@Amit]. The simplest realization of the usual Lifshitz universality class $(N,d,m)$ can be encountered in uniaxial critical (systems $m=1$) behavior. It can be understood in terms of the axial next-nearest-neighbor Ising (ANNNI) model [@Selke1; @Selke2] which corresponds to the usual Ising model including antiferromagnetic exchange interactions among second neighbor spins along a single axis in a cubic lattice. The uniaxial Lifshitz point arises at the confluence of the disordered, a uniformly ordered and a modulated phase. Anisotropic second character $m$-axial points generalize that uniaxial when the competing axes occur along $m$ space directions whenever $m \neq d$. The ANNNI model can be generalized by including further ferromagnetic couplings among third neighbors along the competing axis. Except for a little additional complication in the phase diagrams due to the existence of an additional parameter related with the third neighbor coupling, similar phases can be defined such that the region of intersection of them terminates in a point where the different phases characterizing the system meet, known as the uniaxial Lifshitz point of third character [@Se1]. This reasoning can be extended to contemplate the situation where alternate couplings up to the $L$-th neighbors exist, and the critical point associated to the region of confluence of the several phases of the system is denominated the uniaxial Lifshitz point of $L$-th character. If the system presents competing interactions of this type along $m_{L}\neq d \;(=d)$ space directions, the system is said to have anisotropic (isotropic) critical behavior with $m_{L}$-axial point of $L$-th character [@Se2; @Ni1; @Ni2]. The modulated phases have a distinction when we compare anisotropic and isotropic behaviors. In the former there exist two types of correlation lengths, namely $\xi_{1}$ and $\xi_{L}$, which label two inequivalent subspaces characterized by correlations perpendicular and parallel, respectively, to the $m_{L}$ subspace. In isotropic behaviors only one type of correlation length $\xi_{L}$ characterizes each modulated phase. The anisotropic Lifshitz point of generic $L$th character can also be defined in the description of the most general $d$-dimensional competing system, whenever several types of competing axes show up [*simultaneously*]{}. Let us consider its simplest realization. In that case, there are only nearest neighbor interactions along $m_{1} \equiv (d-m_{2} -...-m_{L})$ (noncompeting) directions, second neighbor competing interactions along $m_{2}$ directions perpendicular to the $m_{1}$ dimensions, competition among third neighbors along $m_{3}$ space directions (orthogonal to the $(m_{1},m_{2})$ subspaces), etc., up to $L$th neighbor alternate couplings along $m_{L}$ directions, with all competition subspaces perpendicular to each other. The model which describes this sort of arbitrary competing systems was proposed a few years ago and named [*competing exchange coupling Ising*]{} ($CECI$) model [@Leite2]. There are $L$ inequivalent correlation lengths owing to the $L$ independent competing axes $m_{n}$ ($n=1,..,L$). This situation allows in principle low temperature $(L-1)$ modulated phases in equilibrium with the uniform ordered phase as well as with the high temperature disordered phase close to the Lifshitz point. In addition, it is also possible that the complex systems can display several low temperature (up to $L-1$) uniformly ordered phases in equilibrium with (at least one) modulated ordered and high temperature disordered phases. The isotropic $m$-axial critical behavior has been experimentally realized in the context of polymers. At first the isotropic behavior was thought of being of purely academic interest. Nevertheless, its theoretical mean-field prediction in copolymer-homopolymer ternary blends [@BF] and subsequent experimental identification in mixtures of block copolymer-homopolymer [@Bates1] caused a certain enthusiasm, but the following paper with a more detailed analysis on the subject showed a microemulsion phase incompatible with the existence of the Lifshitz point [@Bates2]. It was argued there that the fluctuations destroyed this multicritical point, although the associated critical region could be identified with the vicinity of the would be mean field Lifshitz point. The theoretical effect of fluctuations was incorporated immediately afterward using a self-consistent field theory (SCFT), and the ordered lamellar phase previously identified as the modulated phase was understood utilizing a one-component order parameter ($N=1$) [@KM]. This result not only confirmed the previous discussion from Ref.[@Bates2], but also located loci in the mean field phase diagram with third character isotropic Lifshitz point. Later, mean field studies using SCFT indicated the existence of up to $6$th character isotropic critical behavior in blends of diblock copolymers [@OH]. This suggests that the isotropic behaviors of the $CECI$ model might be useful in unveiling properties of these real physical systems, even though the anisotropic realization of this model has not been identified yet. Thence, several experiments have been performed for these polymers. For instance, the susceptibility of a homopolymer-diblock copolymer blend (polybutadiene and polystyrene) has been investigated recently using small angle neutron scattering. Some amplitudes above the Lifshitz temperature were estimated for fixed values of the diblock copolymer composition [@Pipi]. The closest we can get to this system using field theory techniques is to look for universal quantities related to the susceptibility, i.e., the amplitude ratio above and below the Lifshitz critical temperature [@Leite3]. By the same token, the study of a simple property such as the susceptibility for the $CECI$ model could shed light on possible future experiments related with real physical systems manifesting this especial critical behavior. In this paper, the susceptibility amplitude ratio for generic competing systems will be computed using field theory and renormalization group arguments up to one-loop order in a perturbation expansion. The anisotropic behaviors with arbitrary types of competing axes are discussed first. We are going to restrict ourselves to (fields) order parameters of only one component ($N=1$). The results are presented in a manifestly universal form and are shown how to reduce to the ordinary amplitude ratio without competition. We then restrict the number of competition axes to obtaining information on particular universality classes, the most obvious being the $m$-axial anisotropic Lifshitz criticalities. We show that the uniaxial result can be retrieved from the arbitrary anisotropic competing systems in a simple manner. The isotropic amplitude ratios for isotropic critical behaviors are calculated for the first time. The $m$-axial universality class is recognized from the generic situation whenever $n=2$. We show that although the expansion parameter is large for three-dimensional systems, our perturbative results are meaningful for those systems. As an application, we compare our field-theoretic $m=3$ isotropic output with experimental results from homopolymer-diblock copolymer mixtures and show very good agreement among them. The paper is organized as follows. In Section II we present the one-loop effective potential. We highlight a brief explanation of the several subspaces which occur in the problem along with the independent renormalization group transformations in the anisotropic cases. A simpler analogous discussion for calculating the isotropic amplitude ratio is explicited in Section III. Section IV presents the discussion of the results and conclusions. Anisotropic amplitude ratio for generic competing systems ========================================================= We begin with the bare Lagrangian density associated with anisotropic generic competing systems described by the $CECI$ model, which is given by $$\begin{aligned} \label{1} L &=& \frac{1}{2} |\bigtriangledown_{(d- \overset{L}{\underset{n=2}{\sum}} m_{n})} \phi_0\,|^{2} + \sum_{n=2}^{L} \frac{\sigma_{n}}{2} |\bigtriangledown_{m_{n}}^{n} \phi_0\,|^{2} \\ \nonumber && + \sum_{n=2}^{L} \delta_{0n} \frac{1}{2} |\bigtriangledown_{m_{n}} \phi_0\,|^{2} + \sum_{n=3}^{L-1} \sum_{n'=2}^{n-1}\frac{1}{2} \tau_{nn'} |\bigtriangledown_{m_{n}}^{n'} \phi_0\,|^{2} \\ \nonumber &&+ \frac{1}{2} t_{0}\phi_0^{2} + \frac{1}{4!}\lambda_0\phi_0^{4} .\end{aligned}$$ The parameters which correspond to the physical situations are the coefficients of the derivatives of the bare field $\phi_0$ (order parameter of the phase transition), the bare reduced temperature $t_{0} (\propto T-T_{L}$, where $T_{L}$ is the Lifshitz critical temperature) and the bare coupling constant $\lambda_0$. The Lifshitz critical region is defined for particular combinations of the exchange interactions among all the neighbors. This implies the fine-tuning conditions on some parameters, namely, $\delta_{0n} = \tau_{n n'} =0$. At the Lifshitz critical region, the temperature is close but not equal to $T_{L}$. The structure of the field theory considered at this region is such that its momentum dependence on the various competing subspaces is rather peculiar. There are quadratic momenta components along the $m_{1}$-dimensional noncompeting subspace, quartic momenta components along the $m_{2}$-dimensional competing subspace, and so on, up to the $2L$-th power of momenta along the $m_{L}$-dimensional subspace, which are present simultaneously in the free bare critical propagator in momentum space. We can set $\sigma_{n}=1$ provided we perform a dimensional redefinition in the momentum characterizing the $n$-th $m_{n}$-dimensional competition subspace. If $\Lambda$ is a momentum scale, we take the engineering dimension of the competing subspace as $[k_{(n)}]=\Lambda^{\frac{1}{n}}$. The anisotropic behaviors is characterized by $L$ independent correlation lengths $\xi_{n}$, one for each subspace. They induce $L$ independent renormalization group flows in the parameter space of the massless theory. If we use normalization conditions in the definition of the renormalized theory, these flows can be described by $L$ independent sets of normalization conditions, each of them defining a symmetry point $SP_{n}$ ($n=1,..., L$) which simplifies our task of computing universal quantities in this formalism of one-particle irreducible ($1PI$) vertex parts. Typically, The Feynman integrals involved depend on various external momenta scales, namely that characterizing the $(d-m_{2}-...-m_{L})$-dimensional noncompeting subspace, a momentum scale associated to the $m_{2}$ space directions, etc., up to the momentum scale corresponding to the $m_{L}$ competing axes. For example, an explicit integral that shall be used is the one-loop contribution to the four-point function, namely $$\label{2} I(P,K'_{(2)},...,K'_{(L)}) = \int \frac{d^{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})}q \overset{L}{\underset{n=2}{\prod}} d^{m_{n}}k_{(n)}} {[\overset{L}{\underset{n=2}{\sum}}\bigl((k_{(n)} + K_{(n)}^{'})^{2}\bigr)^{n} + (q + P)^{2}] \left(\overset{L}{\underset{n=2}{\sum}}(k_{(n)}^{2})^{n} + q^{2} \right)}\;\;\;.$$ Although this integral should be computed in arbitrary nonvanishing external momenta components, in practice the calculation is simplified when we choose only one subspace, say $m_{n}$, whose momenta are set in the arbitrary value $K^{' 2n}_{(n)}= \kappa_{n}^{2n}$. In case we wish to determine universal quantities associated to vertex parts along the $j$th type of competing axes, we set $\kappa_{n}=0$ for $n \neq j$ maintaining, however, $\kappa_{j} \neq 0$. Unfortunately, the integral cannot be solved exactly for arbitrary external momenta, but can be resolved using the orthogonal approximation which permits to obtain the integral as an homogeneous function of the external momenta. Within this framework, the Feynman integrals of the corresponding field theory can be computed to all loop orders. In normalization conditions, the results for those integrals are independent of the subspace chosen. In order to be precise in our description, we should label the renormalized vertex parts according to the subspace characterized by the nonvanishing momenta scale $\kappa_{n}$ associated to the symmetry point $SP_{n}$. Fortunately, we do not have to employ this label in the present work, since in the context of the orthogonal approximation all renormalization directions possess the same fixed point [@Leite2]. In other words, the susceptibility amplitude ratio is independent of the renormalization group transformation characterized by the variation of the external momenta scale $\kappa_{n}$ in the renormalized vertex parts. In what follows, we shall use for simplicity all vertex parts computed at the symmetry point $SP_{1}\equiv SP$, i.e., the external momenta scale is given by $\kappa_{n}=\delta_{1n}\kappa$ with $P^{2}=\kappa^{2}=1$. With this choice we do not need to employ the orthogonal approximation to perform this integral in the anisotropic case, since the nonvanishing external momenta is contained in the quadratic term of the propagator. This quadratic part can be evaluated to arbitrary external momenta and the resulting expression for this integral is exact. For further details, the reader is advised to consult Ref. [@LeiteII]. The bare quantities can be transformed into renormalized amounts at one-loop level through the renormalization of the bare field and temperature which are given by $t_{0} = Z_{\phi^{2}}^{-1}t, \phi = Z_{\phi}^{-\frac{1}{2}}M$ beside the renormalized coupling constant. We express the latter in the fixed point in terms of the dimensionless entity $u^{*}$ as $g^{*}=u^{*} \kappa^{\epsilon_{L}}$ where $\epsilon_{L} = 4 + \overset{L}{\underset{n=2}{\sum}}\frac{(n-1)}{n} m_{n} - d$ is the perturbation parameter. Using the symmetry point, let us write down the one-loop renormalized Helmholtz free energy density at the fixed point. It is simply the renormalized effective potential at one-loop plus polynomial terms in $t$ used to define additively renormalized vertex parts. For completeness (and anticipating future discussions for the specific heat amplitude ratio as well) we include a term proportional to $t^{2}$ which, however, will have no consequence to our discussion in the present work. Putting those arguments together, we obtain the following expression $$\begin{aligned} \label{3} F(t,M)\;\; =&& \frac {1}{2}t M^{2} + \frac{1}{4!} g^{*} M^{4} + \frac{1}{4}(t^{2} + g^{*} t M^{2} + \frac{1}{4} (g^{*} M^{2})^{2}) I_{SP} \nonumber\\ && + \frac{1}{2} \int d^{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})}q \Bigl[\overset{L}{\underset{n=2}{\prod}} d^{m_{n}}k_{(n)}\Bigr] \Bigl[ln\Bigl(1 + \frac{t+ \frac{1}{2} g^{*} M^{2}}{\left(\overset{L}{\underset{n=2}{\sum}} (k_{(n)}^{2})^{n} + q^{2} \right)}\Bigr) \nonumber\\ &&- \frac{g^{*} M^{2}}{2\left(\overset{L}{\underset{n=2}{\sum}} (k_{(n)}^{2})^{n} + q^{2} \right)}\Bigr]\;\;,\end{aligned}$$ where in the above equation $t, M$ ($t_{0} = Z_{\phi^{2}}^{-1}t, \phi = Z_{\phi}^{-\frac{1}{2}}M$) are the renormalized (bare) reduced temperature and order parameter, respectively, $Z_{\phi^{2}}, Z_{\phi}$ are normalization functions, $g^{*}$ is the renormalized coupling constant at the fixed point, $\vec{q}$ is a $(d-m)$-dimensional wave vector perpendicular to the competing axes, whereas $\vec{k}$ is a $m$-dimensional wave vector whose components are parallel to the competition axes. The integral $I_{SP}$ is defined by: $$\label{4} I_{SP} = \int \frac{d^{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})}q \overset{L}{\underset{n=2}{\prod}} d^{m_{n}}k_{(n)}} {[\overset{L}{\underset{n=2}{\sum}} k_{(n)}^{2n} + (q + P)^{2}] \left(\overset{L}{\underset{n=2}{\sum}} k_{(n)}^{2n} + q^{2} \right)}\;\;\;,$$ where the convenient symmetry point for this integral is defined as above, namely, $P^{2}= \kappa^{2}=1$. This choice has the virtue of transforming the dimensionful coupling constant in its dimensionless version, i. e., $g^{*} = u^{*}$ and is the most effective route to computing universal quantities in the context of the renormalization group strategy. Whenever a loop integral is performed, a typical geometric angular factor is produced, which can be factored out in a redefinition of the coupling constant in a standard way [@Amit]. In our case this factor is given by the expression $[S_{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})} \Gamma(2 - \overset{L}{\underset{n=2}{\sum}}\frac{m_{n}}{2n})(\overset{L}{\underset{n=2}{\prod}} \frac{S_{m_{n}} \Gamma(\frac{m_{n}}{2n})}{2n})]$, such that it is going to be omitted whenever we report the result of any loop integral. The last integral at the symmetry point was already computed in Ref.[@LeiteII] and shown to be given by $I_{SP} = \frac{1}{\epsilon_{L}}(1 + h_{m_{L}}\epsilon_{L})$, where $h_{m_{L}}= 1 + \frac{(\psi(1) - \psi(2- \overset{L}{\underset{n=2}{\sum}}\frac{m_{n}}{2n}))} {2}$ and $\psi(z) = \frac{dln\Gamma(z)}{dz}$. It is worthy to stress that whenever $m_{3} = ...=m_{L}=0$, $h_{m_{2}}=[i_{2}]_{m}$ and the usual anisotropic $m$-axial Lifshitz critical behavior is obtained from this more general competing situation in a rather simple manner. Since we need the value of $M$ in the coexistence curve above and below $T_{L}$, let us compute the renormalized magnetic field, which is given by $$\label{5} H_{R}=\frac{\partial F}{\partial M}= tM + \frac{1}{6} u^{*}M^{3} + \frac{u^{*}M}{2}(t+ \frac{u^{*}M^{2}}{2})[I_{SP}-I],$$ where $$\label{6} I = \int \frac{d^{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})}q \overset{L}{\underset{n=2}{\prod}} d^{m_{n}}k_{(n)}} {[\overset{L}{\underset{n=2}{\sum}}\bigl((k_{(n)})^{2}\bigr)^{n} + q^{2}] \left(\overset{L}{\underset{n=2}{\sum}}(k_{(n)}^{2})^{n} + q^{2} + t + \frac{u^{*}M^{2}}{2} \right)}\;\;\;.$$ Let us compute explicitly this integral. First we use a Feynman parameter to write it as $$\label{7} I = \int_{0}^{1} dx \int \frac{d^{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})}q \overset{L}{\underset{n=2}{\prod}} d^{m_{n}}k_{(n)}} {[\overset{L}{\underset{n=2}{\sum}}\bigl((k_{(n)})^{2}\bigr)^{n} + q^{2} + x(t + \frac{u^{*}M^{2}}{2})]^{2}}\;\;\;.$$ In order to integrate the quadratic momentum out, we employ the identity $$\label{8} \int d^{d}q \frac{1}{[q^{2} + 2kq + m^{2}]^{\alpha}}= \frac{S_{d} \Gamma(\frac{d}{2}) \Gamma(\alpha - \frac{d}{2})}{2\Gamma(\alpha)} (m^{2}-k^{2})^{\frac{d}{2} - \alpha},$$ and get to $$\begin{aligned} \label{9} && I = \frac{1}{2} S_{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})} \Gamma(d - \overset{L}{\underset{n=2}{\sum}}m_{n}) \Gamma(2 -(d - \overset{L}{\underset{n=2}{\sum}}m_{n})) \int_{0}^{1} dx \int \overset{L}{\underset{n=2}{\prod}} d^{m_{n}}k_{(n)} \nonumber\\ && \;\;\;\; \times \frac{1}{[\overset{L}{\underset{n=2}{\sum}}\bigl((k_{(n)})^{2}\bigr)^{n} + x(t + \frac{u^{*}M^{2}}{2})]^{2- \frac{(d - \overset{L}{\underset{n=2}{\sum}}m_{n})}{2}}}\;\;\;.\end{aligned}$$ Now we have to perform the remaining integral with higher power of momentum. Indeed, in the integral $$\label{10} i_{n} = \int \frac{d^{m_{n}}k_{(n)}} {[\overset{L}{\underset{n=2}{\sum}}\bigl((k_{(n)})^{2}\bigr)^{n} + m^{2}]^{\gamma}},$$ first perform the change of variables $r_{(n)}^{2}= k_{1(n)}^{2}+...+k_{m_{n}(n)}^{2}$. Second, change the variables to $z=r_{(n)}^{n}$ and then to $z'=z^{2}$. Collecting together these set of steps, we finally obtain $$\label{11} i_{n} = \frac{1}{2n \Gamma(\gamma)} S_{m_{n}}\Gamma(\frac{m_{n}}{2n}) \Gamma(\gamma - \frac{m_{n}}{2n}) (m^{2})^{-\gamma + \frac{m_{n}}{2n}}.$$ Replacing this identity back into the expression for $I$, we can solve successfully all the integrals in higher powers of momentum along each competition subspace. Using $\epsilon_{L} = 4 + \overset{L}{\underset{n=2}{\sum}}\frac{(n-1)}{n} m_{n} - d$, we find $$\begin{aligned} \label{12} && I = \frac{1}{2} S_{(d-\overset{L}{\underset{n=2}{\sum}} m_{n})} \Gamma(d - \overset{L}{\underset{n=2}{\sum}}m_{n}) \Gamma(2 -(d - \overset{L}{\underset{n=2}{\sum}}m_{n})) (\overset{L}{\underset{n=2}{\prod}} [\frac{S_{m_{n}} \Gamma(\frac{m_{n}}{2n})}{2n})])\nonumber\\ && \;\;\times (t + \frac{u^{*}M^{2}}{2})^{-\frac{\epsilon}{2}} (1+\frac{\epsilon}{2}+O(\epsilon^{2})),\end{aligned}$$ which can be further simplified. In fact, developing the argument of the $\Gamma$-functions, using the identity $\Gamma(a + bx)= \Gamma(a)(1+bx \psi(a))$ and recalling to absorb the angular factor already mentioned above, it is not difficult to show that the integral has the following singular structure $$\label{13} I= \frac{1}{\epsilon_{L}}(1+\epsilon_{L}[h_{m_{L}}-\frac{1}{2}(1 + ln(t+\frac{u^{*}M^{2}}{2}))]).$$ We have to take another derivative of $H_{R}$ with respect to $M$, which will produce the inverse susceptibility $$\label{14} \chi^{-1}= \frac{\partial H_{R}}{\partial M}= t + \frac{u^{*}M^{2}}{2} + \frac{u^{*}}{4}(t + \frac{3u^{*}M^{2}}{2})[1+ln(t + \frac{u^{*}M^{2}}{2})] + \frac{u^{* 2}M^{2}}{4}.$$ For $T>T_{L}$ we substitute $M=0$ and the coupling constant at the fixed point value $u^{*}=\frac{2\epsilon_{L}}{3} + O(\epsilon_{L}^{2})$ into last equation, which produces the result $$\label{15} \chi (T>T_{L}) = (1-\frac{\epsilon_{L}}{6})t^{-(1+\frac{\epsilon_{L}}{6})}.$$ Note that $\gamma_{L}= 1+\frac{\epsilon_{L}}{6}$ and $\chi(T>T_{L})$ above is consistent with scaling in the neighborhood of the critical point. When $T<T_{L}$, we have to use the value of $M$ at the coexistence curve which is defined by the condition $H_{R}=0$, namely $$\label{16} M^{2}= \frac{-6t}{u^{*}} + 3t[1 + ln(-2t)].$$ Replacing this value at expression (14) and neglecting $O(u^{* 2} \sim \epsilon_{L}^{2})$, we can demonstrate that the inverse susceptibility below the Lifshitz temperature has the form $$\label{17} \chi^{-1}= (-2t)(1 + u^{*}(1 + ln[(-2t) + \frac{3u^{*}t}{2}(1+ln(-2t))])) + \frac{3u^{*}t}{2}ln(-2t).$$ Expanding the “logarithm of the logarithm” in the above expression using the expansion $ln(1+x)=x + O(x^{2})$, employing the fixed point $u^{*}=\frac{2\epsilon_{L}}{3} + O(\epsilon_{L}^{2})$ and neglecting $O(\epsilon_{L}^{2})$, we obtain $$\label{18} \chi=(-t)^{-(1+\frac{\epsilon_{L}}{6})}\frac{1}{2}[1 - \frac{\epsilon_{L}}{6} (4 +ln2)].$$ Consequently, the susceptibility amplitude ratio is given by $$\label{19} \frac{C_{+}}{C_{-}}= 2[1+\frac{\epsilon_{L}}{6}(3+ln2)]= 2^{\gamma_{L} -1}\frac{\gamma_{L}}{\beta_{L}},$$ where $\beta_{L}= \frac{1}{2} - \frac{\epsilon_{L}}{6}$. This expression is exact at one-loop level, its functional form in $\epsilon_{L}$ for several universal classes is the same, but the latter encodes [*distinct*]{} universalities since $\epsilon_{L}=\epsilon_{L}(d,m _{2},...,m_{L})$. Amplitude ratio for generic isotropic competing systems ======================================================= There are some minor modifications in the isotropic behaviors, but the trend to obtain the amplitude ratio follows the same script as in the anisotropic case. As there is only one subspace, say along $d= m_{n}$ space directions coupling $n$ neighbors via alternate competing interactions, the bare density Lagrangian is given by $$\begin{aligned} \label{20} L &=& \delta_{0n} \frac{1}{2} |\bigtriangledown_{m_{n}} \phi_0\,|^{2} + \sum_{n'=2}^{n-1}\frac{1}{2} \tau_{nn'} |\bigtriangledown_{m_{n}}^{n'} \phi_0\,|^{2} + \frac{\sigma_{n}}{2} |\bigtriangledown_{m_{n}}^{n} \phi_0\,|^{2} + \nonumber\\ &&+ \frac{1}{2} t_{0}\phi_0^{2} + \frac{1}{4!}\lambda_0\phi_0^{4} .\end{aligned}$$ As before, the isotropic critical region is defined by $\delta_{0n}= \tau_{nn'}=0$ with $T \neq T_{L}$. There is only one renormalization group direction characterized by the $\xi_{n}$ correlation length. We perform a dimensional redefinition in the momentum just as done in the discussion of the anisotropic behavior. The expansion parameter is now $\epsilon_{n}=4n-m_{n}$. The renormalized free energy at one loop can be written in the form $$\begin{aligned} \label{21} F(t,M)\;\; =&& \frac {1}{2}t M^{2} + \frac{1}{4!} g^{*} M^{4} + \frac{1}{4}(t^{2} + g^{*} t M^{2} + \frac{1}{4} (g^{*} M^{2})^{2}) I_{SP} \nonumber\\ && + \frac{1}{2} \int d^{m_{n}}k \Bigl[ln\Bigl(1 + \frac{t+ \frac{1}{2} g^{*} M^{2}}{k^{2n}}\Bigr) \nonumber\\ &&- \frac{g^{*} M^{2}}{2 k^{2n}}\Bigr]\;\;,\end{aligned}$$ and the nomenclature is almost the same as in the anisotropic case, except that now the integral $I_{SP}$ given by $$\label{22} I_{SP}= \Bigl[\int \frac{d^{m_{n}}k}{[(k+K')^{2n}]k^{2n}}\Bigr],$$ is computed at the symmetry point $K^{' 2n}=\kappa^{2n}=1$. Performing a derivative with respect to $M$, we obtain $$\label{23} H_{R}=\frac{\partial F}{\partial M}= tM + \frac{1}{6} u^{*}M^{3} + \frac{u^{*}M}{2}(t+ \frac{u^{*}M^{2}}{2})[I_{SP}-I],$$ with $$\label{24} I = \int \frac{d^{m_{n}}k} {k^{2n} (k^{2n}+ t + \frac{u^{*}M^{2}}{2})}\;\;\;.$$ Let us compute this last integral by employing a Feynman parameter, which gives essentially Eq.(\[7\]) in the absence of the quadratic term. We then discover that the resulting integral has the same pattern as Eq.(\[11\]) and can be solved along the same changes of variables in a identical manipulation which led to Eq.(\[12\]). The geometric angular factor which appears here is just the area of the $m_{n}$-dimensional unity sphere $S_{m_{n}}$ and shall be absorbed in a redefinition of the coupling constant as before. Carrying out this procedure, we get to $$\label{25} I = \frac{1}{\epsilon_{n}}\Bigl[1-\frac{\epsilon_{n}}{2n} ln\Bigl(t + \frac{u^{*}M^{2}}{2}\Bigr)\Bigr].$$ Now, we can calculate the integral $I_{SP}$ either using the orthogonal approximation or exactly. Although we could determine the integral exactly, we would like to know the deviation between the two results. The reason is simple: in the computation of the exponents, only the anomalous dimension of the field had a significant difference: the first term is positive or negative depending the value of $n$ in the exact computation, but it is always positive in the orthogonal approximation. Nevertheless, a numerical analysis proved that the maximal error for increasing space dimension and (arbitraly) fixed $\epsilon_{n}=1$ occurred in the specific heat critical exponent for $n=2$ ($3.9\%$), increased for $n=3$ for the same exponent ($4.1\%$), but decreased for increasing $n$ ($n=4, 2.2\%$; $n=5, 3.4\%$, ...) [@LeiteII]. We have now, for the first time, the opportunity to test the effectiveness of the orthogonal approximation in amplitudes, which in our opinion is worthy analyzing in view of the facts already known from the deviations of critical exponents. Next, let us perform the computation of the amplitude ratio using either the orthogonal approximation or the exact computation of the integral $I_{SP}$. As we shall see in a moment, the amplitudes above and below the Lifshitz temperature change. Amplitude ratio using the orthogonal approximation -------------------------------------------------- According to Ref. [@LeiteII], the integral computed at the symmetry point utilizing the orthogonal approximation was shown to result in the expression $I_{SP}= \frac{1}{\epsilon_{n}}(1+\frac{\epsilon_{n}}{2n})$. Hence, $$\label{26} I_{SP}-I= \frac{1}{2n}\Bigl[1+ln\Bigl(t + \frac{u^{*}M^{2}}{2}\Bigr)\Bigr],$$ which implies $$\label{27} H_{R}=\frac{\partial F}{\partial M}= tM + \frac{1}{6} u^{*}M^{3} + \frac{u^{*}M}{4n}(t+ \frac{u^{*}M^{2}}{2})\Bigl[1+ln\Bigl(t + \frac{u^{*}M^{2}}{2}\Bigr)\Bigr].$$ Taking another derivative with respect to $M$, we find $$\label{28} \chi^{-1}= t + \frac{u^{*}M^{2}}{2} + \frac{u^{*}}{4n}(t + \frac{3u^{*}M^{2}}{2})[1+ln(t + \frac{u^{*}M^{2}}{2})] + \frac{u^{* 2}M^{2}}{4n}.$$ Now, $M=0$ in last equation is the situation corresponding to $T>T_{L} (t>0)$, or in other words $$\label{29} \chi^{-1}(T>T_{L})= t + \frac{u^{*} t}{4n} [1+lnt].$$ Replacing the fixed point value $u^{*}=\frac{2\epsilon_{n}}{3}$, the susceptibility above $T_{L}$ reads $$\label{30} \chi(T>T_{L})= t^{-\gamma_{n}}\Bigl(1-\frac{\epsilon_{n}}{6n}\Bigr),$$ where $\gamma_{n}=1+\frac{\epsilon_{n}}{6n}$ is the isotropic susceptibility exponent. Below $T_{L}$, we determine the value of $M$ in the coexistence curve defined by $H_{R}=0$, which yields $$\label{31} M^{2}= \frac{(-6t)}{u^{*}} + \frac{3t}{n}[1+ln(-2t)].$$ Replacing this value in the expression of $\chi^{-1}$, it leads to $$\label{32} \chi^{-1}(T<T_{L}) = (-2t)[1 + \frac{u^{*}}{4n}ln(-2t) + \frac{u^{*}}{n}],$$ which at the fixed point $u^{*}=\frac{2}{3\epsilon_{n}}$ implies that we can write the susceptibility in the form $$\label{33} \chi(T<T_{L})= (-t)^{\gamma_{n}} \frac{1}{2}[1-\frac{\epsilon_{n}}{6n}(4+ln2)].$$ The susceptibility amplitude ratio which results from the above expressions is written as $$\label{34} \frac{C_{+}}{C_{-}}= 2\Bigl[1 + \frac{\epsilon_{n}}{2n} + \frac{\epsilon_{n}}{6n} ln2 \Bigr]= 2^{\gamma_{n}-1}\frac{\gamma_{n}}{\beta_{n}},$$ where $\beta_{n}= \frac{1}{2} - \frac{\epsilon_{L}}{3n}$ is the magnetization exponent. Exact amplitude ratio --------------------- The main advantage of the isotropic case is that the Feynman integrals can be computed exactly. Thus, we can obtain the susceptibility amplitude ratio without the necessity of using approximations. The $I_{SP}$ integral was already computed in Ref. [@LeiteII] at the symmetry point and was shown to be given by the expression $I_{SP}= \frac{1}{\epsilon_{n}}[1 + D(n)\epsilon_{n}]$, where $D(n) = \frac{1}{2}[\psi(2n) + \psi(1)] - \psi(n)$. First, using Eq. (25) we find $$\label{35} I_{SP}-I= D(n) + \frac{1}{2n} ln\Bigl(t + \frac{u^{*}M^{2}}{2}\Bigr),$$ which turns out to result in the following magnetic field $$\label{36} H_{R}=\frac{\partial F}{\partial M}= tM + \frac{1}{6} u^{*}M^{3} + \frac{u^{*}M}{4n}(t+ \frac{u^{*}M^{2}}{2})\Bigl[2n D(n)+ln\Bigl(t + \frac{u^{*}M^{2}}{2}\Bigr)\Bigr].$$ It is easy to show that the inverse susceptibility which follows can be written as $$\label{37} \chi^{-1}= t + \frac{u^{*}M^{2}}{2} + \frac{u^{*}}{4n}(t + \frac{3u^{*}M^{2}}{2})[2n D(n)+ln(t + \frac{u^{*}M^{2}}{2})] + \frac{u^{* 2}M^{2}}{4n}.$$ Hereafter we are going to use the coupling constant at the fixed point, i.e., $u^{*}=\frac{2\epsilon_{n}}{3}$. Set $M=0$ for $T>T_{L}$ in the above equation in order to find the susceptibility in the following form $$\label{38} \chi(T>T_{L})= t^{-\gamma_{n}}\Bigl(1-\frac{D(n)\epsilon_{n}}{3}\Bigr).$$ For $T<T_{L}$ the value of $M$ in the coexistence curve is given by $$\label{39} M^{2}= \frac{(-6t)}{u^{*}} + \frac{3t}{n}[2n D(n)+ln(-2t)].$$ Substitution of this value into the expression for $\chi^{-1}$ results in the following value for the susceptibility below $T_{L}$ $$\label{40} \chi(T>T_{L})= (-2t)^{-\gamma_{n}} \Bigl[1-\frac{\epsilon_{n}}{2n}\Bigl(1+ \frac{2nD(n)}{3}\Bigr) \Bigr].$$ Using Eqs. (\[38\]) and (\[40\]), we finally obtain $$\label{41} \frac{C_{+}}{C_{-}}= 2\Bigl[1 + \frac{\epsilon_{n}}{2n} + \frac{\epsilon_{n}}{6n} ln2 \Bigr]= 2^{\gamma_{n}-1}\frac{\gamma_{n}}{\beta_{n}},$$ which is the same value obtained using the orthogonal approximation Eq. (\[34\]). As happened to the critical exponents at one-loop order, the orthogonal approximation and the exact computations of the susceptibility amplitude ratio in the isotropic case yield the same value. However, we do expect deviations in both calculations at two-loop order and beyond. Conclusion ========== The obtained anisotropic amplitude ratios maintain the same functional form as its counterpart in the Ising-like universality class, with the parameter $\epsilon_{L}= = 4 + \overset{L}{\underset{n=2}{\sum}}\frac{(n-1)}{n} m_{n} - d$ replacing the ordinary perturbation parameter of noncompeting systems $\epsilon=4-d$. This leads to the property of universality class reduction. This property already appeared in the computation of exponents and is expected to be valid at all loop orders. In fact, if we turn off all the competing interactions, i. e., by setting $m_{3}= ...=m_{L}=0$, keeping just alternate couplings among second neighbors and identify $m_{2}=m$, we obtain the result for the anisotropic $m$-axial universality class. Note that the ratio has the same functional form for all values of $m \neq d$ and reproduces the uniaxial case $m=1$ studied earlier [@Leite3]. For instance, three-dimensional systems have perturbative parameters $\epsilon_{L}= 1 + \frac{m}{2}$ which change at distinct values of $m$. Consequently, they produce different values for the amplitude ratios which is consistent with the universality hypothesis previously stated. Besides, if go on and switch off the competing interactions among second neighbors ($m=0$) we obtain the result of the Ising-like universality class. The isotropic amplitude ratios, on the other hand, possesses its own version of universality class reduction. Different values of the number of neighbors coupled via alternate couplings ($n$) are responsible by the variation of the susceptibility ratio. The case $n=2$ corresponds to the isotropic $m$-axial ($d=m$ close to 8) universality class. Our result is the first computation of isotropic amplitude ratios, perhaps because isotropic systems were found only in lower dimensional systems ($m=d=3$) so far, as proposed in the critical behavior of homopolymer-diblock copolymer mixtures [@BF; @KM; @Pipi; @Natalie], which makes the perturbative parameter rather large ($\epsilon_{2}=5$). Let us try to extract meaningful results from our results for those three-dimensional systems. Although the value of the amplitude above the critical temperature is not universal, let us compare the two values using the approximate and exact results for the isotropic case $n=2$. Using $\epsilon_{2}=5$ in Eq.(\[30\]) we find $C_{+}=0.583$ using the orthogonal approximation, whereas the exact computation from Eq.(\[38\]) using $D(2)=-\frac{1}{12}$ yields $C_{+}=1.13$, and the deviation is huge. Nevertheless, comparing with the Table III from Ref. [@Pipi] both values are allowed. In fact for diblock polymer composition $\Phi_{DB}=0.072$ at temperature $69.2 \pm 0.1 ^{o}C$ the measured amplitude is given by $C_{+}=0.6 \pm 0.04$ which is compatible with the value obtained via the orthogonal approximation. On the other hand, for a slightly change of composition, namely $\Phi_{DB}=0.073$ measured at temperature $69.5 \pm 0.2 ^{o}C$ the amplitude value is $C_{+}=0.94 \pm 0.07$, which is also consistent with the exact amplitude. Notice that even though those authors confirmed the absence of the isotropic Lifshitz point, they considered the Lifshitz critical region [*with the inclusion of the fluctuations*]{} using SCFT (Ref. [@KM]) in their experimental fits of the susceptibility curves, which is quite a different method than the one proposed in the present work [@Smirnova]. Therefore, this is the first solid indication that field theory renormalization group results including the contribution of fluctuations are consistent with experiments in those sort of polymers, in spite of the large value of the perturbative parameter. Though the deviations between the amplitudes are significant and expected from their nonuniversal feature, the experimental results do not rule out the orthogonal approximation result. This is the first experimental ground to test the deviations in both calculations. But we can go on and compare the true universal susceptibility exponent obtained in the experiment with our previous two-loop calculation from Ref.[@LeiteII]. The orthogonal approximation for $(N=1, d=m=3)$ yields $\gamma_{2}=1.90$, whereas the exact exponent is $\gamma_{2}=1.50$. The latter is consistent with the experimental value $\gamma_{2}=1.55 \pm 0.15$ obtained from the isotherm at 69.5$^{o}$C with concentration of diblock copolymer at $\Phi_{DB}=0.071$. It is amazing that the experiment carried out on the homopolymer-diblock copolymer considered by those authors can really be described using the isotropic Lifshitz universality class and its critical region, in spite of the large value of the perturbative parameter for three-dimensional systems. Perhaps the use of other field-theoretic isotropic results already (and to be) developed in other experiments to be performed might be successful in refining our knowledge of the Lifshitz critical region for these systems. Another aspect is the theoretical possibility of occurrence of up to $6th$ character Lifshitz points in $AB/BC$ mixtures of diblock copolymers [@OH]. If this system can be fabricated in the laboratory, our work represents a prevision of results for the susceptibility with increasing values of the perturbation parameter for three-dimensional systems, in analogy to what was studied in Ref. [@Pipi] using small angle neutron scattering. This is rather encouraging an evidence to pursue further universal aspects of this kind of critical behaviors using this field theoretical language, for instance, amplitude ratios above and below the critical temperature. This could shed new light in devising experimental applications to our model in order to measure those effects in a real physical system, with isotropic or (less obvious) anisotropic critical behaviors. The universality class reduction in the isotropic case is even more evident than its anisotropic counterpart. As a matter of fact, $n=1$ corresponds the system without competition and belongs to the Ising-like universality class. Therefore, systems without competition can be understood as special cases either from the anisotropic cases ($m_{n}=0, n=2,...,L$) or from the isotropic cases $n=1$, a property already discovered in the computation of the critical exponents. Since the isotropic ratio can be computed approximately and exactly as well, we calculated the ratio using both procedures for the sake of comparison of the deviations for individual amplitudes and how this deviation could be understood at least in the case $n=2$. The amplitudes themselves are different in both cases, but the ratio is equal. This property also takes place in the determination of critical indices using perturbation theory, but the result is valid only at one-loop level. We expect that both ratios will have deviations at two-loop order. The most interesting extension of the method proposed here is the study of the specific heat amplitude ratio for generic competing systems, generalizing the discussion carried out for the anisotropic $m$-axial critical behavior[@Leite3]. It would be nice to tackle the computation of other universal amplitude ratios either at one-loop level or to extend the method to compute amplitude ratios at two-loop order [@Ber] for generic competing systems. Last but not least, we hope that the present investigation can be significant to motivate experimental techniques in order to determine the susceptibility amplitude ratio in magnetic systems such as $MnP$, $Mn_{0.9}Co_{0.1}P$ [@MnCoP; @Pla], etc.. In addition the new phase encountered in $MnP$ [@new] and $Mn_{0.9}Co_{0.1}P$ [@Pla] might be related to new effects of competition as described in the present work. Acknowledgments =============== CFF would like to thank CAPES for financial support. [99]{} [Ho-Lu-Sh]{} R. M. Hornreich, M. Luban, and S. Shtrikman, [*Phys. Rev. Lett.*]{} [**35**]{} 1678 (1975). [Ho]{} R. M. Hornreich, [*Journ. Magn. Magn. Mat.*]{} [**15-18**]{} 387 (1980). [Be]{} C. C. Becerra, Y. Shapira, N. F. Oliveira Jr., and T. S. Chang, [*Phys. Rev. Lett.*]{} [**44**]{}, 1692 (1980). [Yo1]{} C. S. O. Yokoi, M. D. Coutinho-Filho, and S. R. Salinas, [*Phys. Rev. B*]{} [**24**]{} 5430 (1981). [Yo2]{} C. S. O Yokoi, M. D. Coutinho-Filho, and S. R. Salinas, [*Phys. Rev. B*]{} [**29**]{} 6341 (1984). [Be1]{} C. C. Becerra, V. Bindilatti, and N. F. Oliveira Jr., [*Phys. Rev. B*]{} [**62**]{} 8965 (2000). [Hana]{} T. Hanawa, K. Shinkawa, M. Ishikawa, K. Miyatani, K. Saito, and K. Kohn, [*J. Phys. Soc. Jpn.* ]{} [**63**]{} 2706 (1994). [We]{} H. Weitzel, H. Ehrenberg, C. Heid, H. Fuess, and P. Burlet, [*Phys. Rev. B*]{} [**62**]{} 12146 (2000). [H]{} S. M. Hayden, G. Aeppli, H. Mook, D. Rytz, M. F. Hundley, and Z. Fisk, [*Phys. Rev. Lett.*]{} [**66**]{} 821 (1991). [K]{} B. Keimer, R. J. Birgeneau, A. Cassanho, Y. Endoh, R. W. Erwin, M. A. Kastner, and G. Shirane, [*Phys. Rev. Lett.*]{} [**67**]{} 1930 (1991). [S]{} S. Sachdev, and J. Ye, [*Phys. Rev. Lett.*]{} [**69**]{} 2411 (1992). [Ra]{} S. B. Rananavare, V. G. K. M. Pisipati, and E. W. Wong, [*Phys. Rev. Lett.*]{} [**72**]{} 3558 (1994). [Za]{} B. Zalar, A. Gregorovi$\check{c}$, M. Simsi$\check{c}$, A. Zidan$\check{s}$ec, R. Blinc, S. Keast, and M. Neubert, [*Phys. Rev. Lett.*]{} [**80**]{} 4458 (1998). [Ska]{} M. $\check{S}$karabot, R. Blinc, I. Mu$\check{s}$evi$\check{c}$, A. Rastegar, and Th. Rasing, [*Phys. Rev. E*]{} [**61**]{} 3961 (2000). [MC]{} C. Mergulhão Jr., and C. E. I. Carneiro, [*Phys. Rev. B*]{} [**58**]{} 6047 (1998); [*Phys. Rev. B*]{} [**59**]{}, 13954 (1999). [AL]{} L. C. Albuquerque, and M. M. Leite, [*cond-mat/0208500*]{} (unpublished); L. C. Albuquerque, and M. M. Leite, [*J. Phys. A: Math Gen.*]{} [**34**]{}, L327 (2001). [DS]{} H. W. Diehl, and M. Shpot, [*Phys. Rev. B*]{} [**62**]{} 12338 (2000); M. Shpot, and H. W. Diehl, [*Nucl. Phys. B*]{} [**612**]{}(3) 340 (2001). [L]{} M. M. Leite, [*hep-th/0109037*]{} (unpublished). [Leite1]{} M. M. Leite, [*Phys. Rev.*]{} [**B 67**]{} 104415 (2003). [CL1]{} P. R. S. Carvalho, and M. M. Leite, [*Ann. Phys.*]{} [**324**]{} 178 (2009). [Se1]{} W. Selke, [*Z. Physik B*]{} [**27**]{}, 81 (1977). [Se2]{} W. Selke, [*Phys. Lett. A*]{} [**61**]{} 443 (1977). [Ni1]{} J. F. Nicoll, G. F. Tuthill, T. S. Chang, and H. E. Stanley, [*Phys. Lett. A*]{} [**58**]{} 1 (1976). [Ni2]{} J. F. Nicoll, G. F. Tuthill, T. S. Chang, and H. E. Stanley, [*Physica B*]{} [**86-88**]{} 618 (1976). [Leite2]{} M. M. Leite, [*Phys. Lett.*]{} [**A 326**]{} 281 (2004). [LeiteII]{} M. M. Leite, [*Phys. Rev.*]{}[**B 72**]{} 224432 (2005). [CL2]{} P. R. S. Carvalho, and M. M. Leite, [*Ann. Phys.*]{} [**325**]{} 151 (2010). [Amit]{} For similar field-theoretic conventions in noncompeting systems, see D. J. Amit, and V. Martin-Mayor, [*in Field theory, the renormalization group and critical phenomena*]{}, (World Scientific, Singapore, Third Edition, 2005). [Selke1]{} W. Selke, [*Phys. Rep.*]{} [**170**]{} 213 (1988). [Selke2]{} W. Selke, [*in Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. Lebowitz (Academic Press, London, 1992), Vol.15, 1. [BF]{} D. Broseta, and G. H. Fredrickson, [*J. Chem. Phys.*]{} [**93**]{}, 2927 (1990); R. Holyst, and M. Schick, [*J. Chem. Phys.*]{} [**96**]{}, 7728 (1992). [Bates1]{} F. S. Bates, W. Maurer, T. P. Lodge, M. F. Schulz, M. W. Matsen, K. Almdal, and K. Mortensen, [*Phys. Rev. Lett.*]{} [**75**]{}, 4429 (1995). [Bates2]{} F. S. Bates, W. W. Maurer, P. M. Lipic, M. A. Hillmyer, K. Almdal, K. Mortensen, G. H. Fredrickson, and T. P. Lodge, [*Phys. Rev. Lett.*]{} [**79**]{}, 849 (1997). [KM]{} L. Kielhorn, and M. Muthukumar, [*J. Chem. Phys.*]{} [**107**]{}, 5588 (1997). [OH]{} P. D. Olmstead, and I. W. Hamley, [*Europhys. Lett.*]{} [**45**]{} 83 (1999). [Pipi]{} V. Pipich, D. Schwahn, and L. Willner, [*J. Chem. Phys.*]{} [**123**]{}, 124904 (2005). [Leite3]{} The case $m=1$ was computed in M. M. Leite, [*Phys. Rev. B*]{} [**61**]{} 14691 (2000). [Natalie]{}N. A. Denesyuk, and G. Gompper, [*Macromolecules*]{} [**39**]{} 5497 (2006). [Smirnova]{} Phase diagrams of block copolymer systems can be built using SCFT as discussed by Y. G. Smirnova in [*Microphase separation in two-length-scale multiblock copolymer melts - A theoretical study*]{}, PhD. Thesis, University of Groningen, The Netherlands (2006). [Leite4]{} M. M. Leite, [*Phys. Rev. B*]{} [**68**]{} 052408 (2003). [Ber]{} C. Bervillier, [*Phys. Rev. B*]{} [**14**]{} 4964 (1976). [MnCoP]{} C. C. Becerra, A. Zieba, N. F. Oliveira Jr., and H. F. Jellvag, [*J. App. Phys.*]{} [67]{} 5442 (1990). [Pla]{} T. Plackowski, M. Matusiak, and J. Sznajd, [*Phys. Stat. Solid. (b), DOI: 10.1002/pssb.201147287*]{} (2011) \[[*ArXiv:1105.2701 (cond-mat)*]{}\]. [new]{} C. C. Becerra, [*J. Phys.: Condens. Matter*]{} [**12**]{} 5889 (2000); T. Yamazaki, Y. Tabata, T. Waki, H. Nakamura, M. Matsuura, and N. Aso, [*J. Phys.: Conf. Ser.*]{} [**200**]{} 32079 (2010). [^1]: e-mail:[email protected] [^2]: e-mail:[email protected]

ArXiv

**Flux of radiation from pointlike sources in general relativity**\ Matej Sárený[^1]\ *Faculty of Mathematics, Physics and Informatics, Comenius University, Bratislava, Slovakia*\ Abstract {#abstract .unnumbered} ======== In the paper we study propagation of light in general relativity through a spacetime filled with cold plasma with infinite conductivity. We use the geometric optics based on Synge’s approach. As the main result we provide a formula for calculation of spectral flux of radiation emitted from a pointlike source. The formula employs connecting vectors that are obtained by integrating the ray deviation equation along the reference ray connecting the source and observation event. As a byproduct we formulate Etherington’s reciprocity theorem with the inclusion of plasma, interrelating angular size distance and luminosity distance. We also discuss the Liouville theorem and its formulation in terms of connecting vectors. **Keywords**: relativistic geometric optics, cold plasma, flux of radiation, ray deviation equation, reciprocity theorem, relativistic distances, Liouville theorem Introduction {#introduction .unnumbered} ============ In astrophysics it is often of great importance to know the specific flux of radiation $F_\nu=dE/dtdAd\nu$, especially when dealing with spatially indistinguishable, essentially pointlike, sources. For example, $F_\nu$ is interesting in gravitation lensing events, when the incoming flux from the distant star is briefly increased as a compact object comes into the vicinity of the line of sight. Astronomers plot the dependence of flux on time as the so called “light curves”. Some standard methods to calculate the flux theoretically can be found in textbooks like [@Schneider_Ehlers; @Schneider_Kochanek] and also in the paper [@Cunningham_Bardeen]. These methods, however, seem to lack a certain degree of general applicability, easy implementation in a generic situation, because they are adapted either for situations in which both source and observer are far away from the lens ([@Schneider_Ehlers; @Schneider_Kochanek]) or speciffically to the situation “source close to the black hole and observer far away” in case of [@Cunningham_Bardeen] (to be fair, astrophysically, these are probably the most common situations). In this paper we are going to develop a formula that calculates the flux from a pointlike source in an arbitrarily strong gravitational field, utilizing the ray deviation equation (generalization of the Jacobi equation, see [@Sareny_Balek]). We will start from the definition of the flux in the formalism of kinetic theory, using an approach built upon the foundations provided in textbooks [@MTW; @Thorne_Blandford]. In the process we will discuss the formulation of Liouville theorem in terms of connecting vectors, which will be instrumental to our calculations, and prove a “plasma version” of the Etherington’s reciprocity theorem which interrelates the solid angles and areas placed at source and observer that are connected by light rays. In vacuum, this theorem was discussed in [@Schneider_Ehlers; @Etherington; @Temple; @Ellis], with a nice historical review to be found in [@Ellis_Etherington]. In our article we will study infinitesimally close rays propagating in generally relativistic conditions through cold plasma with infinite conductivity, which has optical influence on the rays. Such setting was previously also studied in [@Schulze-Koops]. For the theory with generally relativistic light propagation in a medium with non-unit index of refraction see Synge’s book [@Synge]. This theory was lately used as a foundation for many works e.g. [@Sareny_Balek; @Schulze-Koops; @Bisnovatyi_2017; @Rogers_2017] etc. (for a recent, more exhaustive, overview on this topic see e.g. paragraph 2 of [@Bisnovatyi_2020]). Our paper is organized into nine paragraphs. In the first paragraph we will display some “naive manipulations” with the flux definition and show the shortcomings of doing so. In the second paragraph we will summarize the basic formulae for the generally relativistic ray optics with the inclusion of plasma. The third paragraph will introduce the definition of the flux in the framework of kinetic theory, and will be followed by the fourth paragraph where the calculation itself will take place. In the fifth paragraph we will show how the flux can be calculated using the connecting vectors. The sixth paragraph will elucidate the connection between the two-ray conservation law accompanying the ray deviation equation and the Liouville theorem, using a nifty formula for determinant discussed in appendix \[app:det\_formula\]. The seventh paragraph will apply this result to provide alternative ways of the flux calculation and in the process of doing so we will prove the reciprocity theorem in a “roundabout” way. A link between two of three proposed methods of calculation is given in appendix \[app:dtdnu\_cons\]. Using a numerical approach on a Kerr black hole surrounded by a disc-like plasma distribution, we present in the eighth paragraph a model calculation of the key quantity that determines the flux. Finally, the nineth paragraph will provide a “concise” proof of the reciprocity theorem and hint on more possibilities of its extension beyond the scope needed here. This paper is a loose follow-up on our previous paper [@Sareny_Balek], but having read that paper is not necessary to understand the present one. In our paper we most closely follow the notation of our previous work [@Sareny_Balek]. We use the “mostly positive” signature $(-,+,+,+)$, with Greek indices assuming values from $\{0..3\}$, lower case Latin indices assuming values from $\{1..3\}$ and capital Latin indices assuming values from $\{1,2\}$, unless explicitly stated otherwise. In the matters of differential geometry we follow the conventions of [@Fecko], most notably with the use of $\nabla$ for metric-compatible, torsion-free covariant derivative and $\omega^\mu_{\ \nu}$ for the connection forms. We also use the notation $\flat/\sharp$ to abstractly denote index lowering/rising operations. We leave 4-vectors and scalars to be distinguished by the context, but we denote 3D vectors w.r.t. some 1-3 split of spacetime by the arrows above them. Orthogonal complement to a span of a set of vectors is denoted as $\{V..W\}^{{\mbox{\scriptsize ort}}}$, or simply $V^{{\mbox{\scriptsize ort}}}$ for a single vector. Finally, we employ the Planck unit convention where $\hbar=c=G=1$. Flux of radiation based on “intuitive” manipulation of differentials {#ch:naive_approach} ==================================================================== Consider a pointlike isotropic source described by its specific luminosity $L_\nu = dE/dtd\nu$, which, somewhere in its history, sends a ray from the event $S$ on its worldline. This light ray is afterwards captured by an observer at an observation event $O$. The spectral flux measured by the observer is $$\begin{aligned} F_{\nu_O}(O)=\frac{\omega_O dN}{dt_OdA_Sd\nu_O}=\frac{\omega_O}{\omega_S}\frac{L_{\nu_S}}{4\pi}\frac{d\Omega_S}{dA_S}\frac{dt_Sd\nu_S}{dt_Od\nu_O}=\frac{\omega_O}{\omega_S}\frac{L_{\nu_S}}{4\pi d_L^2}\nonumber\end{aligned}$$ where $\omega$, $\nu$, $d\nu$ and $dt$ denote angular frequency, ordinary frequency, frequency spread and time spread with which the photons leave the source and arrive to the observer (distinguished by $S$ and $O$ indices for the source and the observer respectively). We have also used the conservation of $dtd\nu$ (we do not know a one-sentence “naive” explanation of this fact, but a detailed investigation is performed in the appendix, utilizing the techniques discussed throughout the main body of this article). The photons that hit the area $dA_S$ at the event $O$ were released into the solid angle $d\Omega_S$ at the source event (they are both tagged with $S$ index to point out that they refer to the bundle of rays with the vertex at $S$). The ratio of the area and the solid angle defines the “(corrected) luminosity distance” $d_L$ via the relation $d_L^2=dA_S/d\Omega_S$ (see chapter 3 of [@Schneider_Ehlers] and paragraph 3 of [@Schulze-Koops] for comparison). If the flux is measured at two different events along the same ray, the ratio of its values is determined by the redshift factor as well as by the ratio of the luminosity distances from the source, $$\begin{aligned} \frac{F_{\nu_1}(O_1)}{F_{\nu_2}(O_2)}=\frac{\omega_1}{\omega_2}\frac{d_{L2}^2}{d_{L1}^2}\nonumber\end{aligned}$$ Furthermore, we can reinterpret the ratio of the luminosity distances squared as the ratio of the areas $dA_1$ and $dA_2$ localized at the events that are hit by the beam of photons coming from the same angle $d\Omega_S$. We can also apply the reciprocity theorem from the paper [@Schulze-Koops], which we will rediscover later in this paper, in equation (\[recip\_thm\_1\]). The flux relations then become $$\begin{aligned} F_{\nu_O}(O)=\frac{\omega_O}{\omega_S}\left(\frac{\omega_Ov_g^O}{\omega_Sv_g^S}\right)^2\frac{L_{\nu_S}}{4\pi d_A^2}\qquad\Rightarrow\qquad\frac{F_{\nu_1}(O_1)}{F_{\nu_2}(O_2)}=\frac{\omega_1}{\omega_2}\left(\frac{\omega_1v_{g1}}{\omega_2v_{g2}}\right)^2\frac{d_{A2}^2}{d_{A1}^2}\nonumber\end{aligned}$$ with $d_A$ being a quantity called “angular size distance” defined by the relation $d_A^2=dA_O/d\Omega_O$, and $v_g$ being the group velocity, which is non-unit if the photons propagate through plasma. Note that the dependence on the wavelength redshift $(\omega_1v_{g1})/(\omega_2v_{g2})$ appears here in addition to the dependence of frequency redshift. Furthermore, we can also reinterpret the ratio of the angular size distances squared as the ratio of the angles $d\Omega_1$ and $d\Omega_2$ under which the observers would see a small source of a finite size located around the source event (if both observers would look at a source of the same size). All these formulae are nice, but a tantalizing question arises on a closer inspection: Given we know the trajectory and the emission characteristics (luminosity) of the source as well as the trajectory of the observer, how do we calculate from these data the values of $d_L$, $d_A$, $d\Omega_S$, $dA_S$, $d\Omega_O$, $dA_O$ and find the flux? If you already know the answer you can stop reading now, but if you, like us, think that this question deserves a deeper contemplation, proceed to the next paragraphs of the paper. Generally relativistic equations for photon propagating through plasma ====================================================================== The relativistic equations for the light propagation in the presence of plasma were discussed to a greater or lesser extent in many papers (see the nice overview in paragraph 2 of [@Bisnovatyi_2020]) with the approach based on Synge’s book [@Synge]. The formulae and notation used in this paper will be most similar to our previous paper [@Sareny_Balek]. We will provide here a review of the most significant parts of this theory which will be used further in the paper. The equations of motion for the light ray travelling through plasma are extremals of the action: $$\begin{aligned} S[x]=\frac{1}{2}\int_{\lambda_1}^{\lambda_2}[p^2-\omega_{pl}^2(x)]d\lambda\nonumber\end{aligned}$$ with $p^\mu=\dot{x}^\mu\equiv dx^\mu/ d\lambda$ being the photon’s 4-momentum, $\lambda$ being the parameter of the ray and $\omega_{{\mbox{\scriptsize pl}}}^2=e^2n(x)/\epsilon_0m_e$ being the electron plasma frequency ($n$ stands for the number density of electrons measured in the rest frame of the medium). Extremalization of the action yields the ray evolution equation (REE) $$\begin{aligned} \label{REE} Dp=-\frac{1}{2}\mathcal{A}\end{aligned}$$ where $D=\nabla_p$ and $\mathcal{A}=\sharp d\omega_{{\mbox{\scriptsize pl}}}^2$. In addition, the solutions are subject to the normalization constraint $$\begin{aligned} \label{REE_nc} p^2=-\omega_{{\mbox{\scriptsize pl}}}^2\end{aligned}$$ Suppose we have two rays propagating infinitesimally close to each other, one designated as the “reference ray” $x^\mu(\lambda)$ and the other being the “neighbouring ray” $y^\mu(\lambda)\approx x^\mu(\lambda)+\epsilon\xi^\mu(\lambda)$ with $\epsilon\ll 1$. The vector $\xi$ is called the connecting vector (or sometimes the separation vector). The solutions for both $x$ and $\xi$ can be found as extremals of the action $$\begin{aligned} \psi [x,\xi]=\int_{\lambda_1}^{\lambda_2}\left( p\cdot D\xi - \frac{1}{2}\xi\cdot \mathcal{A}\right)d\lambda\nonumber\end{aligned}$$ On extremalization, $\psi$ once again yields REE (\[REE\]), but it also leads to the ray deviation equation (RDE), which is a generalization to the equation of geodesic deviation (Jacobi equation) for the case with plasma, $$\begin{aligned} \label{RDE} D^2\xi=R(p,\xi)p-\frac{1}{2}\sharp{\mathcal{H}}(\xi,.)\end{aligned}$$ with $R(A,B)=[\nabla_A,\nabla_B]-\nabla_{[A,B]}$ being the curvature operator[^2] and $\mathcal{H}=\nabla^2\omega_{{\mbox{\scriptsize pl}}}^2$ being the covariant Hesse matrix. The fact that the neighbouring ray is also subject to the normalization condition (\[REE\_nc\]) translates into the constraint $$\begin{aligned} \label{RDE_nc} p\cdot D\xi+\frac{1}{2}\xi\cdot\mathcal{A} = 0\qquad\Leftrightarrow\qquad p\cdot D\xi - \xi\cdot Dp=0\end{aligned}$$ This reduces the space of initial conditions (ICs) for RDE to 7 dimensions. As we said before, the connecting vector links the events of two rays with the same value of $\lambda$. However, which event is assigned the zero value of $\lambda$ remains arbitrary, resulting in equivalence relation between the connecting vectors $$\begin{aligned} \label{reparam} \xi\sim\tilde\xi=\xi+\alpha p\end{aligned}$$ where $\alpha$ is an arbitrary real number. One can easily check that $\xi=p(\lambda)$ is also a solution of RDE – it corresponds to the reference ray pointing at itself. Therefore, the 7-dimensional space of solutions is factorized by this equivalence to 6-dimensional space of unique neighbouring rays. Any two solutions of RDE, $\xi$ and $\xi^\prime$, satisfy the following conservation law: $$\begin{aligned} \label{two_ray_cons} \xi\cdot D\xi^\prime-\xi^\prime\cdot D\xi=const\end{aligned}$$ This can be readily checked by applying $D$ on it and using (\[RDE\]) together with the symmetries of Riemann tensor and covariant Hessian. In the special case when one of the solutions is $p$, the value of the conserving quantity is $0$, due to (\[RDE\_nc\]). This implies that the constant is the same for the whole equivalence class of the connecting vectors, i.e. it is unique for a given neighbouring ray. We will make a heavy use of this observation in the later paragraphs. For further clarification on the topics of this paragraph, we recommend checking out our previous paper [@Sareny_Balek], where we discuss them in more detail. Definition of flux in the kinetic theory ======================================== Let us consider the kinetic theory of photons in plasma. The central object in this theory is the photon number density distribution in the phase space, $\mathcal{N}(x^\mu,p_i,S)=dN/d^3x d^3p$, where $dN$ is the number of particles with 3-D position spread in the cube of the volume $d^3x$ around $\vec{x}=0$ and with momenta spread in the cube of the volume $d^3p$ around $p_i$, all measured in the locally inertial frame $S$ centered on the event $x^\mu$. Although it is not obvious at the first sight, the function is actually independent from the frame[^3] $S$ [@MTW; @Thorne_Blandford]. An additional difficulty present in GR is that the “global $\vec{x}$-space” does not exist. Instead, we choose a spacelike hypersurface[^4] $\Sigma$ in the spacetime. The number of photons $N$ on this hypersurface is obtained by collecting the information from the individual observers on $\Sigma$, whose 4-velocity is orthogonal to $\Sigma$. Mathematically speaking, $$\begin{aligned} \label{ph_num} N=\int d^3\Sigma d^3p \mathcal{N}(x^\mu(\sigma),p_i,S(\sigma))\end{aligned}$$ with $\sigma=(\sigma^1,\sigma^2,\sigma^3)$ being some parametrization of $\Sigma$. The function $\mathcal{N}$ is subject to the Boltzmann equation, which in the simplest case (no scattering, no creation or extinction of photons) reads $$\begin{aligned} \frac{d\mathcal{N}(x^\mu(\lambda),p_i(\lambda),S(\lambda))}{d\lambda}=0\nonumber\end{aligned}$$ where $d/d\lambda$ is the derivative along the solutions of REE (\[REE\]). If such collisionless Boltzmann equation holds, the number of photons does not depend on the choice of $\Sigma$, as long as $\Sigma$ is infinite. One can define the momentum-specific photon number flux density as $$\begin{aligned} \Phi_{\vec{p}}(x^\mu,\vec{n},S)=\frac{dN}{dAdtd^3p}\nonumber\end{aligned}$$ where $dN$ is the number of photons with the momentum from the cube of the volume $d^3p$ around $\vec{p}$, as measured in the system of reference $S$, which cross the area $dA$ with the normal $\vec{n}$, located at $x^\mu$, over the period of time $dt$. For the photons in plasma we have $p^\mu\propto U^\mu$, with $U^\mu$ being the unit-normalized 4-velocity. Thus, supposing the proportionality function varies slowly (the geometric optics approximation), we can express $\Phi_{\vec{p}}$ as $$\begin{aligned} \Phi_{\vec{p}}(x^\mu,\vec{n},S)=\frac{\vec{p}\cdot\vec{n}}{p^0}\mathcal{N}\nonumber\end{aligned}$$ where $p^0=-p\cdot W$ is the frequency[^5] measured in $S$ with $W$ being the 4-velocity of the system $S$. The momentum-specific flux density of any quantity $Q(x^\mu,p_i)$ carried by the photons (it does not need to be conserved) can be then expressed as $F^Q_{\vec{p}}(x^\mu,\vec{n},S)=Q(x^\mu,p_i)\Phi_{\vec{p}}(x^\mu,\vec{n},S)$. Subsequently, the spectral (energy-specific, often called simply “specific”) flux $F_\nu$ is defined as the energy $dE$, measured in the frame $S$, carried by the photons with the frequency from the interval of the length $d\nu$ centered on $\nu$, which cross over the time $dt$ through the area $dA$ located at $x^\mu$ and oriented in the direction $\vec{n}$, with only the crossings in the positive direction relative to $\vec{n}$ counted. Due to the dispersion relation (\[REE\_nc\]), this is obtained from $F^E_{\vec{p}}$ by integration over the momentum direction: $$\begin{aligned} \label{flux_definition} F_\nu=\frac{dE}{dAdtd\nu}=\frac{d\vert\vec{p}\vert}{d\nu}\frac{dE}{dAdtd\vert\vec{p}\vert}=2\pi p^0 \vec{p}^2\int_{\Omega_+}\mathcal{N}\vec{e}_p\cdot\vec{n}d\Omega_p =\int_{\Omega_+}I_\nu(x^\mu,\vec{e}_p,S)\vec{e}_p\cdot\vec{n}\end{aligned}$$ where $\vec{e}_p=\vec{p}/\vert\vec{p}\vert$ and $\Omega_+=\{\vec{e}_p, \vec{e}_p\cdot\vec{n}\geq 0\}$. In this calculation, we have used the approximate differentiation of the dispersion formula (\[REE\_nc\]), $p^0dp^0\approx \vert\vec{p}\vert d\vert\vec{p}\vert$, neglecting the differentiation of the plasma frequency, since it varies slowly. The quantity $I_\nu(x^\mu,\vec{n},S)=dE/dtdAd\nu d\Omega=2\pi p^0 \vec{p}^2\mathcal{N}(x^\mu,\vert\vec{p}\vert (\nu)n_i,S)$ is conventionally called specific (spectral) intensity or brightness [@Rybicki]. Spectral intensity is the flux per solid angle, incoming from the direction $-\vec{n}$ normal to the collecting area. In praxis, the astronomers usually measure the flux when dealing with pointlike sources and the intensity when dealing with extended sources. Flux coming from a pointlike source =================================== A pointlike source produces a 4-parametric family of rays, parametrized e.g. by the proper time of emission $\tau_S$ and the initial momentum of the photon $\vec{p}_S$ measured in the instantaneous rest frame of the source. As a preparation for calculation of the radiation flux from such source, let us first consider the distribution function $\mathcal{N}_1(x^\mu,p_i,S)$ for a single photon propagating along the worldline $X^\mu(\lambda)$, with the momentum $P_i(\lambda)$ measured in a chain of locally inertial systems $S(\lambda)$ set up along that worldline. Let the spacetime be sliced into a 1-parametric family of spacelike hypersurfaces $\Sigma_{\sigma^0}$ by the condition $F(x^\mu)=const\equiv\sigma^0$ for some $F$ such that $\sharp\nabla F$ is timelike everywhere. If these hyperslices are appropriately parametrized by $(\sigma^1,\sigma^2,\sigma^3)$, then $\sigma^\mu$ become new coordinates in the spacetime, therefore $X^\mu(\lambda)$ can be transformed to $\sigma_{\mbox{\scriptsize ph}}^\mu(\lambda)$. The photon number (\[ph\_num\]) should be obviously 1 for each $\Sigma_{\sigma^0}$, therefore $$\begin{aligned} \label{single_ray_N} \mathcal{N}_1(x^\mu,p_i,S)=\frac{\delta(\sigma^1-\hat{\sigma}^1)\delta(\sigma^2-\hat{\sigma}^2)\delta(\sigma^3-\hat{\sigma}^3)}{\sqrt{h}}{\delta^{(3)}}(\vec{p}-\hat{\vec{p}})\equiv {\delta^{(3)}}(\sigma-\hat\sigma){\delta^{(3)}}(\vec{p}-\hat{\vec{p}})\end{aligned}$$ where $h$ is the determinant of the metric tensor $h_{ij}$ induced on $\Sigma$ by its embedding in the spacetime. The hyperslice containing the observer located at $x^\mu$ has $\sigma^0=F(x^\mu)$ and $\hat{\sigma}^i=\sigma_{\mbox{\scriptsize ph}}^i(\lambda_0)$, $\hat{\vec{p}}=\vec{P}(\lambda_0)$, with $\lambda_0$ defined as the value of $\lambda$ at which the ray intersects the observer’s hyperslice, i.e. $\sigma^0=F(X^\mu(\lambda_0))$. Now, to calculate $\mathcal{N}$ for the pointlike source we need to sum over the single ray functions (\[single\_ray\_N\]) of all emitted photons. We will index these photons by $(\vec{p}_S,\tau_S$), so that $$\begin{aligned} \mathcal{N}(x^\mu,p_i,O)=\sum_{\mbox{\scriptsize photons}}\mathcal{N}_1(x^\mu,p_i,O;\vec{p}_S,\tau_S)=\int d^3p_Sd\tau_S\rho(\vec{p}_S,\tau_S)\mathcal{N}_1(x^\mu,p_i,O;\vec{p}_S,\tau_S)\nonumber\end{aligned}$$ where the transition from the sum to the integral introduced the parametric density of photons[^6] $\rho=dN/dp^3_Sd\tau_S$. From now on we will also denote the reference frame of the observer at the event of observation by $O$, reserving $S$ for the reference frame of the source at the event of emission. Plugging this into the flux definition (\[flux\_definition\]) we get $$\begin{aligned} F_{\nu_O}(x^\mu,\vec{n},O)=2\pi p^0_O\vec{p}^2_O \vec{n}\cdot \int d^3p_Sd\tau_S\rho\int_{\Omega_+}d\Omega_p \vec{e}_p {\delta^{(3)}}(\sigma-\hat\sigma){\delta^{(3)}}(\vec{p}-\hat{\vec{p}})\nonumber\end{aligned}$$ Here, $\Sigma$ is chosen in such way that $x^\mu\in\Sigma$ and the observer’s 4-velocity $W_O$ is perpendicular to $\Sigma$ at the event of observation $O$. For clarification, we emphasize that the parametric dependence of the quantities appearing in the formula looks like $\sigma(x^\mu)$, $\vec{p}=\vert \vec{p}\vert(\nu_O)\vec{e}_p$, $\hat\sigma(\vec{p}_S,\tau_S,\Sigma)$, $\hat{\vec{p}}(\vec{p}_S,\tau_S,\Sigma)$, with $\Sigma$ in the brackets indicating the dependence on the choice of $\Sigma$ (the hyperslice drawn through the event $O$). Integral over the momentum direction can be done immediately, yielding $$\begin{aligned} F_{\nu_O}(x^\mu,\vec{n},O)=2\pi p^0_O \vec{n}\cdot \int d^3p_Sd\tau_S \theta(\hat{\vec{p}}\cdot\vec{n})\rho \vec{e}_{\hat p} {\delta^{(3)}}(\sigma-\hat\sigma)\delta(\vert\vec{p}\vert-\vert\hat{\vec{p}}\vert)\nonumber\end{aligned}$$ with the $\theta$-function arising because the halfsphere $\Omega_+$ of the vectors $\vec{e}_p$ is centered on $\vec{n}$. The four remaining $\delta$-functions can be removed by the four remaining integrations. By solving REE (\[REE\]), one can find the ray connecting the event $S$ somewhere on the source’s worldline with the event $O$. This ray, which will now play the role of the reference ray, corresponds to $(\vec{p}_S^{\ {\mbox{\scriptsize ref}}},\tau_S^{{\mbox{\scriptsize ref}}})$ such that the hatted variables with these reference values satisfy the $\delta$-functions. To do the integral, we must first expand the hatted expressions around the reference ray, $$\begin{aligned} \hat{\sigma}^i(\vec{p}_S,\tau_S)\approx \sigma^i+\left.\frac{\partial\hat{\sigma}^i}{\partial s_a}\right\vert_{{\mbox{\scriptsize ref}}}(s_a-s_a^{{\mbox{\scriptsize ref}}})\nonumber\\ \vert\hat{\vec{p}}\vert(\vec{p}_S,\tau_S)\approx \vert\vec{p}\vert +\vec{e}_{\hat p}\cdot \left.\frac{\partial\hat{\vec{p}}}{\partial s_a}\right\vert_{{\mbox{\scriptsize ref}}}(s_a-s_a^{{\mbox{\scriptsize ref}}})\nonumber\end{aligned}$$ where we have for brevity combined $\vec{p}_S$ and $\tau_S$ into $s_a=(\vec{p}_S,\tau_S)$. After using these expansions, the formula for the flux becomes $$\begin{aligned} F_{\nu_O}=2\pi p^0_O \vec{n}\cdot \int d^3p_Sd\tau_S \theta(\hat{\vec{p}}\cdot\vec{n})\rho \vec{e}_{\hat p} {\delta^{(3)}}\left(\left.\frac{\partial\hat{\sigma}^i}{\partial s_a}\right\vert_{{\mbox{\scriptsize ref}}}(s_a-s_a^{{\mbox{\scriptsize ref}}})\right)\delta\left(\vec{e}_{\hat p}\cdot \left.\frac{\partial\hat{\vec{p}}}{\partial s_a}\right\vert_{{\mbox{\scriptsize ref}}}(s_a-s_a^{{\mbox{\scriptsize ref}}})\vert\right)\nonumber\end{aligned}$$ Utilizing the well known formula $\displaystyle\int d^nx\delta(\vec{V}_1\cdot\vec{x})..\delta(\vec{V}_n\cdot\vec{x})=\vert \det(\vec{V}_1..\vec{V}_n)\vert^{-1}$, we obtain $$\begin{aligned} F_{\nu_O}(x^\mu,\vec{n},O)=2\pi f p^0_O \left.\frac{\rho}{\vert\det\mathcal{V}_O\vert\sqrt{h_O}}\right\vert_{{\mbox{\scriptsize ref}}}\nonumber\end{aligned}$$ where we have denoted the cosine factor as $f$, $f=\theta(\vec{n}\cdot\hat{\vec{p}}\vert_{{\mbox{\scriptsize ref}}})\vec{n}\cdot\hat{\vec{p}}\vert_{{\mbox{\scriptsize ref}}}$, and introduced the matrix $\mathcal{V}$ defined as $$\begin{aligned} \label{Vmatrix_1} \mathcal{V}= \left( \begin{array}{cccc} \frac{\partial\hat{\sigma}^1}{\partial p_{1S}}& \frac{\partial\hat{\sigma}^2}{\partial p_{1S}}& \frac{\partial\hat{\sigma}^3}{\partial p_{1S}}& \vec{e}_{\hat p}\cdot\frac{\partial\hat{\vec{p}}}{\partial p_{1S}}\\ \frac{\partial\hat{\sigma}^1}{\partial p_{2S}}& \frac{\partial\hat{\sigma}^2}{\partial p_{2S}}& \frac{\partial\hat{\sigma}^3}{\partial p_{2S}}& \vec{e}_{\hat p}\cdot\frac{\partial\hat{\vec{p}}}{\partial p_{2S}}\\ \frac{\partial\hat{\sigma}^1}{\partial p_{3S}}& \frac{\partial\hat{\sigma}^2}{\partial p_{3S}}& \frac{\partial\hat{\sigma}^3}{\partial p_{3S}}& \vec{e}_{\hat p}\cdot\frac{\partial\hat{\vec{p}}}{\partial p_{3S}}\\ \frac{\partial\hat{\sigma}^1}{\partial \tau_S}& \frac{\partial\hat{\sigma}^2}{\partial \tau_S}& \frac{\partial\hat{\sigma}^3}{\partial \tau_S}& \vec{e}_{\hat p}\cdot\frac{\partial\hat{\vec{p}}}{\partial \tau_S}\\ \end{array} \right)\end{aligned}$$ The function $\rho$ is related to the spectral directional luminosity of the source $L_\nu^{{\mbox{\scriptsize dir}}}$, as one can easily see by a direct calculation, by the formula $$\begin{aligned} L_\nu^{{\mbox{\scriptsize dir}}}=\frac{dE}{d\tau_S d\nu_S d\Omega}=\frac{p^0_SdN}{d\tau_S \vec{p}_S^2d\vert\vec{p}_S\vert d\Omega}2\pi p^0_S\vert\vec{p}_S\vert=2\pi (p^0_S)^2\vert\vec{p}_S\vert\rho\nonumber\end{aligned}$$ Plugging this into the flux formula we obtain $$\begin{aligned} \label{flux_formula} F_{\nu_O}(x^\mu,\vec{n},O)=f \frac{L_\nu^{{\mbox{\scriptsize dir}}}}{\omega_S^3v_g^S}\frac{\omega_O}{\vert\det\mathcal{V}_O\vert\sqrt{h_O}}\end{aligned}$$ where $\omega=p^0$ is the angular frequency (in the units in which $\hbar=1$) and $v_g=\vert\vec{p}\vert/p^0$ is the velocity of the photon (group velocity of the electromagnetic wave, in plasma equal to the index of refraction). Note that the first fraction is evaluated at $S$ and the second one is evaluated at $O$. For an isotropic source, we can also replace the directional luminosity with $L_\nu/(4\pi)$, where $L_\nu=dE/d\tau d\nu$ is the total spectral luminosity. Expression for $\sqrt{h}\vert\det\mathcal{V}\vert$ in terms of connecting vectors ================================================================================= Consider an arbitrary connecting vector $\xi$ linking the reference ray with the infinitesimally close ray according to the formula $x^\mu_{{\mbox{\scriptsize cl}}}(\lambda)=x^\mu_{{\mbox{\scriptsize ref}}}(\lambda)+\epsilon\xi^\mu(\lambda)$. Using the relation $[\xi,p]\vert_{{\mbox{\scriptsize ref}}}=0$ implied by this definition, we can write $$\begin{aligned} (D\xi)^\mu=(\nabla_\xi p)^\mu=\frac{p^\mu_{{\mbox{\scriptsize cl}}}-p^\mu_{{\mbox{\scriptsize ref}}}}{\epsilon}+\omega^\mu_{\ \nu}(\xi)p^\nu\nonumber\end{aligned}$$ where the symmetry property of RLC connection, $\nabla_p\xi-\nabla_\xi p-[p,\xi]=0$, was used. If the indices are considered w.r.t. a locally inertial frame, then the $\omega$-term vanishes and we get $p^\mu_{{\mbox{\scriptsize cl}}}=p^\mu_{{\mbox{\scriptsize ref}}}+\epsilon(D\xi)^\mu$. Let us consider a connecting vector $\xi_1$ which links the reference ray and the ray emitted from the same event, but with the momentum $\vec{p}_{{\mbox{\scriptsize ref}}}+(\Delta p_1,0,0)$ (as measured in the rest frame of the source). From the aforementioned relations it is clear that the values of the connecting vector at the source event must be $\xi_1^\mu=0$ and $(D\xi_1)^i=(\Delta p_1/\epsilon, 0,0)$. The remaining component of $D\xi_1$, $(D\xi_1)^0=(\Delta p_1/\epsilon)(p^i/p^0)$, is determined from the normalization constraint (\[RDE\_nc\]). Vectors $\xi_2$, $\xi_3$ can be defined analogously. Furthermore, the vector $\xi_4$ can be defined for the photon emitted with the same momentum, but at the time $\tau_S^{{\mbox{\scriptsize ref}}}+\Delta\tau$. In this case, since it is displaced in spacetime from the event $S$, the momentum needs to be measured in the “locally inertial instantaneous rest frame” of the source, “LIIRF” for short. Initial values for the resulting connecting vector are $\xi_4^\mu=(\Delta\tau/\epsilon,\vec{0})$, $(D\xi_4)^i=\vec{0}$ and $(D\xi_4)^0=(\Delta\tau/\epsilon)(1/2p^0)W_S\cdot\mathcal{A}$. To use these vectors, one must evolve them to the observation event $O$ according to the RDE (\[RDE\]). At $O$, the connecting vector does not generally lie in $\Sigma$, so it cannot be used directly to extract the coordinate at which the neighbouring ray intersects $\Sigma$. To remedy this, we must force $\xi$ to point at the correct event on the neighbouring ray by shifting it along $p$ in the sense of (\[reparam\]). One can easily determine the correct value of the factor $\alpha$ such that $(\tilde\xi\cdot W_O)(O)=0$. We will denote this procedure as the “hat” operation $$\begin{aligned} \hat{\xi}=\xi-\left.\frac{\xi\cdot W_O}{p\cdot W_O}\right\vert_O p\nonumber\end{aligned}$$ Thus, $\sigma^i_{{\mbox{\scriptsize cl}}}=\sigma^i_{{\mbox{\scriptsize ref}}}+\epsilon\hat{\xi}^i(O)$ and we can calculate e.g. $\partial\hat{\sigma}^i/\partial p_{1S}\vert_{{\mbox{\scriptsize ref}}}$ as $$\begin{aligned} \left.\frac{\partial\hat{\sigma}^i}{\partial p_{1S}}\right\vert_{{\mbox{\scriptsize ref}}}=\frac{\hat{\sigma}^i_{{\mbox{\scriptsize cl}}}-\hat{\sigma}^i_{{\mbox{\scriptsize ref}}}}{\Delta p_1}=\frac{\epsilon\hat{\xi}^i_1(O)}{\Delta p_1}\nonumber\end{aligned}$$ where, naturally, $\hat{\xi}^i=\langle \hat{\xi},d\sigma^i\rangle$. Analogously, the momentum derivatives can be written as $$\begin{aligned} \left.\frac{\partial\hat{p}^i}{\partial p_{1S}}\right\vert_{{\mbox{\scriptsize ref}}}=\frac{\hat{p}^i_{{\mbox{\scriptsize cl}}}-\hat{p}^i_{{\mbox{\scriptsize ref}}}}{\Delta p_1}=\frac{\epsilon D\hat{\xi}^i_1(O)}{\Delta p_1}\nonumber\end{aligned}$$ In the above formula, the components of the momentum must be Cartesian, in accordance with the zero values of $\omega$ in LIIRF w.r.t. $W_O$. It is also worth noting that the derivatives are directly equal to the components of the connecting vectors after putting $\epsilon=\Delta s_a$. In the compactified notation $\chi=(\xi,D\xi)=((\xi^0,\vec{\xi}\ ),(D\xi^0,\vec{D\xi}))$, the initial conditions then become $\chi_i(S)=((0,\vec{0}),(p^i/p^0,\vec{e_i}))$ and $\chi_4(S)=((1,\vec{0}),((1/2p^0)(W_s\cdot\mathcal{A}),\vec{0}))$. The matrix $\mathcal{V}_O$ is expressed by evolving these vectors to $O$ and applying the hat, $$\begin{aligned} \label{Vmatrix_2} \mathcal{V}_O= \left( \begin{array}{cccc} \hat{\xi}_1^1 & \hat{\xi}_1^2 & \hat{\xi}_1^3 & \vec{e}_{\hat p}\cdot \vec{D\hat{\xi}}_1\\ \hat{\xi}_2^1 & \hat{\xi}_2^2 & \hat{\xi}_2^3 & \vec{e}_{\hat p}\cdot \vec{D\hat{\xi}}_2\\ \hat{\xi}_3^1 & \hat{\xi}_3^2 & \hat{\xi}_3^3 & \vec{e}_{\hat p}\cdot \vec{D\hat{\xi}}_3\\ \hat{\xi}_4^1 & \hat{\xi}_4^2 & \hat{\xi}_4^3 & \vec{e}_{\hat p}\cdot \vec{D\hat{\xi}}_4\\ \end{array} \right)(O)\end{aligned}$$ In the matrix, the first three columns depend on the parametrization of $\Sigma$ while the fourth column does not. One can, however, easily show that $\sqrt{h}\vert\det\mathcal{V}\vert$ is invariant w.r.t. the parametrization. Let $\sigma$ and $\tilde{\sigma}$ be two different parametrizations of $\Sigma$ with the Jacobi matrix $J^{\tilde i}_{\ j}=\partial\tilde\sigma^i/\partial\sigma_j$. Then, $h_{ij}=h_{\tilde i \tilde j}J^{\tilde i}_{\ i}J^{\tilde j}_{\ j}$, so that $h= \tilde h(\det J)^2$, and at the same time $\det\mathcal{V}$ transforms like $$\begin{aligned} \label{par_inv_calc} \det\tilde{\mathcal{V}}=\det(\hat{\xi}^{\tilde 1}_a,\hat{\xi}^{\tilde 2}_a,\hat{\xi}^{\tilde 3}_a,*)=J^{\hat 1}_{\ i}J^{\hat 2}_{\ j}J^{\hat 3}_{\ k}\det(\hat{\xi}^i_a,\hat{\xi}^j_a,\hat{\xi}^k_a,*)\nonumber\\ =\epsilon^{ijk}J^{\hat 1}_{\ i}J^{\hat 2}_{\ j}J^{\hat 3}_{\ k}\det(\hat{\xi}^1_a,\hat{\xi}^2_a,\hat{\xi}^3_a,*)=\det J\det\mathcal{V}\end{aligned}$$ where $*$ was used as a “wild card”[^7] symbol to stand for the last column. Putting this together we obtain $\sqrt{h}\vert\det\mathcal{V}\vert=\sqrt{\tilde h}\vert\det\tilde{\mathcal{V}}\vert$, as promised. Since the parametrization is irrelevant, we will from now on choose a parametrization in which $\partial_{\sigma^i}\cdot\partial_{\sigma^j}(O)=\delta_{ij}$, and to signify this we equip $\hat\xi$ with an arrow (one perk is that $h=1$ in this parametrization). On the conservation of phase volume {#par:on_cons_of_ph_vol} =================================== In the traditional Hamiltonian mechanics, the phase volume of a swarm of particles is conserved (Liouville theorem). We will now investigate how this is incorporated into an approach using the connecting vectors. Let us start with the conservation law[^8] (\[two\_ray\_cons\]) for two connecting vectors evolving along the same reference ray. Recall that the vector $\chi_p=(p,-(1/2)\mathcal{A})$ is a special solution of RDE, with the connecting vector (the first part of $\chi$) pointing at a different event on the reference ray. The two-vector conservation law becomes for $\chi_p$ the normalization constraint (\[RDE\_nc\]), with the value of the constant strictly fixed to zero. This implies that “hatting” the connecting vectors does not change value of the constant, $$\begin{aligned} \label{hatted_2_ray_cons} \hat{\xi}\cdot D\hat{\xi}^\prime -\hat{\xi}^\prime\cdot D\hat{\xi}=const=\xi\cdot D\xi^\prime -\xi^\prime\cdot D\xi\end{aligned}$$ The variety of different hat operations is distinguished by the event $P(\lambda)$ and the 4-velocity $W_P$ at that event, since $\hat\xi$ is chosen to be purely spatial w.r.t. $W_P$. The value of the conserved constant does not depend on the selection of $P$ or $W_P$, it stays fixed for a chosen neighbouring ray[^9]. Also note that at the event $P$ arrows can be written above all of the variables, since the scalar product is effectively calculated only in $W_P^{{\mbox{\scriptsize ort}}}$ space. Let us now choose six linearly independent solutions $\chi_a$, $a=1..6$, for the “truly” neighbouring rays, i.e. $\chi_a\neq c\chi_p$, and contract the corresponding conservation laws with the fully antisymmetric symbol: $$\begin{aligned} const=\epsilon^{a_1b_1..a_3b_3}(\vec{\hat{\xi}}_{a_1}\cdot \vec{D\hat{\xi}}_{b_1}-\vec{\hat{\xi}}_{b_1}\cdot \vec{D\hat{\xi}}_{a_1})..(\vec{\hat{\xi}}_{a_3}\cdot \vec{D\hat{\xi}}_{b_3}-\vec{\hat{\xi}}_{b_3}\cdot \vec{D\hat{\xi}}_{a_3}) =-2^3 3! \det(\vec{\hat\chi}_1..\vec{\hat\chi}_6)\nonumber\end{aligned}$$ where we have used the formula (\[det\_formula\]) from the appendix \[app:det\_formula\]. Thus, $\det(\vec{\hat\chi}_1..\vec{\hat\chi}_6)=const$ in two different ways: (i) along the ray, (ii) w.r.t. the change of the observer. Let us elaborate on how this works: at an arbitrary event $P(\lambda_1)$ take the six solutions and calculate the hat operation w.r.t. that exact event and a 4-velocity of your choice, followed by the calculation of the determinant. You will obtain the same number as you would with any other 4-velocity and you will also obtain the same number if you repeat the procedure at another event $P(\lambda_2)$ with any 4-velocity defined there. Finally, let us emphasize that the hatted connection vectors appearing in the determinants computed in this way at two different points are generally obtained by using two different values of $\alpha$, and the vector field $\hat{\xi}(\lambda)$ obtained by collecting the values of $\hat\xi$, going along the reference ray event by event, is therefore generally not a solution of the RDE. One can interpret $\det(\vec{\hat\chi}_1..\vec{\hat\chi}_6)$, recalling that $\vec{\hat\chi}$ hatted at the point $P$ establishes the separation of the neighbouring ray and the reference ray in the phase space set up in LIIRF w.r.t. $W_P$. The determinant gives the volume of the parallelepiped generated by these vectors, which is, up to the $\epsilon$’s, the infinitesimal phase volume that envelopes the corresponding rays. Therefore, its conservation along the ray rightfully deserves the status of (differential) Liouville theorem, stated in terms of the connecting vectors. Flux of radiation and relativistic distances {#par:flux_of_rad} ============================================ In this paragraph we proceed with the calculation of $\det\mathcal{V}$ to connect the spectral flux $F_\nu$ to the angular size distance and luminosity distance. Let us first realize that $$\begin{aligned} \det\mathcal{V}_O=\det(\vec{\hat{\xi}}_{(\mu)},\vec{D\hat{\xi}}_{(\mu)}\cdot \vec{e}_p)\equiv \det(\hat{\chi}_1^{{\mbox{\scriptsize proj}}}..\hat{\chi}_4^{{\mbox{\scriptsize proj}}})\nonumber\end{aligned}$$ where in the last step, the matrix was just transposed and the 4-dimensional subpart of $\hat\chi$ was denoted as the “projected part” (which it indeed is). Volume of such 4-dimensional parallelepiped can be written as the volume of 6-dimensional parallelepiped with “unit heights” in the two remaining directions by adding $\vec{\hat{\chi}}_5=(\vec{0},\vec{\rho}_1)$ and $\vec{\hat{\chi}}_6=(\vec{0},\vec{\rho}_2)$, with $\lbrace \vec{\rho}_1,\vec{\rho}_2,\vec{e}_p\rbrace$ forming an orthonormal triplet. Then $$\begin{aligned} \det(\hat{\chi}_1^{{\mbox{\scriptsize proj}}}..\hat{\chi}_4^{{\mbox{\scriptsize proj}}})(O)=\det(\vec{\hat{\chi}}_1..\vec{\hat{\chi}}_6)(O)\nonumber\end{aligned}$$ as one can also easily check by writing the determinants explicitly in the appropriate basis and applying Laplace expansion. By adding the temporal components to the last two $\chi$’s as $\chi_5(O)=((0,\vec{0}),(0,\vec{\rho}_1))$ and $\chi_6(O)=((0,\vec{0}),(0,\vec{\rho}_2))$ (so they are compatible with the normalization constraint (\[RDE\_nc\])), we can promote these vectors to the connecting vectors and evolve them back to $S$. Then, by the Liouville theorem from the previous paragraph, $$\begin{aligned} \det\mathcal{V}_O=\det(\vec{\hat{\chi}}_1..\vec{\hat{\chi}}_6)(S)=\det \left( \begin{array}{cccc} 0_{3\times 3}&-\vec{p}/p^0 &\vec{\hat \xi}_5 &\vec{\hat \xi}_6\\ I_{3\times 3}&(1/2p^0)\vec{\mathcal{A}} &\vec{D\hat \xi}_5 &\vec{D\hat \xi}_6\\ \end{array} \right)=\det(\vec{p}/p^0,\vec{\hat \xi}_5,\vec{\hat \xi}_6)(S)\nonumber\end{aligned}$$ where we have plugged in the ICs and performed Laplace expansion w.r.t. the columns containing the unit matrix $I$. If we denote $\vec{V}^\perp=\vec{V}-(\vec{V}\cdot\vec{e}_p)\vec{e}_p$, the expression can be further refined into $$\begin{aligned} \label{detV_1} \det\mathcal{V}_O=\frac{\vert\vec{p}_S\vert}{p^0_S}\det(\vec{\hat \xi}^\perp_5,\vec{\hat \xi}^\perp_6)(S)\equiv v_g^S\frac{dA_O}{\epsilon_5\epsilon_6}\end{aligned}$$ where the last equality can be considered the definition of $dA_O$, the area of a small but not pointlike source viewed by the observer under the specifically chosen solid angle (we will give the formal definition of $dA_O$ in paragraph \[ch: reciprocity\]). The quantities $\epsilon_5$, $\epsilon_6$ are small numbers used in the definitions of the vectors $\xi_5$, $\xi_6$, which will drop out of the final formulae. Also, we can define the solid angle under which the small source with the area $dA_O$ would be seen at $O$ as follows: in LIIRF of the observer, take an area of the bundle of at-$O$-converging rays an instant before the convergence occurs, and divide it by the square of its distance from the point of convergence, $$\begin{aligned} \frac{d\Omega_O}{\epsilon_5\epsilon_6}=\frac{\det(\vec{\hat{\xi}}_5^\perp,\vec{\hat{\xi}}_6^\perp)}{dr^2}(\lambda_O-d\lambda)\approx\frac{\det(\vec{D\hat{\xi}}_5^\perp,\vec{D\hat{\xi}}_6^\perp)d\lambda^2}{(\omega_Ov_g^O)^2d\lambda^2}(O)=\frac{1}{(\omega_Ov_g^O)^2}\nonumber\end{aligned}$$ where in the last equality we made use of the ICs, $\xi_{I+4}(O)=0$. Combining the last two formulae we obtain $$\begin{aligned} \det \mathcal{V}_O=\frac{v_g^S}{(\omega_Ov_g^O)^2}d_A^2\nonumber\end{aligned}$$ where $d_A$, given by $d_A^2=dA_O/d\Omega_O$, is the angular size distance, i.e. the distance determined from the angular size of a source with known dimensions by using the Euclidean formula (see also [@Schneider_Ehlers; @Schulze-Koops]). After inserting this result into the flux formula (\[flux\_formula\]), we obtain an expression for the flux in terms of the angular size distance, $$\begin{aligned} \label{flux_ang_dist} F_{\nu_O}(x^\mu,\vec{n},O)=f \frac{L_\nu^{{\mbox{\scriptsize dir}}}}{d_A^2}\frac{\omega_O^3(v_g^O)^2}{\omega_S^3(v_g^S)^2}\end{aligned}$$ We can also choose an alternative approach to calculate $\det\mathcal{V}$ by applying once again the conservation law (\[hatted\_2\_ray\_cons\]) on the $i$-th vector, $i=1,2,3$, and the $5$-th vector: $$\begin{aligned} \vec{\rho}_1 \cdot \vec{\hat\xi}_i(O)=-\vec{e}_i \cdot \vec{\hat\xi}_5(S)\nonumber\end{aligned}$$ where we have used the initial conditions (the hat is calculated easily in these cases, since $\xi=0$ results in $\alpha=0$). It is easy to see that mixing the vectors $\chi_1..\chi_3$ with an arbitrary rotation matrix does not change $\det \mathcal{V}$ (the calculation can be done very similarly to that in (\[par\_inv\_calc\]), so we will refrain from it here). We can choose to rotate these vectors in such way that $\vec{D\xi}_3(S)\propto \vec{p}$, which is equivalent to an appropriate choice of the basis $\vec{e}_i$, such that $\vec{e}_3\propto \vec{p}$, from the beginning. In such an appropriate basis we can write $$\begin{aligned} \label{detV_2} \det(\vec{\hat \xi}^\perp_5,\vec{\hat \xi}^\perp_6)(S)=\det(\vec{e}_I\cdot\vec{\hat \xi}_{J+4})(S)=\det(\vec{\rho}_J\cdot\vec{\hat \xi}_I)(O)=\det(\vec{\hat \xi}^\perp_1,\vec{\hat \xi}^\perp_2)(O)\equiv \frac{dA_S}{\epsilon_1\epsilon_2}\end{aligned}$$ for $I,J\in\{ 1,2\}$. The last equality is the definition of $dA_S$, the area placed at the observer that is swiped perpendicularly by the congruence generated by $\chi_1$, $\chi_2$. At the source, we can also calculate the spatial angle $d\Omega_S$ from which the congruence was sent, analogously to the calculation of $d\Omega_O$, $$\begin{aligned} \label{dOmega_definition} \frac{d\Omega_S}{\epsilon_1\epsilon_2}=\frac{\det(\vec{D\hat{\xi}}_1^\perp,\vec{D\hat{\xi}}_2^\perp)}{(\omega_Sv_g^S)^2}(S)=\frac{1}{(\omega_Sv_g^S)^2}\end{aligned}$$ Putting the results together we can express $\det \mathcal{V}_O$ as $$\begin{aligned} \det \mathcal{V}_O= \frac{1}{\omega_S^2 v_g^S}d_L^2\nonumber\end{aligned}$$ where $d_L$, given by $d_L^2=dA_S/d\Omega_S$, is conventionally called the (corrected) luminosity distance. By plugging this into the flux formula (\[flux\_formula\]), we obtain yet another way to calculate the flux $$\begin{aligned} \label{flux_lum_dist} F_{\nu_O}(x^\mu,\vec{n},O)=f \frac{L_\nu^{{\mbox{\scriptsize dir}}}}{d_L^2}\frac{\omega_O}{\omega_S}\end{aligned}$$ The comparison of the formulae (\[flux\_ang\_dist\]) and (\[flux\_lum\_dist\]) is a roundabout proof of the reciprocity theorem (the “plasma version” of it), which reads $$\begin{aligned} \label{recip_thm_1} d_L^2 =d_A^2\left(\frac{\omega_Sv_g^S}{\omega_Ov_g^O}\right)^2\end{aligned}$$ In section \[ch: reciprocity\] we will carry out the formal proof more concisely, extracting the important parts from the previous calculations. To finish off this section let us summarize – the spectral flux $F_\nu$ from a pointlike source can be calculated by the formula (\[flux\_formula\]) and $\sqrt{h}\det\mathcal{V}$ can be calculated in terms of the connecting vectors. One is free to choose out of three alternatives: 1. use the congruence $A=\{\chi_1..\chi_4\}$ with $\chi_i(S)=((0,\vec{0}),(p^i/p^0,\vec{e_i}))$ and $\chi_4(S)=((1,\vec{0}),((1/2p^0)(W_s\cdot\mathcal{A}),\vec{0}))$, then calculate the determinant according to (\[Vmatrix\_2\]) 2. use the congruence $B_S=\{\chi_1,\chi_2\}$ with $\chi_I(S)=((0,\vec{0}),(0,\vec{e}_I))$ and $\vec{e}_I\cdot \vec{p}=0$ and calculate the determinant by the means of (\[detV\_2\]) combined with (\[detV\_1\]) 3. use the congruence $B_O=\{\chi_5,\chi_6\}$ with $\chi_{I+4}(O)=((0,\vec{0}),(0,\vec{\rho}_I))$ and $\vec{\rho}_I\cdot \vec{p}=0$ and calculate the determinant by the means of (\[detV\_1\]) Note that the alternative (2) uses two of the four vectors from the alternative (1). The reduction in the number of required vectors from 4 to 2 is related to the “$dtd\nu$ conservation” discussed in the appendix \[app:dtdnu\_cons\]. Demonstration of the flux calculation {#par:demo} ===================================== Let us demonstrate the use of the formula (\[flux\_formula\]) on Kerr geometry. We will work in Boyer-Lindquist coordinates $(t,r,\theta,\phi)$ and in the locally non-rotating reference frame (LNRF), denoted as $(e_t,e_r,e_\theta,e_\phi)$. All relevant formulae can be found in the appendix of [@Sareny_Balek]. Around the black hole described by the Kerr metric we put plasma distributed in a disc-like way, $$\begin{aligned} \omega_{{\mbox{\scriptsize pl}}}^2=\frac{K}{r^2+r_c^2}\exp\left[-\frac{(\theta-\pi/2)^2}{s^2}\right]\end{aligned}$$ where $K$, $r_c$ and $s$ are constants of appropriate physical dimensions. To perform the numerical calculation, we need an ODE integrator for the set of 16 ODE (\[REE\]), (\[RDE\]). In fact, the number of equations is less than 16, since it can be reduced by the normalization constraints (\[REE\_nc\]), (\[RDE\_nc\]), and in Kerr geometry also by the conservation laws arising from the cyclic coordinates $(t,\phi)$. The numerical calculation should proceed as follows: 1. Choose the coordinates $x_S$ for the source event, the 4-velocity $W_S$ of the source and the ICs for the reference photon. 2. Construct the ICs for 7 basis solutions of RDE, out of which any solution can be constructed. We have chosen the basis $\chi\in\{Y_0..Y_6\}$ with $Y_0(S)=(p,Dp)$, $Y_i(S)=((0,\vec{e}_i),(-(\vec{e}_i\cdot \vec{Dp})/p^0,\vec{0}))$ and $Y_{i+3}=((0,\vec{0}),(\delta_i^1\vert\vec{p}\vert/p^0,\vec{e}_i))$, where the 1+3 split is done w.r.t. the observer’s frame of reference and $\vec{e}_i\cdot\vec{e_j}=\delta_{ij}$, $\vec{e}_1=\vec{p}/\vert\vec{p}\vert$. Recalculate these ICs into the LNRF basis using the standard formulae for boost (in our calculation this is not needed since the observer is at rest w.r.t. LNRF). 3. Integrate the 7 basis solutions along the reference ray to obtain $Y_a(\lambda)$, $a=0..6$. 4. Choose a set of observers $W_O(\lambda)$ along the reference ray. 5. Calculate $\vert\det\mathcal{V}\sqrt{h}\vert$ according to the alternative (1). The vectors $\chi_1..\chi_3$ are directly $Y_4..Y_6$ and the vector $\chi_4$ is to be found as an appropriate linear combination $\chi_4=\alpha^aY_a$ that satisfies the desired initial conditions. 6 conditions requiring that the observer-spatial components of $\chi_4$ vanish reduce the space of $\alpha$’s from $7D$ to $1D$, the condition $\xi_4^0=1$ fixes the $\alpha$ completely, and the component $D\xi^0_4$ will be automatically correct thanks to the fact that all $Y$’s satisfy (\[RDE\_nc\]). 6. Calculate $\vert\det\mathcal{V}\sqrt{h}\vert$ according to the alternative (2). The vectors $\chi_1, \chi_2$ are directly $Y_5,Y_6$. 7. Calculate $\vert\det\mathcal{V}\sqrt{h}\vert$ according to the alternative (3). The vectors $\chi_5,\chi_6$ must be found as appropriate linear combinations $\chi_{I+4}=\alpha_I^aY_a$. By applying the five “obvious” conditions for the zero vector components (the “hidden” sixth condition $\vec{\rho_I}\cdot\vec{p}=0$ can be safely ignored thanks to (\[RDE\_nc\])), we reduce the space of $\alpha$’s from 7D to 2D. The $\alpha$’s are uniquely fixed by the remaining two conditions (note: the last step can be avoided by choosing any linearly independent $\alpha$’s from the 2D space and using the shape invariance of $dA_O/d\Omega_O$ discussed in chapter \[ch: reciprocity\], which is how we actually proceeded). Note that to avoid the necessity of projections and hats in $\det(\vec{\hat \xi}^\perp_5,\vec{\hat \xi}^\perp_6)(S)$, one can use the trick of adding “unit heights” to the parallelepiped and calculate $(1/\vert\vec{p}\vert)\det(W_S,p,\xi_5,\xi_6)(S)$ instead. Here we illustratively calculate the evolution of $\sqrt{h}\vert\det\mathcal{V}\vert$ along two reference rays. The results are in figures \[fig:ray1\] and \[fig:ray2\]. The spin[^10] of the black hole is $a=0.5$, both rays start at $x_S=(0,2.5,\pi/2+0.1,0)$ and the source is at rest w.r.t. LNRF, as are all the observers along the ray. The reference ray is emitted under the angles $(\bar\theta=\pi/2,\bar\phi=0)$, which are standard spherical coordinates with the north pole direction given by $-e_r$ and the azimuth reference direction given by $e_\phi$. The rays depicted in the two figures differ by the initial frequency, $\omega_S=2\omega_{{\mbox{\scriptsize pl}}}(S)$ for the first ray and $\omega_S=1.37\omega_{{\mbox{\scriptsize pl}}}(S)$ for the second ray. The parameter of the ray $\lambda$ was integrated from[^11] 0 to 100 and the plots were created from 1000 points sampled along the path of integration. In the left panel we have plotted the trajectory of the ray in order to better visualize the situation. When $\sqrt{h}\vert\det\mathcal{V}\vert$ is close to zero, the flux is increased significantly and the light is strongly lensed (some instances of this can be seen in the figures). Also, the value of $\sqrt{h}\vert\det\mathcal{V}\vert$ is always zero at the source event, which is to be expected since flux near a “really point-like source” tends to infinity. The evolution of $\sqrt{h}\vert\det\mathcal{V}\vert$ obtained by the three different alternatives of calculation is depicted in different colors in the right panel. Since the results should be the same in all three cases, our calculation can also be used to test the accuracy of the ODE integration for RDE. From our results it seems that rays with the winding number $\Delta\phi/2\pi\gtrsim 1$, such as the one in fig. \[fig:ray2\], can be problematic for the numerics. In such cases with large numerical errors, the shape of the curves may also change for different choices of $\vec{e}_2,\vec{e}_3$ in the ICs (the freedom w.r.t. rotation around $\vec{p}$). Other ways to test the accuracy of the numerics include testing the conservation laws (\[hatted\_2\_ray\_cons\]) for different pairs of solutions and testing the phase volume conservation. Also, the one solution we know, $Y_0=(p,Dp)$, can be compared with the solution of REE. In this paragraph we have demonstrated how we can calculate for a relatively low price (7 integrations of a system of 16 ODE) the flux seen by any observer along the ray (on the graphs we have only depicted $\sqrt{h}\vert\det\mathcal{V}\vert$ for the LNRF observers, but at each point there exist also different observers parametrized by their velocity $\vec{v}$ w.r.t. the LNRF), and how we obtain simultaneously a credibility check on the numerics by comparing the results obtained from the three different calculations. [2]{} ![Photon with $\omega_S/\omega_{{\mbox{\scriptsize pl}}}(S)=2$: the trajectory (left) and the evolution of $\sqrt{h}\vert\det\mathcal{V}\vert$ (right). The sphere in the left figure is the event horizon and the coordinates $(r, \theta,\phi)$ are projected in the Euclidean way. The magenta dots denote ticks at $\lambda=5,10,15...$ For the calculations, we have used the ODE integrator `lsoda` from the Python library `scipy` [@scipy].[]{data-label="fig:ray1"}](traj_omp=2 "fig:"){width="\linewidth"} ![Photon with $\omega_S/\omega_{{\mbox{\scriptsize pl}}}(S)=2$: the trajectory (left) and the evolution of $\sqrt{h}\vert\det\mathcal{V}\vert$ (right). The sphere in the left figure is the event horizon and the coordinates $(r, \theta,\phi)$ are projected in the Euclidean way. The magenta dots denote ticks at $\lambda=5,10,15...$ For the calculations, we have used the ODE integrator `lsoda` from the Python library `scipy` [@scipy].[]{data-label="fig:ray1"}](detV_omp=2 "fig:"){width="\linewidth"}\ [2]{} ![Photon with $\omega_S/\omega_{{\mbox{\scriptsize pl}}}(S)=1.37$: the trajectory (left) and the evolution of $\sqrt{h}\vert\det\mathcal{V}\vert$ (right).[]{data-label="fig:ray2"}](traj_omp=1_37 "fig:"){width="\linewidth"} ![Photon with $\omega_S/\omega_{{\mbox{\scriptsize pl}}}(S)=1.37$: the trajectory (left) and the evolution of $\sqrt{h}\vert\det\mathcal{V}\vert$ (right).[]{data-label="fig:ray2"}](detV_omp=1_37 "fig:"){width="\linewidth"}\ Reciprocity theorem with the inclusion of plasma {#ch: reciprocity} ================================================ To prove the reciprocity theorem concisely, let us first provide a basic set of definitions, inspired by the above calculations. Consider a fixed reference ray in the spacetime and the set of its connecting vectors. **Def. 1:** *Isofrequential vertex beam* (IVB) w.r.t. the event $S$ and the 4-velocity $W_S$ is the set $B(S,W_S,\chi_1,\chi_2)= \{\chi; \chi=\chi_1\cos\sigma +\chi_2\sin\sigma, \sigma\in[0,2\pi )\}$ with $\chi_I(S)=((0,\vec{0}),(0,\vec{V}_I))$ and $\vec{V}_I\cdot \vec{p}=0$.\ On closer inspection, one finds that the beam defined in this way is indeed isofrequential in LIIRF w.r.t. $W_S$ ($D\xi\cdot W_S (S)=0$) and forms a vertex at $S$ ($\vec{\hat{\xi}}(S)\equiv \vec{\xi}(S)=\vec{0}$). Also, there exists a 4-parametric class of IVBs for a given $S$, $W_S$, parametrized by $GL(2,\mathbb{R})$ matrices $\alpha_{IJ}$, i.e. if $\chi_I$ generate an IVB, then so do $\alpha_{IJ}\chi_J$. **Def. 2:** The *solid angle* $d\Omega_S$ measured by the observer $W_S$ at $S$ on the IVB $B(S,W_S,\chi_1,\chi_2)$ is given by $$\begin{aligned} \frac{d\Omega_S}{\epsilon^2}=\frac{\vert\det (\vec{D\xi}^\perp_1,\vec{D\xi}^\perp_2 )(S)\vert}{(\omega_Sv_g^S)^2}\nonumber\end{aligned}$$ Note that we have used $\epsilon^2$ instead of $\epsilon_1$ and $\epsilon_2$, which can always be done by writing $\epsilon_I=c_I\epsilon$ and merging $c_I$ into $\xi_I$. Also, in this way the whole information about the shape of the beam is encoded in $\chi$’s and no part of it is contained in $\epsilon$’s. Should one wish to repeat the calculations of this paragraph with two different $\epsilon$’s, one would need to include them in the definition of the IVB. **Def. 3:** The *area* $dA_S(O,W_O)$ perpendicular to the ray direction, measured by the observer $W_O$ at the event $O$ on the IVB $B(S,W_S,\chi_1,\chi_2)$, is given by $$\begin{aligned} \frac{dA_S}{\epsilon^2}=\vert\det (\vec{\hat \xi}^\perp_1,\vec{\hat \xi}^\perp_2 )(O)\vert\nonumber\end{aligned}$$ Note that although $dA_S$ and $d\Omega_S$ are dependent on the shape of the IVB, i.e. on the choice of $\chi_1,\chi_2$, they both transform the same way. For $dA_S$ we have $$\begin{aligned} \frac{dA_S^\prime}{\epsilon^2} = \vert\det (\vec{E}_I\cdot\vec{\hat \xi}^\prime_J)(O)\vert=\vert\det (\vec{E}_I\cdot\vec{\hat \xi}_K\alpha_{KJ})(O)\vert =\vert\det\alpha\vert \frac{dA_S}{\epsilon^2}\nonumber\end{aligned}$$ where $\vec{E}_I$ is an orthonormal basis of $(W_O,k)^{{\mbox{\scriptsize ort}}}$, and a similar calculation shows that $d\Omega_S^\prime=\vert\det\alpha\vert d\Omega_S$. Hence, the ratio $dA_S/d\Omega_S$ is a shape-invariant quantity with the dimension of distance squared. If the vertex event is chosen to be $S$ as above, we talk about the luminosity distance, and if the vertex event is $O$, we get the angular size distance. Let us now state the reciprocity theorem, previously also formulated in [@Ellis; @Schneider_Ehlers; @Etherington; @Ellis_Etherington; @Temple] for vacuum and in [@Schulze-Koops] for plasma. **The reciprocity theorem:** Consider an arbitrary spacetime filled with plasma and light propagating in the approximation of geometric optics with the index of refraction $n^2=1-\omega_{{\mbox{\scriptsize pl}}}^2/\omega^2$. Let the reference ray connect the source event $S$ and the observation event $O$ associated with the respective 4-velocities $W_S$, $W_O$. Then, the quantities $dA_S$, $d\Omega_S$ measured on an arbitrary IVB $B(S,W_S,\chi_1,\chi_2)$ and the quantities $dA_O$, $d\Omega_O$ measured on an arbitrary IVB $B^\prime(O,W_O,\chi^\prime_1,\chi^\prime_2)$ are related as $$\begin{aligned} \label{recip_thm_2} \frac{dA_S}{d\Omega_S}=\frac{dA_O}{d\Omega_O}\left(\frac{\omega_Sv_g^S}{\omega_Ov_g^O}\right)^2\qquad\Leftrightarrow\qquad d_L^2=d_A^2\left(\frac{\omega_Sv_g^S}{\omega_Ov_g^O}\right)^2\end{aligned}$$ Proof: The ratio $dA/d\Omega$ is shape-invariant, therefore it is sufficient to prove the relation (\[recip\_thm\_2\]) for special IVBs generated by $\chi_I$, $\chi^\prime_I$ such that $\chi_I(S)=((0,\vec{0}),(0,\vec{e}_I))$ and $\chi^\prime_I(O)=((0,\vec{0}),(0,\vec{e}^{\ \prime}_I))$, where $\vec{e}_I$ form an orthonormal basis of $\vec{k}^{{\mbox{\scriptsize ort}}}$ at $S$ and $\vec{e}^{\ \prime}_I$ do so at $O$. Applying the conservation law (\[hatted\_2\_ray\_cons\]) on one non-primed and one primed vector yields $$\begin{aligned} \vec{e}_I\cdot \vec{\hat \xi}^\prime_J(S)=-\vec{\hat\xi}_I\cdot\vec{e}^{\ \prime}_J(O)\nonumber\end{aligned}$$ which may be used to interrelate the two areas as $$\begin{aligned} \label{dA_relation} \frac{dA_S}{\epsilon^2}=\vert\det(\vec{\hat\xi}_I\cdot\vec{e}_J^{\ \prime})(O)\vert=\vert\det(-\vec{\hat\xi}_J^\prime\cdot\vec{e}_I)(S)\vert=\frac{dA_O}{\epsilon^2}\end{aligned}$$ where we have utilized the standard rules for the sign extraction and transposition in determinants. Thanks to the ICs, the determinants in the definitions of $d\Omega_S$ and $d\Omega_O$ are equal to $1$ and we obtain two formulae $(d\Omega_*/\epsilon^2)(\omega_*v_g^*)^2=1$ (with the “wild card” $*$ replaced by either $S$ or $O$). By placing these $1$’s into the denominator of (\[dA\_relation\]) we obtain the shape-invariant formulae (\[recip\_thm\_2\]), which finishes the proof. Note: There are other beams, apart from IVBs, for which the reciprocity relation holds and can be proved by a method analogous to the one above. As an example we mention the “*isochronously arriving beam*” (IAB), generated by a pair of vectors satisfying $\xi(S)=0$, $\vec{\hat{\xi}}\cdot\vec{p}(O)=0$ (we have termed the beam “isochronously arriving” because the second condition says that the rays belonging to the beam arrive at the collecting area placed at $O$ perpendicularly to the ray at the same instant from the point of view of the observer to whom the hat is referring). We have not investigated what are the most general beams that satisfy the reciprocity relation, since only the IVBs were relevant to our cause. On the other hand, we did not want to hide from the reader that other beams of such kind may exist too (and do exist, as the reader can easily verify on the case of the IABs). Discussion {#discussion .unnumbered} ========== The flux formula obtained in this paper is an alternative to the calculation of the flux within the traditional approach, explained in books on gravitational lensing such as [@Schneider_Ehlers; @Schneider_Kochanek] and in the paper [@Cunningham_Bardeen]. Although these methods are usually good enough for astrophysical situations, their applicability to a generic problem could be questionable. Even in the situations where the usual formulae are applicable, our approach could at least provide an important consistency check. We have not tested the computational efficiency of our approach against the conventional methods, but we think that it might be slightly slower than the conventionally used method due to the need of additional numerical integration. However, this type of calculation could prove especially efficient if one was interested in calculating flux for all possible observers along the given ray, since the alternatives 1 and 2 from the end of paragraph \[par:flux\_of\_rad\] provide the results for all these observers for a relatively small price, by one integration only. The reciprocity theorem with the inclusion of plasma was also investigated in [@Schulze-Koops]. The authors formulate the reciprocity relation (equation (73) of their paper) which looks the same as our equation (\[recip\_thm\_2\]). The difference is, however, that their result is (according to the commentary following equation (73)) only valid for the observers with 4-velocity orthogonal to the chosen Sachs basis (the definition of this term can be found in their paper), and for other observers the relation requires additional calculations. Our results, on the other hand, show that the reciprocity relation is valid for any two observers. Their technique is also vastly different from the calculation presented here. They equip the spacetime with an additional conformally rescaled metric $\tilde g_{\mu\nu}=\omega_{{\mbox{\scriptsize pl}}}^2g_{\mu\nu}$, in which the light rays in plasma are timelike geodesics. Subsequently they construct the Sachs basis, then derive Sachs equations in the tilded metric, and finally carry the equations back into the original metric. Nevertheless, some similar features in our calculations can also be detected, the most noteworthy ones being the involvement of the deviation equation (reducing to the Jacobi equation in their case) and the occurrence of the conservation law (53) which is somewhat similar to our equation (\[hatted\_2\_ray\_cons\]). We believe that the different viewpoints in their work and ours could help an interested reader to gain a deeper insight into the topic. Conclusion {#conclusion .unnumbered} ========== We have investigated the propagation of radiation from a pointlike source. The radiation propagated through cold plasma with infinite conductivity and was considered in the generally relativistic framework and the geometrical optics limit. A natural language for the description of such problem proved to be the kinetic theory. As a main result of the paper, we have provided a systematic derivation of the formula for the spectral flux of radiation of a pointlike source in an arbitrary spacetime. The formula allows us to calculate the flux in terms of connecting vectors evolved along the reference ray that connects the emission event with the observation event. A major advantage of this formula is that one only needs to solve ordinary differential equations and is not forced to deal with partial differential equations, as one might first expect. The procedure is summarized at the end of paragraph \[par:flux\_of\_rad\] and a demonstration of it can be found in paragraph \[par:demo\]. Another topic that emerged naturally in the course of the calculations was Etherington’s reciprocity theorem. The plasma version of it was recently also explored in [@Schulze-Koops] by a completely different technique than ours. We have provided a “roundabout” proof which emerged as part of the flux calculation and also a “concise” proof which only focuses on the concepts important to the theorem in question. We have also pointed out possible extensions from the “isofrequential vertex beams” that we have investigated to other types of beams. Finally, as a part of the discussion we have compared our approach with the approach in [@Schulze-Koops]. Finally, we would like to point out our analysis of the two-ray conservation law (\[hatted\_2\_ray\_cons\]) and its connection to the Liouville theorem known from Hamiltonian mechanics. This analysis leads to a variant of Liouville theorem formulated in terms of connecting vectors, in the form of the conservation law for the determinant specified in section \[par:on\_cons\_of\_ph\_vol\]. In our future work we would like to use the results of this paper in the calculation of a light curve of a gravitational lensing event. The light source could be a distant star, a source orbiting the lens or a source approaching the lens on a hyperbolic trajectory, and the lens could be e.g. Kerr black hole surrounded by plasma (which would make it a “gravitational-optical” lens). Acknowledgements {#acknowledgements .unnumbered} ================ I am thankful to Vladimír Balek for his invaluable insight and advice. Utility formula for even-dimensional matrix determinant {#app:det_formula} ======================================================= Let $\chi_1,..\chi_{2n}$ be $2n$ $2n$-dimensional vectors written as $\chi_k\equiv(\xi_k,\pi_k)$, with $\xi_k$, $\pi_k$ being $n$-dimensional vectors. The following relation is satisfied $$\begin{aligned} \label{det_formula} pf(n)2^nn!\det(\chi_1..\chi_{2n})=\epsilon^{\mu_1\nu_1..\mu_n\nu_n}(\xi_{\mu_1}\cdot\pi_{\nu_1}-\xi_{\nu_1}\cdot\pi_{\mu_1})..(\xi_{\mu_n}\cdot\pi_{\nu_n}-\xi_{\nu_n}\cdot\pi_{\mu_n})\end{aligned}$$ where $\mu_k$, $\nu_k$ run over $1..2n$ and $pf(n)$ is a permutation factor assuming the values $1,-1,-1,1,$ $1,-1,-1,..$ for $n=1,2,3..$ according to the recurrent formula $pf(n)=pf(n-1)(-1)^{n-1}$ with the initial condition $pf(1)=1$. Proof: The proof relies on the philosophy “calculate from both sides and meet somewhere in between”. On the left hand side there appears $$\begin{aligned} \det(\chi_1..\chi_{2n})=\epsilon^{\alpha_1..\alpha_{2n}}\chi_{\alpha_1}^1..\chi_{\alpha_{2n}}^{2n}=\epsilon^{\mu_1..\mu_n\nu_1..\nu_n}\xi^1_{\mu_1}..\xi^n_{\mu_n}\pi^1_{\nu_1}..\pi^n_{\nu_n}\end{aligned}$$ where the definition of determinant was used, followed by the index renaming and separating $\chi$ to its two parts. Note that the component indices are the upper ones and the indices for different $\chi$’s are the lower ones throughout the whole calculation. The right hand side is $$\begin{aligned} \epsilon^{\mu_1\nu_1..\mu_n\nu_n}(\xi_{\mu_1}\cdot\pi_{\nu_1}-\xi_{\nu_1}\cdot\pi_{\mu_1})..(\xi_{\mu_n}\cdot\pi_{\nu_n}-\xi_{\nu_n}\cdot\pi_{\mu_n})=pf(n)2^n\epsilon^{\mu_1..\mu_n\nu_1..\nu_n}\xi_{\mu_1}\cdot\pi_{\nu_1}..\xi_{\mu_n}\cdot\pi_{\nu_n}\nonumber\end{aligned}$$ where the permutation factor arises as the parity of the permutation $\mu_1\nu_1..\mu_n\nu_n\mapsto\mu_1..\mu_n\nu_1..\nu_n$. For $n=1$ it is obviously $1$ and for $n>1$ the whole procedure can be carried out in two steps $\mu_1\nu_1..\mu_n\nu_n\mapsto \mu_1..\mu_{n-1}\nu_1..\nu_{n-1}\mu_n\nu_n\mapsto\mu_1..\mu_n\nu_1..\nu_n$, with the first permutation obviously giving $pf(n-1)$ and the second permutation giving $(-1)^{n-1}$ (since $\mu_n$ must “jump over” $n-1$ of $\nu$’s). Put together this gives the recurrent prescription for $pf$ stated above, which may be read “change the value from the previous one if $n$ is even”. We can now continue the calculation of the right hand side by writing the scalar products explicitly with the indices $i_k=1..n$, $$\begin{aligned} pf(n)2^n\epsilon^{\mu_1..\mu_n\nu_1..\nu_n}\xi^{i_1}_{\mu_1}\pi^{i_1}_{\nu_1}..\xi^{i_n}_{\mu_n}\pi^{i_n}_{\nu_n} =pf(n)2^n\sum_{perms.}\epsilon^{\mu_1..\mu_n\nu_1..\nu_n}\xi^{s_1}_{\mu_1}\pi^{s_1}_{\nu_1}..\xi^{s_n}_{\mu_n}\pi^{s_n}_{\nu_n}\nonumber\\ =pf(n)2^nn!\epsilon^{\mu_1..\mu_n\nu_1..\nu_n}\xi^1_{\mu_1}\pi^1_{\nu_1}..\xi^n_{\mu_n}\pi^n_{\nu_n}\nonumber\end{aligned}$$ where we have replaced, using the antisymmetry, the sum over multiple indices with the sum over the permutations ($s_1..s_n$ being a permutation of $1..n$) in the first step. In the second step, further permutation of each term and renaming of the indices produced $n!$ identical terms. Spotting the common factor at the end of l.h.s. and r.h.s. calculations we can now easily find the desired formula, which finishes the proof. Conservation of $dtd\nu$ {#app:dtdnu_cons} ======================== When calculating a quantity such as flux, the neighbouring photons’ parameters $t_{\mbox{\scriptsize inc}}$ (time of incidence on $dA$), $\Delta\vec{x}^\perp$ (perpendicular shift from the reference photon’s point of incidence) and $\nu_{\mbox{\scriptsize inc}}$ (frequency) are supposed to be from the Cartesian product of intervals determined by $dA$, $d\nu$, $dt$. These photons can be described by various connecting vectors, which form a rectangular 4-dimensional box in the subspace of phase space dealing with position and frequency (wavenumber). One can get into that subspace using the projection $(\xi,D\xi)^{{\mbox{\scriptsize proj}}}=(\vec{\hat\xi},D\hat\xi^\parallel)$, where $\parallel$ denotes the part parallel to $\vec{p}$. The edges of the rectangular box will be of lengths $dA$ (product of two edges), $dL\simeq v_gdt$, $d\vert\vec{p}\vert\simeq 2\pi v_g^{-1}d\nu$ (in the approximation of geometric optics). In this sense, the quantities such as $dA, dt, d\nu$ cannot be defined for one specific infinitesimal bunch of photons, because an originally rectangular 4-D box would be deformed in the course of the evolution of the beam and we lose the ability to resolve it properly into these quantities. How is it then possible that $dtd\nu$ is “conserved” and can be cancelled out of the calculations, as in paragraph \[ch:naive\_approach\]? To properly investigate this issue, let us compare the formula (\[flux\_formula\]) to the definition $F_\nu=dE_O/(dA_S dt_Od\nu_O)$. At the point $S$, take a rectangular box $B_S$ in the phase space and at the point $O$ take also a rectangular box $B_O$: both boxes contain the same number of photons $dN$, but not necessarily the same photons. Then we can write[^12] $$\begin{aligned} \label{app2:flux_analysis} \frac{L_{\nu_S}^{{\mbox{\scriptsize dir}}}}{\omega_S}\omega_O=\frac{dN\omega_O}{d\nu_Sdt_Sd\Omega_S}\qquad\Rightarrow\qquad \omega_S^2v_g^S\vert\det\mathcal{V}\vert=\frac{d\nu_Odt_OdA_S}{d\nu_Sdt_Sd\Omega_S}\end{aligned}$$ The $\det\mathcal{V}$ is calculated in the position-wavenumber subspace of the phase space generated by the previously defined projection $\chi^{{\mbox{\scriptsize proj}}}$. The projected vectors $\chi_1^{{\mbox{\scriptsize proj}}}..\chi_4^{{\mbox{\scriptsize proj}}}$ are, in general, not perpendicular to each other, therefore we must take linear combinations of them that are perpendicular to each other and do not change the value of $\det\mathcal{V}$ (the latter condition is needed in order to avoid changing the flux). There exists a 3-parametric family of possibilities to do this, since a generic 4D rectangular box has 4 free parameters (lengths of the independent edges) and the volume is fixed. Thus, the values of $dA_S$, $dt_O$ and $d\nu_O$ are not uniquely fixed. However, by fixing the choice of $dA_S$ in the way indicated in equation (\[detV\_2\]), the product $dt_Od\nu_O$ becomes also uniquely fixed[^13] and the discussion about its conservation may begin. To extract $dA_S$ from $\det\mathcal{V}$ in a way that conserves it, one needs to create a $2\times 2$ block of zeros in the upper right corner of the matrix $\mathcal{V}$ by changing $\hat{\chi}_3,\hat{\chi}_4$ to $\hat{\tilde{\chi}}_3,\hat{\tilde{\chi}}_4$, where $\hat{\tilde{\xi}}_{I+2}\cdot \vec{e}_J=0$ (the vectors $\vec{e}_J$ span $\vec{p}^{\ {\mbox{\scriptsize ort}}}$). The change is made by the transformation $\tilde{\chi}_{I+2}=\chi_{I+2}+\alpha_{IJ}\chi_J$ with $\alpha_{IJ}$ defined as $$\begin{aligned} \alpha_{IJ}=-(\hat{\xi}_{I+2}\cdot\vec{e}_K)A_{JK}\nonumber\end{aligned}$$ where $A$ is the inverse matrix to $\hat{\xi}_{J}\cdot\vec{e}_K$ (the matrix is invertible iff its determinant is non-zero, ergo iff $dA_S\neq 0$). After the transformation we can write $$\begin{aligned} \det\mathcal{V}=\det(\vec{\hat\xi}_1^\perp,\vec{\hat\xi}_2^\perp)\det \left( \begin{array}{cc} \hat{\tilde\xi}_3^\parallel & \hat{\tilde\xi}_4^\parallel \\ D\hat{\tilde\xi}_3^\parallel & D\hat{\tilde\xi}_4^\parallel \end{array} \right)\qquad\Rightarrow\qquad dt_Od\nu_O\propto\hat{\tilde\xi}_3^\parallel D\hat{\tilde\xi}_4^\parallel- \hat{\tilde\xi}_4^\parallel D\hat{\tilde\xi}_3^\parallel \nonumber\end{aligned}$$ Recovery of the proportionality factor involves evaluation of the remaining terms in the equation (\[app2:flux\_analysis\]) by collecting bits and pieces of information from the main part of the article: use $\epsilon_a=\Delta s_a$ (before the equation (\[Vmatrix\_2\])) for $dt_S=\epsilon_4$, and combine that with the proportionality $\vec{D\xi}_3(S)\propto\vec{p}$ (before the equation (\[detV\_2\])) and with the differential of the dispersion formula (after the equation (\[flux\_definition\])) into $d\nu_S=(2\pi)^{-1}v_g^Sd\vert\vec{p}_S\vert=(2\pi)^{-1}v_g^S\epsilon_3$. Then retrieve $dA_S$ from the equation (\[detV\_2\]) and $d\Omega_S$ from (\[dOmega\_definition\]), stuff all these data into (\[app2:flux\_analysis\]) and after vast cancellations obtain $$\begin{aligned} dt_Od\nu_O=\frac{\omega_S^2v_g^S\vert\det\mathcal{V}\vert d\nu_S dt_Sd\Omega_S}{dA_S}=\frac{\epsilon_3\epsilon_4}{2\pi}\left\vert\hat{\tilde\xi}_3^\parallel D\hat{\tilde\xi}_4^\parallel- \hat{\tilde\xi}_4^\parallel D\hat{\tilde\xi}_3^\parallel\right\vert\end{aligned}$$ The expression in the absolute value can be written as an instance of the conserving quantity (\[hatted\_2\_ray\_cons\]), and we can get rid of parallels, arrows and hats $$\begin{aligned} \hat{\tilde\xi}_3^\parallel D\hat{\tilde\xi}_4^\parallel- \hat{\tilde\xi}_4^\parallel D\hat{\tilde\xi}_3^\parallel=\vec{\hat{\tilde\xi}}_3\cdot D\vec{\hat{\tilde\xi}}_4- \vec{\hat{\tilde\xi}}_4\cdot D\vec{\hat{\tilde\xi}}_3 = \tilde\xi_3\cdot D\tilde\xi_4- \tilde\xi_4\cdot D\tilde\xi_3\nonumber\end{aligned}$$ Note that the conservation law (\[hatted\_2\_ray\_cons\]) is actually a bilinear antisymmetric mapping $\mathcal{F}:\mathcal{S}\times\mathcal{S}\mapsto \mathbb{R}$ acting on two objects from the space $\mathcal{S}$ of the solutions of RDE and out of $\chi_1..\chi_4$, the only pair with a non-zero value is $\mathcal{F}(\chi_3,\chi_4)=v_g^S$, as can be easily calculated from the ICs. Moreover, we can drop the tildes out of the last expression appearing above, since $\mathcal{F}(\tilde\chi_3,\tilde\chi_4)=\mathcal{F}(\chi_3,\chi_4)+\mbox{zero terms}$. The results can be put together into the final equation $$\begin{aligned} \label{app2:dtdnu_formula} dt_Od\nu_O=\frac{\epsilon_3\epsilon_4}{2\pi}\vert\mathcal{F}(\chi_3,\chi_4)\vert=\frac{\epsilon_3\epsilon_4}{2\pi}v_g^S\end{aligned}$$ Note that the expressions for $d\nu_S$ and $dt_S$ we have found earlier give the same result as choosing the event $O$ to be $S$ in the formula (\[app2:dtdnu\_formula\]) does. Thus $dt_Od\nu_O$ is a constant independent of $\lambda$ or the choice of $W_O(\lambda)$ and the conservation law $dt_Sd\nu_S=dt_Od\nu_O$ really holds, Q.E.D. [99]{} Schneider, P.; Ehlers, J.; Falco, E. E.: *Gravitational Lenses*, 1999, Springer Schneider, P.; Kochanek, C.; Wambsganss, J.: *Gravitational Lensing: Strong, Weak and Micro*, 2006, Springer Cunningham, C. T.; Bardeen, J. M.: *The Optical Appearance of a Star Orbiting an Extreme Kerr Black Hole*, 1973, Astrophysical Journal, Volume 183 Sárený, M.; Balek, V.: *Effect of black hole-plasma system on light beams*, 2019, General Relativity and Gravitation, Volume 51, Issue 11 Etherington, I. M. H.: *Republication of: LX. On the definition of distance in general relativity*, 2007, General Relativity and Gravitation, Volume 39, Issue 7 Temple, G.: *New Systems of Normal Co-ordinates for Relativistic Optics*, 1938, Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences, Volume 168, Issue 932 Ellis, G. F. R.: *Relativistic Cosmology*, republication, 2009, General Relativity and Gravitation, Volume 41, Issue 3 Ellis, G. F. R.: *On the definition of distance in general relativity: I. M. H. Etherington (Philosophical Magazine ser. 7, vol. 15, 761 (1933))*, 2007, General Relativity and Gravitation, Volume 39, Issue 7 Schulze-Koops, K.; Perlick, V.; Schwarz, D. J.: *Sachs equations for light bundles in a cold plasma*, 2017, Classical and Quantum Gravity, Volume 34, Issue 21 Synge, J. L.: *Relativity: The general theory*, 1960, Amsterdam: North-Holland Pub. Co. Bisnovatyi-Kogan, G.; Tsupko, O.: *Gravitational Lensing in Presence of Plasma: Strong Lens Systems, Black Hole Lensing and Shadow*, 2017, Universe, vol. 3, issue 3 Rogers, A.: *Escape and trapping of low-frequency gravitationally lensed rays by compact objects within plasma*, 2017, Monthly Notices of the Royal Astronomical Society, Volume 465, Issue 2 Tsupko, O. Y.; Bisnovatyi-Kogan, G. S.: *Hills and holes in the microlensing light curve due to plasma environment around gravitational lens*, 2020, Monthly Notices of the Royal Astronomical Society, Volume 491, Issue 4 Fecko M.: *Differential Geometry and Lie Groups for Physicists*, 2006, Cambridge University Press Misner C. W.; Thorne K. S.; Wheeler J.A.: *Gravitation*, 1973, W. H. Freeman and Company Thorne, K. S.; Blandford, R. D.: *Modern classical physics : optics, fluids, plasmas, elasticity, relativity, and statistical physics*, 2017, Princeton University Press Rybicki G. B.; Lightman, A. P.: *Radiative processes in astrophysics*, 2004, Weinheim: Wiley-VCH Verlag https://docs.scipy.org/doc/scipy/reference/generated/scipy.integrate.ode.html [^1]: [email protected] [^2]: the components of the Riemann tensor are then defined as $R^\mu_{\ \nu\kappa\lambda}=\langle e^\mu,R(e_\kappa,e_\lambda)e_\nu\rangle$, see [@Fecko] for more details [^3]: we leave $S$ in the argument list of $\mathcal{N}$ to make clear what is the frame of reference in which $\vec{p}$ is measured [^4]: meaning its normal is timelike everywhere [^5]: actually the energy, but the two quantities coincide in Planck units [^6]: if scattering, creation or extinction of photons is involved, $\rho$ becomes also a function of the ray parameter $\lambda$ and one should proceed here more carefully, employing once again the “$\Sigma$ slice technique” and taking the value of $\rho$ at the point of intersection [^7]: wild card = (in card games) a card that has no value of its own and takes the value of any card that the player chooses (OALD) [^8]: In the traditional Hamiltonian mechanics, this law can also be viewed as a consequence of the symplectic form conservation along the Hamiltonian flow (see e.g. [@Fecko]). [^9]: In fact, one could even hat the non-primed and the primed solutions with hats w.r.t. two different events or 4-velocities. We will, however, only use the variant with both hats being of the same kind. [^10]: we give all the following quantities in the dimensionless units $G=c=M=1$, with $M$ being the mass of the black hole, as the convention $\hbar=1$ is not needed for numerical calculations [^11]: thanks to the scaling symmetry of the equations, $\omega_S$ can be scaled away, so these values should be multiplied by $\omega_S^{-1}$. For more informations on this issue see the end of §4.2 in [@Sareny_Balek]. [^12]: we use locally Cartesian coordinates in which $h=1$ [^13]: the two quantities are still not uniquely fixed individually, but that is not required

ArXiv

--- abstract: 'Several popular *Ansätze* of lepton mass matrices that contain texture zeros are confronted with current neutrino observational data. We perform a systematic $\chi^2$-analysis in a wide class of schemes, considering arbitrary Hermitian charged lepton mass matrices and symmetric mass matrices for Majorana neutrinos or Hermitian mass matrices for Dirac neutrinos. Our study reveals that several patterns are still consistent with all the observations at 68.27% confidence level, while some others are disfavored or excluded by the experimental data. The well-known Frampton-Glashow-Marfatia two-zero textures, hybrid textures and parallel structures, among others, are considered.' author: - 'Luís M. Cebola' - 'D. Emmanuel-Costa' - 'R. González Felipe' title: Confronting predictive texture zeros in lepton mass matrices with current data --- Introduction {#sec:intro} ============ In the absence of a convincing theory to explain the origin of the lepton flavor structure, different approaches have been pursued to address this question. Among them, the imposition of texture zeros in the lepton mass matrices has been quite popular. The reason is two-fold. The vanishing of some matrix elements obviously reduces the number of free parameters, thus increasing, in some cases, the predictive power of the flavor patterns. Furthermore, texture zeros can naturally appear in theories with an extended scalar sector in the presence of Abelian symmetries [@Grimus:2004hf; @Felipe:2014vka]. Thus, the study of the phenomenological implications of lepton mass matrices with vanishing elements is well motivated on theoretical grounds. During the last years, our knowledge of neutrino masses and leptonic mixing has been enriched thanks to the data accumulated from several solar, atmospheric, reactor and accelerator neutrino experiments [@Capozzi:2013csa; @Forero:2014bxa; @Gonzalez-Garcia:2014bfa], as well as to cosmological observations [@Planck:2015xua]. Furthermore, an improved sensitivity to the Dirac CP phase has emerged from the complementarity of accelerator and reactor neutrino data. It is conceivable that leptonic CP violation is observed in current and next-generation neutrino oscillation experiments, which makes the search for such effects one of the main goals of the future research in neutrino physics [@Branco:2011zb]. It has been known for some time that, in the flavor basis where the charged lepton mass matrix is diagonal, neutrino mass matrices with more than two independent zero entries are not compatible with neutrino oscillation data, while seven patterns with two zeros are viable, as shown by Frampton, Glashow and Marfatia (FGM) in Ref. [@Frampton:2002yf]. The latter contain four complex parameters, from which nine physical quantities should be determined (three neutrino masses, three mixing angles, one Dirac CP phase and two Majorana phases), assuming that light neutrinos are Majorana particles. More recently, the aforementioned two-zero textures have been scrutinized (see e.g. Refs. [@Fritzsch:2011qv; @Meloni:2012sx; @Meloni:2014yea]). Other predictive textures can be envisaged as well in the flavor basis. The so-called hybrid textures [@Kaneko:2005yz], having one texture zero and two equal nonzero elements, contain the same number of physical parameters as the FGM textures. A systematic analysis of such hybrid textures has been presented in Ref. [@Liu:2013oxa], in which the authors concluded that 39 patterns for Majorana neutrinos are compatible with current neutrino oscillation data at the $3\sigma$ confidence level (C.L.). Restrictive patterns for the lepton mass matrices can also be constructed when the charged lepton mass matrix is not diagonal. For instance, one can consider scenarios in which both matrices exhibit a “parallel" structure [@Branco:1999nb; @Branco:2007nn] with the vanishing matrix elements located at the same positions [@Nishiura:1999yt; @Xing:2002sb] (see also Ref. [@Gupta:2012dma] and references therein). Recently, a detailed survey of texture zeros in lepton mass matrices has been performed, for both Dirac and Majorana neutrinos, considering parallel and non-parallel matrix structures [@Ludl:2014axa]. In the latter study, however, the Dirac phase was not included in the numerical $\chi^2$-analysis, which was carried out at the $5\sigma$ C.L.. In this work, we perform a detailed $\chi^2$-analysis of several popular *Ansätze* for lepton mass matrices that contain texture zeros. We aim at determining whether such patterns are consistent or not with current neutrino oscillation data at the $1\sigma\, (68.27\%)$ C.L.. In particular, the well-known FGM two-zero textures, the hybrid textures, as well as parallel structures will be analyzed. In our fitting procedure, we take into account six neutrino observables, namely, the two mass-squared differences, the three mixing angles, and the Dirac CP-violating phase. We also impose the recent cosmological bound on the sum of the neutrino masses [@Planck:2015xua]. The paper is organized as follows. In Sec. \[sec:chianalysis\], we briefly explain our strategy for the numerical analysis and minimization procedure. Then, we proceed in Sec. \[sec:fgmtextures\] to revisit the FGM two-zero textures for Majorana neutrinos. Two-zero textures for the lepton mass matrices in the case of Dirac neutrinos are also considered. Section \[sec:hybridtextures\] is devoted to the systematic $\chi^2$-analysis of hybrid textures containing one-zero texture and two equal nonzero elements, for both Majorana and Dirac neutrinos. Parallel structures with two and three zeros are studied in Sec. \[sec:partextures\]. In Sec. \[sec:nnitextures\], predictive neutrino textures in combination with a charged lepton mass matrix exhibiting the so-called nearest-neighbor-interaction (NNI) form are considered. Finally, our concluding remarks are given in Sec. \[sec:summary\]. Strategy for the numerical analysis {#sec:chianalysis} =================================== Leptonic mixing is described by the Pontecorvo, Maki, Nakagawa and Sakata (PMNS) matrix [@pmns], which, in the standard parametrization, can be written as [@Agashe:2014kda] $$\label{pmns-pdg} \mathbf{U} = \left( \begin{array}{ccc} c_{12}c_{13} & s_{12}c_{13} & s_{13}e^{-i\delta } \\ -s_{12}c_{23}-c_{12}s_{23}s_{13}e^{i\delta } & \quad c_{12}c_{23}-s_{12}s_{23}s_{13}e^{i\delta }\quad & s_{23}c_{13} \\ s_{12}s_{23}-c_{12}c_{23}s_{13}e^{i\delta } & -c_{12}s_{23}-s_{12}c_{23}s_{13}e^{i\delta } & c_{23}c_{13} \end{array} \right) \cdot \mathrm{diag\, }(1,e^{i\alpha_{21}/2}, e^{i\alpha_{31}/2}),$$ where $c_{ij}\equiv \cos \theta _{ij}\ ,\ s_{ij}\equiv \sin \theta _{ij}\ $, with all the angles $\theta _{ij}$ in the first quadrant, $\delta$ is the Dirac CP phase, and $\alpha_{21}, \alpha_{31}$ are two Majorana phases. The unitary matrix $\mathbf{U}$ in Eq.  relates the mass eigenstate neutrinos $\nu_i \ (i = 1,2,3)$ to the flavor eigenstate neutrinos $\nu_f \ (f=e,\mu,\tau)$. For Majorana neutrinos, the neutrino mass matrix $\mathbf{m}_\nu$ is a $3\times3$ complex symmetric matrix, which can be diagonalized by the unitary transformation $\mathbf{U}_{\nu L}^\dagger \mathbf{m}_\nu \mathbf{U}^{\ast}_{\nu L}=\mathrm{diag\,}(m_1,m_2,m_3)$, with $\mathbf{U}_{\nu L}$ a unitary matrix and the neutrino masses $m_i$ real and positive. If neutrinos are Dirac particles, then the corresponding unitary transformation is $\mathbf{U}_{\nu L}^\dagger \mathbf{m}_\nu \mathbf{U}_{\nu R}=\mathrm{diag\,}(m_1,m_2,m_3)$, in analogy to the charged leptons, for which the mass matrix $\mathbf{m}_\ell$ is diagonalized by $\mathbf{U}_{\ell L}^\dagger \mathbf{m}_\ell \mathbf{U}_{\ell R}=\mathrm{diag\,}(m_e,m_\mu,m_\tau)$. The leptonic mixing matrix $\mathbf{U}$ is then given by $\mathbf{U}=\mathbf{U}_{\ell L}^\dagger \mathbf{U}_{\nu L}$, which can always be parametrized in the form of Eq. . The absolute scale of neutrino masses is not yet known and there are two possible orderings of the light neutrino masses: normal ordering (NO) with $m_1<m_2<m_3$ or inverted ordering (IO) with $m_3<m_1<m_2$. The spectrum may vary from hierarchical to quasi-degenerate masses. Nevertheless, cosmological observations place a stringent upper bound on the sum of the masses. Assuming three species of degenerate massive neutrinos and a $\Lambda$CDM model, the Planck collaboration has released the bound [@Planck:2015xua] $$\label{planck} \sum_i m_i < 0.23~\text{eV}\quad \text{(95\% C.L.)},$$ obtained from a combined analysis of data.[^1] Although this bound is not definite and requires confirmation by forthcoming experiments, its inclusion in the analysis of neutrino mass models may lead to important conclusions about the viability of a given model. In this work, we shall perform a $\chi^2$-analysis using the standard $\chi^2$-function $$\label{chisquared} \chi^2(x) = \sum_{i} \frac{(\mathcal{P}_i(x)-\overline{\mathcal{O}}_i)^2}{\sigma_i^2},$$ where $x$ denotes the physical input parameters (in our case, the matrix elements of the lepton mass matrices), $\mathcal{P}_i(x)$ are the predictions of the *Ansätze* for the observables $\mathcal{O}_i$, $\overline{\mathcal{O}}_i$ are the best-fit values of $\mathcal{O}_i$, and $\sigma_i$ are their corresponding $1\sigma$ errors. In our study, we make use of the current neutrino parameters at $1\sigma$, obtained in Ref. [@Forero:2014bxa] from the global fit of neutrino oscillation data. Furthermore, we impose the cosmological constraint on the sum of the neutrino masses given in Eq. . We shall fit the zero-textures of lepton mass matrices taking into account six observables: the mass-squared differences $\Delta m_{21}^2,\, \Delta m_{31}^2$, the mixing angles $\mathrm{sin}^2\theta_{12},\, \mathrm{sin}^2\theta_{23},\, \mathrm{sin}^2\theta_{13}$, and the Dirac CP phase $\delta$. Since the Majorana phases are presently not constrained, we do not include them in the analysis. A given texture is considered to agree well with the experimental data if the model predictions for the physical observables in Eq.  are within the $1\sigma$ interval given in Table \[tab:nudata\]. Thus, $\chi^2_{min}\lesssim 6$ is a necessary condition for a pattern to be consistent with all observations. Parameter Best fit $\pm$ $1\sigma$ ----------------------------------------------------------- -------------------------- $\Delta m^2_{21}\: [10^{-5}\,\text{eV}^2]$ 7.60$^{+0.19}_{-0.18}$ \[2mm\] $|\Delta m^2_{31}|\: [10^{-3}\,\text{eV}^2]$ (NO) 2.48$^{+0.05}_{-0.07}$ (IO) 2.38$^{+0.05}_{-0.06}$ \[2mm\] $\sin^2\theta_{12} / 10^{-1}$ 3.23$\pm$0.16 \[2mm\] $\sin^2\theta_{23} / 10^{-1}$ (NO) 5.67$^{+0.32}_{-1.24}$ (IO) 5.73$^{+0.25}_{-0.39}$ \[2mm\] $\sin^2\theta_{13} / 10^{-2}$ (NO) 2.26$\pm$0.12 (IO) 2.29$\pm$0.12 \[2mm\] $\delta/\pi$(NO) 1.41$^{+0.55}_{-0.40}$ (IO) 1.48$\pm$0.31 : \[tab:nudata\] Neutrino oscillation parameters at $68.27\%$ C.L. taken from Ref. [@Forero:2014bxa]. The upper and lower rows in $\Delta m^2_{31}$, $\sin^2\theta_{23}$, $\sin^2\theta_{13}$,and $\delta$ correspond to normal (NO) and inverted (IO) neutrino mass ordering, respectively. We remark that our approach to the determination of the charged lepton masses slightly differs from that of Ref. [@Ludl:2014axa]. In our search for viable charged-lepton mass matrices, we always require that the eigenvalues of the input mass matrix correctly reproduce the central values of the charged lepton masses [@Agashe:2014kda], i.e. $$\begin{aligned} m_e&=0.510998928~\text{MeV},\;\nonumber\\ m_\mu&=105.6583715~\text{MeV},\;\\ m_\tau&=1776.82~\text{MeV}.\nonumber\end{aligned}$$ The minimization of the $\chi^2$-function is carried out with respect to the 6 neutrino observables using the MINUIT package [@James:1975dr; @root]. To improve the quality of the minima, this procedure is repeated $10^4$ times, with randomly chosen initial charged lepton and neutrino mass matrices. Clearly, in the weak basis where the charged lepton mass matrix is diagonal and real (flavor basis), one has $\mathbf{m}_\ell=\mathrm{diag\,}(m_e,m_\mu,m_\tau)$ and thus this matrix is fixed. Moreover, one can easily show that the absolute value of any matrix element of $\mathbf{m}_\nu$ is always smaller than the largest neutrino mass, i.e. $|(\mathbf{m}_\nu)_{ij}|<\max_k\, (m_k)$. Therefore, the cosmological bound in Eq.  implies $|(\mathbf{m}_\nu)_{ij}|\lesssim 0.08$ eV. FGM textures {#sec:fgmtextures} ============ In this section we revisit the well-known FGM patterns for lepton mass matrices [@Frampton:2002yf], consisting of $3\times3$ Majorana neutrino mass matrices $\mathbf{m}_\nu$ with two zero elements in the charged lepton flavor basis with $\mathbf{m}_\ell=\text{diag}\,(m_e,m_\mu,m_\tau)$. We shall also consider the case of Dirac neutrinos, for which the matrix $\mathbf{m}_\nu$ is Hermitian. For Majorana neutrinos, the mass matrix $\mathbf{m}_\nu$ is a symmetric matrix with six independent complex entries. There are $6!/[n! (6 - n)!]$ different textures, each containing $n$ independent texture zeros. One can show that any pattern of $\mathbf{m}_\nu$ with more than two independent zeros ($n> 2$) is not compatible with current neutrino oscillation data. For $n=2$, there are fifteen two-zero textures of $\mathbf{m}_\nu$, which can be classified into six categories ($\mathbf{A}, \mathbf{B}, \mathbf{C}, \mathbf{D}, \mathbf{E}, \mathbf{F}$), according to their physical predictions: $$\begin{aligned} \label{FGMtextures} \mathbf{A}_1: \begin{pmatrix} 0 & 0 & \ast \\ 0 & \ast & \ast \\ \ast & \ast & \ast \end{pmatrix} \,,\quad \mathbf{A}_2: \begin{pmatrix} 0 & \ast & 0 \\ \ast & \ast & \ast \\ 0 & \ast & \ast \end{pmatrix} \,,\quad \mathbf{B}_1: \begin{pmatrix} \ast & \ast & 0 \\ \ast & 0 & \ast \\ 0 & \ast & \ast \end{pmatrix}\,,\nonumber\\[3mm] \mathbf{B}_2: \begin{pmatrix} \ast & 0 & \ast \\ 0 & \ast & \ast \\ \ast & \ast & 0 \end{pmatrix}\,,\quad \mathbf{B}_3: \begin{pmatrix} \ast & 0 & \ast \\ 0 & 0 & \ast \\ \ast & \ast & \ast \end{pmatrix}\,,\quad \mathbf{B}_4: \begin{pmatrix} \ast & \ast & 0 \\ \ast & \ast & \ast \\ 0 & \ast & 0 \end{pmatrix}\,,\nonumber\\[3mm] \mathbf{C}: \begin{pmatrix} \ast & \ast & \ast \\ \ast & 0 & \ast \\ \ast & \ast & 0 \end{pmatrix}\,,\quad \mathbf{D}_1: \begin{pmatrix} \ast & \ast & \ast \\ \ast & 0 & 0 \\ \ast & 0 & \ast \end{pmatrix}\,,\quad \mathbf{D}_2: \begin{pmatrix} \ast & \ast & \ast \\ \ast & \ast & 0 \\ \ast & 0 & 0 \end{pmatrix}\,,\nonumber\\[3mm] \mathbf{E}_1: \begin{pmatrix} 0 & \ast & \ast \\ \ast & 0 & \ast \\ \ast & \ast & \ast \end{pmatrix}\,,\quad \mathbf{E}_2: \begin{pmatrix} 0 & \ast & \ast \\ \ast & \ast & \ast \\ \ast & \ast & 0 \end{pmatrix}\,,\quad \mathbf{E}_3: \begin{pmatrix} 0 & \ast & \ast \\ \ast & \ast & 0 \\ \ast & 0 & \ast \end{pmatrix}\,,\nonumber\\[3mm] \mathbf{F}_1: \begin{pmatrix} \ast & 0 & 0 \\ 0 & \ast & \ast \\ 0 & \ast & \ast \end{pmatrix} \,,\quad \mathbf{F}_2: \begin{pmatrix} \ast & 0 & \ast \\ 0 & \ast & 0 \\ \ast & 0 & \ast \end{pmatrix}\,,\quad \mathbf{F}_3: \begin{pmatrix} \ast & \ast & 0 \\ \ast & \ast & 0 \\ 0 & 0 & \ast \end{pmatrix}\,.\nonumber\\\end{aligned}$$ Here, the symbol “$\ast$" stands for arbitrary nonzero matrix elements. Clearly, the matrices $\mathbf{F}_i$ can be straightforwardly excluded since they lead to the decoupling of one generation and thus are not experimentally viable. Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ --------------------------- --------------------- ----------------------- ------------------- ----------------------- ----------------------- ----------------------- ----------------------- ----------------------- $\mathbf{A}_1$ $2.92\times10^{-1}$ ($3.81\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark (\times)$ $\checkmark$ $\checkmark (\times)$ $\mathbf{A}_2$ $1.23\times10^{-2}$ ($3.14\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark (\times)$ $\checkmark$ $\checkmark (\times)$ $\mathbf{B}_1$ $8.39\times10^{-1}$ ($4.04\times10^{-3}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_2$ $3.39\times10^{-2}$ ($1.02\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\mathbf{B}_3$ $9.12\times10^{-1}$ ($3.45\times10^{-3}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_4$ $2.10\times10^{-2}$ ($1.11\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\mathbf{C}$ $6.20\times10^{2}$ ($1.04\times10^{-1}$) $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $ \checkmark$ $\mathbf{D}_1$ $1.33\times10^{2}$ ($3.43\times10^{1}$) $\checkmark$ $\checkmark$ $\times$ $\times$ $\checkmark$ $\checkmark (\times)$ $\mathbf{D}_2$ $2.82\times10^{2}$ ($4.88\times10^{1}$) $\checkmark$ $\checkmark$ $\times$ $\times$ $\checkmark$ $\checkmark (\times)$ $\mathbf{E}_1$ $1.40\times10^{1}$ ($1.15\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\mathbf{E}_2$ $1.03\times10^{2}$ ($1.14\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\mathbf{E}_3$ $2.09\times10^{1}$ ($1.17\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ Dirac $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ ------------------------ -------------------- ----------------------- ------------------- ----------------------- ----------------------- ----------------------- ----------------------- ---------------------- $\mathbf{C}$ $6.19\times10^{2}$ ($1.04\times10^{-1}$) $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $ \checkmark$ $\mathbf{E}_1$ $1.40\times10^{1}$ ($1.15\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times (\checkmark)$ $\checkmark$ $\checkmark(\times)$ $\mathbf{E}_2$ $1.03\times10^{2}$ ($1.14\times10^{2}$) $\checkmark$ $\checkmark$ $\times$ $\times (\checkmark)$ $\checkmark$ $\checkmark(\times)$ Our results are presented in Table \[tab:FGM\], in which the minimum of $\chi^2$ for each FGM texture with a normal or inverted neutrino mass ordering is given. The results are obtained using the current neutrino oscillation data of Table \[tab:nudata\] and imposing the upper bound on the sum of neutrino masses given in Eq. . We indicate with a check mark or a cross whether the texture predictions are or not within the $1\sigma$ interval given in Table \[tab:nudata\]. Note that, in order to ease the reading of the table, whenever a given observable is simultaneously compatible (or incompatible) with data for NO and IO, we just indicate it with a single symbol, i.e. with a check mark (or a cross). Henceforth, this notation will be used in all tables. From Table \[tab:FGM\] we conclude that patterns $\mathbf{A}_{1,2}$ and $\mathbf{B}_{1,2,3,4}$ are allowed for NO, while only patterns $\mathbf{B}_{1,3}$ and $\mathbf{C}$ are compatible with neutrino oscillation data for an IO mass spectrum at the $1\sigma$ level.[^2] We remark that, if the stringent upper bound on the sum of neutrino masses given in Eq.  is relaxed, pattern $\mathbf{C}$ is also allowed for a NO neutrino mass spectrum [@Meloni:2014yea]. In the latter case, we obtain $\chi^2_{min} \simeq 0.32$ with $\sum_i m_i < 1$ eV. For completeness, in Figs. \[fig:fig-A1\]-\[fig:fig-C\] of Appendix \[appendix1\], we present the probability distribution of the six neutrino observables, obtained for the seven viable FGM textures $\mathbf{A}_{1,2}$, $\mathbf{B}_{1,2,3,4}$ and $\mathbf{C}$, for both NO and IO mass spectra. We notice that textures in the same category lead in general to similar physical predictions for the observables. We now consider the case of Dirac neutrinos. We analyze again the two-zero textures given in Eq. . These patterns have been recently studied for Dirac neutrinos in Ref. [@Liu:2012axa], where the authors concluded that only the patterns $\mathbf{A}_{1,2}$ and $\mathbf{C}$ are compatible with the oscillation data at the $2\sigma$ level. First we note that by redefining the right-handed neutrino fields we can assume, without loss of generality, that the mass matrix $\mathbf{m}_\nu$ is Hermitian. Furthermore, it is straightforward to show that if one off-diagonal matrix element is zero, then the invariant quantity $\mathcal{J}_{CP}=\text{Im}\, [\mathbf{U}_{12}\mathbf{U}_{23}\mathbf{U}_{13}^\ast \mathbf{U}_{22}^\ast]$ vanishes, and thus CP is conserved in the lepton sector. Therefore, only patterns $\mathbf{C}$, $\mathbf{E}_1$, and $\mathbf{E}_2$ can lead to leptonic CP violation, while $\delta=0$ or $\pi$ for the remaining two-zero patterns. In view of the above, we shall only present the results for patterns $\mathbf{C}$, $\mathbf{E}_1$, and $\mathbf{E}_2$. The minimum of $\chi^2$ is given in Table \[tab:FGM-Dirac\]. As can be seen from the table, there is essentially no difference with respect to the results obtained for Majorana neutrinos. Only pattern $\mathbf{C}$ with an inverted hierarchy is allowed by current data. Relaxing the cosmological bound on the sum of the neutrino masses, we conclude that a normal hierarchical neutrino spectrum is also allowed for pattern $\mathbf{C}$, with $\chi^2_{min} \simeq 0.29$ for $\sum_i m_i < 1$ eV. Notice also that the parameter counting for Hermitian Dirac matrices differs from that of symmetric Majorana matrices, since in the former case the counting depends on the position of the zeros. For two vanishing diagonal matrix entries, the matrix $\mathbf{m}_\nu$ contains at most seven real parameters. Hybrid textures {#sec:hybridtextures} =============== Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ -------------------------------- --------------------- ------------------------ ------------------- ------------------- --------------- ----------------------- ----------------------- -------------- $\widehat{\mathbf{A}}_{1(13)}$ $1.78\times10^{-8}$ ($1.40\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark (\times)$ $\checkmark$ $\widehat{\mathbf{A}}_{1(22)}$ $1.65$ ($7.16\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{1(23)}$ $1.76\times10^{1}$ ($6.44\times10^{-9}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{1(33)}$ $9.11$ ($7.22\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times ($$)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{2(12)}$ $1.29\times10^{-8}$ ($1.55\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\widehat{\mathbf{A}}_{2(22)}$ $3.45$ ($1.99\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{2(23)}$ $2.06\times10^{1}$ ($4.75\times10^{-11}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{2(33)}$ $7.34\times10^{-1}$ ($9.89\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ -------------------------------- --------------------- ------------------------ ------------------- ------------------- ---------------------- ----------------------- --------------- ---------------------- $\widehat{\mathbf{B}}_{1(11)}$ $6.63$ ($3.14\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $ \times$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{1(12)}$ $6.34\times10^{-1}$ ($3.67\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $ \checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{1(23)}$ $2.08\times10^{1}$ ($1.31\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{1(33)}$ $3.42\times10^{-5}$ ($6.83$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{2(11)}$ $3.04\times10^{1}$ ($3.80\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{2(13)}$ $9.64\times10^{-3}$ ($7.34$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{2(22)}$ $7.18\times10^{-1}$ ($1.72\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{2(23)}$ $1.80\times10^2$ ($9.61\times10^{-10}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{3(11)}$ $5.52$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{3(13)}$ $5.59\times10^{-1}$ ($2.69\times10^{-5}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{3(23)}$ $1.77\times10^{1}$ ($1.83\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{3(33)}$ $7.61\times10^{-2}$ ($6.27$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{4(11)}$ $2.32\times10^{1}$ ($1.18\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{4(12)}$ $1.88\times10^{-2}$ ($6.85$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{4(22)}$ $9.72\times10^{-1}$ ($8.36\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{4(23)}$ $1.44\times10^{2}$ ($1.25\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ ------------------------------- --------------------- ------------------------ ------------------- ---------------------- ----------------------- ---------------------- ---------------------- -------------- $\widehat{\mathbf{C}}_{(11)}$ $1.40\times10^{-1}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{C}}_{(12)}$ $3.19\times10^{-1}$ ($3.68\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{C}}_{(13)}$ $1.88\times10^{-6}$ ($3.72$) $\checkmark$ $\checkmark$ $\checkmark $ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{C}}_{(23)}$ $6.20\times10^{2}$ ($3.52\times10^{-11}$) $\checkmark$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\times(\checkmark)$ $\checkmark$ Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ -------------------------------- ---------------------- ----------------------- ------------------- ------------------- ----------------------- ----------------------- --------------- -------------- $\widehat{\mathbf{D}}_{1(11)}$ $1.07\times10^{-8}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{1(12)}$ $4.77\times10^{-8}$ ($2.11\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{1(13)}$ $1.53\times10^{-7}$ ($7.88\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{1(33)}$ $6.88$ ($1.87$) $\checkmark$ $\checkmark$ $\checkmark$ $\times ($$)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{2(11)}$ $1.14\times10^{-8}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{2(12)}$ $6.60\times10^{-10}$ ($1.62\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{2(13)}$ $1.63\times10^{-7}$ ($2.59\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{2(22)}$ $3.12$ ($3.38\times10^{-3}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ -------------------------------- --------------------- ----------------------- ------------------- ----------------------- ----------------------- ----------------------- ---------------------- ---------------------- $\widehat{\mathbf{E}}_{1(12)}$ $1.65\times10^{-6}$ ($1.51\times10^{1}$) $\checkmark$ $\checkmark (\times)$ $\checkmark$ $ \checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\widehat{\mathbf{E}}_{1(13)}$ $7.10\times10^{-7}$ ($1.97\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{1(23)}$ $2.03\times10^{1}$ ($1.33\times10^{-9}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{1(33)}$ $1.40$ ($1.39\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{2(12)}$ $2.55\times10^{-6}$ ($2.58\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{2(13)}$ $7.41\times10^{-9}$ ($9.09$) $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\widehat{\mathbf{E}}_{2(22)}$ $1.88$ ($9.94\times10^{-1}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{2(23)}$ $1.77\times10^{2}$ ($1.33\times10^{-9}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{3(12)}$ $2.90\times10^{-8}$ ($3.24\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{E}}_{3(13)}$ $1.52\times10^{-6}$ ($2.51\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{E}}_{3(22)}$ $1.32\times10^{2}$ ($3.17\times10^{-9}$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{3(33)}$ $2.80\times10^{2}$ ($5.74\times10^{-4}$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ -------------------------------- ---------------------- ----------------------- ------------------- ------------------- ----------------------- ----------------------- ----------------------- -------------- $\widehat{\mathbf{F}}_{1(11)}$ $2.15\times10^{-12}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{1(22)}$ $8.26\times10^{-1}$ ($3.25\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{1(23)}$ $1.88\times10^{1}$ ($2.15\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{1(33)}$ $2.87\times10^{-6}$ ($6.97$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{2(11)}$ $1.68\times10^{1}$ ($3.68\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark(\times)$ $\checkmark$ $\widehat{\mathbf{F}}_{2(13)}$ $7.95\times10^{-7}$ ($1.43\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{2(22)}$ $2.87$ ($3.62\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{2(33)}$ $2.50\times10^{2}$ ($1.42\times10^{-6}$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\widehat{\mathbf{F}}_{3(11)}$ $1.86\times10^{1}$ ($3.07\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{3(12)}$ $6.88\times10^{-8}$ ($9.89\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{3(22)}$ $1.04\times10^{2}$ ($2.94\times10^{-6}$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\widehat{\mathbf{F}}_{3(33)}$ $5.61$ ($4.32\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ Dirac $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ -------------------------------- --------------------- ----------------------- ------------------- ---------------------- ----------------------- ----------------------- ----------------------- ----------------------- $\widehat{\mathbf{A}}_{1(22)}$ $1.16\times10^{1}$ ($3.15\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{1(33)}$ $1.31\times10^{1}$ ($3.09\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{2(22)}$ $5.65$ ($4.35\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{A}}_{2(33)}$ $7.65\times10^{1}$ ($5.01\times10^{1}$) $\checkmark$ $\checkmark$ $\times$ $\times$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{1(11)}$ $6.91$ ($3.15\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{1(33)}$ $1.72\times10^{-2}$ ($5.40\times10^{1}$) $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{2(11)}$ $3.07\times10^{1}$ ($3.80\times10^{2}$) $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{2(22)}$ $9.05\times10^{-1}$ ($4.16\times10^{1}$) $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\widehat{\mathbf{B}}_{3(11)}$ $6.37$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{B}}_{3(33)}$ $6.56\times10^{-3}$ ($1.41\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{4(11)}$ $2.62\times10^{1}$ ( $1.18\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{B}}_{4(22)}$ $9.89\times10^{-1}$ ($3.43\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{C}}_{(11)}$ $2.51\times10^{-1}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{1(11)}$ $4.51\times10^{-1}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{1(33)}$ $1.60\times10^{1}$ ($2.62$) $\checkmark$ $\checkmark$ $\times(\checkmark)$ $\times ($$)$ $\checkmark$ $\times$ $\widehat{\mathbf{D}}_{2(11)}$ $1.49$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{D}}_{2(22)}$ $5.47$ ($5.88\times10^{-1}$) $\checkmark$ $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{1(33)}$ $5.76\times10^{1}$ ($7.34\times10^{-1}$) $\checkmark$ $\checkmark$ $\times(\checkmark)$ $\times(\checkmark)$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{E}}_{2(22)}$ $5.79\times10^{1}$ ($8.78$) $\checkmark$ $\checkmark$ $\times$ $\times $ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{E}}_{3(22)}$ $1.33\times10^{2}$ ($1.92\times10^{-3}$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\times(\checkmark)$ $\widehat{\mathbf{E}}_{3(33)}$ $2.80\times10^{2}$ ($4.02\times10^{-1}$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\times(\checkmark)$ $\widehat{\mathbf{F}}_{1(11)}$ $2.94\times10^{-2}$ ($1.08\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{1(22)}$ $1.37\times10^{1}$ ($2.12$) $\checkmark$ $\checkmark$ $\checkmark$ $\times(\checkmark)$ $\checkmark$ $\checkmark (\times)$ $\widehat{\mathbf{F}}_{1(33)}$ $9.53\times10^{1}$ ($3.50$) $\checkmark$ $\checkmark$ $\checkmark$ $\times$ $\checkmark$ $\times(\checkmark)$ $\widehat{\mathbf{F}}_{2(11)}$ $2.00\times10^{1}$ ($3.68\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark(\times)$ $\checkmark$ $\widehat{\mathbf{F}}_{2(22)}$ $9.31$ ($2.40\times10^{1}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times (\checkmark)$ $\checkmark$ $\checkmark(\times)$ $\widehat{\mathbf{F}}_{2(33)}$ $2.50\times10^{2}$ ($1.34$) $\checkmark$ $\checkmark$ $\times (\checkmark)$ $\times (\checkmark)$ $\times (\checkmark)$ $\checkmark$ $\widehat{\mathbf{F}}_{3(11)}$ $1.86\times10^{1}$ ($3.07\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\times$ $\checkmark$ $\checkmark$ $\widehat{\mathbf{F}}_{3(22)}$ $1.04\times10^{2}$ ($1.22\times10^{1}$) $\checkmark$ $\checkmark$ $\times$ $\times$ $\times (\checkmark)$ $\times$ $\widehat{\mathbf{F}}_{3(33)}$ $1.95\times10^{1}$ ($2.37\times10^{1}$) $\checkmark$ $\checkmark$ $\times$ $\times (\checkmark)$ $\checkmark$ $\checkmark$ Hybrid textures [@Kaneko:2005yz] are particular cases of one-zero textures of the Majorana neutrino mass matrix, which additionally have two equal nonzero elements, and are defined in the flavor basis. There are $(6!/5!) \times 5!/(2!\, 3!)=60$ possible hybrid textures. Among them, it has been shown that only 39 textures are compatible with current neutrino oscillation data at the $3\sigma$ level [@Liu:2013oxa]. To keep a coherent notation, without the need of introducing any new classification scheme, we shall label these matrices as follows. We associate to each FGM matrix $\mathbf{M}$ given in Eq.  a hybrid-type matrix $\widehat{\mathbf{M}}$, in which the two zeros in $\mathbf{M}$ are replaced by equal nonvanishing elements in $\widehat{\mathbf{M}}$. Then, the position of the zero element in the hybrid matrix $\widehat{\mathbf{M}}$ is indicated with a subscript in parenthesis. Consider, for instance, the hybrid texture $$\begin{aligned} \label{hybridexample} \begin{pmatrix} \text{X} & \text{X} & \ast \\ \text{X} & \ast & \ast \\ \ast & \ast & 0 \end{pmatrix}\,,\end{aligned}$$ where “X" stands for equal nonzero elements. Following the definition of the matrix $\mathbf{A}_1$ given in Eq. , the hybrid matrix  would be represented as $\widehat{\mathbf{A}}_{1(33)}$ in our notation. Obviously, for each FGM texture in Eq. , one can construct four different hybrid textures, depending on the position of the zero matrix element. For comparison, below we list the complete set of 39 hybrid textures studied in Ref. [@Liu:2013oxa]: $$\begin{aligned} \label{hybridtextures} &\widehat{\mathbf{A}}_{1\,\{(13),(22),(23),(33)\}}\,, ~\widehat{\mathbf{A}}_{2\,\{(12),(22),(23),(33)\}}\,, \nonumber\\ &\widehat{\mathbf{B}}_{1\,\{(12),(23),(33)\}}\,, \widehat{\mathbf{B}}_{2\,\{(13),(22),(23)\}}\,,\nonumber\\ &\widehat{\mathbf{B}}_{3\,\{(13),(23),(33)\}}\,, ~\widehat{\mathbf{B}}_{4\,\{(12),(22),(23)\}}\,, ~\widehat{\mathbf{C}}_{(11)}\,,\nonumber\\ &\widehat{\mathbf{D}}_{1\,\{(11),(12),(13),(33)\}}\,, ~\widehat{\mathbf{D}}_{2\,\{(11),(12),(13),(22)\}}\,,\\ &\widehat{\mathbf{E}}_{1(33)}\,,~\widehat{\mathbf{E}}_{2(22)}\,, ~\widehat{\mathbf{E}}_{3\,\{(22),(33)\}}\,,\nonumber\\ &\widehat{\mathbf{F}}_{1\,\{(22),(33)\}}\,, ~\widehat{\mathbf{F}}_{2\,\{(22),(33)\}}\,, ~\widehat{\mathbf{F}}_{3\,\{(22),(33)\}}\,,\nonumber\end{aligned}$$ where we have indicated, inside curly brackets, the possible choices for the texture-zero position. The results of the $\chi^2$-minimization are summarized in Tables \[tab:hybridA\]-\[tab:hybridF\]. First we note that all textures given in Eq.  are compatible with data at the $1\sigma$ level either for NO, IO or both types of neutrino mass spectrum. In particular, the patterns $\widehat{\mathbf{A}}_{2(33)}$, $\widehat{\mathbf{B}}_{1(12)}$, $\widehat{\mathbf{B}}_{2(22)}$, $\widehat{\mathbf{B}}_{3(13)}$, $\widehat{\mathbf{B}}_{4(22)}$, $\widehat{\mathbf{C}}_{(12)}$, $\widehat{\mathbf{D}}_{1(12)}$, $\widehat{\mathbf{D}}_{1(13)}$, $\widehat{\mathbf{D}}_{2(12)}$, $\widehat{\mathbf{D}}_{2(13)}$, $\widehat{\mathbf{F}}_{1(22)}$, $\widehat{\mathbf{F}}_{2(13)}$, and $\widehat{\mathbf{F}}_{3(12)}$ turn out to be compatible with experimental data for NO and IO mass spectra. Despite the fact that our analysis is performed at the stringent $1\sigma$ C.L., constraining also the Dirac phase $\delta$, among the sixty possible hybrid patterns for Majorana neutrinos, only six fail in reproducing the data for any hierarchy and can be completely excluded. These are the matrices $\widehat{\mathbf{B}}_{1(11)}$, $\widehat{\mathbf{B}}_{2(11)}$, $\widehat{\mathbf{B}}_{3(11)}$, $\widehat{\mathbf{B}}_{4(11)}$, $\widehat{\mathbf{F}}_{2(11)}$, and $\widehat{\mathbf{F}}_{3(11)}$, all having a zero element in the (1,1) position. We remark that in Ref. [@Liu:2013oxa], only 13 patterns were found compatible with data at the $1\sigma$ level: $\widehat{\mathbf{A}}_{1(22)}$, $\widehat{\mathbf{A}}_{1(23)}$, $\widehat{\mathbf{A}}_{1(33)}$, $\widehat{\mathbf{B}}_{1(23)}$, $\widehat{\mathbf{B}}_{1(33)}$, $\widehat{\mathbf{B}}_{2(22)}$, $\widehat{\mathbf{B}}_{3(13)}$, $\widehat{\mathbf{B}}_{4(23)}$, $\widehat{\mathbf{D}}_{2(11)}$, $\widehat{\mathbf{D}}_{2(13)}$, $\widehat{\mathbf{E}}_{2(22)}$, $\widehat{\mathbf{F}}_{2(22)}$, and $\widehat{\mathbf{F}}_{2(33)}$. The fact that several viable hybrid textures were missed in Ref. [@Liu:2013oxa] could be attributed to the numerical procedure followed by the authors, who performed a simple random scanning of the parameter space instead of the more reliable $\chi^2$-analysis. In the case of Dirac neutrinos, thirty patterns were considered, which are listed in Table \[tab:hybrid-Dirac\]. We include all the Hermitian patterns that do not have any off-diagonal zero element, and thus may lead to Dirac-type CP violation. For the remaining patterns, the Dirac phase $\delta$ is always 0 or $\pi$ and CP is conserved in the lepton sector. Looking at the table we note that only twelve textures are consistent with data either for NO or IO neutrino mass spectrum. These are the matrices $\widehat{\mathbf{B}}_{1(33)}$, $\widehat{\mathbf{B}}_{2(22)}$, $\widehat{\mathbf{B}}_{3(33)}$, $\widehat{\mathbf{B}}_{4(22)}$, $\widehat{\mathbf{C}}_{(11)}$, $\widehat{\mathbf{D}}_{1(11)}$, $\widehat{\mathbf{D}}_{2(22)}$, $\widehat{\mathbf{E}}_{1(33)}$, $\widehat{\mathbf{E}}_{3(22)}$, $\widehat{\mathbf{E}}_{3(33)}$, $\widehat{\mathbf{F}}_{1(11)}$, and $\widehat{\mathbf{F}}_{2(33)}$. None of these matrices is simultaneously allowed for both mass spectra. Parallel textures {#sec:partextures} ================= In this section, we perform a systematic $\chi^2$-analysis of lepton mass matrices that exhibit the same texture, i.e. with $\mathbf{m}_\ell$ and $\mathbf{m}_\nu$ having their zeros located at the same positions. Besides the possibility of implementing a universal flavor structure in the context of grand unified models, there is an additional theoretical motivation for considering parallel structures. It is well known that an attractive and economical framework to generate small neutrino masses is the seesaw mechanism. In its simplest type-I realization, three right-handed neutrinos are added to the standard model particle content. It is then conceivable that the presence of family symmetries enforces texture-zero structures in the Dirac neutrino mass matrix $\mathbf{m}_D$ and the heavy Majorana mass matrix $\mathbf{M}_R$, which, in some cases, could be preserved by the effective neutrino mass matrix $\mathbf{m}_\nu=-\mathbf{m}_D \mathbf{M}_R^{-1} \mathbf{m}_D^\mathsf{T}$.[^3] It is worth noticing that any permutation transformation acting on parallel patterns is allowed, since it leads to textures with the same physical content. Indeed, they can be related by a weak basis transformation, performed by a permutation matrix $\mathbf{P}$, $$\label{wbtransf} \mathbf{m}^{\prime}_{\ell}=\mathbf{P}^{\mathsf{T}}\,\mathbf{m}_{\ell}\,\mathbf{P}\,,\quad \mathbf{m}^{\prime}_{\nu}=\mathbf{P}^{\mathsf{T}}\,\mathbf{m}_{\nu}\,\mathbf{P}\,,$$ which automatically preserves the parallel structure, but changes the position of the zeros. The matrix $\mathbf{P}$ belongs to the group of six permutations matrices, which are isomorphic to the symmetry group $S_3\,$. Two-zero textures ----------------- The FGM-type Ansätze can be classified into four weak basis equivalent classes (or permutation sets) [@Branco:2007nn]: $$\begin{aligned} \label{psets} \begin{split} \text{Class I:} &\quad \mathbf{A}_1, \mathbf{A}_2, \mathbf{B}_3,\mathbf{B}_4, \mathbf{D}_1, \mathbf{D}_2;\\ \text{Class II:} &\quad \mathbf{B}_1, \mathbf{B}_2, \mathbf{E}_3; \\ \text{Class III:} &\quad \mathbf{C}, \mathbf{E}_1, \mathbf{E}_2;\\ \text{Class IV:} &\quad \mathbf{F}_1, \mathbf{F}_2, \mathbf{F}_3. \end{split}\end{aligned}$$ It is clear that class IV is not experimentally viable, since it always leads to the decoupling of one generation. Note also that the weak basis transformations given in Eq.  are not allowed in a scheme with a diagonal and ordered charged lepton mass matrix, as in the texture schemes discussed in previous section. In our $\chi^2$-analysis, all parallel FGM textures with arbitrary complex Hermitian (or real symmetric) $\mathbf{m}_{\ell}$ and complex symmetric $\mathbf{m}_{\nu}$ were found to be viable for both normal and inverted neutrino mass ordering. Similar results were obtained for Dirac neutrinos with Hermitian neutrino mass matrices. We have also considered the feasibility of arbitrary complex Hermitian $\mathbf{m}_{\ell}$ and real symmetric $\mathbf{m}_{\nu}$. In this case, the number of physical parameters is equal to 10 for classes I and II, while for class III there are 11 parameters since, in general, the invariant quantity $\arg\bigl[(\mathbf{m}_{\ell})_{12}(\mathbf{m}_{\ell}^{\ast})_{13}(\mathbf{m}_{\ell})_{23}\bigr]$ does not vanish. As far as the analysis of the neutrino oscillation data is concerned, there is no distinction between Majorana or Dirac neutrinos. The minimum of $\chi^2$ was always found to be much smaller than one, so that all patterns in classes I, II, and III are consistent with neutrino data for any mass hierarchy. Majorana $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ --------------------------- --------------------- ----------------------- ------------------- ------------------- ----------------------- ---------------------- --------------- ---------------------- $\mathbf{A}_{1(13)}$ $1.25\times10^{-7}$ ($1.71\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{1(22)}$ $9.66\times10^{-8}$ ($4.95\times10^{-9}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{1(23)}$ $5.53\times10^{-8}$ ($3.06$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\mathbf{A}_{1(33)}$ $1.34\times10^{-7}$ ($5.13\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{2(22)}$ $7.02\times10^{-8}$ ($9.65\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{2(23)}$ $1.31\times10^{-7}$ ($4.51\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{2(33)}$ $7.88\times10^{-8}$ ($3.38\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{1(23)}$ $5.35\times10^{-7}$ ($3.78\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{1(33)}$ $1.38\times10^{-6}$ ($4.57\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{2(13)}$ $1.52$ ($2.88\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{2(22)}$ $7.93\times10^{-7}$ ($1.54\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{2(23)}$ $1.52$ ($6.29\times10^{-5}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{3(13)}$ $1.84\times10^{-7}$ ($3.05$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\mathbf{B}_{3(23)}$ $1.20\times10^{-6}$ ($4.05\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{4(23)}$ $1.95\times10^{-7}$ ($6.08\times10^{-9}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{C}_{(11)}$ $1.49\times10^{-6}$ ($1.75\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{C}_{(23)}$ $2.69\times10^{-8}$ ($2.43\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{D}_{1(11)}$ $4.99\times10^{-7}$ ($1.61\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{D}_{2(11)}$ $4.21\times10^{-6}$ ($5.29\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark $ $\checkmark$ $\checkmark$ $\mathbf{F}_{1(23)}$ $1.98\times10^{1}$ ($1.83\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times$ $\checkmark$ $\times$ Three-zero textures ------------------- There are only 6 possible three-zero parallel textures that can be constructed for both the charged-lepton and Majorana neutrino mass matrices. Since these matrices are related by weak basis transformations (permutations), they all have the same physical content and thus lead to the same predictions. We denote them by $$\begin{aligned} \label{3zerotextures} \mathbf{A}_{1(33)}\,, ~\mathbf{A}_{2(22)}\,,~\mathbf{B}_{1(33)}\,, ~\mathbf{B}_{2(22)}\,, ~\mathbf{D}_{1(11)}\,, ~\mathbf{D}_{2(11)}\,,\end{aligned}$$ where the subscript in parenthesis refers to the position of the additional zero in the corresponding two-zero texture given in Eq. . Note that the matrix $\mathbf{C}_{(11)}$ is not included in the above list since it is traceless and, therefore, incompatible with the lepton masses. Furthermore, textures with null determinant or those leading to the decoupling of one generation have also been excluded. The texture $\mathbf{A}_{2(22)}$ is known as the nearest-neighbor-interaction pattern [@Fritzsch:1979zq; @Branco:1988iq]. In the context of the standard model, it has been shown that imposing an NNI texture simultaneously in the up- and down-quark sectors simply corresponds to a weak basis choice [@Branco:1988iq]. For non-Hermitian quark mass matrices, this is an example of parallel four-zero textures without any physical content. This is not necessarily true in the lepton sector, unless neutrinos are Dirac particles. For Majorana neutrinos, the assumption of a parallel NNI structure would have physical implications. For an arbitrary complex Hermitian $\mathbf{m}_{\ell}$ and a complex symmetric $\mathbf{m}_{\nu}$ (Majorana neutrinos), the above parallel 3-zero textures contain 9 physical parameters. No viable solution was found either for NO ($\chi^2_{min}\simeq 74$) or IO ($\chi^2_{min}\simeq 182$) neutrino mass spectrum. For a normal ordering of neutrino masses, all the textures fail in reproducing the three mixing angles, while for an inverted spectrum the mixing angles $\theta_{12}$ and $\theta_{23}$, and the phase $\delta$ did not satisfy the required $\chi^2$ criteria. In Fig. \[fig:fig-3zeros-Majorana\] of Appendix \[appendix2\], we present the probability distribution of the neutrino observables, obtained for the textures given in Eq. , for NO and IO mass spectrum, respectively. For Dirac neutrinos, with the matrix $\mathbf{m}_{\nu}$ being Hermitian, similar results were found, thus excluding these patterns for both NO and IO mass spectra.[^4] Predictive neutrino textures with NNI charged lepton mass matrix {#sec:nnitextures} ================================================================ In the previous section, we have considered parallel structures for both lepton sectors, assuming an Hermitian charged lepton mass matrix. In particular, we showed that the parallel NNI texture $\mathbf{A}_{2(22)}$ is not compatible with the current neutrino data. In this section, we shall lift the assumption of Hermiticity on the NNI pattern of the charged lepton mass matrix and look for viable predictive neutrino zero textures. Such patterns are of interest since they contain the same number of physical parameters as the FGM and hybrid textures (assuming that $\mathbf{m}_{\nu}$ has three zeros). From the theoretical viewpoint, NNI lepton structures are also well motivated. For instance, it is possible to conceive flavour symmetries in the two-Higgs doublet model [@Branco:2010tx], in supersymmetric theories [@Babu:2009nn], and in grand unified models based on $\mathsf{SU}(5)$ [@EmmanuelCosta:2011jq; @Emmanuel-Costa:2013gia] that lead to NNI textures. In the latter models, the charged lepton mass matrix $\mathbf{m}_{\ell}$ exhibits an NNI pattern, while the effective neutrino mass matrix $\mathbf{m}_{\nu}$ contains some vanishing elements. We shall assume that the non-Hermitian charged lepton mass matrix $\mathbf{m}_{\ell}$ is described by the NNI form $\mathbf{A}_{2(22)}$ and search for a maximal number of zeros in $\mathbf{m}_{\nu}$ compatible with the data. As before, we take $\mathbf{m}_{\nu}$ as a general complex symmetric matrix for Majorana neutrinos and an Hermitian matrix for Dirac neutrinos. We remark that, without loss of generality, all the non-vanishing matrix elements in $\mathbf{m}_{\ell}$ can be assumed real and positive. Thus, there remain two free parameters in $\mathbf{m}_{\ell}$ after fitting the charged lepton masses. In our $\chi^2$-search, none of the neutrino textures with more than three zeros was found compatible with the observed neutrino data. In Table \[tab:nniMajorana\] and \[tab:nniDirac\] we present the results for three-zero $\mathbf{m}_{\nu}$ textures for Majorana and Dirac neutrinos, respectively. As can be seen from the tables, only the pattern $\mathbf{F}_{1(23)}$ fails in reproducing the data for any neutrino spectra. Moreover, the patterns $\mathbf{A}_{1(23)}$ and $\mathbf{B}_{3(13)}$ are compatible with the data only for NO neutrino masses. The remaining 17 textures are viable at the $1\sigma$ level irrespective of the mass ordering. In particular, once the Hermiticity of $\mathbf{m}_{\ell}$ is lifted, the parallel structure $\mathbf{A}_{2(22)}$ turns out to be consistent with data. We remark that taking an NNI Hermitian $\mathbf{m}_{\ell}$ together with any non-parallel three-zero neutrino pattern does not lead to viable pairs of textures. Therefore, the non-Hermiticity condition of the charged lepton mass matrix is a crucial ingredient in this particular case. Dirac $\mathbf{m}_\nu$ $\chi^2_{min}$ NO (IO) $\Delta m^2_{21}$ $\Delta m^2_{31}$ $\theta_{12}$ $\theta_{23}$ $\theta_{13}$ $\delta$ ------------------------ --------------------- ----------------------- ------------------- ------------------- ----------------------- ---------------------- --------------- ---------------------- $\mathbf{A}_{1(13)}$ $6.44\times10^{-7}$ ($3.51\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{1(22)}$ $3.20\times10^{-8}$ ($1.93\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{1(23)}$ $1.18\times10^{-8}$ ($3.06$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\mathbf{A}_{1(33)}$ $5.08\times10^{-7}$ ($6.03\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{2(22)}$ $5.44\times10^{-8}$ ($2.25\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{2(23)}$ $1.46\times10^{-7}$ ($6.74\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{A}_{2(33)}$ $2.36\times10^{-7}$ ($6.21\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{1(23)}$ $1.31\times10^{-7}$ ($5.63\times10^{-3}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{1(33)}$ $2.01\times10^{-5}$ ($1.27\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{2(13)}$ $1.52$ ($1.51\times10^{-1}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{2(22)}$ $9.22\times10^{-7}$ ($1.64\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{2(23)}$ $1.52$ ($1.28\times10^{-6}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{3(13)}$ $1.43\times10^{-6}$ ($3.05$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark(\times)$ $\checkmark$ $\checkmark(\times)$ $\mathbf{B}_{3(23)}$ $7.99\times10^{-7}$ ($9.56\times10^{-9}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{B}_{4(23)}$ $2.83\times10^{-7}$ ($6.08\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{C}_{(11)}$ $5.62\times10^{-7}$ ($6.20\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{C}_{(23)}$ $9.84\times10^{-8}$ ($1.90\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{D}_{1(11)}$ $1.23\times10^{-6}$ ($1.28\times10^{-7}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark$ $\mathbf{D}_{2(11)}$ $4.23\times10^{-6}$ ($2.14\times10^{-8}$) $\checkmark$ $\checkmark$ $\checkmark$ $\checkmark $ $\checkmark$ $\checkmark$ $\mathbf{F}_{1(23)}$ $1.98\times10^{1}$ ($1.83\times10^{2}$) $\checkmark$ $\checkmark$ $\checkmark (\times)$ $\times$ $\checkmark$ $\times$ Conclusions {#sec:summary} =========== There has been lately a revival of the interest in texture-zero models that aim at explaining the flavor structure observed in lepton mass matrices. In this work, we have confronted various popular texture-zero *Ansätze* of lepton mass matrices with current neutrino data. We have performed a thorough $\chi^2$-analysis in a wide class of schemes, considering Hermitian charged lepton mass matrices in combination with symmetric Majorana or Hermitian Dirac neutrino mass matrices. In our study we included the well-known FGM textures, the so-called hybrid textures, as well as parallel patterns. We concluded that while a significant number of these patterns is still consistent with all the observations at 68.27% C.L., there are several textures that can be excluded or are marginally allowed. We have also considered predictive neutrino zero textures with the assumption that the charged lepton mass matrix has the well-known NNI form. In the latter case, requiring non-Hermiticity of the charged lepton mass matrix is a necessary condition to obtain viable neutrino patterns. Predictive textures were found with a maximum number of three zeros, for both Majorana and Dirac neutrinos. It is well known that texture-zero models have in general a weak predictive power. We have not addressed here the question of the predictability of a given texture. This issue is beyond the scope of the present work. The reader is referred, e.g., to Ref. [@Ludl:2014axa], in which the authors attempted to identify predictive classes of texture zeros by defining numerical measures of predictability. For instance, maximally restrictive Majorana textures can predict, in most cases, the effective neutrino mass parameter $m_{\beta\beta}=|\sum_i \mathbf{U}_{ei}^2\,m_i|$, relevant in neutrinoless double beta decays. From our study of different lepton mass matrix textures, it becomes clear that present neutrino oscillation data does not give an explicit hint on which category of textures is more appropriate to describe the observations. The precise measurements of neutrino oscillation parameters in upcoming experiments (including the determination of the absolute neutrino mass scale and the Dirac CP phase, and the improvement of the bounds on the sum of neutrino masses and the effective mass in $0\nu\beta\beta$ decays) are expected to shed some light on the flavor structure of the neutrino sector. This in turn would allow us to determine, among the plethora of texture-zero patterns, what are the most predictive textures capable of explaining the experimental data, as well as those that are disfavored or excluded at a high confidence level. Acknowledgements {#acknowledgements .unnumbered} ================ We are grateful to M. Nebot and S. Palomares-Ruiz for useful discussions and comments. The work of D.E.C. was supported by *Associação do Instituto Superior Técnico para a Investigação e Desenvolvimento* (IST-ID) and *Fundação para a Ciência e a Tecnologia* (FCT). D.E.C. and R.G.F. acknowledge support from FCT through the projects CERN/FP/123580/2011, PTDC/FIS-NUC/0548/2012 and UID/FIS/00777/2013, and thank CERN Theoretical Physics Unit for hospitality and financial support. Neutrino observables for the FGM two-zero textures {#appendix1} ================================================== For completeness, in this appendix we present the probability distribution of the neutrino observables for the viable FGM patterns $\mathbf{A}_{1,2}$, $\mathbf{B}_{1,2,3,4}$ and $\mathbf{C}$ for NO and IO mass spectra. The results are presented in Figs. \[fig:fig-A1\]-\[fig:fig-C\] and have been obtained using $10^4$ random input neutrino mass matrices. In all figures, the vertical red dashed line denotes the best-fit value of the corresponding observable, taken from Table \[tab:nudata\] for a normal ordering of the neutrino mass spectrum. The histograms reflect the frequency of the values of the six neutrino observables included in the $\chi^2$-analysis, namely, the mass-squared differences $\Delta m_{21}^2,\, \Delta m_{31}^2$, the mixing angles $\mathrm{sin}^2\theta_{12},\, \mathrm{sin}^2\theta_{23},\, \mathrm{sin}^2\theta_{13}$, and the Dirac CP phase $\delta$. {width="0.57\linewidth"} {width="0.57\linewidth"} {width="0.57\linewidth"} {width="0.57\linewidth"} {width="0.57\linewidth"} {width="0.57\linewidth"} {width="0.57\linewidth"} Neutrino observables for three-zero parallel textures ===================================================== Here we present the probability distribution of neutrino observables for the three-zero parallel patterns given in Eq. , in the case of Majorana neutrinos. As in the case of the FGM textures, $10^4$ random input mass matrices for charged leptons and neutrinos were generated. The results are presented in Fig. \[fig:fig-3zeros-Majorana\]. Similar distributions are obtained if neutrinos are Dirac particles. \[appendix2\] {width="0.57\linewidth"} [99]{} W. Grimus, A. S. Joshipura, L. Lavoura and M. Tanimoto, Eur. Phys. J. C [**36**]{} (2004) 227 \[hep-ph/0405016\]. R. González Felipe and H. Serôdio, Nucl. Phys. B [**886**]{} (2014) 75 \[arXiv:1405.4263\]. F. Capozzi, G. L. Fogli, E. Lisi, A. Marrone, D. Montanino and A. Palazzo, Phys. Rev. D [**89**]{} (2014) 9, 093018 \[arXiv:1312.2878 \[hep-ph\]\]. D. V. Forero, M. Tortola and J. W. F. Valle, Phys. Rev. D [**90**]{} (2014) 9, 093006 \[arXiv:1405.7540\]. M. C. Gonzalez-Garcia, M. Maltoni and T. Schwetz, JHEP [**1411**]{} (2014) 052 \[arXiv:1409.5439\]. P. A. R. Ade [*et al.*]{} (Planck Collaboration), “Planck 2015 results. XIII. Cosmological parameters”, arXiv:1502.01589 \[astro-ph.CO\]. G. C. Branco, R. G. Felipe and F. R. Joaquim, Rev. Mod. Phys.  [**84**]{} (2012) 515 \[arXiv:1111.5332\]. P. H. Frampton, S. L. Glashow and D. Marfatia, Phys. Lett. B [**536**]{} (2002) 79 \[hep-ph/0201008\]. H. Fritzsch, Z. -z. Xing and S. Zhou, JHEP [**1109**]{} (2011) 083 \[arXiv:1108.4534\]. D. Meloni and G. Blankenburg, Nucl. Phys. B [**867**]{} (2013) 749 \[arXiv:1204.2706 \[hep-ph\]\]. D. Meloni, A. Meroni and E. Peinado, Phys. Rev. D [**89**]{} (2014) 053009 \[arXiv:1401.3207\]. S. Kaneko, H. Sawanaka and M. Tanimoto, JHEP [**0508**]{} (2005) 073 \[hep-ph/0504074\]. J. Y. Liu and S. Zhou, Phys. Rev. D [**87**]{} (2013) 9, 093010 \[arXiv:1304.2334\]. G. C. Branco, D. Emmanuel-Costa and R. González Felipe, Phys. Lett. B [**477**]{} (2000) 147 \[hep-ph/9911418\]. G. C. Branco, D. Emmanuel-Costa, R. González Felipe and H. Serôdio, Phys. Lett. B [**670**]{} (2009) 340 \[arXiv:0711.1613\]. H. Nishiura, K. Matsuda and T. Fukuyama, Phys. Rev. D [**60**]{} (1999) 013006 \[hep-ph/9902385\]. Z. z. Xing, Phys. Lett. B [**550**]{} (2002) 178 \[hep-ph/0210276\]. M. Gupta and G. Ahuja, Int. J. Mod. Phys. A [**27**]{} (2012) 1230033 \[arXiv:1302.4823 \[hep-ph\]\]. P. O. Ludl and W. Grimus, JHEP [**1407**]{} (2014) 090 \[Erratum-ibid.  [**1410**]{} (2014) 126\] \[arXiv:1406.3546\]. B. Pontecorvo, Sov. Phys. JETP [**7**]{} (1958) 172 \[Zh. Eksp. Teor. Fiz.  [**34**]{} (1957) 247\]; Z. Maki, M. Nakagawa and S. Sakata, Prog. Theor. Phys.  [**28**]{} (1962) 870; B. Pontecorvo, Sov. Phys. JETP [**26**]{} (1968) 984 \[Zh. Eksp. Teor. Fiz.  [**53**]{} (1967) 1717\]. K. A. Olive [*et al.*]{} (Particle Data Group Collaboration), Chin. Phys. C [**38**]{} (2014) 090001. Z. Hou [*et al.*]{}, Astrophys. J.  [**782**]{} (2014) 2, 74 \[arXiv:1212.6267\]. F. James and M. Roos, Comput. Phys. Commun.  [**10**]{} (1975) 343. Rene Brun and Fons Rademakers, ROOT - An Object Oriented Data Analysis Framework, Proceedings AIHENP’96 Workshop, Lausanne, Sep. 1996, Nucl. Inst. & Meth. in Phys. Res. A 389 (1997) 81-86. See also http://root.cern.ch/. X. w. Liu and S. Zhou, Int. J. Mod. Phys. A [**28**]{} (2013) 1350040 \[arXiv:1211.0472\]. H. Fritzsch, Nucl. Phys. B [**155**]{} (1979) 189. G. C. Branco, L. Lavoura and F. Mota, Phys. Rev. D [**39**]{} (1989) 3443. H. Fritzsch, arXiv:1503.07927. G. C. Branco, D. Emmanuel-Costa and C. Simoes, Phys. Lett. B [**690**]{} (2010) 62 \[arXiv:1001.5065\]. K. S. Babu and Y. Meng, Phys. Rev. D [**80**]{} (2009) 075003 \[arXiv:0907.4231\]. D. Emmanuel-Costa and C. Simoes, Phys. Rev. D [**85**]{} (2012) 016003 \[arXiv:1102.3729\]. D. Emmanuel-Costa, C. Simoes and M. Tortola, JHEP [**1310**]{} (2013) 054 \[arXiv:1303.5699\]. [^1]: Similar bounds are inferred from other cosmological observations. For instance, the median value $\sum_i m_i=0.32 \pm 0.11$ eV has been obtained by the South Pole Telescope collaboration, with a $3\sigma$ detection of a positive sum and $\sum_i m_i \in [0.01,0.63]$ eV at 99.7% C.L. [@Hou:2012xq]. [^2]: The seven matrices were previously found to be compatible with neutrino oscillation data at the $1\sigma$ level for NO and IO mass spectrum [@Fritzsch:2011qv]. [^3]: The patterns belonging to the classes I and IV in Eq.  have this property [@Branco:2007nn]. [^4]: Our conclusions do not agree with the result of Ref. [@Fritzsch:2015haa], in which the parallel NNI texture $\mathbf{A}_{2(22)}$ is found to be compatible with the experimental data.

ArXiv

--- abstract: 'We prove that the hyperbolic components of bicritical rational maps having two distinct attracting cycles each of period at least two are bounded in the moduli space of bicritical rational maps. Our arguments rely on arithmetic methods.' address: - 'Einstein Institute of Mathematics, The Hebrew University of Jerusalem' - 'Department of Mathematics, Indiana University' author: - 'Hongming Nie and Kevin M. Pilgrim' bibliography: - 'references.bib' title: Bounded hyperbolic components of bicritical rational maps --- Introduction ============ A degree $d\ge 2$ rational map $f: \mathbb{P}^1 \to \mathbb{P}^1$ defines a dynamical system on the complex projective line. The moduli space $\mathrm{rat}_d:=\mathrm{Rat}_d/\mathrm{Aut}(\mathbb{P}^1)$ is the space of holomorphic conjugacy classes of degree $d$ rational maps, and is naturally an affine variety [@Silverman07 Theorem 4.36]. A rational map is *hyperbolic* if each critical point converges under iteration to an attracting cycle; this property is invariant under conjugacy. The set of hyperbolic maps is an open and conjecturally dense subset of $\mathrm{rat}_d$. A connected component $\mathcal{H}$ of this subset is called a *hyperbolic component*. We are interested in the general question of when a hyperbolic component $\mathcal{H}$ has compact closure in $\mathrm{rat}_d$. In particular, we are interested in the analogous question for natural dynamically defined subvarieties of $\mathrm{rat}_d$. A rational map is *bicritical* if it has exactly two critical points. Equivalently, $f=M\circ z^d \circ N$ for $M, N\in \mathrm{Aut}(\mathbb{P}^1)$. Such maps were studied by Milnor [@Milnor00]. In [@Milnor00 Corollary 1.3], he showed that the moduli space $\mathcal{M}_d\subset\mathrm{rat}_d$ of bicritical rational maps is biholomorphic to $\mathbb{C}^2$. Explicit coordinates are given by $(\sigma_1, \sigma_d)$ where the $\sigma_i$’s are the elementary symmetric functions of the multipliers of the $d+1$ fixed-points [@Milnor00 Remark 2.7]. It follows that the inclusion $\mathcal{M}_d \hookrightarrow \mathrm{rat}_d$ is proper. Hyperbolic components in $\mathcal{M}_d$ are classified as in the quadratic case [@Milnor93]. A hyperbolic component $\mathcal{H}\subset\mathcal{M}_d$ is of *type D* if each map in $\mathcal{H}$ has two distinct attracting cycles; for convenience, we say it is of *strict type D* if neither of the attracting cycles is a fixed point. The properness of the embedding $\mathcal{M}_d \hookrightarrow \mathrm{rat}_d$ implies that $\mathcal{H}$ has compact closure in $\mathcal{M}_d$ if and only if its image has compact closure in $\mathrm{rat}_d$. Our main result is the following, which extends Epstein’s boundedness result for strict type $D$ quadratic rational maps [@Epstein00 Theorem 1]. \[main\] Let $\mathcal{H}\subset\mathcal{M}_d$ be a strict type D hyperbolic component. Then $\mathcal{H}$ has compact closure in $\mathcal{M}_d$. In rough outline, our arguments and Epstein’s are similar. Suppose to the contrary that there exists a degenerating sequence $f_1, f_2, \ldots$ of rational maps in $\mathcal{H}$. Since the sequence lies in a single hyperbolic component $\mathcal{H}$, there is a period $n$ such that each $f_k$ has an attracting cycle of period $n$. In particular, the multiplier of an $n$ cycle remains bounded along this sequence. From this we extract a so-called rescaling limit, $g=\lim g_k$ where $g_k = M_k\circ f^q_k\circ M_k^{-1}$, $M_k \in \mathrm{Aut}(\mathbb{P}^1)$, and $\deg(g) \geq 2$. We analyze the possibilities for the limits of the attracting cycles for the $g_k$ and how such limits relate to the dynamics of $g$. Using the existence of a second cycle of bounded multiplier and a case-by-case analysis, we derive an over-determined set of constraints on the dynamics of critical points of $g$, and hence a contradiction. However, our arguments differ from Epstein’s in key respects. First, Epstein works with sequences $f_k$. Our argument exploits right away the type $D$ hypothesis to conclude that, via the multipliers of the two attracting cycles, $\mathcal{H}$ is algebraically equivalent to $\mathbb{D} \times \mathbb{D}$ and is in particular semi-algebraic [@Milnor12; @Milnor14]. It follows that if $\mathcal{H}$ is unbounded, then we may take the family $f_{k}$ to be of the form $f_{t_k}$ where $f_t$ is a holomorphic one-parameter family, $t \in \mathbb{D}$, with $f_t$ degenerating as $t \to 0$. Second, Epstein derives the existence of rescaling limits and the relationship between the dynamics of $g$ and of the $g_k$ via detailed analytic estimates. In contrast, we rely on much softer arithmetic methods. The holomorphic family $f_t$ induces a dynamical system $\mathbf{f}: \mathbf{P}^1 \to \mathbf{P}^1$ on the Berkovich space associated to a non-Archimedean field that is a completion of the field of Puiseux series [@Kiwi15]. Very loosely, the map $\mathbf{f}$ captures the asymptotics as $t \to 0$ of the family $f_t$ at all possible algebraically parameterized locations and scales. Kiwi [@Kiwi14] gives a thorough analysis of the global structure of these Berkovich dynamical systems in degree two. We begin by generalizing Kiwi’s results from the quadratic to the bicritical case; see section \[Berh-dyn\]. From the existence of a multiplier that remains bounded, we conclude the existence of a rescaling limit $g$–equivalently, a type II repelling cycle for $\mathbf{f}$; see Lemma \[repelling-existence\]. Using a case-by-case analysis, we then derive an over-determined set of constraints on the critical dynamics of $g$. This step uses arithmetic results of Rivera-Letelier (see Theorem \[fp-number\]) and Epstein’s refined version of the Fatou-Shishikura Inequality (see Theorems \[FSI\] and \[refined-FSI\]). Finally, we do not know if Epstein’s original analytic arguments may be extended to the bicritical case. Arguments of Kabelka [@Kabelka10] using a convenient normal form with a distinguished fixed-point should show that boundedness of the multiplier of a cycle yields a nonlinear rescaling limit, which is the first step in both Epstein’s and our argument. When $d>2$ there are more than two other fixed-points and estimating their multipliers is more difficult. It would be interesting to have a close interpretation of Epstein’s original arguments in arithmetic terms. We conclude this introduction with a brief survey of boundedness and unboundedness results for hyperbolic components. Makienko [@Makienko00 Theorem A] gives sufficient general topological-dynamical conditions for unboundedness of hyperbolic components; see also [@Tan02]. Applied to bicritical maps, this immediately yields that components of type A (maps with an attracting fixed-point) and certain components of type B (maps with both critical points in the immediate basin of an attracting cycle of period at least two) are unbounded. However, we do not know if components of type C (so-called capture components, in which both critical points are attracted to an attracting periodic cycle, but only one lies in the immediate basin) can be unbounded, or if they can be bounded. Unlike components of type D, components of types A, B, and C are transcendental objects. If a component $\mathcal{H}$ of type C were unbounded, we do not know how to extract a degenerating holomorphic family $f_t$ such that $f_{t_k} \in \mathcal{H}$ for some sequence $t_k \to 0$.\ [**Acknowledgement.**]{} We thank Jan Kiwi for encouragement and helpful insights. Berkovich dynamics of bicritical rational maps {#Berh-dyn} ============================================== In this section, we first recall some concepts and known results from Berkovich dynamics, and then extend Kiwi’s classification results from the quadratic to the bicritical case. Berkovich space {#berkovich-space .unnumbered} --------------- See [@Baker10; @Kiwi14] for details. The field of Laurent series $\sum_{n \geq N}a_nt^n$ has an algebraic completion whose elements are Puiseux series. It has a further algebraic and metric completion given by a field, here denoted $\mathbb{L}$, defined as follows. The elements of $\mathbb{L}$ are formal series $z=a_{q_0}t^{q_0}+a_{q_1}t^{q_1}+\ldots, q_n \in \mathbb{Q}, q_n \uparrow \infty$. The metric is given by the non-Archimedean absolute value $|z|:=\exp(-q_0)$. With the metric topology induced by the absolute value, the projective space $\mathbb{P}^1_{\mathbb{L}}$ is totally disconnected and not locally compact. To remedy this, it is compactified by the Berkovich projective space over $\mathbb{L}$, denoted $\mathbf{P}^1$, which is a uniquely arcwise connected, compact, Hausdorff topological space. The points in $\mathbb{P}^1_{\mathbb{L}}$ form a dense subset of $\mathbf{P}^1$, which are called *type I* points. We denote by $\mathbb{H}:=\mathbf{P}^1\setminus\mathbb{P}^1_{\mathbb{L}}$. The branch points (those whose complement has three or more components) in $\mathbf{P}^1$ are called *type II* points. A type II point corresponds to a closed ball in $\mathbb{L}$ with radius in the value group of $\mathbb{L}$. The type II point in $\mathbf{P}^1$ corresponding to the closed unit disk in $\mathbb{L}$ is called the *Gauss point* and denoted by $\mathbf{g}$. Suppose $\xi \in \mathbf{P}^1$. Recall that $\mathbf{P}^1$ is a tree. A connected component of $\mathbf{P}^1\setminus \{\xi\}$ is therefore called a *direction* at $\xi$; such a component is also sometimes called an *open Berkovich ball*. Abusing notation and terminology, we denote such a direction by $\vec{v}$. While a direction is determined by the basepoint $\xi$ and any element in the corresponding ball, a direction is not an infinitesimal object. The set of such directions is called the *tangent space* to $\xi$, is denoted $T_{\xi}\mathbf{P}^1$. If $\xi$ is a type II point, then $T_{\xi}\mathbf{P}^1$ is isomorphic to the complex projective space $\mathbb{P}^1$. Dynamics on Berkovich space {#dynamics-on-berkovich-space .unnumbered} --------------------------- In the remainder of this section, $\phi$ denotes an element of $\mathbb{L}(z)$. We are mainly interested in the dynamics of $\phi$, so typically we assume $\deg(\phi) \geq 2$. A rational map $\phi\in\mathbb{L}(z)$ extends uniquely to a map $\phi:\mathbf{P}^1\to\mathbf{P}^1$. In particular, a degree $d$ holomorphic family $f_t, t \in \mathbb{D}$, with $f_t \in \mathrm{Rat}_d\subset \mathbb{P}^{2d+1}$ for $t \neq 0$, induces a map $\mathbf{f}: \mathbf{P}^1\to\mathbf{P}^1$. For each type II point $\xi \in \mathbf{P}^1$, there is an induced map $T_{\xi}\phi: T_{\xi}\mathbf{P}^1\to T_{\phi(\xi)}\mathbf{P}^1$; under the isomorphism of the previous paragraph, it is a rational map of one complex projective line to another, and therefore has a degree, which is a nonnegative integer. A direction $\vec{v}\in T_{\xi}\mathbf{P}^1$ is a *bad direction* of $\phi$ if the image of the corresponding component of $\mathbf{P}\setminus\{\xi\}$ under $\phi$ is the whole space $\mathbf{P}^1$. Otherwise, the image of this component is another such component, and we say $\vec{v}$ is a *good direction* of $\phi$. The properties of ‘bad’ and ‘good’ are not infinitesimal properties. In summary, under $\phi$, good directions map to directions, and bad directions map to all of $\mathbf{P}^1$. A $q$-cycle $\langle\xi\rangle:=\{ \xi_0, \xi_1, \ldots, \xi_{q-1}\} \subset\mathbb{H}$ is *indifferent* if $\deg T_{\xi_0}\phi^q=1$. Otherwise, it is *repelling*. A point $\xi\in\mathbf{P}^1$ belongs to the *(Berkovich) Julia set* $J(\phi)$ if for all neighborhoods $V$ of $\xi$, the set $\cup\phi^n(V)$ omits at most two points. The *(Berkovich) Fatou set* $F(\phi)$ is the complement of $J(\phi)$. Simple maps. {#simple-maps. .unnumbered} ------------ A map $\phi\in\mathbb{L}(z)$ is *simple* if its Julia set $J(\phi)$ is a singleton; equivalently, up to conjugacy, $\phi$ fixes the Gauss point ${\mathbf{g}}$, and the corresponding complex rational map $T_{{\mathbf{g}}}\phi$ has degree $\deg(\phi)$; equivalently, $\phi$ has (potentially) good reduction; see [@Baker10 Lemma 10.53]. For a degree $d$ holomorphic family $f_t$, the induced map $\mathbf{f}$ is simple if and only if $M_t\circ f_t\circ M_t^{-1} \to g \in \mathrm{Rat}_d$ for some holomorphic family $M_t $ of degree $1$ rational maps. Rivera domains {#rivera-domains .unnumbered} -------------- A distinguishing feature of bicritical maps is the presence of special periodic Fatou components, called Rivera domains, in cases of interest. A component $U$ of $F(\phi)$ of period $p$ is a *Rivera domain* if $\phi^p: U \to U$ is a bijection; if $p=1$ we say $U$ is a *fixed* Rivera domain. The boundary of such a Rivera domain consists of a finite set of periodic cycles in $J(\phi)$ [@Rivera03II]. For a fixed Rivera domain $U$ which is not an open Berkovich ball, the convex hull $\mathrm{Hull}(\partial U)$ is an invariant finite simplicial tree. A fixed Rivera domain $U$ is *starlike* if the finite simplicial tree $\mathrm{Hull}(\partial U)$ contains at most one branch point and exactly one fixed point. Let $U\subset\mathbf{P}^1$ be a connected set. For $\xi\in\mathbf{P}^1\setminus U$, denote by $\vec{v}_{\xi}(U)\in T_{\xi}\mathbf{P}^1$ the direction at $\xi$ containing $U$. To ease notation, if $U=\{\xi'\}$ contains exactly one point, instead of $\vec{v}_{\xi}(U)$, we sometimes write $\vec{v}_{\xi}(\xi')$ the direction at $\xi$ containing $U$. The following result is due to Rivera-Letelier, see [@Rivera03II]. For the definition of multiplicity, see [@Baker10]. [@Kiwi14 Theorem 2.4]\[fp-number\] Suppose $\deg(\phi) \geq 2$. Assume that $U$ is a fixed Rivera domain for $\phi$. Let $N(f,U)\ge 0$ be the number of fixed points, counted with multiplicity, in $U\cap\mathbb{P}^1_{\mathbb{L}}$. Then $$N(f,U)= 2+ \sum_{\xi\in\partial U, \phi(\xi)=\xi}(m_\xi(U)-2),$$ where $m_\xi(U)\ge 0$ is the multiplicity of the direction $\vec{v}_{\xi}(U)\in T_{\xi}\mathbf{P}^1$ as a fixed point of $T_\xi\phi$. The ramification locus {#the-ramification-locus .unnumbered} ---------------------- Suppose now $d:=\deg(\phi)\geq 2$. A point $z \in \mathbb{P}^1_{\mathbb{L}}$ is a *critical point* of $\phi$ if $\phi'(z)=0$ when computed in local coordinates; there are $2d-2$ of them, counted with multiplicity. The map $\phi$ is *bicritical* if it has exactly two critical points. Just as for maps defined over $\mathbb{C}$, a bicritical map has the form $\phi(z)=M\circ z^d \circ N$ for some $M, N \in \mathrm{Aut}(\mathbb{P}^1_{\mathbb{L}})$. The *ramification locus* is defined by $$\mathcal{R}_\phi=\{\xi\in\mathbf{P}^1:\deg T_{\xi}\phi\ge 2\}.$$ If some point in $\mathbf{P}^1$ has local degree achieving the maximum possible value $d$, i.e. if $\phi$ is totally ramified, then the ramification locus $\mathcal{R}_\phi$ coincides with the convex hull of the critical points of $\phi$ [@Faber13I Theorem C]. Bicritical rational maps {#bicritical-rational-maps .unnumbered} ------------------------ We next specialize exclusively to bicritical maps $\phi$ of degree $d \geq 2$. The previously mentioned fact and [@Baker10 Theorem 9.42] yield the following, which we state for reference and use repeatedly without explicit mention in what follows. \[ramification\] Let $\phi:\mathbf{P}^1\to\mathbf{P}^1$ be a bicritical rational map. 1. The ramification locus $\mathcal{R}_\phi$ is the segment connecting the two critical points. 2. $\deg T_{\xi}\phi=d$ if $\xi\in\mathcal{R}_\phi$ and equals $1$ otherwise. 3. If $\xi \not\in \mathcal{R}_\phi$, a direction is good if and only if it does not meet $\mathcal{R}_\phi$. If $\xi \in \mathcal{R}_\phi$, every direction is good. Dynamical structure of bicritical maps {#dynamical-structure-of-bicritical-maps .unnumbered} -------------------------------------- Here we generalize Kiwi’s structure result of [@Kiwi14]. In the next result, case (3) will be of primary interest, since rescaling limits correspond to type II repelling periodic points, see [@Kiwi15]. \[Julia-cycle\] Let $\phi:\mathbf{P}^1\to\mathbf{P}^1$ be a bicritical rational map which is not simple. Then exactly one of the following holds: 1. There are no cycles in $J(\phi)\cap\mathbb{H}$. 2. There is exactly one cycle $\mathcal{O}$ in $J(\phi)\cap\mathbb{H}$, and $\mathcal{O}$ is indifferent. 3. There is at least one repelling cycle in $J(\phi)\cap\mathbb{H}$. Using Lemma \[ramification\], the proof of this proposition is exactly the same as the one given in [@Kiwi14 Proposition 3.1] A trichotomy {#a-trichotomy .unnumbered} ------------ The next three propositions give the rough structure of the dynamics in each of these cases, respectively. For completeness, we give the structure results also in cases (1) and (2). They follow from the same arguments in [@Kiwi14 Propositions 4.1 and 4.2], and we omit the proofs. \[attracting\] Let $\phi:\mathbf{P}^1\to\mathbf{P}^1$ be a bicritical rational map which is not simple. Assume that $J(\phi)\cap\mathbb{H}$ contains no periodic points. Then $\phi$ has an attracting fixed point $\xi\in\mathbb{P}^1_{\mathbb{L}}$ and $F(\phi)$ is the immediate basin of $\xi$. \[indifferent\] Let $\phi:\mathbf{P}^1\to\mathbf{P}^1$ be a bicritical rational map which is not simple. Assume that $J(\phi)\cap\mathbb{H}$ contains an indifferent periodic orbit. Then every periodic Fatou component is a fixed Rivera domain. In our main case (3) of interest, there are repelling type II periodic points in $\mathbb{H}$, and we have the following. A degree $d\ge 2$ polynomial $P(z)\in\mathbb{C}[z]$ is *unicritical* if $P(z)$ is conjugate via an element of $\mathrm{Aut}(\mathbb{C})$ to $z^d+c$. \[repelling\] Let $\phi:\mathbf{P}^1\to\mathbf{P}^1$ be a degree $d\ge 2$ bicritical rational map which is not simple. Assume that $J(\phi)\cap\mathbb{H}$ contains a repelling periodic orbit. Then the following statements hold: 1. $\phi$ has exactly one fixed Fatou component $U$ which is a starlike Rivera domain. The boundary of $U$ is a repelling cycle of type II points $\mathcal{O}=\{\xi_0, \cdots, \xi_{q-1}\}$, $q>1$. The set $\mathrm{Hull}(\partial U)$ contains a unique fixed point $\mathbf{c}$ and the map $T_{\mathbf{c}}\phi$ is a rotation of order $q$. The periodic Fatou components of higher periods, if they exist, are open Berkovich balls. 2. $\mathbf{P}^1\setminus U$ is the disjoint union of $q$ closed Berkovich balls $B_0,\cdots, B_{q-1}$ labeled such that $\partial B_j=\{\xi_j\}$. One of these balls, say $B_0$, contains both critical points of $\phi$ and the ramification locus. The local degrees are $\deg_{\xi_0}\phi=d$ and $\deg_{\xi_j}\phi=1$ for $j \in \{1, \ldots, q-1\}$. Moreover, for $j=1,\cdots,q-1$, the bad direction at $\xi_j$ is the direction $\vec{v}_{\xi_j}(U)$. 3. For all $\xi_j$, the first-return map $T_{\xi_j}\phi^q$ is a degree $d$ bicritical rational map with a multiple fixed point at the direction $\vec{v}_{\xi_j}(U)$. 4. $\phi$ has at most $2$ repelling cycles in $\mathbb{H}$. Assume that $\phi$ has another type II repelling $q'$-cycle $\mathcal{O}'\subset\mathbb{H}$. Then $q'>q$ and for any $\xi'\in\mathcal{O}'$, the map $T_{\xi'}\phi^{q'}$ conjugate to a degree $d$ unicritical polynomial. Moreover, there exists a critical point of $\phi$ that belongs to a Fatou component which is an open ball $B'$ with $\partial B'\subset\mathcal{O}'$. Statements $(1)-(2)$ follow from the proofs of conclusion $(4)$ in [@Kiwi14 Theorem 2] and conclusions $(1)-(2)$ in [@Kiwi14 Lemma 5.1]. Note for $\xi_j\in\mathcal{O}$ and $\xi'\in\mathcal{O}'$ (if it exists) the maps $T_{\xi_j}\phi^q$ and $T_{\xi'}\phi^{q'}$ each have exactly two critical points. Then statements $(3)-(4)$ are consequences of [@Arfeux16b Theorem 2] and [@Kiwi15 Proposition 3.4]. To end this section, we illustrate Proposition \[repelling\] by the following example. Below, we let $\xi_{a,|t^r|}\in\mathbf{P}^1$ be the point corresponding to the closed ball in $\{z : |z-a|\leq |t^r|\}\subset\mathbb{L}$. Recall that $\mathbf{P}^1$ is a tree; we use the notation $[x,y]$, $(x,y)$ to denote respectively the closed and open segments joining $x,y \in \mathbf{P}^1$. For $d\ge 2$ consider the bicritical rational map $$\phi(z)=1+\frac{t}{1-z^d+g(t)}\in\mathbb{L}(z),$$ where $g(t)=\sum_{n\ge 1}a_nt^n\in\mathbb{L}$. Put $\xi_0:=\mathbf{g}=\xi_{0, |1|}$ and $\xi_1:=\xi_{1, |t|}$. A computation shows that $\xi_0\mapsto \xi_1 \mapsto \xi_0$ is a repelling $2$-cycle, so here $q=2$ and the boundary of the fixed Rivera domain $U$ is $\partial U=\{\xi_0, \xi_1\}$. Denote $G(z)=T_{{\mathbf{g}}}\phi^2(z)$. Direct computation of $\phi^2(z)$ and then setting $t=0$ and cancelling shows that $$G(z)=\frac{(a_1+1)z^d+d-a_1-1}{a_1z^d+d-a_1}$$ and $G(1)=1$ with $G'(1)=1$. The holes of the degenerate map of degree $d^2$ corresponding to $G$ consist of the indifferent fixed-point $\hat{z}=1$ together with a set $\Lambda$ of $d-1$ other points $h$ satisfying $G(h)=1, G'(h) \neq 0$. We now show that by judicious choice of the coefficients $a_1, a_2, \ldots$, there exists a repelling 3-cycle $\xi_0' \mapsto \xi_1' \mapsto \xi_2' \mapsto \xi_0'$ with $\xi_0' \in [{\mathbf{g}},0]$. To do this, we start with the more modest goal of identifying parameters for which the orbit of the origin $x_0:=0$ under $\phi$ satisfies $x_3:=\phi^3(0)=O(t)$; note that this implies $x_3\in \vec{v}_{{\mathbf{g}}}(0)$. Computations show $$x_1:=\phi(0)=1+t+O(t^2),\;\; x_2:=\phi^2(0)=1+\frac{1}{d+a_1}+O(t).$$ To obtain $x_3\in \vec{v}_{{\mathbf{g}}}(0)$ we must arrange so that $x_2$ lies in a bad direction of $T_{\mathbf{g}}\phi^2$. This corresponds to $1+\frac{1}{d+a_1} \in \Lambda$ and we achieve this by picking some $h \in \Lambda$, solving for $a_1$ in terms of $h$, and setting $a_1$ to this value. To control the next image, we look at the denominator in the defining expression for $\phi$. We find $$1-(x_2)^d+g=ct+O(t^2)$$ where $c$ depends on $a_2$. Then $$x_3=1+\frac{t}{ct+O(t^2)}=1+\frac{1}{c}+O(t)$$ and by appropriate choice of $a_2$ we find $1+\frac{1}{c}=0$ as required. We are looking for an appropriate power $0<r\leq 1$ so that $\xi_0'=\xi_{0, |t^r|}$ is periodic of period $3$ and is deployed as shown in the diagram below, where $\mathrm{Hull}(\partial U)$ is indicated with wiggly edges. Since $x_3=O(t)$, we have that $x_3\in\mathbb{L}$ is in the closed ball corresponding to $\zeta=\xi_{0, |t|}\in \mathbf{P}^1$. In summary, we have the following picture: $$\xymatrix{ x_2\ar@{-}[d] &\infty\ar@2{-}[dd] & &x_1 \ar@{-}[d] \\ \xi_2'& & & \xi_1' \ar@{-}[d] \\ &{\mathbf{g}}=\xi_0\ar@{-}[ul]\ar@{~}[r] \ar@2{-}[d]& {\mathbf{c}}\ar@{~}[r]& \xi_1 \\ &\xi_0'\ar@2{-}[d] & & \\ &\zeta \ar@2{-}[d]& &\\ x_3\ar@{-}[ur] &x_0 =0& & \\ }$$ Assuming $r$ is so chosen, the map $\phi$ induces homeomorphisms of segments $$[\xi_0, \xi_0'] \to [\xi_1, \xi_1'] \to [\xi_0, \xi_2'] \to [\xi_1, \xi_0'].$$ Let $\rho$ denote the hyperbolic length metric on $\mathbb{H}$, and $\ell$ the length of a path with respect to $\rho$, see [@Baker10]. Appropriately normalized we have $\rho(\xi_0, \xi_1)=1$ and $\rho(\xi_{0,1}, \xi_{0, |t^r|})=r$. This metric has the property that $\ell(\phi([a,b])=\ell(a,b)$ if $(a,b)\cap \mathcal{R}_\phi=\emptyset$, and for the bicritical maps considered here, $\ell(\phi([a,b])=d\cdot \ell([a,b])$ if $[a,b] \subset \mathcal{R}_\phi$. Exploiting this and the fact that $$[\xi_1, \xi_0'] = [\xi_1, \xi_0] \cup [\xi_0, \xi_0']$$ we see that the power $r$ must be chosen so that $$d\cdot \ell([\xi_0', \xi_0])=\ell([\xi_0', \xi_0])+\ell([\xi_0, \xi_1])\iff dr+1=r\iff r=\frac{1}{d-1}.$$ At $\xi_0'$, the directions corresponding to $0, \infty$ are invariant under the degree $d$ map $T_{\xi'_0}\phi^3$ and have multiplicity each equal to $d$. It follows that $T_{\xi_0'}\phi^3$ is conjugate to the unicritical polynomial $z^d$. When $d=3$, one may take $$a_1=\frac{7+3\sqrt{3}i}{3+\sqrt{3}i}, \;\; a_2=-\frac{2(49+25\sqrt{3}i)}{9(3+\sqrt{3}i)}, \;\; a_k=0, k>2.$$ Proof of Theorem \[main\] ========================= Unboundedness via a degenerating holomorphic family --------------------------------------------------- Here, we show that an ideal boundary point of a type $D$ hyperbolic component $\mathcal{H}$, if it exists, is accessible through a holomorphic family. A bicritical rational map of type $D$ has two distinct critical points; it also has a non-critical fixed-point. By conjugating so that the two critical points are at zero and infinity and a fixed-point is at $z=1$, It follows that any such map is conjugate to one of finitely many in the two-complex-dimensional algebraic family $$\mathcal{F}:=\left\{ \frac{\alpha z^d + \beta}{\gamma z^d + \delta}: \alpha\delta - \beta\gamma = 1, \; \alpha+\beta=\gamma+\delta \right\}\subset \mathrm{Rat}_d.$$ In suitable affine coordinates on the locus $\alpha+\beta=\gamma+\delta$, the family $\mathcal{F}$ becomes a quadric surface in $\mathbb{C}^3$. A generic line intersects $\mathcal{F}$ in two points. Projection from a generic point on $\mathcal{F}$ onto a generic hyperplane in $\mathbb{P}^3$ yields a birational map $\mathcal{F} \to \mathbb{C}^2$. Explicitly, for $u,v \in \mathbb{C}$, set $$f_{u,v}(z):={\frac {( (u+2) v+{u}^{2}+2\,u+2 ) {z}^{d}+u v+{v}^{2}-1}{ ( -{v}^{2}+uv+ \left( u+1 ) ^{2} \right) {z}^ {d}+ ( u+2\,v+2) v}}=\frac{Az^d+B}{Cz^d+D}.$$ Noting that $AD-BC=\left( v+1 \right)^2 \left( u+v+1 \right)^2$, we conclude $$\mathcal{F}\simeq \{f_{u,v} : v +1 \neq 0, u+v+1\neq 0\}\subset \mathbb{C}^2_{u,v}.$$ We compactify this family as $\overline{\mathcal{F}}:=\{[u:v:r] \in \mathbb{P}^2\}$ by adding to $\mathcal{F}$ the *degeneracy locus* $\Delta$ consisting of the line at infinity and the two lines where the resultant vanishes: $$\Delta:=\{r=0, v+r=0, u+v+r=0\}.$$ Let $\mathcal{M}_d^\sharp$ be the moduli space of critically marked bicritical rational maps. The forgetful map $\mathcal{M}_d^\sharp \to \mathcal{M}_d$ is proper, so a hyperbolic component in $\mathcal{M}_d$ is bounded if and only if its lift to $\mathcal{M}_d^\sharp$ is bounded. Suppose now $\mathcal{H} \subset \mathcal{M}^\sharp_d$ is a hyperbolic component of type $D$ and $\widetilde{\mathcal{H}} \subset \mathcal{F}$ is its lift to the family $\mathcal{F}$. The two critical points in $\mathcal{F}$, being located at the origin and at infinity, are therefore marked, hence so are the two corresponding attracting cycles. The multipliers $\lambda, \mu$ of these cycles are complex algebraic functions which are well-defined on $\widetilde{\mathcal{H}}$, yielding an isomorphism $$\widetilde{\mathcal{H}} \to \mathbb{D}\times\mathbb{D}$$ given by $$f_{uv} \mapsto (\lambda(u,v), \mu(u,v)).$$ We conclude that $\widetilde{\mathcal{H}}$ is a real-algebraic domain in $\mathcal{F}\subset \mathbb{C}_{u,v}$. We now specialize to the case that $\widetilde{\mathcal{H}}$ is strict type $D$, and we suppose, contrary to the conclusion of Theorem \[main\], that the image of $\widetilde{\mathcal{H}}$ in moduli space is unbounded. Then $\widetilde{\mathcal{H}}$ meets one of the lines in $\Delta$ at some point. In suitable complex affine coordinates $(x,y)$, this point is the origin, and $\widetilde{\mathcal{H}}$ is described by two real-algebraic inequalities. The Curve Selection Lemma [@Milnor68 Lemma 3.1] implies that there is a real-analytic curve $t \mapsto (x(t), y(t)), 0 \leq t \leq 1$, such that $x(0)=y(0)=0$ and $(x(t),y(t)) \in \widetilde{\mathcal{H}}$ for $0<t\leq 1$. Complexifying this curve as in [@Nie-P], we conclude: *if $\mathcal{H}$ is unbounded, then there exists a holomorphic family $t \mapsto f_{t}:=f_{u(t), v(t)}$ such that for some sequence of parameters $t_k \to 0$, the corresponding maps $f_{t_k}$ belong to $\widetilde{\mathcal{H}}$ and diverge in the moduli space $\mathrm{rat}_d$ as $k \to \infty$*. Bounded multiplier implies existence of a rescaling limit --------------------------------------------------------- We say a holomorphic family $\{f_t\}$ of rational maps is *truly degenerate* if the induced map $\mathbf{f}\in\mathbb{L}(z)$ is not simple. If a sequence $\{f_{t_k}\}$, arising from a holomorphic family $\{f_t\}$ of degree $d\ge 2$ rational maps, diverges in the moduli space $\mathrm{rat}_d$, that is $[f_{t_k}]\to\infty$, then $\{f_t\}$ is truly degenerate. Here, we show that existence of a bounded multiplier for a truly degenerating family implies the existence of a rescaling limit. \[repelling-existence\] Let $\{f_t\}$ be a truly degenerate holomorphic family of bicritical rational maps. Assume that there is an $n$-cycle $\langle z(t_k)\rangle$ of $f_{t_k}$ with period $n\ge 2$ and bounded multiplier as $t_k\to 0$. Then the induced map $\mathbf{f}$ has a type II repelling cycle. We remark that Proposition \[repelling\] shows that the period $q$ of this repelling cycle satisfies $1<q$; in Corollary \[Rivera-boundary-period\] below we refine this to establish $1<q\leq n$, though this fact is not needed in our proof. The elements of an $n$-cycle of $f_t$ are algebraic functions of the parameter $t$ and define elements of $\mathbb{P}^1_{\mathbb{L}}$ comprising an $n$-cycle $\langle \mathbf{z}\rangle$ of $\mathbf{f}$. Now suppose the multiplier of some $n$-cycle of $f_{t_k}$ is bounded as $k \to \infty$. Passing to a subsequence, we may assume these cycles are parameterized by a common algebraic function $\langle z(t)\rangle$; we let $\lambda(t)$ denote the multiplier of this cycle. Since $\lambda(t)$ is algebraic and $\lambda(t_k)$ is bounded as $t_k \to 0$, the function $\lambda(t)$ is bounded as $t \to 0$. It follows that $\langle\mathbf{z}\rangle\subset\mathbb{P}^1_{\mathbb{L}}$ is an attracting or indifferent $n$-cycle of $\mathbf{f}$. For each of these two possibilities, we consider the three cases in the conclusion of Proposition \[Julia-cycle\], and we will rule out the first two. Thus the third–the existence of a type II repelling cycle and, hence, of a rescaling limit–must hold. Suppose $\langle\mathbf{z}\rangle$ is attracting. The cycle $\langle\mathbf{z}\rangle$ is in the Fatou set of $\mathbf{f}$. If the first case of Proposition \[Julia-cycle\] holds–that is, there are no periodic cycles in $\mathbb{H}$–then Proposition \[attracting\] implies that the entire Fatou set of $\mathbf{f}$ is the immediate basin of an attracting fixed point of $\mathbf{f}$; which is impossible since $n \geq 2$. If the second case holds, we have an indifferent cycle $\langle \xi \rangle \subset \mathbb{H}$, and Proposition \[indifferent\] implies that each Fatou component of $\mathbf{f}$ is a fixed Rivera domain. As in our argument for the first case, this contradicts the existence of an attracting $n$-cycle. Now suppose $\langle\mathbf{z}\rangle$ is indifferent. The cycle $\langle\mathbf{z}\rangle$ is in the Fatou set of $\mathbf{f}$; see [@Benedetto10 Proposition 4.24] and [@Baker10 Theorem 10.67]. The first case is ruled out by arguing as in the previous paragraph using Proposition \[attracting\]. We conclude $J(\phi)\cap\mathbb{H}$ contains a cycle. Next, suppose we are in the second case, so that say $\langle \xi \rangle$ is the unique cycle in $J(\phi)\cap\mathbb{H}$ and is indifferent. Let $U$ be a Fatou component of $\mathbf{f}$ which contains a point in $\langle\mathbf{z}\rangle$. Then $U$ is periodic, and hence a fixed Rivera domain by Proposition \[indifferent\]. So $\langle\mathbf{z}\rangle\subset U$. The boundary $\partial U$ is also a cycle in $J(\phi) \cap \mathbb{H}$ and so since there is exactly one such cycle we have $\partial U = \langle \xi \rangle$ is indifferent. Hence for each point $\xi\in\partial U$, the degree $\deg T_{\xi}\mathbf{f}^n=1$, and hence the multiplicity of the direction $\vec{v}_{\xi}(U)\in T_{\xi}\mathbf{P}^1$ as a fixed point of $T_\xi\mathbf{f}^n$ is at most $1$. Now consider the number $N(\mathbf{f}^n,U)$ of fixed points of $\mathbf{f}^{n}$ in $U\cap\mathbb{P}^1_\mathbb{L}$. Thus by Theorem \[fp-number\], we have $N(\mathbf{f}^n,U)<2$. On the other hand, since each point in $\langle\mathbf{z}\rangle$ is a fixed point of $\mathbf{f}^n$, it follows that $N(\mathbf{f}^n,U)\ge n\ge 2$. This is impossible. Coarse structure of the Berkovich dynamics ------------------------------------------ Lemma \[repelling-existence\] provides the existence of a rescaling limit. Our next result, Lemma \[lemma:coarse-structure\], applies Proposition \[repelling\] and some basic combinatorial arguments to describe how the Rivera domain $U$ is deployed relative to other dynamical features. Though Lemma \[lemma:coarse-structure\] describes two cases, our subsequent arguments do not distinguish between them. In the figures below, the ramification locus is indicated with doubled edges, and the edges comprising the simplicial tree $\mathrm{Hull}(\partial U)$ are indicated with wiggly edges. \[lemma:coarse-structure\] Let $\{f_t\} \subset\overline{\mathcal{F}}$ be a truly degenerate holomorphic family of bicritical rational maps. Assume that there is an $n$-cycle $\langle z(t_k)\rangle$ of $f_{t_k}$ with period $n\ge 2$ and bounded multiplier as $t_k\to 0$. Then the induced map $\mathbf{f}$ has a fixed starlike Rivera domain $U$ with center ${\mathbf{c}}$ a type II indifferent fixed point, and $\partial U = \langle {\mathbf{g}}=\xi_0, \xi_1, \ldots, \xi_{q-1}\rangle$ a type II repelling $q$-cycle. Furthermore, either 1. ${\mathbf{g}}\in ({\mathbf{c}}, 1)$: $$\xymatrix{ & \xi_1\ar@{~}[d] & \infty \ar@2{-}[d]& \\ \xi_2 \ar@{~}[r] & {\mathbf{c}}\ar@{~}[r] \ar@{~}[d]&{\mathbf{g}}\ar@{-}[r] \ar@2{-}[d]& 1\\ & \xi_{q-1}& 0 & \\ }$$ or 2. ${\mathbf{c}}\in ({\mathbf{g}}, 1)$: $$\xymatrix{ \infty \ar@2{-}[d] & \xi_1 & p\\ {\mathbf{g}}\ar@2{-}[d] \ar@{~}[r] & {\mathbf{c}}\ar@{~}[d] \ar@{-}[r] \ar@{~}[u] \ar@{-}[ur]& 1\\ 0 & \xi_{q-1}\\ }$$ In this case, we have 1. $({\mathbf{g}}, 1] \subset U$; 2. $\mathbf{f}$ sends the direction $\vec{v}_{{\mathbf{c}}}(1)$ onto itself via a homeomorphism; 3. the set $U\cap\mathbb{P}^1_{\mathbb{L}}$ contains exactly $2$ fixed points of $\mathbf{f}$, the point $1$ and a point $p \neq 1$, each of multiplicity $1$; 4. ${\mathbf{c}}\in(p,1)$ and $[{\mathbf{c}}, p] \cap [{\mathbf{c}}, 1] = \{{\mathbf{c}}\}$. Since $f_t$ is assumed truly degenerate, the induced map $\mathbf{f}$ is not simple. The bounded multiplier hypothesis implies, by Lemma \[repelling-existence\], the existence of a type II repelling cycle. Proposition \[repelling\] then gives the existence of a fixed starlike Rivera domain $U$; we adopt the notation in the statement for the center and boundary of $U$. We first show ${\mathbf{g}}=\xi_0 \in \partial U$. By Proposition \[repelling\](3), there is a unique point, denoted by ${\mathbf{x}}$, in the intersection of $\partial U$ and the ramification locus $\mathcal{R}_{\mathbf{f}}=[0, \infty]$. Suppose ${\mathbf{x}}\neq {\mathbf{g}}$. The center ${\mathbf{c}}$ cannot lie on the segment $[{\mathbf{x}},{\mathbf{g}}]\cup[{\mathbf{g}},1]$. We have the following configuration in $\mathbf{P}^1$, with the double arrow indicating a subsegment of the ramification locus: $$\xymatrix{ \mathbf{f}({\mathbf{x}})={\mathbf{y}}\ar@{-}[r] & {\mathbf{c}}\ar@{-}[r] & {\mathbf{x}}\ar@2{-}[r]&{\mathbf{g}}\ar@{-}[r]&1=\mathbf{f}(1). }$$ Recall $T_{\mathbf{x}}\mathbf{P}^1, T_{\mathbf{y}}\mathbf{P}^1 \simeq \mathbb{P}^1$. The map $T_{\mathbf{x}}\mathbf{f}: T_{\mathbf{x}}\mathbf{P}^1\to T_{\mathbf{y}}\mathbf{P}^1$ is rational and has degree $d$. By Lemma \[ramification\](3), each direction at ${\mathbf{x}}$ is a good direction. Since $\mathbf{f}(1)=1$, the direction $\vec{v}_{{\mathbf{x}}}(1)$ maps to the direction $\vec{v}_{{\mathbf{y}}}(1)$ with local degree $d$ under $T_x\mathbf{f}$. But by Proposition \[repelling\](2), $\mathbf{f}$ maps the segment $[{\mathbf{c}},{\mathbf{x}}]$ to the segment $[{\mathbf{c}},{\mathbf{y}}]$. So the direction $\vec{v}_{{\mathbf{x}}}({\mathbf{c}})$ also maps to the direction $\vec{v}_{{\mathbf{y}}}({\mathbf{c}})=\vec{v}_{{\mathbf{y}}}(1)$. This implies the rational map $T_{{\mathbf{x}}}\mathbf{f}: T_{{\mathbf{x}}}\mathbf{P}^1\to T_{{\mathbf{y}}}\mathbf{P}^1$ has at least $d+1$ preimages of the point corresponding to the direction $\vec{v}_{{\mathbf{y}}}(1)$, which is impossible. Now suppose we are not in Case (1), and that to the contrary we have ${\mathbf{c}}\not\in ({\mathbf{g}}, 1)$. Let ${\mathbf{v}}$ be the projection of ${\mathbf{c}}$ onto $({\mathbf{g}},1)$ and let ${\mathbf{w}}=\mathbf{f}({\mathbf{v}})$: $$\xymatrix{ & {\mathbf{c}}\ar@{~}[d]\\ {\mathbf{g}}\ar@{~}[r] & {\mathbf{v}}\ar@{-}[r]& 1 }$$ In the starlike Rivera domain $U$, the intervals $({\mathbf{c}}, \xi_i], ({\mathbf{c}},\xi_j]$ are disjoint when $i\neq j$. The restriction $\mathbf{f}: [{\mathbf{c}},1] \to [{\mathbf{c}}, 1]$ is a homeomorphism fixing each endpoint, since this segment is outside the ramification locus and the endpoints are fixed. The restriction $\mathbf{f}: ({\mathbf{c}},\xi_0={\mathbf{g}}] \to ({\mathbf{c}}, \xi_1]$ is also homeomorphism. In particular it is a homeomorphism from $({\mathbf{c}},{\mathbf{v}}]$ to its image $({\mathbf{c}}, {\mathbf{w}}]$. Thus ${\mathbf{w}}\in ({\mathbf{c}}, 1)$. The segments $({\mathbf{c}}, {\mathbf{g}}=\xi_0]$ and $({\mathbf{c}}, \xi_1]$ must then overlap in a nonempty subsegment of $({\mathbf{c}},{\mathbf{v}}]$; this is impossible. It remains to verify the claims *(a)-(d)* in Case (2). The direction $\vec{v}_{\mathbf{c}}(1)$ is a good direction. The segment $[{\mathbf{c}}, 1]$ has fixed endpoints and does not meet the ramification locus, hence it maps to itself by a homeomorphism, verifying *(b)*. This observation also shows $\emph{(a)}$: otherwise, some $\xi_i \in \partial U \cap ({\mathbf{c}},1)$, $i \neq 0$, whence $\xi_i$ cannot iterate to ${\mathbf{g}}= \xi_0$. By Theorem \[fp-number\], the map $\mathbf{f}$ has exactly $2$ fixed points in $U\cap\mathbb{P}^1_{\mathbb{L}}$. Note $T_{\mathbf{c}}\mathbf{f}$ is a rotation of order $q\ge 2$. Then $T_{\mathbf{c}}\mathbf{f}$ fixes $2$ distinct directions, and *(c),(d)* follow. Structure of the rescaling limit -------------------------------- In this subsection, we let $\{f_t\}$ be a holomorphic family such that $f_t\in\mathcal{F}$ if $t\not=0$ and satisfies the assumptions in the Lemma \[repelling-existence\]. Let $U$ be the fixed Rivera domain given by Lemma \[lemma:coarse-structure\]. Then the Gauss point $\mathbf{g}=\xi_0$ lies in the cycle $\partial U=\langle \xi_0, \ldots, \xi_{q-1}\rangle$ of $\mathbf{f}$ and so is fixed under $\mathbf{f}^q$. Since this cycle is repelling, the reduction of $\mathbf{f}^q$ at $\mathbf{g}$ has degree at least two. It follows that $f_t^q$ converges locally uniformly away from a finite set to a nonlinear rational map $G$; see [@Kiwi15 Proposition 3.4]. In this subsection, we analyze the structure of $G$. There is a unique isomorphism $T_{{\mathbf{g}}}\mathbf{P}^1\to \mathbb{P}^1$ sending the directions at ${\mathbf{g}}$ corresponding to $0, \infty, 1 \in \mathbb{P}^1_{\mathbb{L}}$ to the points $0, \infty, 1 \in \mathbb{P}^1$, respectively. We denote by $G(z)=T_{{\mathbf{g}}}\mathbf{f}^q(z)$ the corresponding rational map. With these normalizations, $G \in \mathcal{F}$. Lemma \[lemma:coarse-structure\] and Proposition \[repelling\](3) imply the direction $\vec{v}_{{\mathbf{g}}}(U)$ corresponds to a multiple fixed point $\hat{z}$ of $G$. So $\hat{z}\in \mathbb{P}^1, G(\hat{z})=\hat{z}, G'(\hat{z})=1$. Lemma \[lemma:coarse-structure\] gives the location of $\hat z$. Indeed, in case (1) of Lemma \[lemma:coarse-structure\], $\hat{z}\neq 1$; in case (2), $\hat{z}=1$. Expressing $f_t^q$ and $G$ in projective coordinates as a pair of homogeneous degree $d$ polynomials, we have $f_t^q \to H\cdot G$ where $H$ is a homogeneous polynomial whose roots–called the *holes* of the limit–correspond to points where the convergence $f_t^q \to G$ fails to be locally uniform; see [@DeMarco05 Lemma 4.2]. By Proposition \[repelling\] (2), the bad directions of $\mathbf{f}^q$ at the Gauss point ${\mathbf{g}}$ are the directions corresponding to points in $G^{-1}(\hat z)$. By [@Faber13I Lemma 3.17], the set of holes of the degenerate rational map corresponding to $G$ is the set $G^{-1}(\hat z)$. Recalling that $\hat{z}$ is fixed, we write the set of holes as $G^{-1}(\hat{z})=\{\hat{z}\}\cup \Lambda$, so that $\Lambda:=G^{-1}(\hat{z})-\{\hat{z}\}$. Our first result asserts that the holes are disjoint from the critical points of $G$. \[preimage-simple\] Let $G$ and $\hat z$ be as above. Then $$G^{-1}(\hat z)\cap\{0, \infty\}=\emptyset.$$ Note $\hat z\in G^{-1}({\hat z})$ since $\hat z$ is a fixed point of $G$. By Proposition \[repelling\], $G'(\hat z)=1$, so $\hat z\not\in\{0,\infty\}$ since $0,\infty$ are critical points. Suppose to the contrary that $G(0)=\hat{z}$. Then by Proposition \[repelling\], the directions $\vec{v}_{{\mathbf{g}}}(0)$ and $\vec{v}_{{\mathbf{g}}}(\hat z)$ both map to $\vec{v}_{\xi_1}({\mathbf{g}})$ with degree $d$ and $1$, respectively. Thus under $G$ we have $0\mapsto \hat{z}$ by degree $d$ and $\hat{z}\mapsto \hat{z}$ by degree $1$. This is impossible since $\deg(G)=d$. The case $G(\infty)=\hat{z}$ is ruled out similarly. Limits of cycles with bounded multipliers ----------------------------------------- We continue the setup and notation of the previous subsection. By assumption, the multiplier $\lambda(t_k)$ of $f_{t_k}$ associated to the cycle $\langle z({t_k})\rangle$ remains bounded as $t_k \to 0$. The proof of Lemma \[repelling-existence\] shows that $\lambda(t)$ in fact converges, say to $\lambda(0)$, as does the cycle itself, say $\langle z(t)\rangle \to \Gamma \subset \mathbb{P}^1$. Elementary arguments show easily that $G(\Gamma) \subset \Gamma$. In general, equality need not hold. The next lemma gives the possibilities in our setting. \[limit-infinity\] Let $G, \hat z, \Lambda, \Gamma$ be as above. Then the limit $\Gamma$ of the cycle $\langle z(t)\rangle$ either - contains a preperiodic critical point that iterates under $G$ to $\hat{z}$ - contains a cycle disjoint from $\hat{z}$, or - collapses to the singleton $\{\hat{z}\}$. More precisely, $\hat z\in\Gamma$ and exactly one of the following holds. 1. $\Gamma\cap \Lambda\not=\emptyset$ Then $J(\mathbf{f})$ contains two distinct type II repelling cycles, and the set $\{0,\infty\}\cap\Gamma$ consists of a single point $c'$. The point $c'$ is a critical point of $G$ and $G^\ell(c')=\hat{z}$ for some $1<\ell< n$. 2. $\Gamma\cap \Lambda=\emptyset$ and $\Gamma\not=\{\hat z\}$. Then $\Gamma\setminus\{\hat z\}$ is a cycle of $G$ and $q\mid n$. 3. $\Gamma=\{\hat z\}$. Recall that by Proposition \[repelling\] (2), the bad directions of $\mathbf{f}^q$ at the Gauss point ${\mathbf{g}}$ are the directions corresponding to points in $G^{-1}(\hat z)$. We begin with a preliminary result. Suppose $\Gamma\cap\Lambda\not=\emptyset$. We show that under this assumption, $\mathbf{f}$ has a second repelling cycle $\mathcal{O}'$ distinct from that containing $\mathbf{g}$. Let $w\in\Gamma\cap\Lambda$. The direction $\vec{v}_{{\mathbf{g}}}(w)$ is bad for $\mathbf{f}^q$ and so ${\mathbf{g}}$ has preimages under $\mathbf{f}^q$ in the direction $\vec{v}_{{\mathbf{g}}}(w)$. Hence the Julia set $J(\mathbf{f})$ meets the direction $\vec{v}_{{\mathbf{g}}}(w)$ since ${\mathbf{g}}\in J(\mathbf{f})$. The cycle $\langle\mathbf{z}\rangle\subset\mathbb{P}^1_{\mathbb{L}}$ is nonrepelling. Hence $\langle\mathbf{z}\rangle$ is in the classical Fatou set of $\mathbf{f}$ ([@Benedetto10 Proposition 4.24]) and hence in the Berkovich Fatou set of $\mathbf{f}$ [@Baker10 Theorem 10.67]. Thus the map $\mathbf{f}$ has periodic Fatou components which are not fixed. By Proposition \[repelling\] (1), these periodic components are open balls. It follows that the corresponding boundary points of these Fatou components form a periodic cycle $\mathcal{O}'\subset J(\mathbf{f})$ for which $\mathbf{g}\not\in\mathcal{O}'$. Therefore, $J(\mathbf{f})\cap\mathbb{H}$ contains at least two cycles. Then by Proposition \[Julia-cycle\], $\mathcal{O'}$ is repelling cycle distinct from that containing $\mathbf{g}$. Now we claim that $\Gamma\setminus\Lambda\not=\emptyset$. For otherwise, $\Gamma\cap\Lambda=\Gamma\not=\emptyset$, and the hypothesis of the setup of the previous paragraph is satisfied. By the conclusion of the previous paragraph, we have a second repelling cycle $\mathcal{O}'$. Recall that the ramification locus is $\mathcal{R}_{\mathbf{f}}=[0, \infty]$. We have $$\mathcal{O}'\subset\bigcup_{w\in G^{-1}(\hat{z})}\vec{v}_{{\mathbf{g}}}(w).$$ By Lemma \[preimage-simple\], it follows that $\mathcal{O}'$ is disjoint with the ramification locus $\mathcal{R}_{\mathbf{f}}$. Thus $\mathcal{O}'$ is indifferent, which contradicts the fact that $\mathcal{O}'$ is repelling. The preceding paragraph implies there exists $z\in\Gamma\setminus\Lambda$. Then either $z=\hat z$, and hence $\hat z\in\Gamma$, or the direction $\vec{v}_{{\mathbf{g}}}(z)$ is a good direction. In the latter case, since $z\in\Gamma$, there exists $\mathbf{z}_i\in\langle\mathbf{z}\rangle$ such that $\mathbf{z}_i\in\vec{v}_{{\mathbf{g}}}(z)$. Note $\mathbf{f}({\mathbf{g}})$ is contained in the direction $\vec{v}_{{\mathbf{g}}}(\hat z)$, see Proposition \[repelling\]. It follows that $\mathbf{z}_{i+1}\in\vec{v}_{{\mathbf{g}}}(\hat z)$ and so $\hat z\in\Gamma$ in this case too. Finally, we prove the second part of conclusion (1), and conclusion (2). Suppose $\Gamma\cap\Lambda\not=\emptyset$. Since the cycle $\mathcal{O}'$ is repelling, it contains an element, say $\xi'$, that lies in the ramification locus. Proposition \[repelling\] (4) implies there is a critical point $c'\in\{0,\infty\}$ such that the direction $\vec{v}_{\xi'}(c')$ is a good direction of $\mathbf{f}^n$. Note there exists $\mathbf{z}_i\in\langle\mathbf{z}\rangle$ such that $\xi'\in(\mathbf{z}_i, {\mathbf{g}})$. It follows that $\mathbf{z}_i$ is in the direction $\vec{v}_{{\mathbf{g}}}(c')$. Hence $c'\in\Gamma$. Since $\hat z$ is a multiple fixed point, there exists a critical point $c$ of $G$ which is attracted to $\hat z$. Then the direction $\vec{v}_{{\mathbf{g}}}(c)$ is contained in the Fatou set $F(\mathbf{f})$ and $\langle\mathbf{z}\rangle\cap\vec{v}_{{\mathbf{g}}}(c)=\emptyset$. Hence $c\not\in\Gamma$. Since $G \in \mathcal{F}$, the set of critical points of $G$ is $\{0,\infty\}$, so $\{0,\infty\}\cap\Gamma=\{c'\}$. Now we claim that $c'$ iterates under $G$ to $\hat z$. To prove the claim, we show there exists $k\ge 1$ such that $\vec{v}_{{\mathbf{g}}}(c')$ maps to $\vec{v}_{\xi_1}(U)$ under the tangent map $T_{{\mathbf{g}}}\mathbf{f}^{kq+1}$. First note $\vec{v}_{{\mathbf{g}}}(c')$ intersects with the Julia set $J(\mathbf{f})$ since $\vec{v}_{{\mathbf{g}}}(c')\cap\mathcal{O}'\not=\emptyset$. Suppose there exists no such $k$. Then by Proposition \[repelling\] (2), for any $k\ge 0$, the direction $T_{{\mathbf{g}}}\mathbf{f}^{kq+1}(\vec{v}_{{\mathbf{g}}}(c'))$ is a good direction at $\xi_1$. Again by Proposition \[repelling\] (1) and (2), it follows that $T_{{\mathbf{g}}}\mathbf{f}^{m}(\vec{v}_{{\mathbf{g}}}(c'))$ is a good direction at $\mathbf{f}^m({\mathbf{g}})$ for all $m\ge 0$. Therefore, $U\bigcap\bigcup_{m\ge 0}\mathbf{f}^m(\vec{v}_{{\mathbf{g}}}(c'))=\emptyset$, and hence $\vec{v}_{{\mathbf{g}}}(c')$ is contained in $F(\mathbf{f})$, which is a contradiction. Noting that $T_{{\mathbf{g}}}\mathbf{f}^{(k+1)q}(\vec{v}_{{\mathbf{g}}}(c'))=\vec{v}_{{\mathbf{g}}}(U)$, we conclude $G^{k+1}(c')=\hat z$. Now choose the smallest such $k$ and set $\ell=k+1$. Then $G^{\ell}(c')=\hat z$. Note $qk<n$ since $\mathbf{f}^{qk}(\mathbf{z}_i)$ is not in the direction $\vec{v}_{{\mathbf{g}}}(c')$. We have $1<\ell=k+1\le qk<n$ since $q\ge 2$. If $\Gamma\cap\Lambda=\emptyset$ and $\Gamma\not=\{\hat z\}$, we may assume that $\mathbf{z}_0\in\vec{v}_{{\mathbf{g}}}(z_0)$ for some $z_0\in\mathbb{C}\setminus\{0,\hat z\}$, and assume $\mathbf{z}_1\in\vec{v}_{{\mathbf{g}}}(\hat z)$. For $k \in \mathbb{N}$ let $z_{kq}\in\Gamma$ be such that $\mathbf{z}_{kq}\in\vec{v}_{{\mathbf{g}}}(z_{kq})$. Note the directions $\vec{v}_{{\mathbf{g}}}(z_{kq})$ are away from the bad directions of $\mathbf{f}^q$ at ${\mathbf{g}}$. Then $G(z_{kq})=z_{(k+1)q}$, see [@Kiwi15 Lemma 3.2], and $q\mid n$. It follows that the $z_{kq}$s comprise a cycle of $G$. The conclusion that $\{z_{kq}\}=\Gamma\setminus\{\hat z\}$ then follows by Proposition \[repelling\](1). Now we do a more elaborate analysis for the case (1) in the above result. Mainly, we focus on the repelling cycle $\mathcal{O}'\not=\partial U$. \[cycle-period\] Fix the notations as before. Suppose $\Gamma\cap \Lambda\not=\emptyset$ and let $q'$ be the period of $\mathcal{O}'$. Then $q'\mid n$ and $\mathcal{O}'$ consists of the boundaries of the Fatou components containing $\langle\mathbf{z}\rangle$. In particular, let $\xi'\in\mathcal{O}'\cap[0,\infty]$ and let $\mathbf{z}_i\in\langle\mathbf{z}\rangle$ be in the direction $\vec{v}_{{\mathbf{g}}}(c')$; then $\mathbf{f}^{kq'}(\mathbf{z}_i)$ is not contained in the direction $\vec{v}_{{\mathbf{g}}}(c)$, where $c\in\{0,\infty\}$ and $c\not=c'$. By Lemma \[limit-infinity\] (1), renumbering the points in $\langle\mathbf{z}\rangle$, we assume $\mathbf{z}_0$ and $c'$ are in the same direction at ${\mathbf{g}}$. Note $\langle\mathbf{z}\rangle$ is a nonrepelling cycles of $\mathbf{f}$ and hence is contained in the Fatou set $F(\mathbf{f})$. By Proposition \[repelling\] (1), the Fatou component $\Omega(\mathbf{z}_0)$ containing $\mathbf{z}_0$ is an open ball. Then the boundary $\partial\Omega(\mathbf{z}_0)$ is a periodic point in the Julia set $J(\mathbf{f})$. By Proposition \[Julia-cycle\] and Lemma \[repelling-existence\], $\partial\Omega(\mathbf{z}_0)$ is repelling. Now we claim $\partial\Omega(\mathbf{z}_0)\in \mathcal{O'}$. Since $G$ has a parabolic fixed point $\hat z$, a critical point is attracted to $\hat z$. So this critical point is $c\not=c'$ and there is no periodic point in $({\mathbf{g}},c)$. Thus by Proposition \[repelling\] (4), $\mathcal{O'}\cap({\mathbf{g}},c')\not=\emptyset$ contains exactly one point $\xi'$. Again by Proposition \[repelling\] (4), $\mathbf{f}$ has exactly two repelling cycles $\mathcal{O}$ and $\mathcal{O'}$ in $\mathbb{H}$. If $\partial\Omega(\mathbf{z}_0)\notin \mathcal{O'}$, then $\partial\Omega(\mathbf{z}_0)\in \mathcal{O}$. Lemma \[limit-infinity\] (1) implies that $\Omega(\mathbf{z}_0)$ are contained in the direction $\vec{v}_{\mathbf{g}}(c')$. Since $\Omega(\mathbf{z}_0)$ is an open ball, we have $\Omega(\mathbf{z}_0)=\vec{v}_{\mathbf{g}}(c')$, which is impossible because $\xi'$ is contained in $\vec{v}_{\mathbf{g}}(c')$. Now by Proposition \[repelling\](4), the map $T_{\xi'}\mathbf{f}^{q'}$ is conjugate to a polynomial of degree $d$. Then by Lemma \[ramification\] (3), for any $m\ge 2$, the bad directions for $\mathbf{f}^m$ at $\xi'$ is $\vec{v}_{\xi'}(c)$. Therefore, $\mathbf{z}_0$ is contained in a good direction of $\mathbf{f}^n$ at $\xi'$. Then the conclusion follows. \[Rivera-boundary-period\] Under the assumption of Lemma \[repelling-existence\], we have $1<q\le n$. Proposition \[repelling\] implies $q>1$. To show $q\le n$, we apply Lemmas \[limit-infinity\] and \[cycle-period\]. If $\Gamma=\{\hat z\}$, by Proposition \[repelling\] (1), $\langle\mathbf{z}\rangle\subset U$, and hence $q\mid n$. If $\Gamma\cap\Lambda\not=\emptyset$, then by Lemma \[cycle-period\], it follows that $q<n$ since $q<q'\leq n$ by Proposition \[repelling\] (4). If $\Gamma\cap\Lambda=\emptyset$ and $\Gamma\not=\{\hat z\}$, it holds by Lemma \[limit-infinity\] (2). Since we will focus on the case that $\langle z(t)\rangle$ is an attracting cycle in next following subsections, we sharpen the conclusion of case (2) in Lemma \[limit-infinity\] as follows. We must be careful when examining derivatives of degenerating families. Written in terms of ratios of homogeneous polynomials, if $f_t \to HG \in \mathbb{P}^{2d+1}$ where $G$ is in lowest terms, then $f_t' \to H^2G'$ in $\mathbb{P}^{4d+1}$; however, $G'$ need not be in lowest terms. The holes of $H^2G'$ are the holes of $HG$ together with a subset of the zeros of the denominator of $G$, i.e. the poles of $G$. \[nonrepelling-cycle\] Fix the notations as before. Assume $\langle z(t)\rangle$ is an attracting cycle. If $\Gamma\cap \Lambda=\emptyset$ and $\Gamma\not=\{\hat z\}$. Then $\Gamma\setminus\{\hat z\}$ is a nonrepelling cycle of $G$. By Lemma \[limit-infinity\] (2), $\Gamma\setminus\{\hat z\}$ is a cycle of $G$. We show this cycle is nonrepelling. If $\infty\in\Gamma\setminus\{\hat z\}$, then $\Gamma\setminus\{\hat z\}$ is superattracting since $\infty$ is a critical point of $G$. Now we deal with the case that $\infty\not\in\Gamma\setminus\{\hat z\}$. Since $f_t^q$ converges to $G$ locally uniformly on $\mathbb{P}^1$ off finitely many points, there exists a homogeneous polynomial $H$ such that $f_t^q\to HG$ in $\mathbb{P}^{2d^q+1}$. It is easy to check that $(f_t^q)'\to H^2G'$ in $\mathbb{P}^{4d^{q}+1}$. Thus the holes of $H^2G'$ are contained in $\Lambda\cup\{\hat z\}$ and the poles of $G$. For any point $x\in\Gamma\setminus\{\hat z\}$, let $x(t)\in\langle z(t)\rangle$ such that $x(t)$ converges to $x$ as $t\to 0$. Since $\Gamma\cap\Lambda=\emptyset$ and $\infty\not\in\Gamma\setminus\{\hat z\}$, then $x$ is not a hole of $H^2G'$. Now regard $(f_t^q)'$ and $G'$ as rational maps in lowest terms. Then $(f_t^q)'(x(t))\to G'(x)$ as $t\to 0$, see [@DeMarco07 Lemma 2.6]. Now renumber the points in $\langle z(t)\rangle$ such that $z_0(t)\to z_0\in\Gamma\setminus\{\hat z\}$. Then by Lemma \[limit-infinity\] and the chain rule, $$(f_t^{n})'(z_0(t))=(f_t^q)'(z_{n/q-1}(t))\cdots(f_t^q)'(z_{q}(t))(f_t^q)'(z_{0}(t)).$$ It follows that $(f_t^{n})'(z_0(t))\to (G^{n/q})'(z_0)$. Since $|(f_t^{n})'(z_0(t))|<1$, we have that $|(G^{n/q})'(z_0)|\le 1$. Note $G^{n/q}(z_0)=z_0$ since $f_t^n(z_0(t))=z_0(t)$. Thus the cycle $\Gamma\setminus\{\hat z\}$ is nonrepelling. Limits of a pair of attracting cycles ------------------------------------- We continue the setup and notation of the previous two subsections. However, from now on, we assume in addition that the family $f_t$ has a degenerate sequence $f_{t_k}$ possessing two distinct *attracting* cycles of periods at least $2$. Lemma \[limit-infinity\] gives three possibilities for the limit of each these cycles. With two indistinguishable cycles, we get a priori six possibilities to analyze in total. The following result constrains the limiting map $G(z)$ in certain cases, and serves to rule out some possibilities. \[infinity-infinity\] Let $\{f_t\}$ and $\{f_{t_k}\}$ be as above. Assume that $\langle z(t_k)\rangle$ and $\langle w(t_k)\rangle$ are two distinct attracting cycles of periods at least $2$. Suppose $\langle z(t_k)\rangle\to\{\hat z\}$ as $k \to \infty$. Then $\langle w(t_k)\rangle\to\{\hat z\}$ also. Moreover, both $\langle z(t_k)\rangle$ and $\langle w(t_k)\rangle$ have period $q$, and the point $\hat{z}$ is a parabolic fixed point $G(z)$ with multiplicity $3$. For the unique fixed Rivera domain $U$ of $\mathbf{f}$, Theorem \[fp-number\] implies that $\mathbf{f}$ has $2$ fixed points in $U\cap\mathbb{P}^1_{\mathbb{L}}$. Let $n$ be the period of the cycle $\langle z(t_k)\rangle$. Now consider the number $N(\mathbf{f}^{n},U)$ of the fixed points of $\mathbf{f}^{n}$ in $U\cap\mathbb{P}^1_{\mathbb{L}}$. By Theorem \[fp-number\], we have on the one hand $$N(\mathbf{f}^{n},U)=2+q(m-2),$$ where $m$ is the multiplicity of $\hat z\in\mathbb{C}\setminus\{0\}$ as a fixed point of $G(z)$. Since $\langle z(t_k)\rangle\to\{\hat z\}$, it follows that $\langle\mathbf{z}\rangle\subset U$. For otherwise, Proposition \[repelling\](2) implies there are points distinct from $\hat z$ in the limit of $\langle z(t_k)\rangle$. Note each point in $\langle\mathbf{z}\rangle$ is a fixed point of $\mathbf{f^n}$. So we have on the other hand that $N(\mathbf{f}^{n},U)\ge 2+n$. Hence $m\ge 3$ and so the parabolic fixed point $\hat z$ attracts in its basin at least two critical points of $G$. Since $G$ has exactly two critical points, it follows that $m=3$, equivalently $n=q$. We now consider the three possibilities for the limit of the cycle $\langle w(t_k) \rangle$ given by Lemma \[limit-infinity\]. The previous paragraph shows that each critical point of $G$ has infinite forward orbit and converges to the multiplicity $3$ parabolic fixed-point $\hat{z}$. If the limit contains a cycle, it must be non-repelling, by the proof of Lemma \[nonrepelling-cycle\]. By [@Fatou; @Mane93], to a non-repelling cycle is associated yet another critical point of $G$, and so this is impossible; to see this, consider each of the attracting, parabolic, Siegel, and Cremer cases. The limit cannot contain a preperiodic critical point of $G$ either. Hence this limit must collapse to $\hat{z}$. We conclude that both cycles collapse and hence collide to $\hat{z}$. Fatou-Shishikura Inequality --------------------------- To finally prove our main result, we will apply the Fatou-Shishikura inequality (FSI) and its version refined by A. Epstein (refined FSI) to obtain impossible on constraints the number of nonrepelling cycles of the limiting map $G$. [@Shishikura87 Corollary 1]\[FSI\] Let $f \in \mathrm{Rat}_d$, $d \geq 2$. Then $f$ has at most $2d-2$ nonrepelling cycles. [@Epstein99 Theorem 1]\[refined-FSI\] Let $f \in \mathrm{Rat}_d$, $d \geq 2$. For a cycle $\langle z\rangle\subset\mathbb{P}^1$, define $$\gamma_{\langle z\rangle}:= \begin{cases} 0&\text{if}\ \langle z\rangle\ \text{is\ repelling\ or\ superattracting},\\ 1&\text{if}\ \langle z\rangle\ \text{is\ attracting\ or\ irrationally indifferent},\\ \nu &\text{if}\ \langle z\rangle\ \text{is\ parabolic-repelling},\\ \nu+1& \text{if}\ \langle z\rangle\ \text{is\ parabolic-attracting\ or\ parabolic-indifferent}, \end{cases}$$ where $\nu$ is the corresponding degeneracy if $\langle z\rangle$ is parabolic. Set $$\gamma(f)=\sum_{\langle z\rangle\subset\mathbb{P}^1}\gamma_{\langle z\rangle},$$ and let $\delta(f)$ be the number of infinite tails of critical orbits. Then $$\gamma(f)\le\delta(f).$$ Hyperbolic components of strict type D are bounded -------------------------------------------------- Recall that $\mathcal{M}_d$ is the moduli space of bicritical rational maps of degree $d\ge 2$. Theorem \[main\] is a consequence of the following result. \[bounded-family\] Let $\mathcal{H}\subset\mathcal{M}_d$ be a hyperbolic component possessing two distinct attracting cycles of periods at least $2$. Let $\{f_t\}_{t \in \mathbb{D}}\subset\overline{\mathcal{F}}$ be a holomorphic family such that $[f_{t_k}]\in\mathcal{H}$ for some sequence $t_k\to 0$. Then $[f_t]$ is bounded in $\mathcal{M}_d$. Suppose that $[f_t]\to\infty$ in $\mathcal{M}_d$ as $t\to\infty$. Then the induced map $\mathbf{f}$ is not simple. Let $\langle z(t_k)\rangle$ and $\langle w(t_k)\rangle$ be the two distinct attracting cycles of periods at least $2$ of $f_{t_k}$. Assume that $\langle z(t)\rangle\to\Gamma^1$ and $\langle w(t)\rangle\to\Gamma^2$ as $t \to 0$. Then by Lemma \[limit-infinity\], $\hat z\in\Gamma^1\cap\Gamma^2$. Recall that $U$ denotes the unique fixed Rivera domain of $\mathbf{f}$, and $q\ge 2$ the period of a point in $\partial U$. Lemma \[limit-infinity\] gives three possibilities for $\Gamma^1$, which we treat in turn. **Case 1: $\Gamma^1=\{\hat z\}$.** Then by Lemma \[infinity-infinity\], we have $\Gamma^2=\{\hat z\}$. Moreover, both $\langle z(t)\rangle$ and $\langle w(t)\rangle$ have period $q$ and the cycles $\langle\mathbf{z}\rangle$ and $\langle\mathbf{w}\rangle$ are in the Rivera domain $U$. It follows that $$N(\mathbf{f}^q,U)\ge 2+2q.$$ On the other hand, in this case again by Lemma \[infinity-infinity\] the point $\hat z\in\mathbb{C}\setminus\{0\}$ is a parabolic fixed point $G(z)$ with multiplicity $3$. Hence, by Theorem \[fp-number\], $$N(\mathbf{f}^q,U)=2+q.$$ It is impossible since $q\ge 2$. **Case 2: $\Gamma^1\cap\Lambda\not=\emptyset$.** By Proposition \[repelling\](4) and Lemma \[limit-infinity\], the map $\mathbf{f}$ has a type II repelling cycle $\mathcal{O}'$ of period $q'>q$ and there exists $c'\in\{0,\infty\}$ such that $c'\in\Gamma^1$ and is preperiodic under $G$. Since $G$ also has a parabolic fixed point at $\hat{z}\in\mathbb{C}\setminus\{0\}$, Lemmas \[limit-infinity\], \[infinity-infinity\] and the refined FSI imply that $\Gamma^2\cap\Lambda\not=\emptyset$ and $c'\in\Gamma^2$ too. By Proposition \[repelling\](4), at the point $\xi'\in\mathcal{O'}\cap[0,\infty]$, the map $T_{\xi'}\mathbf{f}^{q'}$ is conjugate to a degree $d$ unicritical polynomial. Now choose a suitable holomorphic family $M_t$ of degree $1$ rational map such that for the induced map $\mathbf{M}\in\mathrm{Aut}(\mathbb{P}^1_{\mathbb{L}})$, $\mathbf{M}(\xi')={\mathbf{g}}$ and the map $T_{{\mathbf{g}}}(\mathbf{M}\circ\mathbf{f}^{q'}\circ\mathbf{M}^{-1})$ is a unicritical polynomial. Recall that there exist exactly two repelling cycles in $J(\mathbf{f})$. Since $\xi'\in\mathcal{O'}$, by Lemma \[cycle-period\], renumbering the points in $\langle\mathbf{z}\rangle$ and $\langle\mathbf{w}\rangle$, we obtain $\mathbf{z}_0\in\langle \mathbf{z}\rangle$ and $\mathbf{w}_0\in\langle \mathbf{w}\rangle$ such that the boundaries of the corresponding periodic Berkovich Fatou components $\Omega(\mathbf{z}_0)$ and $\Omega(\mathbf{w}_0)$ are the point $\xi'$. Note $\{\mathbf{M}(\mathbf{z}_{kq'})\}_{k\ge 0}$ and $\{\mathbf{M}(\mathbf{w}_{kq'})\}_{k\ge 0}$ are two nonrepelling cycles of $\mathbf{M}\circ\mathbf{f}^{q'}\circ\mathbf{M}^{-1}$ in $\mathbb{L}$ and hence in $\mathbf{P}^1$. Lemma \[cycle-period\] implies that $\{M_t(z_{kq'}(t))\}$ and $\{M_t(w_{kq'}(t))\}$ converge to two (not necessarily distinct) cycles in $\mathbb{C}$ of the corresponding polynomial. Moreover, noting the holes of the limit of $M_t\circ f_t^{q'}\circ M_t^{-1}$ are at $\infty$ and applying the proof of Lemma \[nonrepelling-cycle\], we have that these cycles are nonrepelling. The FSI implies that a unicritical polynomial has at most one nonrepelling cycle in $\mathbb{C}$. Hence $\{M_t(z_{kq'}(t))\}$ and $\{M_t(w_{kq'}(t))\}$ converge to the same cycle. We next appeal to two elementary lemmas from complex analysis regarding limits of cycles under a locally uniformly convergent sequence of maps. By [@Nie-P Lemma 2.5] and [@Epstein00 Lemma 1], we know this cycle is parabolic-attracting or parabolic-indifferent. The refined FSI implies that a unicritical polynomial cannot have such a cycle. **Case 3: $\Gamma^1\cap\Lambda=\emptyset$ and $\Gamma^1\not=\{\hat z\}$.** By Lemma \[nonrepelling-cycle\], $\Gamma^1\setminus\{\hat z\}$ is a nonrepelling cycle. By symmetry, it only remains to consider the case that $\Gamma^2\setminus\{\hat z\}$ is also a nonrepelling cycle. Note $G(z)$ has a parabolic fixed point. The FSI implies these two nonrepelling cycles collide, and hence they are same, which is again a parabolic-attracting or parabolic-indifferent cycle of $G$. Again, it contradicts to the refined FSI.

ArXiv

--- abstract: | The “noncommutative geometry” of complex algebraic curves is studied. As first step, we clarify a morphism between elliptic curves, or complex tori, and $C^*$-algebras $T_{\theta}=\{u,v~|~vu=e^{2\pi i\theta}uv\}$, or noncommutative tori. The main result says that under the morphism isomorphic elliptic curves map to the Morita equivalent noncommutative tori. Our approach is based on the rigidity of the length spectra of Riemann surfaces. [*Key words and phrases: elliptic curve, noncommutative torus*]{} author: - | Igor  Nikolaev\ Department of Mathematics\ 2500 University Drive N.W.\ Calgary T2N 1N4 Canada\ [[email protected]]{} title: On complex and noncommutative tori --- Introduction {#introduction .unnumbered} ============ Noncommutative geometry is a branch of algebraic geometry studying “varieties” over noncommutative rings. The noncommutative rings are usually taken to be rings of operators acting on a Hilbert space [@MvN]. The rudiments of noncommutative geometry can be traced back to F. Klein [@Kle1], [@Kle2] or even earlier. The fundamental modern treatise [@C] gives an account of status and perspective of the subject. The noncommutative torus $T_{\theta}$ is a $C^*$-algebra generated by linear operators $u$ and $v$ on the Hilbert space $L^2(S^1)$ subjected to the commutation relation $vu=e^{2\pi i\theta}uv, ~\theta\in {\Bbb R}-{\Bbb Q}$ [@Rie]. The classification of noncommutative tori was given in [@EfS], [@PiV], [@Rie]. Recall that two such tori $T_{\theta}, T_{\theta'}$ are Morita equivalent if and only if $\theta,\theta'$ lie in the same orbit of the action of group $GL(2,{\Bbb Z})$ on irrational numbers by linear fractional transformations. It is remarkable that “moduli problem” for $T_{\theta}$ looks as such for the complex tori $E_{\tau}={\Bbb C}/({\Bbb Z}+\tau {\Bbb Z})$, where $\tau$ is complex modulus. Namely, complex tori $E_{\tau},E_{\tau'}$ are isomorphic if and only if $\tau,\tau'$ lie in the same orbit of the action of $SL(2,{\Bbb Z})$ on complex numbers by linear fractional transformations. It was observed by some authors, e.g. [@Man], [@Soi], that it might be [*not*]{} just a coincidence. This note is an attempt to show that it is indeed so: there exists a general morphism between Riemann surfaces and $C^*$-algebras. Let us give rough idea of our approach. Given Riemann surface $S$, there is a function $S\to {\Bbb R}^{\infty}_+$ which maps the (discrete) set of closed geodesics of $S$ to a discrete subset of real line by assigning each closed geodesic its riemannian length. If $T_g(S)$ is the space of all Riemann surfaces of genus $g\ge0$, then function $$\label{eq1} {\goth W}: T_g(S)\longrightarrow {\Bbb R}^{\infty}_+,$$ is finitely-to-one and “generically” one-to-one [@Wol]. In the case $g=1$, function ${\goth W}$ is always one-to-one. It is known also that restriction ${\goth W}_{syst}: T_g(S)\to {\Bbb R}_+$ of ${\goth W}$ to the shortest closed geodesic of $S$ (called [*systole*]{}) is a $C^0$ Morse function on $T_g(S)$ [@SS], §5. Below we focus on the case $g=1$, i.e. $T_1\simeq E_{\tau}$. Recall that $T_{\theta}$ has unique state $s_0$ (which is actually a tracial state) [@Rie]. Any positive functional on $T_{\theta}$ has form $\omega s_0$, where $\omega>0$ is a real number. Let $\Theta=\{T_{\theta}~|~\theta\in {\Bbb R}-{\Bbb Q}\}$ and $\Omega=\{\omega\in {\Bbb R}~|~\omega>0\}$. We define a map $$\label{eq2} {\goth V}:\Theta\times\Omega\longrightarrow {\Bbb R}^{\infty}_+$$ by formula $(T_{\theta},\omega) \mapsto \{f_n(\omega)\ln tr (A_n)\}_{n=0}^{\infty}$, where $$\begin{aligned} A_0 &=& \left(\small\matrix{a_0 & 1\cr 1 & 0}\right),\nonumber\\ A_1 &=& \left(\small\matrix{a_0 & 1\cr 1 & 0}\right) \left(\small\matrix{a_1 & 1\cr 1 & 0}\right),\nonumber\\ &\vdots& \\ A_n &=& \left(\small\matrix{a_0 & 1\cr 1 & 0}\right) \left(\small\matrix{a_1 & 1\cr 1 & 0}\right)\dots \left(\small\matrix{a_n & 1\cr 1 & 0}\right),\nonumber\end{aligned}$$ are integer matrices whose entries $a_i>0$ are partial denominators of continued fraction expansion of $\theta$ and $f_n$ are monotone $C^0$ functions of $\omega$. Assuming that functions ${\goth W}, {\goth V}$ have common range, one gets a mapping ${\goth W}{\goth V}^{-1}: E_{\tau}\to (T_{\theta},\omega)$. Morphisms between $E_{\tau}$ and $T_{\theta}$ have been studied in [@Man], [@PoS], [@Ram], [@SoV]. The works [@Man], [@Ram] and [@SoV] treat noncommutative tori as “quantum compactification” of the space of elliptic curves. This approach deals with an algebraic side of the subject. In particular, Manin [@Man] suggested to use “pseudolattices” (i.e. $K_0$-group of $T_{\theta}$) to solve the multiplication problem for real number fields. This problem is part of the Hilbert 12th problem. In [@PoS] a functor from derived category of holomorphic vector bundles over $T_{\theta}$ to the Fukaya category of such bundles over $E_{\tau}$ was constructed. In this note we prove the following results. \[thm1\] Let $E_{\tau}$ be complex torus of modulus $\tau, Im~\tau>0$ and $(T_{\theta},\omega)$ pair consisting of noncommutative torus with an irrational Rieffel’s parameter $\theta$ and a positive functional $T_{\theta}\to {\Bbb C}$ of norm $\omega$. Then there exists a one-to-one mapping $E_{\tau}\to (T_{\theta},\omega)$. The action of modular group $SL(2,{\Bbb Z})$ on the complex half plane $\{\tau\in {\Bbb C}~|~Im~\tau>0\}$ is equivariant with: \(i) the action of group $GL(2,{\Bbb Z})$ on irrationals $\{\theta\in {\Bbb R}-{\Bbb Q}~|~\theta>0\}$ by linear fractional transformations; \(ii) a discrete action on positive reals $\{\omega\in {\Bbb R}~|~\omega>0\}$. In particular, isomorphic complex tori map to the Morita equivalent noncommutative tori, and vice versa. \[dfn1\] The irrational number $\theta$ of mapping $E_{\tau}\to (T_{\theta},\omega)$ we call a projective curvature of the elliptic curve $E_{\tau}$. \[thm2\] Projective curvature of an elliptic curve with complex multiplication is a quadratic irrationality. [Acknowledgements.]{} It is my pleasure to thank G. A. Elliott, Yu. I. Manin and M. Rieffel for their interest in the subject of present note. I am greatful to the referee for critical remarks and helpful suggestions. Proofs ====== The proof of both theorems is based on the rigidity of length spectrum of complex torus, cf Wolpert [@Wol]. A preliminary information on complex and noncommutative tori can be found in Section 2. Proof of Theorem 1 ------------------ Let us review main steps of the proof. By the rigidity lemma (Lemma \[lm1\]) the length spectrum $Sp~ E$ defines conformal structure of $E$. In fact, this correspondence as a bijection. Under isomorphisms of $E$ the length spectrum can acquire a real multiple or get a “cut of finite tail” (Lemma \[lm2\]). We attach to ${\Bbb C}/L$ a continued fraction of its projective curvature $\theta$ as specified in Introduction. Then isomorphic tori ${\Bbb C}/L$ will have continued fractions which differ only in a finite number of terms. In other words, one can attach a Morita equivalence class of noncommutative tori to every isomorphism class of complex tori. \[lm1\] [**(Rigidity of length spectrum)**]{} Let $Sp~E$ be length spectrum of a complex torus $E={\Bbb C}/L$. Then there exists a unique complex torus with the spectrum. This correspondence is a bijection. [*Proof.*]{} See McKean [@McK]. $\square$ Let $Sp~X=\{l_1,l_2,\dots\}$ be length spectrum of a Riemann surface $X$. Let $a>0$ be a real number. By $aSp~X$ we understand the length spectrum $\{al_1,al_2,\dots\}$. Similarly, for any $m\in {\Bbb N}$ we denote by $Sp_m~X$ the length spectrum $\{l_m,l_{m+1},\dots\}$, i.e. the one obtained by deleting the first $(m-1)$-geodesics in $Sp~X$. \[lm2\] Let $E\sim E'$ be isomorphic complex tori. Then either: \(i) $Sp~E'=|\alpha|Sp~E$ for an $\alpha\in {\Bbb C}^{\times}$, or \(ii) $Sp~E'=Sp_m~E$ for a $m\in {\Bbb N}$. [*Proof.*]{} (i) The complex tori $E={\Bbb C}/L, E'={\Bbb C}/M$ are isomorphic if and only if $M=\alpha L$ for a complex number $\alpha\in {\Bbb C}^{\times}$. It is not hard to see that closed geodesic of $E$ are bijective with the points of the lattice $L=\omega_1{\Bbb Z}+\omega_2{\Bbb Z}$ in the following way. Take a segment of straight line through points $0$ and $\omega$ of lattice $L$ which contains no other points of $L$. This segment represents a homotopy class of curves through $0$ and a closed geodesic of $E$. Evidently, this geodesic will be the shortest in its homotopy class with the length $|\omega|$ equal to absolute value of complex number $\omega$. Thus, $|\omega|$ belongs to the length spectrum of $E$. Let now $Sp~E=\{|\omega_1|,|\omega_2|,\dots\}$ with $\omega_i\in L$. Since $M=\alpha L$, one gets $Sp~E'=\{|\alpha||\omega_1|,|\alpha||\omega_2|,\dots\}$ and $Sp~E'=|\alpha|Sp~E$. Item (i) follows. \(ii) Note that according to (i) the length spectrum $Sp~X=\{l_0,l_1,l_2,\dots\}$ can be written as $Sp~X=\{1,l_1,l_2,\dots\}$ after multiplication on $1/l_0$, where $l_0$ is the length of shortest geodesic. Note also that shortest geodesic of complex torus has homotopy type $(1,0)$ or $(0,1)$ (standard generators for $\pi_1E$). Let $a,b,c,d$ be integers such that $ad-bc=\pm 1$ and let $$\begin{aligned} \label{eq5} \omega_1' &=& a\omega_1+b\omega_2,\nonumber\\ \omega_2' &=& c\omega_1+d\omega_2,\end{aligned}$$ be an automorphism of the lattice $L=\omega_1{\Bbb Z}+\omega_2{\Bbb Z}$. This automorphisms maps standard generators $(1,0)$ and $(0,1)$ of $L$ to the vectors $\omega_1=(a,b),\omega_2=(c,d)$. Let their lengths be $l_m,l_{m+1}$, respectively. As we showed earlier, $l_m,l_{m+1}\in Sp~E$ and it is not hard to see that there are no geodesics of the intermediate length. (This gives a justification for the notation chosen.) Note that $\omega_1,\omega_2$ are standard generators for the complex torus $E'\sim E$ and therefore one of them is the shortest closed geodesics of $E'$. One can normalize it to the length $1$. On the other hand, there are only finite number of closed geodesics of length smaller than $l_n$ (McKean [@McK]). Thus $Sp~E\cap Sp~E'=\{l_m,l_{m+1},\dots\}$ for a finite number $m$ and since (\[eq5\]) is automorphism of the lattice $L$. In other words, $Sp~E'=Sp_m~E$. Item (ii) follows. $\square$ To finish the proof of item (i) of Theorem \[thm1\], one needs to combine Lemmas \[lm1\],\[lm2\] with the fact that two noncommutative tori $T_{\theta},T_{\theta'}$ are Morita equivalent if and only if their continued fractions differ only in a finite number of terms (Section 2.1). To prove item (ii) of Theorem \[thm1\], let to the contrary the action of $SL(2,{\Bbb Z})$ be non-discrete, i.e. having limit points in $\Omega$. Let $p=\lim_{n\to\infty} (T_{\theta_n},\omega_n)$, where $\theta_n$ lie in the same orbit of $GL(2,{\Bbb Z})$. Let $E_p$ be corresponding complex torus such that $E_p\not\cong E_{p_n}$ are non-isomorphic. By continuity of the systole function ${\goth W}_{syst}$ (see Introduction) $Sp~E_p=\lim Sp ~E_{p_n}$. Then by the rigidity of length spectra, $E_p\cong E_{p_n}$ are isomorphic. The contradiction proves item (ii) of theorem. $\square$ Proof of Theorem 2 ------------------ Let us outline the idea of the proof. If $E$ admits complex multiplication, then its complex modulus $\tau$ lies in an imaginary quadratic field $K$. In fact, up to an isogeny, the ring of endomorphisms $End~E={\cal O}_K$, where ${\cal O}_K$ is the ring of integers of field $K$. It can be shown that $L$ is an ideal in ${\cal O}_K$ (Section 2.3). The length spectrum $Sp~E$ of elliptic curve with complex multiplication is a “geometric progression” with the growth rate $|\alpha|$, where $\alpha\in~End~E$ (Lemma \[lm5\]). One can use Klein’s lemma (Lemma \[lm5b\]) to characterize length spectra in terms of continued fractions. In particular, length spectrum with asymptotically geometric growth correspond to periodic continued fractions. Thus, projective curvature converges to quadratic irrationality. Length spectrum $Sp~E$ of an elliptic curve $E={\Bbb C}/L$ is called $\alpha$-multiplicative, if there exists a complex number $\alpha\in {\Bbb C}^{\times}$ with $|\alpha|>1$ such that $$Sp~E=\{l_1,\dots,l_N,|\alpha|l_1,\dots,|\alpha|l_N,\dots,|\alpha|^nl_1,\dots, |\alpha|^nl_N,\dots\},$$ for a $N\in{\Bbb N}$. \[lm5\] Let $E$ be an elliptic curve with complex multiplication. Then its length spectrum $Sp~E$ is $\alpha$-multiplicative for an $\alpha\in {\Bbb C}^{\times}$. [*Proof.*]{} Let $E={\Bbb C}/L$ be complex torus which admits non-trivial endomorphisms $z\mapsto\alpha z,\alpha\in K={\Bbb Q}(\sqrt{-d})$. It is known that $End~E$ is an order in the field $K$. In fact, up to an isogeny of $E$, $End~E\simeq {\cal O}_K$, where ${\cal O}_K$ is the ring of integers of imaginary quadratic field $K$ (Section 2.3). Lattice $L$ in this case corresponds to an ideal in ${\cal O}_K$. Let $l_1$ be minimal length of closed geodesic of $E$. For an endomorphism $\alpha: E\to E, \alpha\in {\Bbb C}^{\times}$, consider the set of geodesics whose lengths are less than $|\alpha|l_1$. By the properties of $Sp~E$ mentioned in Section 1.1, such a set will be finite. Let us denote the lengths of geodesics in this set by $l_1,\dots,l_N$. Since every geodesic in $Sp~E$ is a complex number $\omega_i$ lying in the ring $L\subseteq {\cal O}_K$, one can consider the set of geodesics $\alpha\omega_1,\dots,\alpha\omega_N$. The length of these geodesics will be $|\alpha|l_1,\dots,|\alpha|l_N$, respectively. It is not hard to see that by the choice of number $N$, the first $2N$ elements of $Sp~E$ are presented by the following growing sequence of geodesics: $l_1,\dots,l_N, |\alpha|l_1,\dots, |\alpha|l_N$. We proceed by iterations of $\alpha$, until all closed geodesics of $E$ are exhausted. The conclusion of Lemma \[lm5\] follows. $\square$ We shall need the following statement regarding geometry of the regular continued fractions [@Kle1],[@Kle2]. It is valid for any regular fraction, not necessarily periodic. \[lm5b\] [**(F. Klein)**]{} Let $$\label{eq9} \omega=\mu_1+{1\over\displaystyle \mu_2+ {\strut 1\over\displaystyle \mu_3\displaystyle +\dots}}$$ be a regular continued fraction. Let us denote the convergents of $\omega$ by: $$\label{eq10} {p_{-1}\over q_{-1}}={0\over 1}, \quad {p_{0}\over q_{0}}={1\over 0}, \quad {p_{1}\over q_{1}}={\mu_1\over 1},\dots, {p_{\nu}\over q_{\nu}}={\mu_{\nu}p_{\nu-1}+p_{\nu-2}\over \mu_{\nu}q_{\nu-1}+q_{\nu-2}}.$$ For any lattice $L$ in ${\Bbb C}$, consider a segment $I$ with ends in the points $(p_{\nu-2},q_{\nu-2})$ and $(p_{\nu},q_{\nu})$. Then the segment $J$ which joins $0$ with the point $p_{\nu-1},q_{\nu-1}$ is parallel to $I$ and $$\label{eq10b} |I|=\mu_{\nu}|J|,$$ where $|\bullet|$ denotes the length of the segment. [*Proof.*]{} We refer the reader to [@Kle2]. $\square$ \[cor1\] Let $\omega_{\nu}=(p_{\nu},q_{\nu})$ be lattice points mentioned in Lemma \[lm5b\]. Then the length of vector $\omega_{\nu}$ can be evaluated with the help of the following asymptotic formula: $$\label{eq10c} |\omega_{\nu}|\approx |\omega_{\nu-2}|+\mu_{\nu}|\omega_{\nu-1}|.$$ [*Proof.*]{} Indeed, using notation of Lemma \[lm5b\], one can write $|(p_{\nu},q_{\nu})|\approx $$ |(p_{\nu-2},q_{\nu-2})| +|I|$. But according to equation (\[eq10b\]), $|I|=\mu_{\nu}|(p_{\nu-1},q_{\nu-1})|$. Corollary \[cor1\] follows. $\square$ Note that according to the recurrent formula (\[eq10c\]) the length spectrum $\{|\omega_{\nu}|\}$ coming from continued fraction (\[eq9\]) is completely determined by the first two values: $|\omega_1|$ and $|\omega_2|$. Using (\[eq10c\]), one can easily deduce the following asymptotic formula for $|\omega_{\nu}|$ as function of $|\omega_1|,|\omega_2|$: $$\label{eq10d} |\omega_{\nu}|\approx |\omega_2|\prod_{k=3}^{\nu}\mu_k+ |\omega_1|\prod_{k=4}^{\nu}\mu_k + O(\nu).$$ Fix $N$ a positive integer. It follows from equation (\[eq10d\]) that: $$\begin{aligned} \label{eq10e} \lim_{\nu\to\infty}{|\omega_{\nu+N}|\over |\omega_{\nu}|} &=& \mu_{\nu+1}\dots\mu_{\nu+N} \lim_{\nu\to\infty}\left({\mu_{\nu}\dots\mu_3|\omega_2|+ \mu_{\nu}\dots\mu_4|\omega_1|+O(\nu)\over \mu_{\nu}\dots\mu_3|\omega_2|+ \mu_{\nu}\dots\mu_4|\omega_1|+O(\nu)}\right)=\nonumber\\ &=& \mu_{\nu+1}\dots\mu_{\nu+N}. \end{aligned}$$ Let $E$ be an elliptic curve with complex multiplication. Then by Lemma \[lm5\] its length spectrum $Sp~E$ is $\alpha$-multiplicative. In other words, $$\label{eq10f} {l_{\nu+N}\over l_{\nu}}=|\alpha|=~Const,$$ for an $N\in {\Bbb N}$ and any $\nu~mod~N$. Note that $|\alpha|$ is a rational integer. Thus, by formula (\[eq10e\]) we have $\mu_{\nu+1}\dots\mu_{\nu+N}=~Const$, for any $\nu~mod~N$. The last requirement can be satisfied if and only if continued fraction (\[eq9\]) is $N$-periodic. Theorem \[thm2\] is proven. $\square$ Background information ====================== In present section we briefly review noncommutative and complex tori. The excellent source of information on noncommutative torus are papers [@EfS], [@Rie] and monograph of [@RLL]. The literature on complex torus is fairly vast. We recommend for the reference Ch. VI of [@S]. Noncommutative torus -------------------- By the $C^*$-algebra one understands a noncommutative Banach algebra with an involution [@RLL]. Namely, a $C^*$-algebra $A$ is an algebra over $\Bbb C$ with a norm $a\mapsto ||a||$ and an involution $a\mapsto a^*, a\in A$, such that $A$ is complete with respect to the norm, and such that $||ab||\le ||a||~||b||$ and $||a^*a||=||a||^2$ for every $a,b\in A$. If $A$ is commutative, then the Gelfand theorem says that $A$ is isometrically $*$-isomorphic to the $C^*$-algebra $C_0(X)$ of continuous complex-valued functions on a locally compact Hausdorff space $X$. For otherwise, $A$ represents a “noncommutative” topological space $X$.  Given a $C^*$-algebra, $A$, consider new $C^*$-algebra $M_n(A)$, i.e. the matrix algebra over $A$. There exists a remarkable semi-group, $A^+$, connected to the set of projections in algebra $M_{\infty}=\cup_{n=1}^{\infty} M_n(A)$. Namely, projections $p,q\in M_{\infty}(A)$ are Murray-von Neumann equivalent $p\sim q$ if they can be presented as $p=v^*v$ and $q=vv^*$ for an element $v\in M_{\infty}(A)$. The equivalence class of projections is denoted by $[p]$. The semi-group $A^+$ is defined to be the set of all equivalence classes of projections in $M_{\infty}(A)$ with the binary operation $[p]+[q]=[p\oplus q]$. The Grothendieck completion of $A^+$ to an abelian group is called a [*$K_0$-group of $A$*]{}. The functor $A\to K_0(A)$ maps the unital $C^*$-algebras into the category of abelian groups so that the semi-group $A^+\subset A$ corresponds to a “positive cone” $K_0^+\subset K_0(A)$ and the unit element $1\in A$ corresponds to the “order unit” $[1]\in K_0(A)$. The ordered abelian group $(K_0,K_0^+,[1])$ with the order unit is called a [*dimension (Elliott) group*]{} of $A$. The dimension (Elliott) group is complete invariant of the $AF$ $C^*$-algebras.  Fix $\theta$ irrational and consider a linear flow $\dot x=\theta,\dot y=1$ on the torus. Let $S^1$ be a closed transversal to our flow. The [*noncommutative torus*]{} $T_{\theta}$ is a norm-closed $C^*$-algebra generated by the unitary operators in the Hilbert space $L^2(S^1)$: Uf(t)=z(t)f(t),Vf(t)=f(t-), which are multiplication by a unimodular function $z(t)$ and rotation operators. It could be easily verified that $UV=e^{2\pi i\alpha}VU$. As an “abstract” algebra, $T_{\theta}$ is a crossed product $C^*$-algebra $C(S^1)\rtimes_{\phi}{\Bbb Z}$ of (commutative) $C^*$-algebra of complex-valued continuous functions on $S^1$ by the action of powers of $\phi$, where $\phi$ is a rotation of $S^1$ through the angle $2\pi\alpha$. $T_{\theta}$ is not $AF$, but can be embedded into an $AF$-algebra whose dimension group is $P_{\theta}$ (to be specified below); the latter is known to be intimately connected with the arithmetic of the irrational numbers $\theta$’s. The following beautiful result is due to the efforts of many mathematicians [^1] (Effros, Elliott, Pimsner, Rieffel, Shen, Voiculescu, etc): [**(Classification of noncommutative tori)**]{} Let $T_{\theta}$ be a noncommutative torus. Suppose that the $\theta$ has a continued fraction expansion =a\_0+ 1a\_1+ 1a\_2+… def= \[a\_0,a\_1,a\_2,…\]. Let $\varphi_n$ be a composition of isometries of the lattice ${{\Bbb Z}}^2\subset{{\Bbb R}}^2:\quad\varphi_n=\left( \small\matrix{ a_0 & 1\cr 1 & 0}\right)$$\dots\left(\small\matrix{ a_n & 1\cr 1 & 0}\right)$. Then $T_{\theta}$ can be embedded into an $AF$-algebra whose dimension group is a direct limit of the ordered abelian groups: $P_{\theta}=\lim_{n\to\infty}({{\Bbb Z}}^2,\varphi_n)$. Moreover, if $\theta=[a_0,a_1,\dots]$ and $\theta'=[b_0,b_1,\dots]$ are two irrational numbers, then $P_{\theta}$ and $P_{\theta'}$ are isomorphic (i.e. noncommutative tori $T_{\theta}$ and $T_{\theta'}$ are Morita equivalent) if and only if $a_{m+k}=b_m$ for an integer number $k\in{{\Bbb Z}}$. In other words, the irrational numbers $\theta$ and $\theta'$ are modular equivalent: $\theta'={a\theta+b\over c\theta+d},\quad ad-bc=\pm 1$, where $a,b,c,d\in{{\Bbb Z}}$ are integer numbers. [*Proof.*]{} An algebraic proof of this fact can be found in [@EfS]. $\square$ Complex torus ------------- Let $L$ denote a lattice in the complex plane ${\Bbb C}$. Attached to $L$, there are following classic Weierstrass function $\wp(z;L)$ and Eisenstein series $G_k(L)$: $$\begin{aligned} \wp(z;L) &=& {1\over z^2}+\sum_{\omega\in L^{\times}} \left\{{1\over (z+\omega)^2}-{1\over\omega^2}\right\},\\ G_k(L) &=& \sum_{\omega\in L^{\times}}^{k\ge2}{1\over\omega^{2k}}.\end{aligned}$$ $\wp(z;L)$ is analytic and $G_k(L)$ is convergent for any lattice $L$ [@S]. There exists a duality between lattices $L$ and cubic curves $E$ given by the following theorem. \[thm4\] Let $L$ be a lattice in $\Bbb C$. Then the map $z\mapsto (\wp(z;L),{1\over 2}\wp'(z;L))$ is an analytic isomorphism from complex torus ${\Bbb C}/L$ to elliptic cubic $E=E({\Bbb C})$: $$E({\Bbb C})=\{(x,y)\in {\Bbb C}^2~|~y^2=x^3-15G_4(L)x-35G_6(L)\}.$$ Conversely, to any cubic in the Weierstrass normal form $y^2=x^3+ax+b$ there corresponds a unique lattice $L$ such that $a=-15G_4(L)$ and $b=-35G_6(L)$. [*Proof.*]{} We refer the reader to [@S] for a detailed proof of this fact. $\square$  Let $L$ be a lattice in $\Bbb C$. The Riemann surface ${\Bbb C}/L$ is called a [*complex torus*]{}. Let $f:{\Bbb C}/L\to {\Bbb C}/M$ be holomorphic and invertible map (isomorphism) between two complex tori. Since $f$ is covered by a linear map $z\to\alpha z$ on $\Bbb C$, one can easily conclude that $\alpha L=M$ for an $\alpha\in {\Bbb C}^{\times}$. On the other hand, lattice $L$ can always be written as $L=\omega_1{\Bbb Z}+\omega_2{\Bbb Z}$, where $\omega_1,\omega_2\in {\Bbb C}^{\times}$ and $\omega_2\ne k\omega_1$ for a $k\in {\Bbb R}$. The complex number $\tau={\omega_2\over\omega_1}$ is called a [*complex modulus*]{} of lattice $L$. \[lm6\] Two complex tori are isomorphic if and only if their complex moduli $\tau$ and $\tau'$ satisfy the equation: $$\label{eq17} \tau'={a\tau+b\over c\tau+d} \qquad ad-bc=\pm 1,\qquad a,b,c,d\in {\Bbb Z}.$$ [*Proof.*]{} The proof of this fact can be found in [@S]. $\square$ Elliptic curves with complex multiplication ------------------------------------------- Let $E={\Bbb C}/L$ be an elliptic curve. Consider the set $End~E$ of analytic self-mappings of $E$. Each $f\in End~E$ is covered on the complex plane by map $z\mapsto\alpha z$ for an $\alpha\in {\Bbb C}$. It is not hard to see that $End~E$ has the structure of a ring under the pointwise addition and multiplication of functions. The set $End~E$ is called an [*endomorphism ring*]{} of elliptic curve $E$. By the remarks above, $End~E$ can be thought of as a subring of complex numbers: $$\label{eq20} End~E=\{\alpha\in {\Bbb C}~|~\alpha L\subset L\}.$$ There exists a fairly complete algebraic description of such rings. Roughly speaking, they are either “rational integers” $\Bbb Z$ or integers ${\cal O}_K$ of an algebraic number field $K$. The following lemma is true. \[lm8\] Let $\alpha\in End~E$ be a complex number. Then either: \(i) $\alpha$ is a rational integer, or \(ii) $\alpha$ is an algebraic integer in an imaginary quadratic number field $K={\Bbb Q}(\sqrt{-d})$. [*Proof.*]{} See [@S]. $\square$  If $End~E $ is different from $\Bbb Z$, $E$ is said to be [*elliptic curve with complex multiplication*]{}. If $E$ admits complex multiplication, then its ring $End~E$ is an order in an imaginary quadratic field $K$. In fact, $E$ admits an isogeny (analytic homomorphism) to a curve $E'$ such that $End~E'\simeq {\cal O}_K$, where ${\cal O}_K$ is the ring of integers of field $K$ [@S]. Thus, by property $\alpha L\subseteq L$, lattice $L$ is an ideal in ${\cal O}_K$. Denote by $h_K$ the class number of field $K$. It is well known, that there exist $h_K$ non-isomorphic ideals in ${\cal O}_K$. Therefore, elliptic curves $E_1={\Bbb C}/L_1,\dots, E_{h_K}={\Bbb C}/L_{h_K}$ are pairwise non-isomorphic, but their endomorphism ring is the same [@S]. [100]{} A. Connes, Noncommutative Geometry, Academic Press, 1994. E. G. Effros and C. L. Shen, Approximately finite $C^*$-algebras and continued fractions, Indiana Univ. Math. J. 29 (1980), 191-204. F. Klein, Über eine geometrische Auffassung der gewöhnlichen Kettenbruchentwicklung, Nachrichten der Kgl. Gesellschaft der Wissenschaften zu Göttingen, Math.-phys. Klasse (1895), Heft 3. F. Klein, Ausgewählte Kapitel der Zahlentheorie, Math. Annalen 48 (1896/97). Yu. I. Manin, Real multiplication and noncommutative geometry, arXiv:math.AG/0202109 v1, Preprint, 46 p. H. P. McKean, Selberg’s trace formula as applied to a compact Riemann surface, Comm. Pure Appl. Math. 25, (1972), 225-246. F. J. Murray and J. von Neumann, On rings of operators, Annals of Math. 37 (1936), 116-229. M. Pimsner and D. Voiculescu, Imbedding the irrational rotation $C^*$-algebra into an $AF$-algebra, J. Operator Theory 4 (1980), 201-210. A. Polishchuk and A. Schwarz, Categories of holomorphic vector bundles on noncommutative two-tori, Commun. Math. Phys. 236 (2003), 135-159. N. Ramachandran, [*Private communication.*]{} M. A. Rieffel, $C^*$-algebras associated with irrational rotations, Pacific J. of Math. 93 (1981), 415-429. M. Rørdam, F. Larsen and N. Laustsen, An introduction to $K$-theory for $C^*$-algebras, LMS Student Texts, 49, Cambridge University Press, 2000. P. Schmutz Schaller, Geometry of Riemann surfaces based on closed geodesics, Bull. Amer. Math. Soc. 35 (1998), 193-214. J. H. Silverman, The Arithmetic of Elliptic Curves, Springer, 1986. Y. Soibelman, Quantum tori, mirror symmetry and deformation theory, Lett. Math. Phys. 56 (2001), 99-125. Y. Soibelman and V. Vologodsky, Non-commutative compactifications and elliptic curves, Preprint, math.AG/0205117. S. Wolpert, The length spectra as moduli for compact Riemann surfaces, Annals of Math. 109 (1979), 323-351. [^1]: The author apologizes for possible erroneous credits regarding history of the problem. Classification of noncommutative tori seems to be an old problem; early results in this direction can be found in the works of Klein [@Kle1],[@Kle2].

ArXiv

--- abstract: 'We apply heavy-quark effective theory to separate long- and short-distance effects of heavy quarks in lattice gauge theory. In this approach, the inverse heavy-quark mass and the lattice spacing are treated as short distances, and their effects are lumped into short-distance coefficients. We show how to use this formalism to match lattice gauge theory to continuum QCD, order by order in the heavy-quark expansion. In this paper, we focus on heavy-light currents. In particular, we obtain one-loop results for the matching factors of lattice currents, needed for heavy-quark phenomenology, such as the calculation of heavy-light decay constants, and heavy-to-light transition form factors. Results for the Brodsky-Lepage-Mackenzie scale $q^*$ are also given.' address: | $^1$Department of Physics, Hiroshima University, Higashi-Hiroshima 739-8526, Japan\ $^2$High Energy Accelerator Research Organization (KEK), Tsukuba 305-0801, Japan\ $^3$Center for Computational Physics, University of Tsukuba, Tsukuba 305-8577, Japan\ $^4$Theoretical Physics Department, Fermi National Accelerator Laboratory, Batavia, Illinois 60510\ $^5$Yukawa Institute for Theoretical Physics, Kyoto University, Sakyo-ku, Kyoto 606-8502, Japan author: - 'Junpei Harada,$^1$ Shoji Hashimoto,$^2$ Ken-Ichi Ishikawa,$^{2,3}$ Andreas S. Kronfeld,$^{3,4}$ Tetsuya Onogi,$^{1,5}$[@TO] and Norikazu Yamada$^2$' date: 30 December 2004 title: | Application of heavy-quark effective theory to lattice QCD:\ II. Radiative corrections to heavy-light currents --- \#1[[R]{}\[\#1\]]{} \#1 Introduction {#sec:intro} ============ A key ingredient in flavor physics is the calculation of hadronic matrix elements of the electroweak Hamiltonian. For example, one would like to calculate, from first principles, quantities such as leptonic decay constants, semi-leptonic form factors, and the amplitudes for neutral-meson mixing. Numerical calculations with lattice QCD offer a way to obtain these quantities, eventually with well-controlled estimates of the numerical uncertainties [@Kronfeld:1993jf]. The properties of $B$ and $D$ mesons are especially interesting, but the relatively large $b$ and $c$ quark masses make it difficult, with today’s computers, to carry out lattice calculations in the limit $m_Q a\to 0$ for which lattice QCD was first developed. (Here $m_Q$ is the $b$ or $c$ quark mass, and $a$ is the lattice spacing.) One can, however, use the simplifying features of the heavy-quark limit of QCD to make lattice calculations tractable. As $m_Q$ is increased far above the typical scale of the wave function, $\Lambda_{\text{QCD}}$, the hadrons’ wave functions depend less and less on $m_Q$. As $m_Q\to\infty$ the wave functions become flavor and spin symmetric [@Isgur:1989vq]. For quarkonia similar simplifications occur, including spin symmetry [@Caswell:1986ui]. In this paper we construct vector and axial vector currents with one quark heavy and the other light. These currents are needed to obtain the decay constants of heavy-light mesons, and the form factors for decays of the form $H\to Ll\nu_l$, where $H$ is a charmed or $b$-flavored hadron (*e.g.*, $B$, $D$; $\Lambda_b$, $\Lambda_c$), decaying to a light hadron $L$ (*e.g.*, $\pi$, $K$, $\rho$; $p$, etc.) and a lepton $l$ and its neutrino $\nu_l$. In particular, we provide a way to treat radiative and power corrections consistently. This paper is a sequel to Ref. [@Kronfeld:2000ck], which focussed on power corrections. Here we discuss the case of heavy-light bilinears in detail, and we compute explicitly the matching factors for the currents introduced in the “Fermilab” formalism [@El-Khadra:1997mp]. Heavy-heavy bilinears are considered in a companion paper [@Kronfeld:1999tk]. To interpret lattice calculations when $m_Qa\ll1$, it is convenient to describe cutoff effects with the Symanzik local effective Lagrangian (LE${\cal L}$) and expand the LE${\cal L}$’s short-distance coefficients in powers of $m_Qa$ [@Symanzik:1979ph; @Symanzik:1983dc; @Luscher:1996sc; @Jansen:1996ck]. When $m_Qa\not\ll1$, however, one should realize that it is not lattice gauge theory that breaks down but rather the Symanzik description, especially its expansion in $m_Qa$. If $m_Qa$ is large because $m_Q\gg\Lambda_{\text{QCD}}$, then the simplifying features of the heavy-quark limit provide an alternative. Instead of matching lattice gauge theory directly to continuum QCD, one can match to the heavy-quark effective theory (HQET) or, for quarkonia, to non-relativistic QCD (NRQCD). In this approach, the inverse heavy-quark mass and the lattice spacing are both treated as short distances, and a simple picture arises, in which heavy-quark discretization effects are lumped into short-distance coefficients. Heavy-quark cutoff effects are systematically reducible, by adjusting the heavy-quark expansion for lattice gauge theory to agree term-by-term with continuum QCD. Such application of HQET to lattice QCD was started in Ref. [@Kronfeld:2000ck], building on Ref. [@El-Khadra:1997mp]. In this paper we extend the formalism to heavy-light currents. We use the heavy-quark expansion, as generated by HQET, to derive matching conditions, which are valid for all $m_Qa$ and to all orders in the gauge coupling. Our derivation is explicit for dimension-four currents, which is the next-to-leading dimension, but generalization to higher-dimension operators should be clear. We also present explicit results for the one-loop radiative corrections to the normalization of the current. These calculations show that the temporal and spatial components of the current do not have the same radiative corrections. This feature has been found already [@Kuramashi:1998tt; @Ishikawa:1997xh], and the HQET formalism shows why it arises. In deriving these results we have found a compact way of arranging the Dirac algebra, which may be useful for calculations with actions, such as highly improved actions, that are not considered here. As expected, the coefficients have a strong mass dependence. Most of this dependence can be handled non-perturbatively [@Hashimoto:2000yp; @Simone:2000nv; @El-Khadra:2001rv]. For equal mass, it is simple to normalize the temporal vector current, for all masses. One can then form ratios of renormalization factors, from which the dominant mass dependence drops out. Results for these combinations are also presented, in Sec. \[sec:loop\]. Our one-loop results extend those of Ref. [@Ishikawa:1997xh], which considered heavy-light currents with the Sheikholeslami-Wohlert (SW) action [@Sheikholeslami:1985ij] for Wilson fermions [@Wilson:1975hf] and also with non-relativistic QCD (NRQCD). Results for the Wilson action [@Wilson:1975hf] have been obtained first by Kuramashi [@Kuramashi:1998tt]. In Refs. [@Kuramashi:1998tt; @Ishikawa:1997xh] a term in the currents, the so-called rotation term [@Kronfeld:1995nu; @El-Khadra:1997mp], which is needed for tree-level improvement at order $1/m_Q$, was omitted. Here we include the rotation, obtaining the algebraic expression of the Feynman diagrams for the full Fermilab action. We present numerical results for the Wilson action (without rotation) and the SW action (with and without rotation). These results are appropriate for recent calculations of decay constants [@Aoki:1998ji; @El-Khadra:1998hq; @Bernard:1998xi; @AliKhan:2000eg; @Bernard:2000nv], which used the radiative corrections calculated in Refs. [@Kuramashi:1998tt; @Ishikawa:1997xh]. Our new results have been used in a recent calculation of the form factors for the decays $B\to\pi l\nu_l$ and $D\to\pi l\nu_l$ [@El-Khadra:2001rv]. We also have obtained results for the Fermilab action on anisotropic lattices [@Harada:2001ei]. Our formalism should be useful for computing matching factors (beyond one-loop) also in lattice NRQCD [@Lepage:1987gg]. Applied to the static limit [@Eichten:1987xu], it generalizes the formalism of Eichten and Hill [@Eichten:1990zv]. At one-loop order, similar methods have been developed to calculate the heavy-light matching coefficients for lattice NRQCD [@Davies:1993ec; @Morningstar:1998ep]. As in the Symanzik program [@Symanzik:1979ph; @Symanzik:1983dc; @Luscher:1996sc], the advantage of introducing a continuum effective field theory is that the formalism provides a clear definition of the matching coefficients at every order in perturbation theory (in the gauge coupling). Indeed, it may also provide a foundation for a non-perturbative improvement program. This paper is organized as follows. Section \[sec:match\] discusses three ways to separate long and short distance physics with (continuum) effective field theories. The first is Symanzik’s description of lattice spacing effects; we also discuss its breakdown when $m_Qa\not\ll1$. The second is the HQET description of heavy quarks, applied to continuum QCD. The third is the HQET description of heavy quarks on the lattice, which applies when $m_Q\gg\Lambda_{\text{QCD}}$, for all $m_Qa$. In particular, we obtain a definition of the matching factors for the vector and axial-vector heavy-light currents. Section \[sec:match\] also shows how the HQET matching procedure is related to the Symanzik procedure in the regime where both apply. Then, the Fermilab action is reviewed in Sec. \[sec:lattice\], and in Sec. \[sec:loop\] we present one-loop results for the matching factors. Some concluding remarks are made in Sec. \[sec:conclusions\]. Three appendices contain details of the one-loop calculation, including an outline of a method to obtain compact expressions, and explicit results for the one-loop Feynman integrands for the renormalization factors with the full Fermilab action. Instead of printing tables of the numerical results in Sec. \[sec:loop\], we are making a suite of programs freely available[@p:epaps]. This suite includes programs for the heavy-heavy currents treated in our companion paper [@Kronfeld:1999tk]. Matching to Continuum Field Theories {#sec:match} ==================================== In this section we discuss how to interpret the physical content of lattice field theories by matching to continuum field theories. First, the standard Symanzik formalism for describing cutoff effects is reviewed, and we recall how this description breaks down for heavy quarks. After reviewing the HQET description of (continuum) QCD, we adapt HQET to describe lattice gauge theory. Comparison of the two then yields a matching procedure that is valid whenever $m_Q\gg\Lambda_{\text{QCD}}$, and for all $m_Qa$. In the limit $m_Qa\ll1$ both the HQET and the Symanzik descriptions should apply, so we are able to derive relations between the some of the matching coefficients. Symanzik Formalism {#subsec:sym} ------------------ The customary way to define matching factors for lattice gauge theory is to apply Symanzik’s formalism. Then the short-distance lattice artifacts are described by a local effective Lagrangian (LE${\cal L}$) and local effective operators. For the Lagrangian of any lattice field theory one can write [@Symanzik:1979ph; @Symanzik:1983dc] $${\cal L}_{\text{lat}} \doteq {\cal L}_{\text{Sym}}, \label{eq:lat=Sym}$$ where the symbol $\doteq$ can be read “has the same on-shell matrix elements as”. The left-hand side is a lattice field theory, and the right-hand side is a continuum field theory, whose ultraviolet behavior is regulated and renormalized completely separately from the lattice of the left-hand side. The LE${\cal L}$ is the Lagrangian of the corresponding continuum field theory, plus extra terms to describe discretization effects. For lattice QCD $${\cal L}_{\text{Sym}} = {\cal L}_{\text{QCD}} + {\cal L}_I, \label{eq:LEL}$$ where ${\cal L}_{\text{QCD}}$ is the renormalized, continuum QCD Lagrangian. We focus on the quarks, so for our purposes $${\cal L}_{\text{QCD}} = - \bar{q}\left( {\kern+0.1em /\kern-0.65em D} + m_q\right)q. \label{eq:QCD}$$ Lattice artifacts are described by higher-dimension operators, $${\cal L}_I = aK_{\sigma\cdot F} \bar{q} i\sigma_{\mu\nu}F^{\mu\nu} q + \cdots, \label{eq:artifacts}$$ where $a$ is the lattice spacing and $K_{\sigma\cdot F}$ is a short-distance coefficient that depends on details of the lattice action [@Luscher:1996sc]. The lattice artifacts in ${\cal L}_I$ can be treated as a perturbation. In this way a series can be developed, with matrix elements in the (continuum) eigenstates of ${\cal L}_{\text{QCD}}$. Equation (\[eq:LEL\]) omits dimension-five operators of the form $\bar{q}R({\kern+0.1em /\kern-0.65em D} + m_q)q$ or $\bar{q}(-\loarrow{\kern+0.1em /\kern-0.65em D} + m_q)Rq$, for arbitrary $R$, which make no contribution to on-shell matrix elements, owing to the equations of motion implied by Eq. (\[eq:QCD\]). The vector and axial vector currents can be described in a similar way. Consider, for example, the flavor-changing transition $s\to u$. Then one may write [@Luscher:1996sc] $$\begin{aligned} V^\mu_{\text{lat}} & \doteq & Z_V^{-1} {\cal V}^\mu - aK_V \partial_\nu \bar{u} \sigma^{\mu\nu}s + \cdots , \label{eq:LEV} \\ A^\mu_{\text{lat}} & \doteq & Z_A^{-1} {\cal A}^\mu + aK_A \partial^\mu \bar{u}i \gamma_5 s + \cdots , \label{eq:LEA}\end{aligned}$$ where $$\begin{aligned} {\cal V}^\mu & \equiv & \bar{u}i\gamma^\mu s , \label{eq:V} \\ {\cal A}^\mu & \equiv & \bar{u}i\gamma^\mu\gamma_5 s , \label{eq:A}\end{aligned}$$ are the vector and axial vector currents in QCD. Further dimension-four operators are omitted, because they are linear combinations of those listed and others that vanish by the equations of motion. Like the terms of dimension five and higher in ${\cal L}_I$, the dimension-four currents can be treated as perturbations. Matrix elements of $Z_VV^\mu_{\text{lat}}$ and $Z_AA^\mu_{\text{lat}}$ then give those of continuum QCD, at least in the limit $a\to 0$. The short-distance coefficients—$K_{\sigma\cdot F}$, $K_J$, and $Z_J$ ($J=V$, $A$)—are, in general, functions of the gauge coupling and the quark masses (in lattice units), and they depend on the renormalization scheme of the LE$\cal L$. For $m_qa\ll1$ ($q=u$, $d$, $s$), it is consistent and satisfactory to replace $K_{\sigma\cdot F}$ and $K_J$ with their values at $m_qa=0$, and to replace the $Z_J$ with the first two terms of the Taylor expansion around $m_qa=0$. For example, with Wilson fermions [@Wilson:1975hf; @Sheikholeslami:1985ij] and conventional bilinears for the lattice currents, one finds $K_V^{[0]}=K_A^{[0]}=0$, and $$\begin{aligned} K_{\sigma\cdot F}^{[0]} & = & \case{1}{4}(1-c_{\text{SW}}) + O(ma), \label{eq:Ctree} \\ Z_V^{[0]} = Z_A^{[0]} & = & 1 + \case{1}{2}(m_u+m_s)a + O(m^2a^2), \label{eq:Ztree}\end{aligned}$$ where the superscript “$[0]$” denotes the tree level, and $c_{\text{SW}}$ is the clover coupling of the SW action [@Sheikholeslami:1985ij] (cf. Sec. \[sec:lattice\]). Moreover, in the hands of the [*Alpha*]{} Collaboration [@Luscher:1996sc; @Jansen:1996ck], Eqs. (\[eq:lat=Sym\])–(\[eq:LEA\]) are the foundation of a non-perturbative procedure for adjusting $K_{\sigma\cdot F}$, $K_V$, and $K_A$ to be of order $aM_p$, where $M_p$ is a (light) hadronic mass scale, and also for computing $Z_V$ and $Z_A$ non-perturbatively (through order $M_pa$). Then all lattice artifacts in the mass spectrum, decay constants, and form factors are of order $a^2$. For a heavy flavor $Q$, however, it is not practical to keep $m_Qa$ small enough so that this program straightforwardly applies. Recent work that uses the fully $O(a)$-improved action and currents has chosen the heavy-quark mass $m_Qa$ to be as large as 0.7 or so. Thus, $(m_Qa)^2$ is not small,[^1] and one should check whether contributions of order $(m_Qa)^2$ are under control. Indeed, if one keeps the full mass dependence in the coefficients, one finds that the simple description of Eqs. (\[eq:LEL\])–(\[eq:LEA\]) breaks down. The relation between energy and momentum becomes [@El-Khadra:1997mp; @Mertens:1998wx] $$E^2(\bbox{p}) = m_1^2 + \frac{m_1}{m_2}\bbox{p}^2 + O(p^4a^2), \label{eq:Emp}$$ where, for the Wilson and SW actions, $$\begin{aligned} m^{[0]}_1a & = & \ln(1+m_0a), \label{eq:m1} \\ \frac{1}{m^{[0]}_2a} & = & \frac{2}{m_0a(2+m_0a)} + \frac{1}{1+m_0a}, \label{eq:m2Wilson}\end{aligned}$$ and $m_0$ is the bare lattice mass. Generalizations valid at every order in perturbation theory also have been derived [@Mertens:1998wx]. In a similar vein, the spatial and temporal components of the currents no longer take the same matching coefficients, as shown by explicit one-loop calculations [@Kuramashi:1998tt; @Ishikawa:1997xh]. The energy-momentum relation in Eq. (\[eq:Emp\]) is obtained for $\bbox{p}a\ll1$ but $m_Qa\not\ll1$. It can be described by modifying the standard LE${\cal L}$ to $${\cal L}_{\text{lat}} \doteq - \bar{q}\left(\gamma_4 D_4 + \sqrt{\frac{m_1}{m_2}} \bbox{\gamma}\cdot\bbox{D} + m_1\right)q + {\cal L}'_I, \label{eq:LELmod}$$ that is, temporal and spatial directions must be treated asymmetrically in the dimension-four Lagrangian, and also in the higher-dimension terms ${\cal L}'_I$. From the tree-level formulas, Eqs. (\[eq:m1\]) and (\[eq:m2Wilson\]), $$\frac{m_1}{m_2} = 1 - \case{2}{3}m_1^2a^2 + \case{1}{2}m_1^3a^3 + \cdots, \label{eq:m1m2}$$ so one sees that the deviation from the standard description is of order $(ma)^2$. (At the one-loop order [@Mertens:1998wx], and at every order in $g^2$, Eq. (\[eq:m1m2\]) still has no term linear in $ma$.) One can arrive at Eqs. (\[eq:LELmod\]) and (\[eq:m1m2\]) also by starting with Eq. (\[eq:LEL\]), including higher-dimension terms, and eliminating $\gamma_4D_4^3$ and $D_4^4$, etc., by applying the equations of motion. In any case, deviations of $m_1/m_2$—and similar ratios—from 1 are present in lattice calculations. With the Wilson or SW actions $1-m_1/m_2$, for example, is 10 percent or greater for $m_1a>0.6$. Although this numerical estimate is made at the tree level, it is implausible that radiative corrections or bound-state effects could wash the error away. In summary, the description of Eqs. (\[eq:LEL\])–(\[eq:LEA\]) is no longer accurate when $m_Qa\not\ll1$. There are several possible remedies. One is to do numerical calculations with $a$ so small that, even for the $b$ quark, $m_ba\ll1$. Despite the exponential growth in computer power, this remedy will not be available for many years. Another remedy is to add a parameter to the lattice action, which can be tuned to set $m_1=m_2$ [@El-Khadra:1997mp]. An example of this is an action with two hopping parameters. Then, the continuum description can again take the form in Eq. (\[eq:LEL\]), starting with the continuum ${\cal L}_{\text{QCD}}$, although it is still useful to describe the higher-dimension terms asymmetrically. A third remedy is to realize that it is the *description*, rather than the underlying lattice gauge theory, that has broken down. Since lattice gauge theory with Wilson fermions has a well-behaved heavy-quark limit [@El-Khadra:1997mp], it is possible to use heavy-quark effective theory (HQET) or NRQCD to describe short-distance effects, including the lattice artifacts of the heavy quark [@Kronfeld:2000ck]. This last remedy is explained in detail in Sec. \[subsec:hqetlat\], where we show also how all three strategies are connected. HQET description of QCD {#subsec:hqetcont} ----------------------- The breakdown of the standard Symanzik description of cutoff effects for Wilson fermions arises because the kinematics of heavy hadron decays single out a vector, namely, the heavy hadron velocity. But, since the heavy-quark mass is also much larger than the spatial momenta of the problem, the dynamics simplify. In continuum QCD, this has led to the development of the effective field theories HQET [@Eichten:1987xu; @Eichten:1990zv; @Grinstein:1990mj; @Georgi:1990um; @Eichten:1990vp] and NRQCD [@Caswell:1986ui; @Lepage:1987gg]. These two effective theories are useful for generating an expansion in $\bbox{p}/m_Q$. They share a common effective Lagrangian, but the power in $\bbox{p}/m_Q$ assigned to any given operator is not necessarily the same. In HQET the power can be deduced immediately from the dimension, whereas in NRQCD it is deduced by counting powers of the relative velocity of the $\bar{Q}Q$ system. The discussion in this paper will follow the counting of HQET, but the logic could be repeated with the counting of NRQCD. Our aim is to show, for the case of heavy-light currents, how to use HQET to extend the standard Symanzik program into the region where $m_Qa$ is no longer small. This program was started in Ref. [@Kronfeld:2000ck], building on Ref. [@El-Khadra:1997mp]. The formalism holds for all $m_Qa$, but, like the usual HQET, it requires $$m_Q \gg \bbox{p},\, \Lambda_{\text{QCD}}. \label{eq:when}$$ First, in this subsection, we recall the HQET description of continuum QCD, paralleling the discussion in Sec. \[subsec:sym\]. Then, in Sec. \[subsec:hqetlat\], we explain what changes are needed to describe the cutoff effects of lattice NRQCD and of lattice gauge theory with Wilson fermions. The HQET conventions are the same as those given Sec. III of Ref. [@Kronfeld:2000ck]. The velocity needed to construct HQET is $v$. The fourth Euclidean component $v_4=iv^0$, so in the rest frame $v=(i,\bbox{0})$. The metric is taken to be $\diag(\pm 1,1,1,1)$, with the upper (lower) sign for Euclidean (Minkowski) spacetime. In either case, $v^2=-1$. The heavy quark field is called $h_v$, and it satisfies the constraint $\case{1}{2}(1-i\kern+0.1em /\kern-0.55em v)h_v = h_v$, or $$\kern+0.1em /\kern-0.55em v h_v = ih_v, \quad \bar{h}_v\kern+0.1em /\kern-0.55em v = i\bar{h}_v. \label{eq:Pvh}$$ Physically Eq. (\[eq:Pvh\]) means that $h_v$ describes only quarks, but not anti-quarks. The tensor $\eta^\mu_\nu=\delta^\mu_\nu+v^\mu v_\nu$ projects onto components orthogonal to $v$. For a vector $p$, the component orthogonal to $v$ is $p_\perp^\mu=\eta^\mu_\nu p^\nu=p^\mu+v^\mu{(v\cdot p)}$; in the rest frame, these are the spatial components. HQET describes the dynamics of heavy-light bound states with an effective Lagrangian built from $h_v$. So, for these states, one can say $${\cal L}_{\text{QCD}} \doteq {\cal L}_{\text{HQET}}, \label{eq:QCD=HQET}$$ where $${\cal L}_{\text{HQET}} = {\cal L}^{(0)} + {\cal L}^{(1)} + {\cal L}^{(2)} + \cdots. \label{eq:L}$$ For HQET ${\cal L}^{(s)}$ contains terms of dimension $4+s$. Note that the ultraviolet regulator and renormalization scheme of the two sides of Eq. (\[eq:QCD=HQET\]) need not be the same, although dimensional regularization and the $\overline{\rm MS}$ scheme are usually used for both. For this paper it is enough to consider the first two terms, ${\cal L}^{(0)}$ and ${\cal L}^{(1)}$. The leading, dimension-four term is $${\cal L}^{(0)} = \bar{h}_v(iv\cdot D - m)h_v. \label{eq:L0}$$ The choice of $v$ is somewhat arbitrary. If $v$ is close to the heavy quark’s velocity,[^2] then ${\cal L}^{(0)}$ is a good starting point for the heavy-quark expansion, which treats the higher-dimension operators as small. The most practical choice is the containing hadron’s velocity. The mass term in ${\cal L}^{(0)}$ is often omitted. By heavy-quark symmetry, it has an effect neither on bound-state wave functions nor, consequently, on matrix elements. It does affect the mass spectrum, but only additively. Including the mass obscures the heavy-quark flavor symmetry, but only slightly [@Kronfeld:2000ck]. When the mass term is included, higher-dimension operators are constructed with ${\cal D}^\mu=D^\mu-imv^\mu$ [@Falk:1992fm]. To describe on-shell matrix elements one may omit operators that vanish by the equation of motion, $-iv\cdot{\cal D}h_v=0$, derived from Eq. (\[eq:L0\]). In practice, therefore, higher-dimension operators are constructed from ${\cal D}^\mu_\perp=D^\mu_\perp$ and $[{\cal D}^\mu,{\cal D}^\nu]=[D^\mu,D^\nu]=F^{\mu\nu}$. The dimension-five interactions are $${\cal L}^{(1)} = {\cal C}_2 {\cal O}_2 + {\cal C}_{\cal B} {\cal O}_{\cal B}, \label{eq:L1}$$ where ${\cal C}_2$ and ${\cal C}_{\cal B}$ are short-distance coefficients, and $$\begin{aligned} {\cal O}_2 & = & \bar{h}_vD_\perp^2 h_v, \label{eq:O2} \\ {\cal O}_{\cal B} & = & \bar{h}_v s_{\alpha\beta}B^{\alpha\beta}h_v, \label{eq:OB}\end{aligned}$$ with $s_{\alpha\beta}=-i\sigma_{\alpha\beta}/2$ and $B^{\alpha\beta}=\eta^\alpha_\mu\eta^\beta_\nu F^{\mu\nu}$. In Eq. (\[eq:L0\]) one should think of the quark mass $m$ as a short-distance coefficient. By reparametrization invariance [@Luke:1992cs], the same mass appears in the denominator of the kinetic energy ${\cal C}_2{\cal O}_2$, namely, $${\cal C}_2 = \frac{1}{2m}.$$ If operator insertions in HQET are renormalized with a minimal subtraction in dimensional regularization, then $m$ is the (perturbative) pole mass. With other ultraviolet regulators, the operator and the mass $m$ could become $\mu$-dependent. Even in mass-independent schemes, the chromomagnetic operator ${\cal O}_{\cal B}$ depends on the renormalization point $\mu$ of the HQET, and that dependence is canceled by $${\cal C}_{\cal B}(\mu) = \frac{z_{\cal B}(\mu)}{2m},$$ with $2m$ appearing so that $z_{\cal B}$ is unity at the tree level. The description of electroweak flavor-changing operators proceeds along the same lines. The flavor-changing vector current for a $b\to q$ transition, defined to be ${\cal V}^\mu=\bar{q}i\gamma^\mu b$ as in Eq. (\[eq:V\]), is described in HQET by $${\cal V}^\mu \doteq C_{V_\parallel} v^\mu \bar{q}h_v + C_{V_\perp} \bar{q}i\gamma^\mu_\perp h_v - \sum_{i=1}^6 B_{Vi} {\cal Q}^\mu_{Vi} + \cdots, \label{eq:VcontHQET}$$ where $h_v$ is the HQET field, which satisfies Eq. (\[eq:L\]) and whose dynamics are given by ${\cal L}_{\text{HQET}}$. The dimension-four operators are $$\begin{aligned} {\cal Q}^\mu_{V1} & = & - v^\mu \bar{q} {\kern+0.1em /\kern-0.65em D}_\perp h_v, \label{eq:Q1} \\ {\cal Q}^\mu_{V2} & = & \bar{q}i\gamma^\mu_\perp {\kern+0.1em /\kern-0.65em D}_\perp h_v, \label{eq:Q2} \\ {\cal Q}^\mu_{V3} & = & \bar{q}iD^\mu_\perp h_v, \label{eq:Q3} \\ {\cal Q}^\mu_{V4} & = & + v^\mu \bar{q} {\kern+0.1em /\kern-0.65em \loarrow{D}}_\perp h_v, \label{eq:Q4} \\ {\cal Q}^\mu_{V5} & = & \bar{q} {\kern+0.1em /\kern-0.65em \loarrow{D}}_\perp i\gamma^\mu_\perp h_v, \label{eq:Q5} \\ {\cal Q}^\mu_{V6} & = & \bar{q}i\loarrow{D}^\mu_\perp h_v. \label{eq:Q6}\end{aligned}$$ Further dimension-four operators are again omitted, because they are linear combinations of those listed and others that vanish by the equations of motion. For example, $\bar{q}(iv\cdot\loarrow{D}) v^\mu h_v = \bar{q}({\kern+0.1em /\kern-0.65em \loarrow{D}}_\perp - {\kern+0.1em /\kern-0.65em \loarrow{D}})v^\mu h_v = {\cal Q}^\mu_{V4}-m_qv^\mu\bar{q}h_v$, where the Dirac equation is used for the last step. The axial vector current ${\cal A}^\mu=\bar{q}i\gamma^\mu\gamma_5b$ has a completely analogous description, $${\cal A}^\mu \doteq C_{A_\perp} \bar{q}i\gamma^\mu_\perp \gamma_5h_v - C_{A_\parallel} v^\mu \bar{q}\gamma_5h_v - \sum_{i=1}^6 B_{Ai} {\cal Q}_{Ai}^\mu + \cdots, \label{eq:AcontHQET}$$ where each operator ${\cal Q}_{Ai}^{\mu}$ is obtained from ${\cal Q}_{Vi}^{\mu}$ by replacing $\bar{q}$ with $-\bar{q}\gamma_5$. The short-distance coefficients of HQET depend on the heavy-quark mass $m$, as well as $\mu/m$ and $m_q/m$, where $\mu$ is the the renormalization scale and $m_q$ is the light quark mass. They are not explicitly needed in this paper, but it may be instructive to give the coefficients of the dimension-three terms through one-loop order, with $m_q=0$ ($J=V$, $A$) [@Eichten:1990zv]: $$\begin{aligned} C_{J_\parallel} & = & 1 + \frac{g^2C_F}{16\pi^2}\left( \gamma_h\ln(m^2/\mu^2) - 2 \right), \label{eq:Cparallel} \\ C_{J_\perp} & = & 1 + \frac{g^2C_F}{16\pi^2}\left( \gamma_h\ln(m^2/\mu^2) - 4 \right), \label{eq:Cperp}\end{aligned}$$ where the anomalous dimension $\gamma_h=3/2$. The $\mu$-independent part of $C_{A_\parallel}$ and $C_{A_\perp}$ given here assumes that the axial current is renormalized in a chirally symmetric way [@Trueman:1995ca]. The coefficients of the dimension-four currents are $$\begin{aligned} B_{Ji}^{[0]} & = & \frac{1}{2m},\quad i\le 2, \label{eq:B1B2} \\ B_{Ji}^{[0]} & = & 0, \quad i\ge 3, \label{eq:B3-6}\end{aligned}$$ at the tree level, but all $B_{Ji}$ become non-trivial when radiative corrections are included. HQET description of lattice gauge theory {#subsec:hqetlat} ---------------------------------------- HQET provides a systematic way to separate the short distance $1/m$ from the scale $\Lambda_{\text{QCD}}$ in heavy-light matrix elements, as long as the condition (\[eq:when\]) holds. The formalism can also be applied to lattice gauge theory, again as long as condition (\[eq:when\]) holds (and $\bbox{p}a\ll1$). When lattice NRQCD is used for heavy-light systems, this is because ${\cal L}_{\text{HQET}}$ is just the Symanzik LE${\cal L}$ for lattice NRQCD. When Wilson fermions are used for heavy quarks, one may also apply HQET, because they have the same particle content and heavy-quark symmetries [@Kronfeld:2000ck]. In both cases bilinears of lattice fermions fields are introduced to approximate the continuum QCD currents. One field corresponds to the light quark, and the other to the heavy quark. An explicit construction, through order $1/m$, is in Ref. [@Morningstar:1998ep] for lattice NRQCD, and a similar construction for Wilson fermions is in Sec. \[sec:lattice\]. Lattice artifacts stemming from the light quark can be described as in Sec. \[subsec:sym\], but lattice artifacts of the heavy quark should be lumped into the HQET short-distance coefficients. Some of the operators needed to describe heavy-quark discretization effects do not appear in the usual HQET description of continuum QCD. For example, the dimension-seven operator $\sum_i\bar{h}_vD_i^4h_v$ (written here in the rest frame) appears in ${\cal L}^{(3)}$ to describe the breaking of rotational invariance on the lattice. Similarly, at and beyond dimension five there are HQET current operators to describe violations of rotational symmetry in the lattice currents. Because of the high dimension, these effects lie beyond the scope of this paper, which concentrates on operators of leading and next-to-leading dimension. In this way, the preceding description of continuum QCD can be repeated for lattice gauge theory with the same logic and structure. Instead of Eq. (\[eq:lat=Sym\]), one introduces a relation like Eq. (\[eq:QCD=HQET\]), $${\cal L}_{\text{lat}} \doteq {\cal L}_{\text{HQET}},$$ where ${\cal L}_{\text{lat}}$ is a lattice Lagrangian for NRQCD or Wilson quarks, and ${\cal L}_{\text{HQET}}$ is an HQET Lagrangian with the same operators as in Eqs. (\[eq:L0\]) and (\[eq:L1\]), but modified coefficients. In the dimension-four HQET Lagrangian ${\cal L}^{(0)}$, one must now replace $m$ with the heavy quark rest mass $m_1$. The other coefficients will be denoted ${\cal C}_i^{\text{lat}}$. In particular, in ${\cal L}^{(1)}$ the coefficient of the kinetic energy becomes $${\cal C}_2^{\text{lat}} = \frac{1}{2m_2}.$$ If operator insertions of ${\cal O}_2$ continue to be defined by dimensional regularization with minimal subtraction, then both the rest mass $m_1$ and the kinetic mass $m_2$ generalize the perturbative pole mass. Like the usual pole mass, they are properties of the pole in the perturbative quark propagator [@Mertens:1998wx], and they are infrared finite and gauge independent [@Kronfeld:1998di]. The lattice breaks Lorentz (or Euclidean) invariance, so reparametrization invariance no longer requires $m_2$ to be the same as $m_1$. Similarly, a heavy-light lattice (axial) vector current $V^\mu_{\text{lat}}$ ($A^\mu_{\text{lat}}$) can be described by $$\begin{aligned} V^\mu_{\text{lat}} & \doteq & C_{V_\parallel}^{\text{lat}} v^\mu \bar{q}h_v + C_{V_\perp}^{\text{lat}} \bar{q}i\gamma^\mu_\perp h_v - \sum_{i=1}^6 B_{Vi}^{\text{lat}} {\cal Q}^\mu_{Vi} + \cdots, \label{eq:VlatHQET} \\ A^\mu_{\text{lat}} & \doteq & C_{A_\perp}^{\text{lat}} \bar{q}i\gamma^\mu_\perp\gamma_5h_v - C_{A_\parallel}^{\text{lat}} v^\mu \bar{q}\gamma_5h_v - \sum_{i=1}^6 B_{Ai}^{\text{lat}} {\cal Q}^\mu_{Ai} + \cdots, \label{eq:AlatHQET}\end{aligned}$$ but there are two important changes from Eq. (\[eq:VcontHQET\]). First, the light quarks (and gluons) are now also on the lattice, so they are described by their usual Symanzik LE${\cal L}$s. Second, the short-distance coefficients of HQET are modified, because the lattice modifies the dynamics at short distances. The coefficients $C_{J_\parallel}^{\text{lat}}$, $C_{J_\perp}^{\text{lat}}$, and $B_{Ji}^{\text{lat}}$ now depend on the lattice spacing $a$, *i.e.*, on $ma$, in addition to $m$, $\mu/m$, and $m_q/m$. A heavy-light lattice axial vector current has an analogous description. On the other hand, in Eqs. (\[eq:VcontHQET\]) and (\[eq:VlatHQET\]) the HQET operators are the same. As a rule, the ultraviolet regulator of an effective theory does not have to be the same as that of the underlying theory. (The standard Symanzik program works this way.) Thus, when describing lattice gauge theory one is free to regulate HQET just as one would when describing continuum QCD. Moreover, since Eqs. (\[eq:Q1\])–(\[eq:Q6\]) give a complete set of dimension-four HQET currents, the coefficients $C_{J_\parallel,J_\perp}^{\text{lat}}$ and $B_{Ji}^{\text{lat}}$ contain short-distance effects from both the light and the heavy sectors. By comparing the HQET descriptions of lattice and continuum QCD, one can see how lattice matrix elements differ from their continuum counterparts. The continuum matrix element of $v\cdot{\cal V}$, for example, is $$\begin{aligned} \langle L|v\cdot {\cal V}|B\rangle & = & - C_{V_\parallel} \langle L|\bar{q}h_v|B_v^{(0)}\rangle - B_{V1} \langle L|v\cdot {\cal Q}_{V1}|B_v^{(0)}\rangle - B_{V4} \langle L|v\cdot {\cal Q}_{V4}|B_v^{(0)}\rangle \nonumber \\ & - & {\cal C}_2 C_{V_\parallel} \int d^4x \langle L|T\,{\cal O}_2(x) \bar{q}h_v|B_v^{(0)}\rangle^\star - {\cal C}_{\cal B} C_{V_\parallel} \int d^4x \langle L|T\,{\cal O}_{\cal B}(x) \bar{q}h_v|B_v^{(0)}\rangle^\star \nonumber \\ & + & O(\Lambda^2/m^2), \label{eq:VHQE}\end{aligned}$$ where $L$ is any light hadronic state, including the vacuum. (The $\star$-ed $T$ product is defined in Ref. [@Kronfeld:2000ck]; this detail is unimportant here.) On the left-hand side $B$ denotes a $b$-flavored hadron, and on the right-hand side $B_v^{(0)}$ denotes the corresponding eigenstate of the leading effective Lagrangian ${\cal L}^{(0)}$. Similarly, the lattice matrix element is [@Kronfeld:2000ck] $$\begin{aligned} \langle L|v\cdot V_{\text{lat}}|B\rangle & = & - C^{\text{lat}}_{V_\parallel} \langle L|\bar{q}h_v|B_v^{(0)}\rangle - B^{\text{lat}}_{V1} \langle L|v\cdot {\cal Q}_{V1}|B_v^{(0)}\rangle - B^{\text{lat}}_{V4} \langle L|v\cdot {\cal Q}_{V4}|B_v^{(0)}\rangle \nonumber \\ & - & {\cal C}^{\text{lat}}_2 C^{\text{lat}}_{V_\parallel} \int d^4x \langle L|T\,{\cal O}_2(x) \bar{q}h_v|B_v^{(0)}\rangle^\star - {\cal C}^{\text{lat}}_{\cal B} C^{\text{lat}}_{V_\parallel} \int d^4x \langle L|T\,{\cal O}_{\cal B}(x) \bar{q}h_v|B_v^{(0)}\rangle^\star \nonumber \\ & - & K_{\sigma\cdot F} C^{\text{lat}}_{V_\parallel} \int\! d^4x \langle L|T\,\bar{q}i\sigma Fq(x) \bar{q}h_v|B_v^{(0)}\rangle^\star + O(\Lambda^2a^2b(ma)).\end{aligned}$$ Compared to Eq. (\[eq:VHQE\]), the short-distance coefficients are modified to depend on $ma$, there is an extra term from the Symanzik LE${\cal L}$ of the light quark, and the next power corrections can, in general, be multiplied by a (bounded) function of $ma$. The matrix elements on the right-hand sides are, however, identical, because in both cases they are defined with ${\cal L}^{(0)}$ describing the heavy quark and ${\cal L}_{\text{QCD}}$ describing the light quark (and gluons). Similar equations hold for matrix elements of ${\cal V}_\perp$ and ${V_\perp}_{\text{lat}}$, and for the axial vector current. If one multiplies the equations for the lattice matrix elements with $$\begin{aligned} Z_{J_\parallel} & = & \frac{C_{J_\parallel}}{C_{J_\parallel}^{\text{lat}}}, \label{eq:defZVpara} \\ Z_{J_\perp} & = & \frac{C_{J_\perp}}{C_{J_\perp}^{\text{lat}}}, \label{eq:defZVperp}\end{aligned}$$ and subtracts the result from the continuum equations, one finds that the difference can be traced solely to the mismatch of the short-distance coefficients, or $$\begin{aligned} \delta{\cal C}_i = {\cal C}_i^{\text{lat}} & - & {\cal C}_i, \label{eq:deltaC} \\ \delta B_{Ji} = Z_{Ji} B_{Ji}^{\text{lat}} & - & B_{Ji}, \label{eq:deltaB}\end{aligned}$$ where the normalization factors $Z_{Ji}$ are $Z_{J_\parallel}$ for $i=1$, 4, and $Z_{J_\perp}$ for $i=2$, 3, 5, 6. In Eqs. (\[eq:deltaC\]) and (\[eq:deltaB\]) a picture emerges, where *heavy-quark lattice artifacts are isolated* into $\delta{\cal C}_i$ and $\delta B_{Ji}$. Furthermore, the analysis presented here makes no explicit reference to any method for computing the short-distance coefficients, so it applies at every order in perturbation theory (in $g^2$) and, presumably, at a non-perturbative level as well. The matching factors $Z_{J_\parallel}$ and $Z_{J_\perp}$ play the following role, sketched in Fig. \[fig:matching\]. (200,120)(50,0) (102,100)[lattice]{} (115,95)[(0,-1)[73]{}]{} (206,68)[(-2,1)[68]{}]{} (200,54)[HQET]{} (206,49)[(-2,-1)[68]{}]{} (105,10)[QCD]{} (75,54)[$C/C^{\text{lat}}$]{} (177,88)[$C^{\text{lat}}$]{} (177,23)[$C$]{} In each case, the denominator converts a lattice-regulated scheme to a renormalized HQET scheme, and the numerator converts the latter to a renormalized (continuum) QCD scheme. As long as the same HQET scheme is used, HQET drops out of the calculation of $Z_{J_\parallel}$ and $Z_{J_\perp}$. Moreover, changes in continuum renormalization conventions modify only the numerator, and changes in the lattice action or currents modify only the denominator. In a similar way, dependence on the HQET renormalization scheme drops out when computing $\delta{\cal C}_i$ and $\delta B_{Ji}$. One can derive a connection between the matching coefficients of the HQET and the Symanzik descriptions when $ma\ll 1$ and $m\gg\bbox{p}$, so that both formalisms apply. With the Lagrangian, one applies HQET to Eqs. (\[eq:LEL\])–(\[eq:artifacts\]) and identifies the short-distance coefficients with $m_1$, ${\cal C}_2^{\text{lat}}$, and ${\cal C}_{\cal B}^{\text{lat}}$. Then one finds, $$\begin{aligned} m_{1b} & = & m_b + O(a^2), \label{eq:m1=m} \\ m_{2b} & = & m_b + O(a^2), \label{eq:m2=m} \\ z_{\cal B}^{\text{lat}} & = & z_{\cal B} - 4 m_ba K_{\sigma\cdot F} C_{\sigma\cdot F}, \label{eq:zBlat=zB}\end{aligned}$$ where the short-distance coefficient $C_{\sigma\cdot F}$ appears in the relation $$\bar{b}i\sigma^{\mu\nu}F_{\mu\nu}b \doteq -2 C_{\sigma\cdot F} {\cal O}_{\cal B}.$$ At the tree level, $C_{\sigma\cdot F}^{[0]}=1$. For the \[axial\] vector current, one inserts Eq. (\[eq:VcontHQET\]) \[Eq. (\[eq:AcontHQET\])\] into Eq. (\[eq:LEV\]) \[Eq. (\[eq:LEA\])\], neglects terms of order $m^2a^2$, and compares with Eq. (\[eq:VlatHQET\]) \[Eq. (\[eq:VlatHQET\])\]. One also must match the tensor and pseudoscalar bilinears to HQET at the dimension-three level, $$\begin{aligned} \bar{q}i\sigma^{\mu\nu}b & \doteq & C_{T_+} \eta^\mu_\alpha \eta^\nu_\beta \bar{q}i\sigma^{\alpha\beta}h_v - C_{T_-} \bar{q}(v^\mu i\gamma_\perp^\nu - v^\nu i\gamma_\perp^\mu)h_v, \\ \bar{q}i\gamma_5b & \doteq & C_P \bar{q}i\gamma_5h_v,\end{aligned}$$ with short-distance coefficients $C_{T_\pm}$ and $C_P$. At the tree level, $C_{T_\pm}^{[0]}=C_P^{[0]}=1$. After carrying out these steps, one finds that $$\begin{aligned} Z_{V_\parallel} & = & Z_V , \label{eq:ZVpara} \\ Z_{V_\perp}^{-1} & = & Z_V^{-1} + (m_q+m_b)a K_VC_{T_-}/C_{V_\perp} , \label{eq:ZVperp} \\ Z_{V_\parallel} B_{V1}^{\text{lat}} & = & B_{V1} + a Z_V K_VC_{T_-} , \label{eq:ZVB1} \\ Z_{V_\parallel} B_{Vi}^{\text{lat}} & = & B_{Vi} + a Z_V K_VC_{T_+} , \quad i=2,6 , \label{eq:ZVB26} \\ Z_{V_\parallel} B_{V3}^{\text{lat}} & = & B_{V3} - a Z_V K_VC_{T_+} , \label{eq:ZVB3} \\ Z_{V_\parallel} B_{V4}^{\text{lat}} & = & B_{V4} - a Z_V K_VC_{T_-} , \label{eq:ZVB4} \\ Z_{V_\parallel} B_{V5}^{\text{lat}} & = & B_{V5} - a Z_V K_V(C_{T_+}-C_{T_-}) \label{eq:ZVB5}\end{aligned}$$ from matching the vector current, and $$\begin{aligned} Z_{A_\perp} & = & Z_A , \label{eq:ZAperp} \\ Z_{A_\parallel}^{-1} & = & Z_A^{-1} + (m_q+m_b)a K_AC_P/C_{A_\parallel} , \label{eq:ZApara} \\ Z_{A_\perp} B_{Ai}^{\text{lat}} & = & B_{Ai} + O(a^2), \quad i=1,2,5 , \label{eq:ZAB125} \\ Z_{A_\perp} B_{Ai}^{\text{lat}} & = & B_{Ai} + a Z_A K_AC_P , \quad i=3,6 , \label{eq:ZAB36} \\ Z_{A_\perp} B_{A4}^{\text{lat}} & = & B_{A4} - a Z_A K_AC_P \label{eq:ZAB4}\end{aligned}$$ from matching the axial vector current. Of course, these relations hold only when describing the same lattice currents $V^\mu_{\text{lat}}$ and $A^\mu_{\text{lat}}$, and then only to order $a^2$. Considering similar relations for the whole tower of higher-dimension operators, one sees $$\begin{aligned} \lim_{a\to 0} {\cal C}_{\cal O}^{\text{lat}} & = & {\cal C}_{\cal O}, \\ \lim_{a\to 0} Z_{Ji} B_{Ji}^{\text{lat}} & = & B_{Ji}.\end{aligned}$$ Eqs. (\[eq:ZVperp\])–(\[eq:ZVB5\]) and (\[eq:ZApara\])–(\[eq:ZAB4\]) illustrate for the next-to-leading dimension operators how the limit is accelerated for standard $O(a)$ improvement, with $K_{\sigma\cdot F}$, $K_V$, and $K_A$ themselves of order $a$. Equations (\[eq:ZVpara\])–(\[eq:ZAB4\]) show that HQET matching connects smoothly to Symanzik matching in the limit where both apply. HQET matching is, therefore, a natural and attractive extension into the more practical region where $ma$ is not very small. Continuum QCD still can be approximated well, but now order by order in the heavy-quark expansion. The remainder of this paper pursues this program in perturbation theory. One-loop corrections to the rest mass $m_1$ and the kinetic mass $m_2$ have been considered already in Ref. [@Mertens:1998wx]. The one-loop correction to ${\cal C}_{\cal B}$ would require a generalization of the calculation of $K_{\sigma\cdot F}$ [@Wohlert:1987rf] to incorporate the full mass dependence of the quark-gluon vertex. In this paper we focus on heavy-light currents. We construct lattice currents suitable for matching through order $1/m_Q$ in the heavy quark expansion. We then calculate the matching factors $Z_{J_\parallel}$ and $Z_{J_\perp}$ at the one-loop level, which are needed to fix the overall normalization of the heavy-light currents. Currents suitable for heavy-to-heavy transitions $b\to c$ are considered in a companion paper [@Kronfeld:1999tk]. Lattice Action and Currents {#sec:lattice} =========================== In this section our aim is to define heavy-light currents with Wilson fermions that are suited to the HQET matching formalism. Because Wilson fermions have the right particle content and obey the heavy-quark symmetries, the descriptive part of the formalism applies in any case. To use HQET to match lattice gauge theory to continuum QCD, however, we would like to ensure that $\delta{\cal C}_i$ and $\delta B_{Ji}$ \[cf. Eqs. (\[eq:deltaC\]) and (\[eq:deltaB\])\] remain bounded in the infinite-mass limit. Good behavior is attained by mimicking the structure of Eqs. (\[eq:Q1\])–(\[eq:Q6\]), so that improvement terms are guaranteed to remain small. Then we would like to adjust free parameters in the currents so that $\delta{\cal C}_i$ and $\delta B_{Ji}$ (approximately) vanish. We show how to do so in perturbation theory, obtaining $B^{\text{lat}}_{Ji}$ at the tree level and, in Sec. \[sec:loop\], the matching factors $Z_{J_\parallel}$ and $Z_{J_\perp}$ at the one-loop level. A suitable lattice Lagrangian was introduced in Ref. [@El-Khadra:1997mp]. It is convenient to write the lattice Lagrangian ${\cal L}_{\text{lat}}={\cal L}_0+{\cal L}_B+{\cal L}_E$. The first term is $$\begin{aligned} {\cal L}_0 &=& - (m_0+m_{0\text{cr}}) \bar{\psi}(x)\psi(x) - \case{1}{2} \bar{\psi}(x)\left[ (1+\gamma_4){D_4^-}_{\text{lat}} - (1-\gamma_4){D_4^+}_{\text{lat}}\right]\psi(x) \label{eq:S0} \\ &-& \zeta \bar{\psi}(x) \bbox{\gamma}\cdot\bbox{D}_{\text{lat}} \psi(x) + \case{1}{2} r_s\zeta a \bar{\psi}(x)\triangle^{(3)}_{\text{lat}}\psi(x). \nonumber\end{aligned}$$ The mass counterterm $m_{0\text{cr}}$ is included here so that, by definition, $m_0=0$ for massless quarks. The covariant difference operators ${D_4^\pm}_{\text{lat}}$, $\bbox{D}_{\text{lat}}$, and $\triangle^{(3)}_{\text{lat}}$, are defined in Ref. [@El-Khadra:1997mp]. They carry the label “lat” to distinguish them from the continuum covariant derivatives in Secs. \[subsec:sym\] and \[subsec:hqetcont\]. The symbol $\psi$ is reserved in this paper for lattice fermion fields. The temporal kinetic term is conventionally normalized, but the spatial kinetic term is multiplied with the coupling $\zeta$. The coupling $r_s$ is, in the technical sense, redundant [@El-Khadra:1997mp], but is included to solve the doubling problem[@Wilson:1975hf]. For ${\cal L}_0$ the tree-level relations between its couplings and the coefficients in the ${\cal L}_{\text{HQET}}$ are well known. By matching the kinetic energy, one finds (for $\bbox{v}=\bbox{0}$) $${{\cal C}^{\text{lat}}_2}^{[0]} = \frac{1}{2m^{[0]}_2a} = \frac{\zeta^2}{m_0a(2+m_0a)} + \frac{r_s\zeta}{2(1+m_0a)}. \label{eq:m2}$$ At higher orders in perturbation theory, ${\cal C}_2$ remains (for $\bbox{v}=\bbox{0}$) the kinetic mass of the quark, which is expressed in terms of the self energy in Ref. [@Mertens:1998wx]. ${\cal L}_0$ has cutoff artifacts, which are described by dimension-five and -higher operators in ${\cal L}_{\text{Sym}}$ (if $m_qa\ll1$) or ${\cal L}_{\text{HQET}}$ (if $m_Q\gg\Lambda_{\text{QCD}}$). The dimension-five effect can be reduced by adding $$\begin{aligned} {\cal L}_B & = & \case{i}{2} a c_B\zeta \, \bar{\psi}(x)\bbox{\Sigma}\cdot\bbox{B}_{\text{lat}}(x)\psi(x), \label{eq:SB} \\ {\cal L}_E & = & \case{1}{2} a c_E\zeta \, \bar{\psi}(x)\bbox{\alpha}\cdot\bbox{E}_{\text{lat}}(x)\psi(x), \label{eq:SE}\end{aligned}$$ and suitably adjusting of $c_B$ and $c_E$. The lattice chromomagnetic and chromoelectric fields, $\bbox{B}_{\text{lat}}$ and $\bbox{E}_{\text{lat}}$, are those given in Ref. [@El-Khadra:1997mp]. By matching the gluon-quark vertex, one finds $${{\cal C}^{\text{lat}}_{\cal B}}^{[0]} = \frac{1}{2m^{[0]}_{\cal B}a} = \frac{\zeta^2}{m_0a(2+m_0a)} + \frac{c_B\zeta}{2(1+m_0a)}. \label{eq:mB}$$ Higher-order corrections to ${\cal C}^{\text{lat}}_{\cal B}$ have not been obtained. By comparing Eqs. (\[eq:m2\]) and (\[eq:mB\]) one sees, however, that $c_B=r_s+O(g^2)$ is needed to adjust ${\cal C}^{\text{lat}}_{\cal B}$ to its continuum counterpart ${\cal C}_{\cal B}=z_{\cal B}/2m_2$. The Euclidean action is $S=-a^4\sum_x{\cal L}(x)$. Special cases are the Wilson action [@Wilson:1975hf], which sets $r_s=\zeta=1$, $c_B=c_E=0$; and the Sheikholeslami-Wohlert action [@Sheikholeslami:1985ij], which sets $r_s=\zeta=1$, $c_B=c_E\equiv c_{\text{SW}}$. But to remove lattice artifacts for arbitrary masses, the couplings $r_s$, $\zeta$, $c_B$ and $c_E$ must be taken to depend on $m_0a$ [@El-Khadra:1997mp]. Our analytical results for the integrands of Feynman diagrams, given in Appendix \[app:dirac\], are for arbitrary choices of these couplings. Indeed, our expressions allow the heavy and light quarks to have different values of all couplings. Heavy-light currents are defined in an essentially similar way. For convenience, first define a “rotated” field [@Kronfeld:1995nu; @El-Khadra:1997mp] $$\Psi_q = \left[1 + ad_1 \bbox{\gamma}\cdot\bbox{D}_{\text{lat}} \right] \psi_q, \label{eq:rotate}$$ where $\psi_q$ is the field in ${\cal L}_0$ of flavor $q$, and $\bbox{D}_{\text{lat}}$ is again the symmetric covariant difference operator. Simple bilinears with the right quantum numbers are $$\begin{aligned} V^\mu_0 & = & \bar{\Psi}_qi\gamma^\mu \Psi_b, \label{eq:Vlat} \\ A^\mu_0 & = & \bar{\Psi}_qi\gamma^\mu\gamma_5 \Psi_b. \label{eq:Alat}\end{aligned}$$ The subscript “0” implies that, as with ${\cal L}_0$, some improvement is desired. To ensure a good large-$ma$ limit, one should pattern the improved current after the right-hand side of Eq. (\[eq:VlatHQET\]). Thus, we take $$\begin{aligned} V^\mu_{\text{lat}} & = & V^\mu_0 - \sum_{i=1}^6 b_{Vi} Q^\mu_{Vi}, \label{eq:Vimp} \\ A^\mu_{\text{lat}} & = & A^\mu_0 - \sum_{i=1}^6 b_{Ai} Q^\mu_{Ai}, \label{eq:Aimp}\end{aligned}$$ where the $b_{Ji}$ are adjustable, and the dimension-four lattice operators are $$\begin{aligned} Q^\mu_{V1} & = & - v^\mu \bar{\psi}_q i{\kern+0.1em /\kern-0.55em v} {\kern+0.1em /\kern-0.65em D_\perp}_{\text{lat}} \psi_b, \label{eq:Q1lat} \\ Q^\mu_{V2} & = & \bar{\psi}_qi\gamma^\mu_\perp {\kern+0.1em /\kern-0.65em D_\perp}_{\text{lat}} \psi_b, \label{eq:Q2lat} \\ Q^\mu_{V3} & = & \bar{\psi}_q i {D^\mu_\perp}_{\text{lat}} \psi_b, \label{eq:Q3lat} \\ Q^\mu_{V4} & = & - v^\mu \bar{\psi}_q {\kern+0.1em /\kern-0.65em \loarrow{D}_\perp}_{\text{lat}} i{\kern+0.1em /\kern-0.55em v} \psi_b, \label{eq:Q4lat} \\ Q^\mu_{V5} & = & \bar{\psi}_q {\kern+0.1em /\kern-0.65em \loarrow{D}_\perp}_{\text{lat}} i\gamma^\mu_\perp \psi_b, \label{eq:Q5lat} \\ Q^\mu_{V6} & = & \bar{\psi}_qi {\loarrow{D}^\mu_\perp}_{\text{lat}} \psi_b, \label{eq:Q6lat}\end{aligned}$$ and each lattice operator $Q^\mu_{Ai}$ is obtained from $Q^\mu_{Vi}$ by replacing $\bar{\psi}_q$ with $-\bar{\psi}_q\gamma_5$. Lattice quark fields do not satisfy Eq. (\[eq:Pvh\]), so ${\kern+0.1em /\kern-0.55em v}$ appears explicitly. In practice, one uses the rest frame here, $v=(i,\bbox{0})$, as in Eq. (\[eq:rotate\]). An analogous construction for lattice NRQCD has been given by Morningstar and Shigemitsu [@Morningstar:1998ep]. It is worthwhile to emphasize the difference between Eqs. (\[eq:VlatHQET\]) and (\[eq:Vimp\]). Equation (\[eq:VlatHQET\]) is a general HQET description of any heavy-light lattice current. Equation (\[eq:Vimp\]) is a definition of a specific lattice current, namely the one used in this paper (and in calculations of $f_B$ and other hadronic matrix elements). In the same vein, the ${\cal Q}_{Ji}$ in Eqs. (\[eq:Q1\])–(\[eq:Q6\]) are HQET operators, whereas the $Q_{Ji}$ in Eqs. (\[eq:Q1lat\])–(\[eq:Q6lat\]) are lattice operators. Finally, the coefficients $B^{\text{lat}}_{Ji}$ are the output of a matching calculation: they depend on the $b_{Ji}$, which must be adjusted to make $\delta B_{Ji}$ vanish. To illustrate, let us consider the calculation of the coefficients $B^{\text{lat}}_{Ji}$ at the tree level. One computes on-shell matrix elements such as $\langle q|J_{\text{lat}}|b\rangle$ and $\langle 0|J_{\text{lat}}|\bar{q}b\rangle$ in lattice gauge theory and compares them to the corresponding matrix elements in HQET. Then one finds $$\begin{aligned} {C^{\text{lat}}_{J_\parallel}}^{[0]} & = & {C^{\text{lat}}_{J_\perp}}^{[0]} = e^{-(m^{[0]}_{1q}+m^{[0]}_{1b})a/2} , \label{eq:matchC[0]} \\ {B^{\text{lat}}_{Ji}}^{[0]} & = & e^{-(m^{[0]}_{1q}+m^{[0]}_{1b})a/2} \left( \frac{1}{2m^{[0]}_3} + b^{[0]}_{Ji}\right), \quad i\le2 \label{eq:B1B2lat} \\ {B^{\text{lat}}_{Ji}}^{[0]} & = & e^{-(m^{[0]}_{1q}+m^{[0]}_{1b})a/2} b^{[0]}_{Ji}, \quad i\ge 3 \label{eq:B3-6lat}\end{aligned}$$ where $$m_1^{[0]}a = \ln(1 + m_0a)$$ and, for our lattice Lagrangian and currents, $$\frac{1}{2m^{[0]}_3a} = \frac{\zeta(1+m_0a)}{m_0a(2+m_0a)} - d_1. \label{eq:m3}$$ Since (continuum QCD’s) $C_J^{[0]}=1$ there already is a non-trivial matching factor at the tree level relating the lattice and continuum currents, $Z_{J_\parallel}^{[0]}=Z_{J_\perp}^{[0]}=e^{(m_{1q}+m_{1b})a/2}$. After comparing Eqs. (\[eq:B1B2lat\])–(\[eq:B3-6lat\]) with Eqs. (\[eq:B1B2\])–(\[eq:B3-6\]), one sees that one can take $b_{Ji}^{[0]}=0$ for all six operators, if $d_1$ is adjusted correctly. At the tree level, the way to adjust $d_1$ is to set $m^{[0]}_3$ equal to the (tree-level) heavy-quark mass. In the effective Lagrangian there are two quark masses, the rest mass $m_1$ and the kinetic mass $m_2$. The former has no effect on matrix elements (and a trivial, additive effect on the mass spectrum). As discussed above, heavy-quark cutoff effects in matrix elements are reduced if ${\cal C}^{\text{lat}}_2={\cal C}_2$, which means one should identify the continuum quark mass with the kinetic mass. Thus, one should set $m^{[0]}_3=m^{[0]}_2$, which is obtained if one adjusts $$d_1 = \frac{\zeta(1+m_0a-\zeta)}{m_0a(2+m_0a)} - \frac{r_s\zeta}{2(1+m_0a)}. \label{eq:d1(m)}$$ The same rotation also improves heavy-heavy currents at the tree level. Beyond the tree level, it is convenient to define $d_1$ for the spatial component of the degenerate-mass, heavy-heavy current [@Kronfeld:1999tk]. Then the corrections heavy-heavy current analogous to $Q_{V2}$ and $Q_{V5}$ would be superfluous, but for unequal masses they are still required. For equal mass currents it is possible to compute $Z_{V_\parallel}$ nonperturbatively for all masses $m_b$. One may therefore prefer to write [@Hashimoto:2000yp; @Simone:2000nv; @El-Khadra:2001rv] $$Z_{J_{\parallel,\perp}^{ub}} = \sqrt{ Z_{V_\parallel^{uu}} Z_{V_\parallel^{bb}} } \rho_{J_{\parallel,\perp}^{ub}} \label{eq:rho}$$ and compute only the factor $\rho_{J_{\parallel,\perp}^{ub}}$ in perturbation theory. To calculate the pre-factor $Z_{V_\parallel^{bb}}$ appearing in Eq. (\[eq:rho\]), one must have a massive quark in the final state. The definition of the heavy-heavy matching factor is given in our companion paper [@Kronfeld:1999tk], along with a calculation of its one-loop level contribution. We give the results for heavy-light $\rho_{J_{\parallel,\perp}}$ in Sec. \[sec:loop\]. For a light quark, with $m_qa\ll 1$, the right hand side of Eq. (\[eq:d1(m)\]) vanishes linearly in $m_0a$. Therefore, $1-m_{3q}/m_{2q}$ is $O(m_q^2a^2)$, and the distinction between $m_{3q}$ and $m_{2q}$ is negligible. For this reason, and to simplify calculation, we set $m_q=0$. Then Eq. (\[eq:d1(m)\]) implies $d_1=0$ for the light quark. One-Loop Results {#sec:loop} ================ In this section we present results for the matching factors at the one-loop level in perturbation theory. The one-loop contributions are known for the Wilson [@Kuramashi:1998tt] and Sheikholeslami-Wohlert (SW) actions [@Ishikawa:1997xh]. Both these works omit the rotation term in the current [@Kronfeld:1995nu; @El-Khadra:1997mp], which is needed to obtain $1/m_3$ correctly. In this section we complete the work started in Ref. [@Ishikawa:1997xh] and report results with the clover term and with the rotation. For comparison we also present our results without the rotation, both with and without the clover term. The computer code for generating these results is freely available [@p:epaps]. The matching factors $Z_J$ ($J=V_\parallel$, $V_\perp$, $A_\parallel$, and $A_\perp$) are simply the ratios of the lattice and continuum radiative corrections: $$Z_J = \frac{ \left[Z_{2h}^{1/2}\Lambda_J Z_{2l}^{1/2}\right]^{\rm cont}}{ \left[Z_{2h}^{1/2}\Lambda_J Z_{2l}^{1/2}\right]^{\rm lat~}}, \label{eq:ZJGamma}$$ where $Z_{2h}$ and $Z_{2l}$ are wave-function renormalization factors of the heavy and light quarks, and the vertex function $\Lambda_J$ is the sum of one-particle irreducible three-point diagrams, in which one point comes from the current $J$ and the other two from the external quark states. The expression relating $Z_2$ to the lattice self energy, for all masses and gauge couplings, can be found in Ref. [@Mertens:1998wx]. Its dominant mass dependence is $$Z_2 \propto e^{-m_1a},$$ where $m_1$ is the all-orders rest mass (of the heavy quark). This mass dependence is not present in the vertex function or the continuum part of Eq. (\[eq:ZJGamma\]). Consequently, we write $$e^{-m^{[0]}_1a/2} Z_{J_\Gamma} = 1 + \sum_{l=1}^\infty g_0^{2l} Z_{J_\Gamma}^{[l]}, \label{eq:ZJ[1]}$$ so that the $Z_{J_\Gamma}^{[l]}$ are only mildly mass dependent. (A slightly different convention was used in presenting results for $Z_2$ in Ref. [@Mertens:1998wx].) By construction, this mass dependence in $\rho_{J_\Gamma}$ cancels out in a gauge-invariant, all orders way. So, we write $$\rho_{J_\Gamma} = 1 + \sum_{l=1}^\infty g_0^{2l} \rho_{J_\Gamma}^{[l]}. \label{eq:rhoJ[1]}$$ This rest of this section is split into two subsections. In the first, we present our results for the full mass dependence of $Z_{J_\Gamma}^{[1]}$ and $\rho_{J_\Gamma}^{[1]}$. In the second, we discuss the related calculation of the Brodsky-Lepage-Mackenzie scale $q^*$. In both cases, we discuss fully a range of checks on our calculations. $Z_J^{[1]}$ and $\rho_J^{[1]}$ ------------------------------ The combinations of wave-function and vertex renormalization in $Z_J$ are gauge invariant and ultraviolet and infrared finite. For vanishing light quark mass there is a collinear divergence (which can be regulated by an infinitesimally small mass), but it is common to lattice and continuum functions. In the desired ratio (\[eq:ZJGamma\]), the divergence cancels, and the result is independent of the scheme for regulating the collinear singularity. For large $ma$ a remnant of this cancellation appears. The lattice theory approaches its static limit, where its ultraviolet behavior is non-logarithmic. But the region of momentum $a^{-1}<q<m$ in the continuum diagrams generates logarithms. At the one-loop level one must find $3\ln(ma)$, with the same anomalous dimension as in Eqs. (\[eq:Cparallel\]) and (\[eq:Cperp\]). At higher loops the usual polynomial in $\ln(ma)$ will arise. We have calculated the one-loop Feynman diagrams for the action specified in Eqs. (\[eq:S0\])–(\[eq:SE\]), with arbitrary $m_0$, $r_s$, $\zeta$, $c_B$, and $c_E$ for the incoming heavy quark, $m'_0=0$, $r'_s$, $\zeta'$, $c'_B$, and $c'_E$ for the outgoing light quark. The needed Feynman rules are in Ref. [@Mertens:1998wx], apart from three new rules for the current itself, which are in Appendix \[app:feynman\]. As shown in Appendix \[app:dirac\], we have found a simple way to incorporate the rotation into the Dirac algebra. The resulting analytical expressions are surprisingly compact, and they are given explicitly in Appendices \[app:dirac\] and \[app:useful\]. We have evaluated these expressions for $r_s=\zeta=1$ and $c_E=c_B\equiv c_{\text{SW}}$. Thus, the numerical results correspond to the SW action ($c_{\text{SW}}=1$) and to the Wilson action ($c_{\text{SW}}=0$). Figure \[fig:ZV\] plots the full mass dependence of the matching factors for the vector current, (a) $Z_{V_\parallel}$, (b) $Z_{V_\perp}$, (c) $\rho_{V_\parallel}$, and (d) $\rho_{V_\perp}$. \(a) ![Full mass dependence of the one-loop coefficients of the matching factors of the vector current (a) $Z^{[1]}_{V_\parallel}$, (b) $Z^{[1]}_{V_\perp}$, (c) $\rho^{[1]}_{V_\parallel}$, and (d) $\rho^{[1]}_{V_\perp}$. Filled (open) symbols denote the SW (Wilson) action; solid (dotted) lines connecting squares (circles) indicate the rotation is included (omitted).[]{data-label="fig:ZV"}](ZV4m1.eps "fig:"){width="44.00000%"} (b) ![Full mass dependence of the one-loop coefficients of the matching factors of the vector current (a) $Z^{[1]}_{V_\parallel}$, (b) $Z^{[1]}_{V_\perp}$, (c) $\rho^{[1]}_{V_\parallel}$, and (d) $\rho^{[1]}_{V_\perp}$. Filled (open) symbols denote the SW (Wilson) action; solid (dotted) lines connecting squares (circles) indicate the rotation is included (omitted).[]{data-label="fig:ZV"}](ZVjm1.eps "fig:"){width="44.00000%"}\ (c) ![Full mass dependence of the one-loop coefficients of the matching factors of the vector current (a) $Z^{[1]}_{V_\parallel}$, (b) $Z^{[1]}_{V_\perp}$, (c) $\rho^{[1]}_{V_\parallel}$, and (d) $\rho^{[1]}_{V_\perp}$. Filled (open) symbols denote the SW (Wilson) action; solid (dotted) lines connecting squares (circles) indicate the rotation is included (omitted).[]{data-label="fig:ZV"}](rhoV4m1.eps "fig:"){width="44.00000%"} (d) ![Full mass dependence of the one-loop coefficients of the matching factors of the vector current (a) $Z^{[1]}_{V_\parallel}$, (b) $Z^{[1]}_{V_\perp}$, (c) $\rho^{[1]}_{V_\parallel}$, and (d) $\rho^{[1]}_{V_\perp}$. Filled (open) symbols denote the SW (Wilson) action; solid (dotted) lines connecting squares (circles) indicate the rotation is included (omitted).[]{data-label="fig:ZV"}](rhoVjm1.eps "fig:"){width="44.00000%"} These numerical results are for the SW action with rotation (solid lines) and also for the SW and Wilson actions without the rotation (dotted lines). Figure \[fig:ZA\] plots the full mass dependence of the matching factors for the axial vector current, (a) $Z_{A_\parallel}$, (b) $Z_{A_\perp}$, (c) $\rho_{A_\parallel}$, and (d) $\rho_{A_\perp}$. \(a) ![Full mass dependence of the one-loop coefficients of the matching factors of the axial vector current (a) $Z^{[1]}_{A_\parallel}$, (b) $Z^{[1]}_{A_\perp}$, (c) $\rho^{[1]}_{A_\parallel}$, and (d) $\rho^{[1]}_{A_\perp}$.[]{data-label="fig:ZA"}](ZA4m1.eps "fig:"){width="44.00000%"} (b) ![Full mass dependence of the one-loop coefficients of the matching factors of the axial vector current (a) $Z^{[1]}_{A_\parallel}$, (b) $Z^{[1]}_{A_\perp}$, (c) $\rho^{[1]}_{A_\parallel}$, and (d) $\rho^{[1]}_{A_\perp}$.[]{data-label="fig:ZA"}](ZAjm1.eps "fig:"){width="44.00000%"}\ (c) ![Full mass dependence of the one-loop coefficients of the matching factors of the axial vector current (a) $Z^{[1]}_{A_\parallel}$, (b) $Z^{[1]}_{A_\perp}$, (c) $\rho^{[1]}_{A_\parallel}$, and (d) $\rho^{[1]}_{A_\perp}$.[]{data-label="fig:ZA"}](rhoA4m1.eps "fig:"){width="44.00000%"} (d) ![Full mass dependence of the one-loop coefficients of the matching factors of the axial vector current (a) $Z^{[1]}_{A_\parallel}$, (b) $Z^{[1]}_{A_\perp}$, (c) $\rho^{[1]}_{A_\parallel}$, and (d) $\rho^{[1]}_{A_\perp}$.[]{data-label="fig:ZA"}](rhoAjm1.eps "fig:"){width="44.00000%"} These and the following figures are plotted against $m_1^{[0]}a$ because this variable conveniently covers the whole mass range: for small mass $m_1\approx m_2$, and for large mass $m_1a\approx\ln m_2a$. We have carried out several checks on our calculations. In each case, identical numerical results have been obtained with two or more completely independent programs. The results for $Z_{J_{\parallel,\perp}}$ agree with those previously obtained, for $c_{\text{SW}}=0$ [@Kuramashi:1998tt] and for $c_{\text{SW}}=1$, $d_1=0$ [@Ishikawa:1997xh]. We have also reproduced limiting cases, as we briefly discuss below. For $ma=0$ our calculation reduces to the usual matching calculation for massless quarks. We find (with $C_F=4/3$) $$\begin{aligned} Z_{V_\parallel}^{[1]} = Z_{V_\perp}^{[1]} & = & \left\{ \begin{array}{ll} - 0.129423(6), & c_{\text{SW}} = 1, \\ - 0.174073(7), & c_{\text{SW}} = 0, \end{array} \right. \\ Z_{A_\parallel}^{[1]} = Z_{A_\perp}^{[1]} & = & \left\{ \begin{array}{ll} - 0.116450(5), & c_{\text{SW}} = 1, \\ - 0.133365(5), & c_{\text{SW}} = 0, \end{array} \right.\end{aligned}$$ in excellent agreement with previous work for $c_{\text{SW}}=1$ [@Gabrielli:1991us; @Luscher:1996ax; @Capitani:2001xi] and $c_{\text{SW}}=0$ [@Capitani:2001xi; @Martinelli:1983mw; @Bernard:1999sx]. (Reference [@Capitani:2001xi] gives precise results as a polynomial in $c_{\text{SW}}$.) As the mass tends to infinity, these actions and currents all lead, up to an unphysical factor, to the same vertices and quark propagator—a Wilson line. Perturbative corrections to the vertex functions must respect this universal static limit, and, therefore, they must tend to a universal value. As $ma\to\infty$, one expects the $Z$ factors for a massive quark to approach those for the static limit, namely $$Z_J^{[1]} = \frac{C_F}{16\pi^2} \left[\gamma_h\ln(m_2a)^2 + z_J^{[1]}\right],$$ where the constant $z_J^{[1]}$ depends on the current $J$ and on $c_{\text{SW}}$ (of the light quark). Since $\ln(m_2a)\approx m_1a$ in this region one expects the linear behavior seen in Figs. \[fig:ZV\] and \[fig:ZA\]. The static limit is also shown in Figs. \[fig:ZV\] and \[fig:ZA\] with $$\begin{aligned} z_{V_\parallel}^{[1]} & = & \left\{ \begin{array}{rl} - 10.248, & c_{\text{SW}} = 1, \\ - 7.929, & c_{\text{SW}} = 0, \end{array} \right. \\ z_{A_\perp}^{[1]} & = & \left\{ \begin{array}{rl} - 12.248, & c_{\text{SW}} = 1, \\ - 9.929, & c_{\text{SW}} = 0, \end{array} \right. \\ z_{V_\perp}^{[1]} & = & \left\{ \begin{array}{rl} - 18.414, & c_{\text{SW}} = 1, \\ - 24.379, & c_{\text{SW}} = 0, \end{array} \right. \\ z_{A_\parallel}^{[1]} & = & \left\{ \begin{array}{rl} - 16.414, & c_{\text{SW}} = 1, \\ - 22.379, & c_{\text{SW}} = 0. \end{array} \right. \end{aligned}$$ We have obtained these constants ourselves [@Ishikawa:1998rv]. They agree with previous (less precise) results for $c_{\text{SW}}=1$ [@Borrelli:1992fy] and $c_{\text{SW}}=0$ [@Eichten:1990zv]. As one can see from looking at Figs. \[fig:ZV\] and \[fig:ZA\], the static result is a good approximation for $m_1^{[0]}a>5$ or, equivalently, $m_0a\approx m_2a>150$. Some of the points at the highest masses have large error and lie nearly one $\sigma$ off the curve. The origin of this behavior is that the lattice and continuum integrals are dominated by different momenta: the continuum integral is dominated by the region $k\sim m_2\gg a^{-1}$, whereas the lattice integral is dominated by the region $k\sim a^{-1}$. This mass region is not of much practical interest, since here one has an essentially static quark. Equations (\[eq:ZVpara\])–(\[eq:ZAB4\]) allow us to check the small (heavy-quark) mass limit against the work of Sint and Weisz [@Sint:1997jx]. In our conventions the matching factors $Z_V$ and $Z_A$ are functions of gauge coupling and quark mass. Thus, $$\begin{aligned} Z_V(m_qa, m_ba) & = & Z_V \left[1 + \case{1}{2}(m_q+m_b)ab_V \right], \\ Z_A(m_qa, m_ba) & = & Z_A \left[1 + \case{1}{2}(m_q+m_b)ab_A \right],\end{aligned}$$ where, on the right-hand side, we adopt the notation of Refs. [@Luscher:1996sc; @Jansen:1996ck; @Sint:1997jx], and the $Z$s and $b$s do not depend on mass. Here only the mass dependence is displayed; all quantities depend also on the gauge coupling. If we omit the rotation, our currents and those considered by Sint and Weisz coincide, apart from one-loop counterterms. Thus, in one-loop calculations the slopes of our mass-dependent matching factors must agree with them. Setting $m_q=0$, and using Eqs. (\[eq:ZVpara\])–(\[eq:ZAB4\]), $$\begin{aligned} \frac{\partial Z^{[1]}_{V_\parallel}}{\partial m_{1b}} & = & \case{1}{2}b^{[1]}_V, \\ \frac{\partial Z^{[1]}_{V_\perp}}{\partial m_{1b}} & = & \case{1}{2}b^{[1]}_V - K^{[1]}_V, \\ \frac{\partial Z^{[1]}_{A_\parallel}}{\partial m_{1b}} & = & \case{1}{2}b^{[1]}_A - K^{[1]}_A, \\ \frac{\partial Z^{[1]}_{A_\perp}}{\partial m_{1b}} & = & \case{1}{2}b^{[1]}_A.\end{aligned}$$ To extract these slopes, we form a combination of integrands with three different (small) values of $m_ba$, yielding $b_J^{[1]}$ and $K_J^{[1]}$ up to $O(m_ba)^2$. In this way we find (for $c_{\text{SW}}=1$) $$\begin{aligned} b_V^{[1]} & = & C_F\times 0.114929(10) = 0.153239(14) \\ &{\rm vs.}\;& C_F\times 0.11492(4) \cite{Sint:1997jx}, \nonumber \\ b_A^{[1]} & = & C_F\times 0.114142(10) = 0.152189(14) \\ &{\rm vs.}\;& C_F\times 0.11414(4) \cite{Sint:1997jx}, \nonumber \\ K_V^{[1]} & = & C_F\times 0.0122499(6) = 0.016332(7) \\ &{\rm vs.}\;& C_F\times 0.01225(1) \cite{Sint:1997jx}, \nonumber \\ K_A^{[1]} & = & C_F\times 0.0056806(11) = 0.0075741(15) \\ &{\rm vs.}\;& C_F\times 0.005680(2) \cite{Sint:1997jx}, \nonumber\end{aligned}$$ which agrees perfectly with Ref. [@Sint:1997jx]. These results have also been checked by Taniguchi and Ukawa [@Taniguchi:1998pf]. We also obtain $$b_V^{[1]} - b_A^{[1]} = C_F\times 0.0007833(11) = 0.0010444(16) \label{eq:bV-bA}$$ by subtracting the integrands first, and then integrating. In taking the difference, large contributions from the self energy cancel, but, even so, the near equality of $b_V^{[1]}$ and $b_A^{[1]}$ is a bit astonishing. Comparing the slopes of Figs. \[fig:ZV\](a) and \[fig:ZA\](b) one sees that $b_V^{[1]}-b_A^{[1]}$ for the Wilson action is not so small. Although these checks are reassuring, the main result of this section is to obtain the full mass dependence of the matching factors. The results at intermediate mass, with $m_1a<3$ or, equivalently, $m_0a<1.5$, are needed for realistic calculations of $B$ meson properties. This region is neither particularly close to the massless limit, nor to the logarithmic behavior of the static limit. BLM scales $q^*$ ---------------- It is well-known that perturbation theory in the bare coupling $g_0^2(1/a)$ converges poorly. Therefore, we calculate the ingredients needed to determine the Brodsky-Lepage-Mackenzie (BLM) scale [@Brodsky:1983gc; @Lepage:1993xa]. For a coupling in scheme $S$, we denote the BLM expansion parameter $g_S^2(q^*_S)$. The BLM scale $q^*_S$ is given by $$\ln(q^*_Sa)^2 = - b_S^{(1)} + \frac{\int d^4k\,\ln(ka)^2\,f(ka)}{\int d^4k\,f(ka)}, \label{eq:qstar}$$ where $k$ is the gluon momentum, and $f(k)$ is the integrand of the quantity of interest, e.g., $\int d^4k\,f(k)=Z_J^{[1]}$. The constant $b_S^{(1)}$ is the $\beta_0$-dependent part of the one-loop conversion from the arbitrary scheme $S$ to the “$V$ scheme”, namely $$\frac{(4\pi)^2}{g^2_S(q)} = \frac{(4\pi)^2}{g^2_V(q)} + \beta_0 b_S^{(1)} + b_S^{(0)} + O(g^2), \label{eq:bS}$$ where for $n_f$ light quarks $\beta_0=11-2n_f/3$, and $b_S^{(0)}$ is independent of $n_f$. The $V$-scheme coupling $g^2_V(q)$ is defined so that the Fourier transform of the heavy-quark potential reads $V(q)=-C_Fg^2_V(q)/q^2$. Equation (\[eq:qstar\]) shows that the definitions of $q^*$ in Refs. [@Brodsky:1983gc] and [@Lepage:1993xa] are identical in the $V$ scheme. For our matching factors it is straightforward to weight the integrands with $\ln(ka)^2$ to obtain $$\ln(q^*_Va)^2 = \frac{{}^*Z^{[1]}}{Z^{[1]}}, \label{eq:qStar}$$ because the integration over $d^4k$ has no divergences. The denominators are the one-loop coefficients given above, and the numerators are presented now. Figure \[fig:starZV\] plots the full mass dependence of the numerators for the vector current, (a) ${}^*Z^{[1]}_{V_\parallel}$, (b) ${}^*Z^{[1]}_{V_\perp}$, (c) ${}^*\rho^{[1]}_{V_\parallel}$, and (d) ${}^*\rho^{[1]}_{V_\perp}$. \(a) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the vector current (a) ${}^*Z^{[1]}_{V_\parallel}$, (b) ${}^*Z^{[1]}_{V_\perp}$, (c) ${}^*\rho^{[1]}_{V_\parallel}$, and (d) ${}^*\rho^{[1]}_{V_\perp}$.[]{data-label="fig:starZV"}](starZV4m1.eps "fig:"){width="44.00000%"} (b) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the vector current (a) ${}^*Z^{[1]}_{V_\parallel}$, (b) ${}^*Z^{[1]}_{V_\perp}$, (c) ${}^*\rho^{[1]}_{V_\parallel}$, and (d) ${}^*\rho^{[1]}_{V_\perp}$.[]{data-label="fig:starZV"}](starZVjm1.eps "fig:"){width="44.00000%"}\ (c) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the vector current (a) ${}^*Z^{[1]}_{V_\parallel}$, (b) ${}^*Z^{[1]}_{V_\perp}$, (c) ${}^*\rho^{[1]}_{V_\parallel}$, and (d) ${}^*\rho^{[1]}_{V_\perp}$.[]{data-label="fig:starZV"}](starrhoV4m1.eps "fig:"){width="44.00000%"} (d) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the vector current (a) ${}^*Z^{[1]}_{V_\parallel}$, (b) ${}^*Z^{[1]}_{V_\perp}$, (c) ${}^*\rho^{[1]}_{V_\parallel}$, and (d) ${}^*\rho^{[1]}_{V_\perp}$.[]{data-label="fig:starZV"}](starrhoVjm1.eps "fig:"){width="44.00000%"} As before, these numerical results are for the SW action with rotation (solid lines) and also for the SW and Wilson actions without the rotation (dotted lines). Figure \[fig:starZA\] plots the full mass dependence of the numerator of Eq. (\[eq:qStar\]) for the axial vector current, (a) ${}^*Z^{[1]}_{A_\parallel}$, (b) ${}^*Z^{[1]}_{A_\perp}$, (c) ${}^*\rho^{[1]}_{A_\parallel}$, and (d) ${}^*\rho^{[1]}_{A_\perp}$. \(a) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the axial vector current (a) ${}^*Z^{[1]}_{A_\parallel}$, (b) ${}^*Z^{[1]}_{A_\perp}$, (c) ${}^*\rho^{[1]}_{A_\parallel}$, and (d) ${}^*\rho^{[1]}_{A_\perp}$.[]{data-label="fig:starZA"}](starZA4m1.eps "fig:"){width="44.00000%"} (b) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the axial vector current (a) ${}^*Z^{[1]}_{A_\parallel}$, (b) ${}^*Z^{[1]}_{A_\perp}$, (c) ${}^*\rho^{[1]}_{A_\parallel}$, and (d) ${}^*\rho^{[1]}_{A_\perp}$.[]{data-label="fig:starZA"}](starZAjm1.eps "fig:"){width="44.00000%"}\ (c) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the axial vector current (a) ${}^*Z^{[1]}_{A_\parallel}$, (b) ${}^*Z^{[1]}_{A_\perp}$, (c) ${}^*\rho^{[1]}_{A_\parallel}$, and (d) ${}^*\rho^{[1]}_{A_\perp}$.[]{data-label="fig:starZA"}](starrhoA4m1.eps "fig:"){width="44.00000%"} (d) ![Full mass dependence of the estimated $\beta_0g^4$ terms of the matching factors of the axial vector current (a) ${}^*Z^{[1]}_{A_\parallel}$, (b) ${}^*Z^{[1]}_{A_\perp}$, (c) ${}^*\rho^{[1]}_{A_\parallel}$, and (d) ${}^*\rho^{[1]}_{A_\perp}$.[]{data-label="fig:starZA"}](starrhoAjm1.eps "fig:"){width="44.00000%"} We have carried out several checks on our calculations. Once again, identical numerical results have been obtained with two or more completely independent programs. Also, at $m_ba=0$ we reproduce the results, for the Wilson action, of Ref. [@Bernard:1999sx]. For ${}^*Z_J^{[1]}$ and ${}^*\rho_J^{[1]}$ the limit of large $ma$ also has distinctive features. In that case $${}^*Z_J^{[1]} = \frac{C_F}{16\pi^2}\left[ \case{1}{2} \gamma_h \ln^2(m_2a)^2 + \gamma'_h \ln(m_2a)^2 + {}^*z_J \right], \label{eq:starZinf}$$ where $\gamma'_h$ is related to the two-loop anomalous dimension. A similar expression holds for ${}^*\rho_J^{[1]}$, with a different constant. Note that—in both cases—the one-loop anomalous dimension appears multiplying $\ln^2(m_2a)$. The growth expected from Eq. (\[eq:starZinf\]) is seen in Figs. \[fig:starZV\] and \[fig:starZA\]. As a consequence, one finds $q^*a\propto\sqrt{m_2a}$ as $ma\to\infty$. Square root behavior is typical of cases with an anomalous dimension. For the $Z$ factors, the resulting values for $q^*a$ are relatively constant in the “low mass” region, $q^*a\sim 2.7$–2.9. Figure \[fig:qstar\] shows how $q^*a$ depends on the heavy quark mass in the region $m_1a\le 2$, which is the one most relevant to calculations of decay constants and form factors. \(a) ![Full mass dependence of the BLM scale $q^*$, for (a) $Z_{V_\parallel}$, (b) $Z_{V_\perp}$, (c) $Z_{A_\parallel}$, and (d) $Z_{A_\perp}$.[]{data-label="fig:qstar"}](qstarZV4m1.eps "fig:"){width="44.00000%"} (b) ![Full mass dependence of the BLM scale $q^*$, for (a) $Z_{V_\parallel}$, (b) $Z_{V_\perp}$, (c) $Z_{A_\parallel}$, and (d) $Z_{A_\perp}$.[]{data-label="fig:qstar"}](qstarZVjm1.eps "fig:"){width="44.00000%"}\ (c) ![Full mass dependence of the BLM scale $q^*$, for (a) $Z_{V_\parallel}$, (b) $Z_{V_\perp}$, (c) $Z_{A_\parallel}$, and (d) $Z_{A_\perp}$.[]{data-label="fig:qstar"}](qstarZA4m1.eps "fig:"){width="44.00000%"} (d) ![Full mass dependence of the BLM scale $q^*$, for (a) $Z_{V_\parallel}$, (b) $Z_{V_\perp}$, (c) $Z_{A_\parallel}$, and (d) $Z_{A_\perp}$.[]{data-label="fig:qstar"}](qstarZAjm1.eps "fig:"){width="44.00000%"} At larger masses $Z^{[1]}$ goes through zero, at which point the original BLM prescription breaks down. A prescription for $q^*$ in this case is given in Ref. [@Hornbostel:2001ey]. For the Wilson action the zero in $Z_{V_\parallel}^{[1]}$ is at a smaller than usual mass \[see Fig. \[fig:ZV\](a)\], which explains its behavior for the BLM $q^*a$ seen in Fig. \[fig:qstar\](a). For the $\rho$ factors the denominator $\rho^{[1]}$ is small over most of the interesting region, as seen in Figs. \[fig:ZV\](c)–(d) and \[fig:ZA\](c)–(d). It is also interesting to see how $q^*$ changes under tadpole improvement. If one introduces the tadpole-improved matching factors $$\tilde{Z}_J = Z_J/u_0,$$ where the mean link $u_0$ is any tadpole-dominated short-distance quantity, the arguments of Ref. [@Lepage:1993xa] suggest that the perturbative series for $\tilde{Z}_J$ has smaller coefficients. In analogy with Eq. (\[eq:ZJGamma\]) we write $$e^{-\tilde{m}^{[0]}_1a/2} \tilde{Z}_J = 1 + \sum_{l=1}^\infty g_0^{2l} \tilde{Z}_J^{[l]}, \label{eq:tildeZJ[1]}$$ where $$\tilde{m}_1^{[0]}a = \ln[1 + m_0a/u_0]$$ is the tadpole-improved rest mass. Then $$\tilde{Z}_J^{[1]} = Z_J^{[1]} - \frac{1}{2} \left(1 + \frac{1}{1+m_0a} \right) u_0^{[1]},$$ and because $Z_J^{[1]}<0$ and $u_0^{[1]}<0$ one sees that the one-loop coefficients are reduced. Similarly, for computing the BLM scale $${}^*\tilde{Z}_J^{[1]} = {}^*Z_J^{[1]} - \frac{1}{2} \left(1 + \frac{1}{1+m_0a} \right) {}^*u_0^{[1]}.$$ To illustrate, we take $u_0$ from the average plaquette, so $u_0^{[1]}=-C_F/16$ and $^*u_0^{[1]}=-0.204049(1)$. Figure \[fig:tadqstar\] shows that, as a rule, $q^*$ is significantly reduced, which means that tadpole improvement has removed some of the most ultraviolet contributions. \(a) ![Tadpole-improved $q^*$ for (a) $\tilde{Z}_{V_\parallel}$, (b) $\tilde{Z}_{V_\perp}$, (c) $\tilde{Z}_{A_\parallel}$, and (d) $\tilde{Z}_{A_\perp}$, with Fig. \[fig:qstar\] in gray.[]{data-label="fig:tadqstar"}](tadqstarZV4m1.eps "fig:"){width="44.00000%"} (b) ![Tadpole-improved $q^*$ for (a) $\tilde{Z}_{V_\parallel}$, (b) $\tilde{Z}_{V_\perp}$, (c) $\tilde{Z}_{A_\parallel}$, and (d) $\tilde{Z}_{A_\perp}$, with Fig. \[fig:qstar\] in gray.[]{data-label="fig:tadqstar"}](tadqstarZVjm1.eps "fig:"){width="44.00000%"}\ (c) ![Tadpole-improved $q^*$ for (a) $\tilde{Z}_{V_\parallel}$, (b) $\tilde{Z}_{V_\perp}$, (c) $\tilde{Z}_{A_\parallel}$, and (d) $\tilde{Z}_{A_\perp}$, with Fig. \[fig:qstar\] in gray.[]{data-label="fig:tadqstar"}](tadqstarZA4m1.eps "fig:"){width="44.00000%"} (d) ![Tadpole-improved $q^*$ for (a) $\tilde{Z}_{V_\parallel}$, (b) $\tilde{Z}_{V_\perp}$, (c) $\tilde{Z}_{A_\parallel}$, and (d) $\tilde{Z}_{A_\perp}$, with Fig. \[fig:qstar\] in gray.[]{data-label="fig:tadqstar"}](tadqstarZAjm1.eps "fig:"){width="44.00000%"} With a lower scale, the coupling $g^2_V(q^*)$ becomes a bit larger with tadpole improvement. For $Z_{V_\parallel}$ and $Z_{A_\perp}$, however, the denominator $\tilde{Z}_J^{[1]}$ already vanishes for $m_1a\approx1.5$–2.0, leading to rapid growth in the BLM $q^*$ for the Wilson action, and a zero in the BLM $q^*$ for the SW action. One should again define $q^*$ in a more robust way [@Hornbostel:2001ey]. Another choice for the mean field is $u_0=8\kappa_{\text{crit}}$. It gives coefficients and BLM scales that lie between the unimproved and tadpole-improved cases [@Harada:2001xm]. Our method also allows us to obtain the BLM scale for the improvement coefficients in the [*Alpha*]{} program. Then we are in a position to compare BLM perturbation theory with non-perturbative determinations of these coefficients. We will give these results for $q^*$ and the mentioned comparison in another publication [@Harada:2001zz]. Conclusions {#sec:conclusions} =========== In this paper we have set up a matching procedure, based on HQET, for heavy-light currents. It is valid for all $ma$, where $m$ is the heavy quark’s mass and $a$ is the lattice spacing, and to all orders in the gauge coupling. It could be applied to lattice NRQCD, although here it is applied to Wilson fermions. In the latter case, HQET matching agrees with Symanzik matching when $ma\ll1$. In this way, HQET matching is a natural and attractive extension into the regime $ma\not\ll1$, which is needed for heavy-quark phenomenology. Our one-loop results for the SW action are of immediate value for lattice calculations of $f_B$ and of form factors for the semi-leptonic decay $B\to\pi l\nu$. Indeed, our earlier one-loop results [@Ishikawa:1997xh] (which omitted the “rotation” terms in the current) were used for $f_B$ in Refs. [@Aoki:1998ji; @El-Khadra:1998hq; @Bernard:1998xi; @AliKhan:2000eg; @Bernard:2000nv], and our results were used for semi-leptonic form factors in Ref. [@El-Khadra:2001rv]. In particular, we have obtained the BLM scale $q^*$ for the matching factors, which should reduce the uncertainty of one-loop calculations. Similarly, computing part of the normalization factor, namely $\sqrt{Z_{V_\parallel^{qq}}Z_{V_\parallel^{bb}}}$, non-perturbatively reduces the normalization uncertainty even further [@Hashimoto:2000yp; @Simone:2000nv; @El-Khadra:2001rv]. (The heavy-heavy normalization factor $Z_{V_\parallel^{bb}}$ is defined in our companion paper for heavy-heavy currents [@Kronfeld:1999tk].) An outstanding problem at this time is the one-loop calculation of the coefficients $B_{Ji}^{\text{lat}}$ of the dimension-four terms in the HQET description. A calculation of these coefficients, and the subsequent adjustment of the parameters $b_{Ji}$ in the lattice currents, would eliminate uncertainties of order $\alpha_s\bar{\Lambda}/m$ and $\alpha_s\bar{\Lambda}a$ in (future) calculations of heavy-quark matrix elements. The algebra quickly becomes voluminous, making this problem well-suited to automated techniques [@Nobes:2001pt]. A.S.K. would like to thank Akira Ukawa and the Center for Computational Physics for hospitality while part of this work was being carried out, and the Aspen Center for Physics for a stimulating atmosphere while part of the paper was being written. S.H., A.S.K., and T.O. would also like to thank the Institute for Nuclear Theory at the University of Washington for hospitality while this paper was being finished. S.H. and T.O. are supported by Grants-in-Aid of the Japanese Ministry of Education (Nos. 11740162 and 12640279, respectively). K.-I.I. and N.Y. are supported by JSPS Research Fellowships. Fermilab is operated by Universities Research Association Inc., under contract with the U.S. Department of Energy. Feynman Rules {#app:feynman} ============= The needed propagators and vertices for quark-gluon interactions are given in Ref. [@Mertens:1998wx]. Here we give the additional Feynman rules induced by the rotation term of the heavy quark. The additional rules are easy to derive by expressing the covariant difference operator as [@Kronfeld:1985zv] $$D^\mu_{\text{lat}} = \left[T_{+\mu} - T_{-\mu} \right]/(2a),$$ where $$T_{\pm\mu} = t_{\pm\mu/2} e^{\pm g_0aA_\mu} t_{\pm\mu/2},$$ and $t_{\pm\mu/2}$ translates fields to its right by one-half lattice spacing in the $\pm\mu$ direction. There are three rules to give, with 0, 1, and 2 gluons emerging from the vertex. Let the Dirac matrix of the current be $\Gamma$. Then, $$\begin{aligned} \mbox{0-gluon} & = & \Gamma\left[1 + id_1\sum_r\gamma_r\sin(p_r)\right] , \label{eq:0} \\ \mbox{1-gluon} & = & g_0 t^a \; d_1 \, \Gamma \gamma_i \cos(p+\case{1}{2} k)_i , \label{eq:1} \\ \mbox{2-gluon} & = & i g_0^2 \case{1}{2}\{t^a, t^b\} \delta_{ij} \; d_1\, \Gamma \gamma_i \sin(p+\case{1}{2} k+\case{1}{2}\ell)_i , \label{eq:2}\end{aligned}$$ where momentum $p$ is quark momentum flowing into vertex, and $k$ and $\ell$ are gluon momentum flowing into vertex. As in Ref. [@Mertens:1998wx], the matrices $t^a$ are anti-Hermitian, *i.e.*, $U_\mu=\exp\left(g_0t^aA^a_\mu\right)$, $\sum_{aj} t^a_{ij} t^a_{jk} = - C_F \delta_{ik}$, and $\tr t^at^b=-\case{1}{2}\delta^{ab}$. Dirac Algebra {#app:dirac} ============= To compute the vertex function, there are four diagrams to consider, depicted in Fig. \[fig:Feyn\]: \(a) ![Feynman diagrams for calculating the vertex function. The $\bullet$ on each side of the $\otimes$ indicates the rotation.[]{data-label="fig:Feyn"}](Feyn_a.eps "fig:"){width="20.00000%"} (b) ![Feynman diagrams for calculating the vertex function. The $\bullet$ on each side of the $\otimes$ indicates the rotation.[]{data-label="fig:Feyn"}](Feyn_b.eps "fig:"){width="20.00000%"} (c) ![Feynman diagrams for calculating the vertex function. The $\bullet$ on each side of the $\otimes$ indicates the rotation.[]{data-label="fig:Feyn"}](Feyn_c.eps "fig:"){width="20.00000%"} (d) ![Feynman diagrams for calculating the vertex function. The $\bullet$ on each side of the $\otimes$ indicates the rotation.[]{data-label="fig:Feyn"}](Feyn_d.eps "fig:"){width="20.00000%"} the usual vertex diagram (with the rotation inside), Fig. \[fig:Feyn\](a); two diagrams with the gluon connected to the incoming rotation, Fig. \[fig:Feyn\](b) and (c); and a tadpole diagram connected to the incoming rotation \[using rule (\[eq:2\])\], Fig. \[fig:Feyn\](d). The tadpole diagram, Fig. \[fig:Feyn\](d), vanishes for zero external three-momentum, because $\ell=-k$ and $p_i=0$. For each non-vanishing diagram, Figs. \[fig:Feyn\](a–c), define the integral $$I_\Gamma^{\text{(a,b,c)}} = - g_0^2 C_F \int \frac{d^4k}{(2\pi)^4} \frac{1}{\hat{k}^2}{\cal I}_\Gamma^{\text{(a,b,c)}},$$ where $k$ is the momentum of the gluon in the loop, and $\hat{k}_\mu=2\sin(\case{1}{2} k_\mu)$. Let the incoming massive quark have couplings $m_0$, $r_s$, $\zeta$, $c_B$, and $c_E$, and external momentum $p$. Similarly, let the outgoing massless quark have couplings $m'_0=0$, $r'_s$, $\zeta'$, $c'_B$, and $c'_E$, and external momentum $p'$. The internal quark lines carry momentum $p+k$ in and $p'+k$ out. The integrals $I$ are obtained directly from the loop diagrams. Then $$Z_J^{[1]} = \frac{1}{2} \left( {Z_{2h}^{[1]}}_{\rm cont} - {Z_{2h}^{[1]}}_{\rm lat} + {Z_{2l}^{[1]}}_{\rm cont} - {Z_{2l}^{[1]}}_{\rm lat} \right) + \sum_{\rm d} \left( {I_\Gamma^{\rm d}}_{\rm cont} - {I_\Gamma^{\rm d}}_{\rm lat} \right),$$ from Eq. (\[eq:ZJGamma\]). The relation between the current $J$ and its Dirac matrix $\Gamma$ is contained in Table \[tab:s\]. $J$ $\Gamma$ $s_\Gamma$ -- ------- -------------------- ---------------- -- $V_4$ $\gamma_4$ $-1$ $A_4$ $\gamma_4\gamma_5$ $+1$ $V_j$ $\gamma_j$ $-\case{1}{3}$ $A_j$ $\gamma_j\gamma_5$ $+\case{1}{3}$ : The factor $s_\Gamma$, defined by $\case{1}{3}\sum_r \gamma_r\Gamma\gamma_r=s_\Gamma \Gamma$.[]{data-label="tab:s"} The expression relating ${Z_2^{[1]}}_{\rm lat}$ to lattice self-energy functions is in Ref. [@Mertens:1998wx]. The most onerous task in evaluating the diagrams is the manipulation of the Dirac matrices. A convenient method is to treat each quark line separately, starting from the initial- or final-state spinor. Then the spinor, the propagator, and the vertices can be written out in $2\times 2$ block diagonal form, with Pauli matrices appearing in the blocks. Once the Feynman rules are as complicated as in the present calculation, it is easier to manipulate $2\times 2$ matrices of Pauli matrices than to manipulate Dirac matrices. A special advantage of this organization is that the rotation bracket in Eq. (\[eq:0\]) merely “rotates” the rest of the leg. We also obtain ${Z_2^{[1]}}_{\rm lat}$ in this way, with much less effort than in Ref. [@Mertens:1998wx]. A further advantage is that the vertex corrections can be expressed compactly. The diagram with a gluon going from the incoming leg to the rotation, Fig. \[fig:Feyn\](b), is $${\cal I}_\Gamma^{\text{(b)}} = d_1 \frac{\zeta}{D} \left[(3-\case{1}{4} \hat{\bbox{k}}^2) L + \case{1}{2} \zeta \sum_r K_r S_r^2 \right],$$ where $S_r=\sin k_r$, and the functions $D$, $L$, and $K_r$ are given in Appendix \[app:useful\]. The diagram with a gluon going from the outgoing leg to the rotation, Fig. \[fig:Feyn\](c), is $${\cal I}_\Gamma^{\text{(c)}} = s_\Gamma d_1 \frac{\zeta'}{D'} \left[(3-\case{1}{4} \hat{\bbox{k}}^2) \maybebar{L}' + \case{1}{2} \zeta' \sum_r K'_r S_r^2 \right],$$ where the functions $D'$, $\maybebar{L}'$, and $K'_r$ are given in Appendix \[app:useful\], and $s_\Gamma$ is given in Table \[tab:s\]. The unbarred function $L$ (barred function $\bar{L}$) is for $\Gamma=\gamma_4$ and $\gamma_j\gamma_5$ ($\Gamma=\gamma_j$ and $\gamma_4\gamma_5$). The vertex diagram, Fig. \[fig:Feyn\](a), is complicated. We find ${\cal I}_\Gamma^{\text{(a)}}=N_\Gamma^{\text{(a)}}/DD'$, with numerator $$N_\Gamma^{\text{(a)}} = (\pm) \left( \maybebar{U}'_0 \R{U_0} - s_\Gamma \maybebar{L}'_0 \R{L_0} \bbox{S}^2 \right) - \zeta\zeta' X_\Gamma.$$ The upper sign and unbarred functions (lower sign and barred functions) are for $\Gamma=\gamma_4$ and $\gamma_j\gamma_5$ ($\Gamma=\gamma_j$ and $\gamma_4\gamma_5$). The part $X_\Gamma$ comes from spatial gluon exchange: $$\begin{aligned} X_\Gamma & = & - s_\Gamma (3-\case{1}{4} \hat{\bbox{k}}^2) \maybebar{L}' \R{L} + s_\Gamma^2 (3-\case{1}{4} \hat{\bbox{k}}^2) \maybebar{V}' \R{V} \bbox{S}^2 \nonumber \\ & & + \case{1}{2} \left( \maybebar{V}'\R{U} - s_\Gamma \maybebar{L}' \R{\zeta}\right) \sum_r K_rS_r^2 \nonumber \\ & & + \case{1}{2} \left( \maybebar{U}'\R{V} - s_\Gamma \zeta' \R{L}\right) \sum_r K'_rS_r^2 \\ & & + \case{1}{4}\left( \maybebar{U}'\R{U} - s_\Gamma \bbox{S}^2 \zeta' \R{\zeta}\right) \sum_r K'_rK_r\hat{k}_r^2 \nonumber \\ & & + \case{1}{8} (1-s_\Gamma^2) \left(\hat{\bbox{k}}^2\bbox{S}^2-3\sum_r\hat{k}_r^2S_r^2\right) \maybebar{V}' \R{V}, \nonumber\end{aligned}$$ where the last term is absent for $V_4$ and $A_4$ (*i.e.*, when $s_\Gamma^2=1$). The rotation enters in the “rotated” functions $$\begin{aligned} \R{U_0} & = & U_0 + d_1 \bbox{S}^2 L_0, \\ \R{L_0} & = & L_0 - d_1 U_0, \\ \R{U} & = & U + d_1 \bbox{S}^2 \zeta, \\ \R{\zeta} & = & \zeta - d_1 U, \\ \R{V} & = & V + d_1 L, \\ \R{L} & = & L - d_1 \bbox{S}^2 V.\end{aligned}$$ Although the vertex diagram is not easy to write down, the rotation modifies it in a fairly simple way, when using the $2\times 2$ Pauli matrix method described above. We have verified that these expressions are correct by completely independent calculation with more common methods for the Dirac algebra. Useful Functions {#app:useful} ================ In this appendix we list the functions appearing in Appendix \[app:dirac\] for the action and currents given in Sec. \[sec:lattice\]. First, let $$\begin{aligned} \mu & = & 1 + m_0 + \case{1}{2} r_s \zeta \hat{\bbox{k}}^2 ,\\ \mu' & = & 1 + m'_0 + \case{1}{2} r'_s \zeta' \hat{\bbox{k}}^2 .\end{aligned}$$ from now on a prime means to replace incoming couplings and momenta with corresponding outgoing couplings and momenta. When the quark propagator is rationalized it has the denominator $$D = 1 - 2\mu\cos(k_4+im^{[0]}_1) + \mu^2 + \zeta^2 \bbox{S}^2,$$ where $m_1^{[0]}=\ln(1+m_0)$. In this calculation, the heavy quark has zero three-momentum, so its spinor consists only of upper components. Depending on the matrix $\Gamma$ the heavy quark couples either to the upper or lower components of the light quark. With the upper components the unbarred functions arise, and with the lower components (of the light quark) the barred functions arise. To express the useful functions compactly, it is convenient to introduce first $$\begin{aligned} U & = & \mu - e^{-m_1^{[0]}+ik_4} , \\ \bar{U} & = & \mu - e^{+m_1^{[0]}-ik_4} ,\end{aligned}$$ because these combinations appear in the other functions. Then $$\begin{aligned} U_0 & = & U e^{+m_1^{[0]} - ik_4/2} - \case{1}{2} \zeta^2 c_E \cos(\case{1}{2} k_4) \bbox{S}^2 , \\ L_0 & = & \zeta\left[ e^{+m_1^{[0]} - ik_4/2} + \case{1}{2} c_E \cos(\case{1}{2} k_4) \bar{U} \right] , \\ V & = & \zeta\left[1 + \case{i}{2} c_E\sin(k_4)\right] + \case{1}{2} c_B U , \\ L & = & - \bar{U} \left[1 + \case{i}{2} c_E\sin(k_4)\right] + \case{1}{2} c_B \zeta \bbox{S}^2 , \\ K_r & = & r_s - c_B \cos^2(\case{1}{2} k_r) = (r_s - c_B) + \case{1}{4} c_B \hat{k}_r^2 ,\end{aligned}$$ and $$\begin{aligned} \bar{U}_0 & = & \bar{U} e^{-m_1^{[0]} + ik_4/2} - \case{1}{2} \zeta^2 c_E \cos(\case{1}{2} k_4) \bbox{S}^2 , \\ \bar{L}_0 & = & \zeta\left[ e^{-m_1^{[0]} + ik_4/2} + \case{1}{2} c_E \cos(\case{1}{2} k_4) U \right] , \\ \bar{V} & = & \zeta\left[1 - \case{i}{2} c_E\sin(k_4)\right] + \case{1}{2} c_B \bar{U} , \\ \bar{L} & = & - U \left[1 - \case{i}{2} c_E\sin(k_4)\right] + \case{1}{2} c_B \zeta \bbox{S}^2 .\end{aligned}$$ The barred functions are obtained from unbarred counterparts by putting $k\to-k$ and $m^{[0]}_1\to-m^{[0]}_1$, so there is no need to introduce $\bar{K}_r=K_r$. In the present calculation the barred functions arise only for the outgoing massless quark, for which $m'_0={m'_1}^{[0]}=0$. [99]{} Present address: Yukawa Institute, Kyoto University. A. S. Kronfeld and P. B. Mackenzie, Annu. Rev. Nucl. Part. Sci. [**43**]{}, 793 (1993) \[hep-ph/9303305\]. For recent status, see S. Hashimoto, Nucl. Phys. Proc. Suppl. [**83**]{}, 3 (2000) \[hep-lat/9909136\]; S. Aoki, in [*Proceedings of the XIX International Symposium on Lepton and Photon Interactions at High Energy*]{}, edited by J. A. Jaros and M. E. Peskin, eConfC [**990809**]{}, 657 (2000) \[hep-ph/9912288\]; S. Ryan, hep-lat/0111010. N. Isgur and M. B. Wise, Phys. Lett. B [**232**]{}, 113 (1989); [**237**]{}, 527 (1990). W. E. Caswell and G. P. Lepage, Phys. Lett. B [**167**]{}, 437 (1986); G. T. Bodwin, E. Braaten, and G. P. Lepage, Phys. Rev. D [**46**]{}, 1914 (1992) \[hep-lat/9205006\]. A. S. Kronfeld, Phys. Rev.  D [**62**]{}, 014505 (2000) \[hep-lat/0002008\]. A. X. El-Khadra, A. S. Kronfeld, and P. B. Mackenzie, Phys. Rev. D [**55**]{}, 3933 (1997) \[hep-lat/9604004\]. J. Harada, S. Hashimoto, A. S. Kronfeld, and T. Onogi, hep-lat/0112045; A. S. Kronfeld and S. Hashimoto, Nucl. Phys. Proc. Suppl.  [**73**]{}, 387 (1999) \[hep-lat/9810042\]. K. Symanzik, in *Recent Developments in Gauge Theories*, edited by G. ’t Hooft *et al*. (Plenum, New York, 1980). K. Symanzik, in *Mathematical Problems in Theoretical Physics*, edited by R. Schrader *et al*. (Springer, New York, 1982); Nucl. Phys. B [**226**]{}, 187, 205 (1983). M. Lüscher, S. Sint, R. Sommer, and P. Weisz, Nucl. Phys. B [**478**]{}, 365 (1996) \[hep-lat/9605038\]. K. Jansen [*et al.*]{}, Phys. Lett. B [**372**]{}, 275 (1996) \[hep-lat/9512009\]; M. Lüscher, S. Sint, R. Sommer, P. Weisz, and U. Wolff, Nucl. Phys. B [**491**]{}, 323 (1997) \[hep-lat/9609035\]. Y. Kuramashi, Phys. Rev. D [**58**]{}, 034507 (1998) \[hep-lat/9705036\]. K.-I. Ishikawa *et al.*, Nucl. Phys. Proc. Suppl. [**63**]{}, 344 (1998) \[hep-lat/9711005\]; Ishikawa, T. Onogi, and N. Yamada, Nucl. Phys. Proc. Suppl. [**83**]{}, 301 (2000) \[hep-lat/9909159\]; S. Hashimoto *et al.*, Phys. Rev.  D [**61**]{}, 014502 (2000) \[hep-ph/9906376\]. J. N. Simone *et al.*, Nucl. Phys. Proc. Suppl.  [**83**]{}, 334 (2000) \[hep-lat/9910026\]; S. Hashimoto *et al.*, hep-ph/0110253. A. X. El-Khadra, A. S. Kronfeld, P. B. Mackenzie, S. M. Ryan, and J. N. Simone, Phys. Rev. D [**64**]{}, 014502 (2001) \[hep-ph/0101023\]. B. Sheikholeslami and R. Wohlert, Nucl. Phys.  B [**259**]{}, 572 (1985). K. G. Wilson, in [*New Phenomena in Subnuclear Physics*]{}, edited by A. Zichichi (Plenum, New York, 1977). A. S. Kronfeld, Nucl. Phys. Proc. Suppl.  [**42**]{}, 415 (1995) \[hep-lat/9501002\]. S. Aoki [*et al.*]{} \[JLQCD Collaboration\], Phys. Rev. Lett.  [**80**]{}, 5711 (1998). A. X. El-Khadra *et al.*, Phys. Rev.  D [**58**]{}, 014506 (1998) \[hep-ph/9711426\]. C. Bernard [*et al.*]{}, Phys. Rev. Lett.  [**81**]{}, 4812 (1998) \[hep-ph/9806412\]. A. Ali Khan [*et al.*]{} \[CP-PACS Collaboration\], Nucl. Phys. Proc. Suppl.  [**83**]{}, 331 (2000) \[hep-lat/9909052\], Phys. Rev. D [**64**]{}, 034505 (2001) \[hep-lat/0010009\]. C. Bernard [*et al.*]{} \[MILC Collaboration\], Nucl. Phys. Proc. Suppl.  [**83**]{}, 289 (2000) \[hep-lat/9909121\]; Nucl. Phys. Proc. Suppl. [**94**]{}, 346 (2001) \[hep-lat/0011029\]. J. Harada, A. S. Kronfeld, H. Matsufuru, N. Nakajima, and T. Onogi, Phys. Rev. D [**64**]{}, 074501 (2001) \[hep-lat/0103026\]. G. P. Lepage and B. A. Thacker, Nucl. Phys. Proc. Suppl. [**4**]{}, 199 (1987); B. A. Thacker and G. P. Lepage, Phys. Rev. D [**43**]{}, 196 (1991); G. P. Lepage, L. Magnea, C. Nakhleh, U. Magnea, and K. Hornbostel, *ibid.* [**46**]{}, 4052 (1992) \[hep-lat/9205007\]. E. Eichten, Nucl. Phys. Proc. Suppl. [**4**]{}, 170 (1987). E. Eichten and B. Hill, Phys. Lett. B [**234**]{}, 511 (1990); [**240**]{}, 193 (1990). C. T. H. Davies and B. A. Thacker, Phys. Rev. D [**48**]{}, 1329 (1993). C. J. Morningstar and J. Shigemitsu, Phys. Rev. D [**57**]{}, 6741 (1998) \[hep-lat/9712016\]; [**59**]{}, 094504 (1999) \[hep-lat/9810047\]. For the program, see http://theory.fnal.gov/people/kronfeld/LatHQ2QCD/ or the EPAPS Document accompanying this paper and Ref. [@Kronfeld:1999tk]. With the latter method, go to the EPAPS home page http://www.aip.org/pubservs/epaps.html , or ftp.aip.org in the directory /epaps/. See the EPAPS home page for more information. (Of course, EPAPS is available only after publication.) B. P. G. Mertens, A. S. Kronfeld, and A. X. El-Khadra, Phys. Rev. D [**58**]{}, 034505 (1998) \[hep-lat/9712024\]. A. F. Falk, M. Neubert and M. Luke, Nucl. Phys. B [**388**]{}, 363 (1992) \[hep-ph/9204229\]. R. Wohlert, DESY report DESY 87/069 (unpublished). B. Grinstein, Nucl. Phys. B [**339**]{}, 253 (1990). H. Georgi, Phys. Lett. B [**240**]{}, 447 (1990). E. Eichten and B. Hill, Phys. Lett. B [**243**]{}, 427 (1990). M. Luke and A. V. Manohar, Phys. Lett. B [**286**]{}, 348 (1992) \[hep-ph/9205228\]; M. Neubert, Phys. Lett. B [**306**]{}, 357 (1993) \[hep-ph/9302269\]; R. Sundrum, Phys. Rev. D [**57**]{}, 331 (1998) \[hep-ph/9704256\]. T. L. Trueman, Z. Phys. C [**69**]{}, 525 (1996) \[hep-ph/9504315\]. A. S. Kronfeld, Phys. Rev. D [**58**]{}, 051501 (1998) \[hep-ph/9805215\]. E. Gabrielli *et al.*, Nucl. Phys. B [**362**]{}, 475 (1991). M. Lüscher, S. Sint, R. Sommer, and H. Wittig, Nucl. Phys. B [**491**]{}, 344 (1997) \[hep-lat/9611015\]. S. Capitani *et al.*, Nucl. Phys. B [**593**]{}, 183 (2001) \[hep-lat/0007004\]. G. Martinelli and Y. Zhang, Phys. Lett. B [**123**]{}, 433 (1983). C. Bernard, M. Golterman, and C. McNeile, Phys. Rev. D [**59**]{}, 074506 (1999) \[hep-lat/9808032\]. A. Borrelli and C. Pittori, Nucl. Phys. B [**385**]{}, 502 (1992). S. Sint and P. Weisz, Nucl. Phys. B [**502**]{}, 251 (1997) \[hep-lat/9704001\]. Y. Taniguchi and A. Ukawa, Phys. Rev. D [**58**]{}, 114503 (1998) \[hep-lat/9806015\]. S. J. Brodsky, G. P. Lepage, and P. B. Mackenzie, Phys. Rev.  D [**28**]{}, 228 (1983). G. P. Lepage and P. B. Mackenzie, Phys. Rev.  D [**48**]{}, 2250 (1993) \[hep-lat/9209022\]. K. Hornbostel, G. P. Lepage and C. Morningstar, Nucl. Phys. Proc. Suppl.  [**94**]{}, 579 (2001) \[hep-lat/0011049\]. J. Harada *et al.*, hep-lat/0110085. J. Harada, S. Hashimoto, A. S. Kronfeld, and T. Onogi, in preparation. M. Nobes, H. Trottier, G. P. Lepage, and Q. Mason, hep-lat/0110051. A. S. Kronfeld and D. M. Photiadis, Phys. Rev. D [**31**]{}, 2939 (1985). K. I. Ishikawa, T. Onogi and N. Yamada, Phys. Rev. D [**60**]{}, 034501 (1999) \[hep-lat/9812007\]. [^1]: Also, the lowest chosen values of $m_Q$ are around 1 GeV, which may be too small to be considered “heavy”. [^2]: In NRQCD applications the relative velocity between the heavy quark and heavy anti-quark should not be confused with the velocity $v$ introduced here.

ArXiv

--- abstract: 'If it is commonly agreed that the presence of a (moderately) close stellar companion affects the formation and the dynamical evolution of giant planets, the frequency of giant planets residing in binary systems separated by less than 100 AU is unknown. To address this issue, we have conducted with VLT/NACO a systematic adaptive optics search for moderately close stellar companions to 130 nearby solar-type stars. According to the data from Doppler surveys, half of our targets host at least one planetary companion, while the other half show no evidence for short-period giant planets. We present here the final results of our survey, which include a new series of second-epoch measurements to test for common proper motion. The new observations confirm the physical association of two companion candidates and prove the unbound status of many others. These results strengthen our former conclusion that circumstellar giant planets are slightly less frequent in binaries with mean semimajor axes between 35 and 100 AU than in wider systems or around single stars.' --- The NACO survey =============== To probe the impact of stellar duplicity on the frequency of giant planets, we have conducted with VLT/NACO an adaptive optics search for stellar companions to $\sim$60 planet-host stars and to $\sim$70 non-planet-host stars (hereafter control stars). This survey revealed 95 companion candidates near 33 targets (Fig. \[fig1\]). Using two-epoch astrometry we identified 19 true companions, 2 likely bound objects, and 34 background stars ([@Eggenberger07 Eggenberger et al. 2007]). The companionship of the remaining 40 companion candidates could not be constrained due to the lack of a second-epoch astrometric measurement. Assuming that all but two of these 40 objects were unbound, we showed that giant planets seem slightly less frequent in $\sim$35-100 AU binaries than around single stars ([@Eggenberger08 Eggenberger et al. 2008]). New observations ================ We recently performed the second-epoch measurements that were missing previously. The new observations show that the companion candidates we detected near the planet-host stars HD 76700, HD 83443, HD 162020, and HD 330075 are all unrelated background stars. On the other hand, the new data confirm the physical association of the companion candidates to the control stars HD 82241 and HD 134180. Statistical analysis ==================== Figure \[fig1\] (right) shows the difference in binary fraction between the control and the planet-host subsamples. According to the updated statistical analysis, the difference in binary fraction is $13.2\pm 5.1$% for semimajor axis below 100 AU, and $-1.5 \pm 2.9$ for semimajor axis between 100 and 200 AU. The positive difference seen for mean semimajor axis $\lesssim$100 AU suggests that giant and intermediate-mass planets are slightly less frequent in moderately close binaries than in wider systems or around single stars. If confirmed with a larger sample, this result would support the idea that the presence of a moderately close stellar companion affects the formation of giant planets, but does not completely stop the process. AE acknowledges support from the Swiss National Science Foundation through a fellowship for advanced researchers. 2008, *ASP-CS*, 398, 179 2007, *A&A*, 474, 273

ArXiv

--- abstract: 'The Boltzmann-Gibbs density, a central result of equilibrium statistical mechanics, relates the energy of a system in contact with a thermal bath to its equilibrium statistics. This relation is lost for non-thermal systems such as cold atoms in optical lattices, where the heat bath is replaced by the laser beams of the lattice. We investigate in detail the stationary phase-space probability for Sisyphus cooling under harmonic confinement. In particular, we elucidate whether the total energy of the system still describes its stationary state statistics. We find that this is true for the center part of the phase-space density for deep lattices, where the Boltzmann-Gibbs density provides an approximate description. The relation between energy and statistics also persists for strong confinement and in the limit of high energies, where the system becomes underdamped. However, the phase-space density now exhibits heavy power-law tails. In all three cases we find expressions for the leading order phase-space density and corrections which break the equivalence of probability and energy and violate energy equipartition. The non-equilibrium nature of the steady state is confounded by explicit violations of detailed balance. We complement these analytical results with numerical simulations to map out the intricate structure of the phase-space density.' author: - Andreas Dechant - Shalom Tzvi Shafier - 'David A. Kessler' - Eli Barkai bibliography: - 'bib.bib' title: 'Heavy-tailed phase-space distributions beyond Boltzmann-Gibbs and equipartition: Statistics of confined cold atoms' --- Introduction \[sec:introduction\] ================================= For a particle in contact with a thermal bath, the stationary phase-space probability density is the Boltzmann-Gibbs density [@lan80], which is given by $P_\text{BG}(x,p) \propto e^{-H(x,p)/({k_\text{B}}T)}$, where $H(x,p)$ is the Hamiltonian in terms of the position $x$ and momentum $p$ of the particle and $T$ is the temperature of the heat bath. This relates the system’s equilibrium statistics (the phase-space density) to its energy (the Hamiltonian) : In thermal equilibrium, states that have the same energy occur with the same probability. If we replace the thermal bath by a non-thermal one, the system is generally not in equilibrium and the connection between probability and energy is lost. An important example of such a non-thermal system is laser cooling [@coh90; @cas90; @bar94; @met99]. This term describes a multitude of techniques that are used to cool atoms or small particles in laboratories all over the world. The common feature of all these techniques is that the surrounding heat bath is replaced with the light field of the laser, which generally does not constitute a thermal equilibrium bath. Nevertheless, it is often convenient to assign an effective temperature, in term of the average kinetic energy, to this optical bath and thus to the atom or particle and treat its statistics as if the bath were thermal. This effective temperature can be very low, which, along with the ease of tuneability makes laser cooling highly attractive from an experimental point of view. For the particular case of a Sisyphus cooling lattice [@dal89; @coh90; @cas90], this allows temperatures of a few $\mu$K to be reached. Despite assigning a temperature, it is important to note that atoms cooled by the Sisyphus mechanism are not in a thermal state. Notably, their momentum probability density exhibits heavy power-law tails [@mar96; @lut03; @dou06] and their dynamics is governed by long-ranged temporal correlations and superdiffusion [@kat97; @sag12]. For a free (aside from the cooling lattice) particle these effects have been well understood from a semiclassical description of the atoms’ dynamics [@dal89; @coh90; @cas90; @lut13], which will be discussed in more detail below. Here, we go a step beyond the free particle by introducing an additional confining potential acting on the atoms in the lattice [@dec15]. For the free particle, the only degree of freedom contributing to the energy is its momentum, so its stationary momentum probability density can trivially be related to the total energy. Introducing confinement yields the position of the particle (with its corresponding potential energy) as an additional degree of freedom. In this situation the relation between total energy and stationary state statistics indeed becomes nontrivial. In order to answer the question, whether and under what conditions the energy of the system determines the stationary state phase-space probability density, we determine the latter from the semiclassical Fokker-Planck equation description of the system. We do so analytically in three limiting cases, and show that, while the leading order results depend only on the energy, higher order corrections violate the 1-to-1 correspondence of probability and energy. We begin with a brief review of Boltzmann-Gibbs statistics of a Brownian particle in a potential and some immediate consequences in Section \[sec:BG-statistics\]. Then, in Section \[sec:optical-lattice\] we introduce the semiclassical description of Sisyphus cooling and mention some consequences for the statistics of free cooled atoms. Adding the confining potential, we arrive at the central equation of this work in Section \[sec:basic-equations\], the Fokker-Plank equation for the phase-space probability density. In Section \[sec:small-D\], we start analyzing the latter by showing how the Boltzmann-Gibbs density serves as a starting point for an expansion describing the center part of the density for deep lattices. The resulting expression gives us corrections to the Boltzmann-Gibbs result that violate energy equipartition. The next part of the analysis in Section \[sec:large-freq\] focuses on an underdamped approximation. We show that for strong confinement, the total energy is approximately conserved. Thus to leading order, the resulting phase-space density depends only on energy. However, we find heavy power-law tails for high energies, in stark contrast to the exponential energy dependence of the Boltzmann-Gibbs density. As we show in Section \[sec:large-E\], these high-energy tails persist even if the confinement is weak, which is a consequence of the peculiar structure of the Sisyphus cooling force. We obtain an explicit expression for the large-energy behavior of the phase-space density including corrections that, once again, violate energy equipartition. In Section \[sec:detailed-balance\] we proceed to show that our system is indeed in a non-equilibrium stationary state in that it violates detailed balance, and relate these violations to the probability current. Finally, in Section \[sec:experimental\], we relate our findings to the experimental systems of Sisyphus cooling of confined atoms and discuss the relevant parameter regime and how the deviations from Boltzmann-Gibbs statistics might be observed in experiments. Throughout the analysis, we support our findings with numerical simulations, which we then use to explore the regime where the expansions break down and show that the results are fully consistent with the behavior emerging from the limiting cases. This paper complements our Letter [@dec15] with detailed derivations and a much extended discussion of the results while also providing several novel aspects. Boltzmann-Gibbs statistics \[sec:BG-statistics\] ================================================ The paradigmatic model that leads to Boltzmann-Gibbs statistics in a natural way is a Brownian particle in a confining potential. This situation can be cast into a Langevin equation [@cof04], $$\begin{aligned} \dot{p} &= - \gamma p - U'(x) + \sqrt{2 D_p}\xi \nonumber \\ \dot{x} &= \frac{p}{m} \label{langevin-brownian} .\end{aligned}$$ For simplicity, we restrict our discussion to the one-dimensional case and denote by $x$ the position of the particle and by $p$ its momentum. Equation describes a particle of mass $m$ moving in the conservative force field $F(x) = -U'(x)$, where $U(x)$ is a confining potential. In addition, the particle is immersed in a thermal environment, which is responsible for the Stokes friction $F_\text{fric}(p) = -\gamma p$ and the fluctuating force $F_\text{fluc} = \sqrt{2 D_p} \xi$. These two forces describe in an effective manner the collisions between the particle and the constituent particles of the environment. Since the environment is thermal, the damping coefficient $\gamma$ and the momentum diffusion constant $D_p$ are related via the temperature $T$ by the fluctuation-dissipation theorem, $D_p = m \gamma {k_\text{B}}T$. $\xi$ is a Gaussian white noise of unit magnitude, i. e.  $\langle \xi(t) \xi(t') \rangle = \delta(t-t')$. Here $\langle \ldots \rangle$ denotes an average over the ensemble of realizations of the process $\xi(t)$. Equivalently, the system may be described via a Kramers-Fokker-Planck equation for the phase-space probability density [@ris86], $$\begin{aligned} \partial_t &P(x,p,t) \nonumber \\ & = \bigg[ - \frac{p}{m} \partial_x + \partial_p \Big[ \gamma p + U'(x) + D_p \partial_p \Big] \bigg] P(x,p,t) . \label{KFP-brownian}\end{aligned}$$ It is then easy to show that the stationary state solution $\partial_t P(x,p,t) = 0$ is precisely given by the Boltzmann-Gibbs density [@ris86], $$\begin{aligned} P_\text{BG}(x,p) = Z^{-1} e^{-\frac{H(x,p)}{{k_\text{B}}T}}, \label{BG-density}\end{aligned}$$ where the Hamiltonian gives the total energy of the particle as a function of its position and momentum, $H(x,p) = p^2/(2 m) + U(x)$, and $Z = \int \text{d}x \ \text{d}p \ e^{-H(x,p)/({k_\text{B}}T)}$ is the normalizing partition function. The Boltzmann-Gibbs density , which is readily generalized to higher dimensions, has a number of important properties and consequences. It depends on the coordinate $x$ and momentum $p$ only through the Hamiltonian. This means that equal-energy surfaces in phase-space are also surfaces of equal probability. An immediate consequence is the equipartition theorem, which relates equilibrium expectation values of the Hamiltonian to the temperature [@tol18; @hua87], $$\begin{aligned} \langle p \ \partial_p H(x,p) \rangle_\text{eq} = \langle x \ \partial_r H(x,p) \rangle_\text{eq} = {k_\text{B}}T. \label{equipartition-general}\end{aligned}$$ For the canonical harmonic potential, $U(x) = m \omega^2 x^2/2$, this yields the more familiar relation $$\begin{aligned} \frac{{k_\text{B}}T}{2} = \langle E_k \rangle_\text{eq} = \langle E_p \rangle_\text{eq} , \label{equipartition-harmonic}\end{aligned}$$ where $E_k = p^2/(2 m)$ is the kinetic and $E_p = m \omega^2 x^2/2$ the potential energy. Thus every quadratic degree of freedom contributes ${k_\text{B}}T/2$ to the average energy. Whenever the phase-space probability density is a function of $H(x,p)$ only, equality holds between the moments. However, the right hand side of Eq.  is only given by the temperature in the Boltzmann-Gibbs case. Furthermore, the Boltzmann-Gibbs density is exponential in the Hamiltonian, and thus additive contributions to the system’s energy factorize in terms of the phase-space density. In particular, the kinetic and the potential energy term separate, $$\begin{aligned} P_\text{BG}(x,p) &= P_x(x) P_p(p) \\ \text{with} \; P_x(x) &= Z_x^{-1} e^{-\frac{U(x)}{{k_\text{B}}T}}, \quad P_p(p) = Z_p^{-1} e^{-\frac{p^2}{2 m {k_\text{B}}T}} \nonumber .\end{aligned}$$ This means that, in equilibrium, the kinetic and potential degrees of freedom are independent of each other and can be described separately. In particular, the equilibrium average kinetic energy is not affected by introducing or changing the potential. Semiclassical description of Sisyphus cooling \[sec:optical-lattice\] ===================================================================== In the above discussion, we took into account the effect of the bath through a dissipative friction term and a fluctuating random force which causes diffusion. As it turns out, a similar description can be employed to describe laser cooling of atoms [@met99]. Here the dissipation and diffusion describe changes in the atomic momentum due to the interactions with the photons of the light field by scattering, absorption and emission. For Sisyphus cooling [@dal89; @coh90; @cas90], the atom interacts with a standing light wave. As long as the atom is not localized in the resulting lattice potential, its trajectory can be described as a classical particle with momentum $p$ and position $x$. If further the momentum of the atom is large compared to the recoil momentum by interacting with a single photon, we can treat the interactions with the light field as a continuous process rather than individual events. Under these conditions, it can be shown that the motion of the atoms can be described by a semiclassical Langevin equation [@cas90], $$\begin{aligned} \dot{p} &= - \frac{\gamma p}{1 + \frac{p^2}{p_c^2}} - U'(x) + \sqrt{2 D_p} \odot \xi \nonumber \\ \dot{x} & = \frac{p}{m} \label{langevin-lattice} .\end{aligned}$$ The interactions between the atom and the cooling lattice are taken into account via an effective friction and a fluctuating force. We stress that $U(x)$ here is an additional spatial potential, which is distinct from the cooling lattice. The Langevin equation is quite similar to the one for a Brownian particle , the apparent difference being the nonlinearity of the friction force $$\begin{aligned} F_\text{fric} = -\gamma p /(1+p^2/p_c^2) , \label{friction}\end{aligned}$$ The latter is Stokes-like only for small momenta $|p| \ll p_c$ but decreases with momentum as $F_\text{fric} \propto - 1/p$ for large momenta $|p| \gg p_c$. Beyond that, the diffusion coefficient now depends on the momentum, $D_p = D_0 + D_1/(1+p^2/p_c^2)$. This implies that we need to specify the interpretation of as a stochastic integral [@ris86; @gar96], here, anti-It[ō]{} or endpoint interpretation, denoted by $\odot$ is the appropriate one, leading to the correct Kramers-Fokker-Planck equation . For a detailed explanation of why the nonlinear friction and the momentum-dependent diffusion coefficient appear, the reader is referred to [@coh90; @cas90]. We note that the four parameters $p_c$, $\gamma$, $D_0$ and $D_1$ can be expressed in terms of the experimental control parameters like the intensity and detuning of the cooling lasers, see Section \[sec:experimental\]. Very roughly, the cooling lattice consists of two superimposed standing waves with orthogonal polarization. The splitting of the atomic Zeemann-sublevels and the transition rates between them thus depend on the atoms’ position in the lattice. Conversely, the potential seen by the atom depends on which state it is in. If the frequency of the lattice beams is chosen in the right way, the atom will climb a potential barrier, transition to a different state for which the same position is a potential minimum and then has to climb yet another barrier. Repeating this process, the atom dissipates kinetic energy via the emission of photons. It is intuitively clear that how well this process works depends crucially on the speed of the atoms: atoms at rest remain in the lowest energy state corresponding to their position, while very fast atoms cannot distinguish between going uphill and downhill in the optical lattice. This explains why the friction force is weak if the atom is either very slow or very fast. The momentum-dependent part of the diffusion coefficient $D_1/(1+p^2/p_c^2)$ reflects the fact that the transitions between different atomic states are probabilistic and thus the atom will occasionally transition at the wrong time. This part of the diffusion coefficient satisfies a fluctuation-dissipation relation with the friction force. The first, momentum-independent part of the diffusion coefficient $D_0$, however, is due to spontaneous emission events and has no counteracting friction term, thereby driving the system out of equilibrium. We remark that deriving the semiclassical equations of motion relies on averaging the dynamics over one wavelength of the cooling lattice [@cas90] and thus the former do not depend on the spatial modulation of the laser field. It can be shown that this approximation does not change the qualitative dynamics [@hol15], however, as discussed in Section \[sec:experimental\], it requires that the scale of the confining potential should be much larger than the lattice wavelength. In the absence of the confining potential $U(x)=0$, the resulting dynamics have been discussed in detail in Refs. [@mar96; @lut03; @kes10; @kes12; @dec12]. The fact that the friction force is weak for fast atoms means that those atoms dissipate little energy and thus can stay fast for a very long time. This induces a broad stationary state momentum distribution [@lut03], $$\begin{aligned} P_p(p) &= Z^{-1} \bigg(1 + \frac{D_0}{D_0+D_1} \frac{p^2}{p_c^2} \bigg)^{-\frac{1}{2 D}} \label{momentum-PDF-free},\end{aligned}$$ long-ranged temporal correlations [@mar96] in the atomic momentum, which lead to superdiffusion for shallow lattices [@kes12; @dec12]. Here we stress that the temporal correlations [@kat97], the asymptotic power-law tails of the momentum density [@dou06] and the superdiffusive motion [@sag12] have all been observed in experiments. In general, the dynamics are controlled by the dimensionless parameter $D \equiv D_0/(\gamma p_c^2)$. This parameter is related to the depth $U_0$ of the cooling lattice by $D = c E_r/U_0$ [@kes12], where $E_r$ is the photon recoil energy and $c \sim 20$ is a constant whose precise value depends on the details of the experiment. Briefly summarizing the qualitative results for the free case, for $D < 1/5$ the momentum correlations decay sufficiently fast in time that the diffusion is normal. For $D > 1/5$, superdiffusion sets in and for $D>1$ there is no longer a stationary momentum probability density [@kes10]. Since without the confining potential, the only degree of freedom entering the Hamiltonian is the momentum, $H(p) = p^2/(2 m)$, we can trivially write the stationary density Eq.  as a function of the Hamiltonian, $$\begin{aligned} P_p(p) = Z^{-1} \bigg(1 + \frac{H(p)}{E_c} \bigg)^{-\frac{1}{2D}},\end{aligned}$$ with $E_c = p_c^2(1+D_1/D_0)/(2 m)$ and $H(p) = p^2/(2 m)$. The above form of the probability density is formally equivalent to a Tsallis distribution [@tsa88; @tsa95; @lut03]. As shown below, this equivalence no longer holds for the confined system [@dec15]. Equivalently, we find the probability distribution of energy, $$\begin{aligned} P_E(E) = \tilde{Z}^{-1} \sqrt{\frac{E_c}{E}} \bigg(1 + \frac{E}{E_c} \bigg)^{-\frac{1}{2D}}. \label{energy-PDF-free}\end{aligned}$$ We note that the parameter $D$ depends only on $D_0$. By contrast, $D_1$ appears in the asymptotic power-law behavior of the momentum density , $P_p(p) \propto p^{-1/D}$ only as a prefactor [@mar96]. Given this, for the sake of simplicity, we will consider the case $D_1 = 0$ going forward, which does not change the qualitative results. We will briefly discuss the effect of nonzero $D_1$ in Section \[sec:experimental\]. Formally, Eq.  with $p$ interpreted as the position is related to the diffusion of a particle in a logarithmic potential [@hir11; @dec11; @hir12], which can be used to model a range of physical systems [@man69; @bra00; @fog07]. In these cases, however, it is not clear what the physical equivalent of the potential $U(x)$ is. Taking into account the confining potential with the Boltzmann-Gibbs result in mind immediately raises the question whether the phase-space density may be written as a function of the Hamiltonian. In this case, the average kinetic energy would be independent of the confinement, energy equipartition would hold and some effective temperature could be assigned to the system. In the following, we will address this question and investigate the similarities between the thermal Boltzmann-Gibbs case and Sisyphus cooling. Confinement and Sisyphus cooling \[sec:basic-equations\] ======================================================== As long as the semiclassical picture is valid, the motion of the atoms can be described by Eq. , where we consider a pure harmonic potential, $U(x) = m \omega^2 x^2/2$. In the following, we will focus only on the steady state, $\partial_t P(x,p,t) = 0$. The stationary phase-space density $P(x,p)$ is then given in terms of a Kramers-Fokker-Planck equation [@ris86], $$\begin{aligned} \Bigg[ \Omega \bigg( - p \partial_{x} + x \partial_{p} \bigg) + \partial_{p} \bigg(\frac{{p}}{1+{p}^2} + D \partial_{p} \bigg) \Bigg] P(x,p) = 0 . \label{KFP-lattice}\end{aligned}$$ In order to simplify the notation, we have changed to dimensionless position and momentum variables $x = m \omega \tilde{x}/p_c$, $p = \tilde{p}/p_c$. Since going forward, we will mostly use the dimensionless variables, we from now on refer to the dimensionful variables as $\tilde{x}$ and $\tilde{p}$. We have further re-introduced the dimensionless parameter $D \equiv D_0/(\gamma p_c^2)$ and defined $\Omega \equiv \omega/\gamma$. In terms of this dimensionless description, it is clear that the properties of the system are governed by the two parameters $D$ and $\Omega$. Equation constitutes the main object of investigation of this work. The first term on the left-hand side of describes the Hamiltonian part of the evolution, the oscillation of a particle in an harmonic well. The second term contains the friction and the noise, which are effects of the “bath”. At this point, we also introduce two equivalent ways of writing Eq. , which will turn out to be convenient for the following discussion. The first alternate representations follows from a simple rescaling of position and momentum, $z = x/\sqrt{D}$ and $u = p/\sqrt{D}$, $$\begin{aligned} \Bigg[ \Omega \bigg( - u \partial_z + z \partial_u \bigg) + \partial_u \bigg(\frac{u}{1+D u^2} + \partial_u \bigg) \Bigg] P_D(z,u) = 0, \label{KFP-lattice-D}\end{aligned}$$ with $P(x,p) = P_D[x/\sqrt{D}, p/\sqrt{D}]/D$. Clearly reduces to the equation for the Boltzmann-Gibbs case in the limit $D \rightarrow 0$. We will use this in the next section to perform a systematic expansion around this limit. Secondly, we change to a polar representation of the phase space, by introducing the energy $\varepsilon = (x^2+p^2)/2$ and the phase-space angle $\alpha = \arctan(p/x)$, $0 \leq \alpha < 2 \pi$. In terms of these coordinates, Eq.  reads $$\begin{aligned} \Bigg[ &\underbrace{\Omega \partial_\alpha}_{\begin{array}{c} \text{ \scriptsize Hamiltonian} \\[-1 ex]\text{ \scriptsize evolution} \end{array}} + \underbrace{\mathcal{L}_{\varepsilon,\alpha}}_{\begin{array}{c}\text{ \scriptsize friction} \\[-1 ex] \text{\scriptsize and noise} \end{array}} \Bigg] P_P(\varepsilon,\alpha) = 0 \quad \text{with} \label{KFP-lattice-E} \\ \mathcal{L}_{\varepsilon,\alpha} &= \partial_\alpha \frac{\sin(\alpha)\cos(\alpha)}{1+2 \varepsilon \sin^2(\alpha)} + \partial_\varepsilon \frac{2 \varepsilon \sin^2(\alpha)}{1+2 \varepsilon \sin^2(\alpha)} \nonumber \\ & + D \bigg[ 2 \partial_\alpha \sin(\alpha) \cos(\alpha) \partial_\varepsilon + \frac{1}{2 \varepsilon} \partial_\alpha \cos^2(\alpha) \partial_\alpha \nonumber \\ & + (\sin^2(\alpha) - \cos^2(\alpha)) \partial_\varepsilon + 2 \sin^2(\alpha) \partial_\varepsilon \varepsilon \partial_\varepsilon \bigg] \nonumber ,\end{aligned}$$ with $P(x,p) = P_P[(x^2+p^2)/2,\arctan(p/x)]$, since the Jacobian of the transformation is unity. The advantage of this representation is that the Hamiltonian part of the evolution consists of just a derivative with respect to $\alpha$, whereas the friction and noise terms are now more complicated. We will use Eq.  extensively in Sections \[sec:large-freq\] and \[sec:large-E\] to find the behavior at large frequencies and large energies, respectively. Note that the dimensionless energy $\varepsilon$ is related to the physical energy $E$ via $E = \varepsilon p_c^2/m$. In the following Sections \[sec:small-D\], \[sec:large-freq\] and \[sec:large-E\], we will discuss the mathematical properties of the solution of Eqs. , and and the physical implications. This solution can be obtained analytically in terms of expansions in certain limits. The general features elucidated in these expansions, however, persist even beyond the validity of the expansions themselves, as we will establish via numerical simulations. In Section \[sec:experimental\], we will come back to the actual experimental system of Sisyphus cooling with added confinement and discuss the relevant parameter regime, as well as other physical effects that need to be taken into account. Deep lattices: Deviations from Boltzmann-Gibbs \[sec:small-D\] ============================================================== The dimensionless parameter $D$ is inversely proportional to the depth of the cooling lattice (see Section \[sec:optical-lattice\]). For deep lattices, $D$ is thus small. As noted before, Eq.  reduces to a Boltzmann-Gibbs-like equation in the limit of $D \rightarrow 0$. In particular, the normalized solution for $D = 0$ is $$\begin{aligned} P_D^{(0)}(z,u) = \frac{1}{2 \pi} e^{-\frac{z^2+u^2}{2}} , \label{small-d-0order}\end{aligned}$$ which is precisely the Boltzmann-Gibbs density . Starting from this, we define an auxiliary function $g(z,u)$ via $$\begin{aligned} P_D(z,u) = P_D^{(0)}(z,u) g(z,u) .\end{aligned}$$ Plugging this into Eq. , we obtain an equation for $g(z,u)$, $$\begin{aligned} \Big[\mathcal{L}_0 &+ D \mathcal{L}_1 + D^2 \mathcal{L}_2\Big] g(z,u) = 0 \label{small-d-auxilary} \\ \mathcal{L}_0 &= \Omega(z \partial_{u} - u \partial_{z}) - u \partial_{u} + \partial_{u}^2 \nonumber \\ \mathcal{L}_1 &= 2 \Omega (z u^2 \partial_{u} - u^3 \partial_{z}) - 3 u^2 + u^4 - 3 u^3 \partial_{u} + 2 u^2 \partial_{u}^2 \nonumber \\ \mathcal{L}_2 &= \Omega (z u^4 \partial_{u} - u^5 \partial_{z}) - u^4 + u^6 - 2 u^5 \partial_{u} + u^4 \partial_{u}^2 \nonumber .\end{aligned}$$ We see that the equation for the function $g(z,u)$ contains three terms multiplied by different powers of $D$. For $D = 0$, the resulting equation $\mathcal{L}_0 g(z,u) = 0$, obviously has the solution $g(z,u) = 1$, which yields the Boltzmann-Gibbs result . To proceed, we assume that we can expand $g(z,u)$ for small $D$, $$\begin{aligned} g(z,u) = 1 + D g^{(1)}(z,u) + \mathcal{O}(D^2) . \label{small-d-auxilary-expansion}\end{aligned}$$ Plugging this into Eq. , and equating orders in $D$ we obtain an equation for $g^{(1)}(z,u)$, $$\begin{aligned} \mathcal{L}_0 g^{(1)}(z,u) - 3 u^2 + u^4 = 0 . \label{small-d-auxilary-1order}\end{aligned}$$ Examining the operator $\mathcal{L}_0$, we can already guess what the solution to this equation might be. If $g^{(1)}(z,u)$ is a polynomial in $z$ and $u$ of total order $N$, then $\mathcal{L}_0$ leaves the total order unchanged. The inhomogeneous terms in , on the other hand, are of order $4$, so if there exists a polynomial solution, we necessarily have $N=4$, $$\begin{aligned} g^{(1)}(z,u) = \sum_{k=0}^{4} \sum_{l=0}^{4-k} a^{(1)}_{k l} z^{k} u^{l} \label{small-d-polynomial} .\end{aligned}$$ Plugging this into Eq.  then gives us a set of linear equations for the coefficients $a^{(1)}_{k l}$. For the first order expansion, we explicitly give the coefficients in Appendix \[app:small-d\]. We can continue the expansion in terms of $D$, Eq. , to higher orders. Before we proceed to do so, however, let us remark on the validity of the expansion. We implicitly assumed that $D g^{(1)}(z,u)$ is small. Since we know that $g^{(1)}(z,u)$ is a polynomial of total order $4$ in $z$ and $u$, we need not only for $D$ to be small but also terms of the form $D z^k u^l$ with $k + l \leq 4$. This means that our expansion is valid for small $D$ in the center part of the phase-space density. Up to first order in $D$, we find for the normalized phase-space probability density $$\begin{aligned} P^{(1)}_D(z,u) = \frac{e^{-\frac{z^2+u^2}{2}}}{2 \pi} \bigg[ 1 + \frac{D}{4 (3+4\Omega^2)} \Big[ 3 u^4 + 18 z^2 - 27 + \Big( 4 u^3 z - 12 u z \Big) \Omega + \Big(3(u^2+z^2)^2 - 24 \Big) \Omega^2 \Big] \bigg] . \label{small-d-1order}\end{aligned}$$ Comparing this first order result to the Boltzmann-Gibbs one, we note that, in contrast to the latter, Eq.  does not depend on $u$ and $z$ as a function of only the Hamiltonian $H(u,z) = (z^2 + u^2)/2$. Consequently, it has less symmetry than the zero-order Boltzmann-Gibbs approximation: It is not symmetric with respect to interchanging $z$ and $u$, or reversing $z \rightarrow -z$ or $u \rightarrow -u$ individually. However, we note that all these symmetries are restored in the limit $\Omega \gg 1$, i.e.  strong confinement, where to leading order in $\Omega$, we find $$\begin{aligned} P^{(1)}_D(z,u) \simeq \frac{e^{-H(z,u)}}{2 \pi} \bigg[ 1 + \frac{D}{4} \Big[ \Big(3 H^2(z,u) - 8 \Big) \Big] \bigg]. \label{small-d-1order-omega}\end{aligned}$$ A similar behavior is found for any $D$ and large $\Omega$ in Section \[sec:large-freq\]. The equation for the second-order correction $g^{(2)}(z,u)$ is easy to derive, $$\begin{aligned} \mathcal{L}_0 g^{(2)}(z,u) + \mathcal{L}_1 g^{(1)}(z,u) - u^4 + u^6 = 0.\end{aligned}$$ Since the term $\mathcal{L}_1 g^{(1)}(z,u)$ contains contributions proportional to $p^8$, $g^{(2)}(z,u)$ is a polynomial of total degree $N=8$. Generally, we can write the expansion of $g(z,u)$ up to order $D^M$ as $$\begin{aligned} g(z,u) = 1 + \sum_{n = 1}^{M} D^{n} \sum_{k=0}^{4 n} \sum_{l=0}^{4 n - k} a^{(n)}_{k,l} z^k u^l ,\end{aligned}$$ thus $g^{(M)}(z,u)$ is a polynomial of order $N = 4 M$. For the following analysis, we perform this expansion up to third order $M=3$, where we use Mathematica to handle the cumbersome algebra. ![Phase-space probability density for $\Omega = 0.5$ and $D = 0.1$ plotted using the position and momentum $(x,p)$ (top) respectively energy and phase-space angle $(E,\alpha)$ (bottom). The colored surfaces are the results of numerical Langevin simulations, the black lines represent the third-order small-$D$ expansion. For the Boltzmann-Gibbs case $D \rightarrow 0$, these plots would display perfect cirlces (top) respectively straight lines (bottom). The slightly tilted elliptical shape (top) respectively wavy lines (bottom) are due to the deviations from Boltzmann-Gibbs for finite $D$. \[fig:jdist\]](jdist_small_d.png "fig:"){width="47.00000%"}\ ![Phase-space probability density for $\Omega = 0.5$ and $D = 0.1$ plotted using the position and momentum $(x,p)$ (top) respectively energy and phase-space angle $(E,\alpha)$ (bottom). The colored surfaces are the results of numerical Langevin simulations, the black lines represent the third-order small-$D$ expansion. For the Boltzmann-Gibbs case $D \rightarrow 0$, these plots would display perfect cirlces (top) respectively straight lines (bottom). The slightly tilted elliptical shape (top) respectively wavy lines (bottom) are due to the deviations from Boltzmann-Gibbs for finite $D$. \[fig:jdist\]](jdist_small_d_angle.png "fig:"){width="47.00000%"} The resulting phase-space density is shown in Fig. \[fig:jdist\]. Plotting the latter using the energy-angle coordinates introduced in Eq.  clearly exhibits the deviations from the Boltzmann-Gibbs density, which is independent of the angle $\alpha$. This density has a number of features which distinguish it from the Boltzmann-Gibbs density, which we discuss in the following. In Section \[sec:BG-statistics\], we saw that a phase-space density that only depends on the Hamiltonian leads to energy equipartition, which in our rescaled variables corresponds to $\langle z^2 \rangle = \langle u^2 \rangle = 1$ for $D=0$. In order to quantify the deviations induced by finite $D$, we define the equipartition ratio $$\begin{aligned} \chi = \langle z^2 \rangle/\langle u^2 \rangle = \langle x^2 \rangle/\langle p^2 \rangle = \langle \varepsilon_p \rangle/\langle \varepsilon_k \rangle, \label{equipart}\end{aligned}$$ where $\varepsilon_p = x^2/2$ and $\varepsilon_k = p^2/2$ are the dimensionless potential and kinetic energy. Obviously $\chi = 1$ corresponds to energy equipartition. From the third order expansion, we find $$\begin{aligned} \chi^{(3)} = 1 + \frac{6}{3+4\Omega^2} D^2 - \frac{6 (38 \Omega^2 - 21)}{(3+4\Omega^2)^2} D^3 . \label{equipart-order3}\end{aligned}$$ We see that there is no linear term in this expansion, meaning that the first order result does induce any violation of energy equipartition, even though it is not a function of the Hamiltonian only. Further, for large $\Omega$, the deviations from equipartition are also of order $\Omega^{-2}$, which agrees with the finding that the phase-space density becomes a function of the Hamiltonian in this limit. ![Equipartition ratio $\chi$ as a function of the trap strength $\Omega$ for different values of $D$. The solid lines are the third-order result Eq. , the circles are the result of numerical Langevin simulations. For comparison, we also show the results of a fourth-order expansion in $D$ at $\Omega = 0.1$, which are closer to the numerical results. \[fig:equipart\]](equi_small_d.png){width="47.00000%"} ![Equipartition ratio $\chi$ as a function of the parameter $D$ for $\Omega = 0.1$. The solid line is the third-order result Eq. , the circles are the result of numerical Langevin simulations. For small $D$, the agreement between numerics and theory is good, at larger $D$, the two results deviate. In particular, the numerically obtained equipartion ratio exhibits a turnover and decreases for larger $D$. \[fig:equipart2\]](equi_d.png){width="47.00000%"} ![Average potential and kinetic energy from the third-order small-$D$ expansion as a function of $\Omega$ for $D = 0.1$. Both have been divided by the average kinetic energy of the free particle. Note that the kinetic energy is reduced compared to the latter. \[fig:equipart3\]](vxmom.png){width="47.00000%"} The average potential energy is always larger than the average kinetic energy. This is reflected in the slightly elongated shape of the phase-space density along the $x$-axis, see Fig. \[fig:jdist\]. These deviations from energy equipartition are strongest for small trapping strengths $\Omega$ and increase with increasing $D$, see Fig. \[fig:equipart\]. Not surprisingly, the results of the small-$D$ expansion deviate from the numerical ones for larger values of $D$, where the expansion breaks down. Nevertheless, the trends with respect to $\Omega$ and $D$ are captured correctly. When plotted as a function of $D$ for $\Omega = 0.1$, we see that the deviations from equipartition are maximal for $D \approx 0.3$ where the potential energy is about $20 \%$ larger than the kinetic one see Fig. \[fig:equipart2\]. The reason for the decrease at larger $D$ is due to the fact that the average energy in the stationary state diverges for $D > 1/2$. This is discussed in more detail in Section \[sec:large-E\]. Instead of looking at the equipartition ratio, it is worthwhile to also discuss the behavior of the potential and kinetic energy individually. The results from the third-order expansion are shown in Fig. \[fig:equipart3\]. Here, we normalized the energy to the kinetic energy of a free particle at the same value of $D$. Even for large frequencies, where the phase-space density can be expressed as a function of the Hamiltonian, both the kinetic and potential energy are less than the kinetic energy of a free particle. This underlines the nontrivial interplay between the confining potential and the nonlinear friction force. Since the friction force is strongest for particles of moderate momentum, the slowing down near the turning points of the oscillatory motion in the potential increases the dissipation. This stays true in the limit of small $\Omega$, where the average kinetic energy is still reduced compared to the free case, even though the potential energy is now larger. Due to the perturbative nature of the above results, the overall effect shown here is small. However, it can be quite substantial at realistic parameter values beyond the validity of the perturbation expansion, where we numerically find an average potential energy that is several times larger than the kinetic one, see Section \[sec:experimental\]. At first glance it might seem disconcerting that even as $\Omega \rightarrow 0$, we do not recover the free particle result. However, here we are only discussing the stationary behavior of the system. For very small $\Omega$, it will take longer and longer to reach this stationary state, as the oscillation period in the potential grows as $1/\Omega$ until at $\Omega = 0$ there is no stationary state at all. The stationary state limit $t \rightarrow \infty$ and the limit $\Omega \rightarrow 0$ thus do not commute. The expansion for the phase-space density for $D \neq 0$ also differs from the Boltzmann-Gibbs case in that it is no longer symmetric with respect to $u \rightarrow -u$ or $z \rightarrow -z$; only the combination of both operations is a symmetry of the problem: $P_D(z,u) = P_D(-z,-u)$. This can be conveniently visualized by defining the antisymmetric and symmetric part of the probability density, $$\begin{aligned} P_D^{\text{sy}}(z,u) &= \frac{1}{2} \Big( P_D(z,u) + P_D(-z,u) \Big) \label{sym-asym} \\ \Rightarrow &P_D^{\text{sy}}(z,u) = P_D^{\text{sy}}(-z,u) = P_D^{\text{sy}}(z,-u) \nonumber \\ P_D^{\text{as}}(z,u) &= \frac{1}{2} \Big( P_D(z,u) - P_D(-z,u) \Big) \nonumber \\ \Rightarrow &P_D^{\text{as}}(z,u) = -P_D^{\text{as}}(-z,u) = -P_D^{\text{as}}(z,-u) . \nonumber\end{aligned}$$ ![Antisymmetric part Eq.  of the phase-space probability density for $\Omega = 0.5$ and $D = 0.1$. The colored surfaces are the results of numerical Langevin simulations, the black lines represent the third-order small-$D$ expansion. \[fig:asdist\]](asdist_small_d.png){width="47.00000%"} For the zeroth-order Boltzmann-Gibbs result, the antisymmetic part of the phase-space density vanishes. For non-zero $D$, however, this antisymmetric part highlights a feature that is hard to discern in the total density: the probability of position and momentum having opposite signs is increased with respect to equal signs; a particle to the right of the center of the potential is more likely to be moving left, and vice versa, see Fig. \[fig:asdist\]. This kind of behavior can be encoded in a single quantity by defining $$\begin{aligned} \eta &= \frac{\int_{-\infty}^{0} \text{d}z \int_{0}^{\infty}\text{d}u P_D(z,u) + \int_{0}^{\infty} \text{d}z \int_{-\infty}^{0}\text{d}u P_D(z,u)}{\int_{-\infty}^{0} \text{d}z \int_{-\infty}^{0}\text{d}u P_D(z,u) + \int_{0}^{\infty} \text{d}z \int_{0}^{\infty}\text{d}u P_D(z,u)} \nonumber \\ &= 1-2 \frac{\int_{0}^{\infty} \text{d}z \int_{0}^{\infty} \text{d}u P_D^{\text{as}}(z,u)}{\int_{0}^{\infty} \text{d}z \int_{0}^{\infty} \text{d}u \Big[P_D^\text{sy}(z,u)+P_D^{\text{as}}(z,u)\Big]}, \label{asym-ratio}\end{aligned}$$ where we split the phase-space density into symmetric and antisymmetric part. ![Antisymmetry ratio $\eta$ Eq.  as a function of $\Omega$ for different values of $D$. The solid lines are the third-order result Eq. , the circles are the result of numerical Langevin simulations and the dashed lines are the large-frequency expansion discussed in Section \[sec:large-freq\]. \[fig:asmeas\]](asypar_small_d.png){width="47.00000%"} The parameter $\eta$ is precisely ratio of the probabilities of having position and momentum in opposite and same directions. Contrary to the equipartition ratio, $\eta$ has a non-monotonous behavior as a function of frequency, see Fig. \[fig:asmeas\]. While for both small and large $\Omega$, having position and momentum in the same and opposite directions is equally likely, for intermediate values of $\Omega$, opposite directions are more likely. For large $\Omega$ this agrees with the observation that the phase-space density is a function of the Hamiltonian only in this limit and thus has rotational symmetry in phase-space. In this limit the motion of the particle can be described as oscillations in the potential with a slowly changing energy due to dissipation and diffusion, see Section \[sec:large-freq\] for a more detailed discussion. For small $\Omega$, on the other hand, the kinetic energy of the particle typically varies on much shorter time scales than the potential one. Thus position and momentum are less correlated and $\eta$, which essentially measures these correlations, is small. Strong confinement: Underdamped approximation \[sec:large-freq\] ================================================================ In the previous section, we noticed that the center part of the phase-space probability density can be written as a function of the Hamiltonian for $\Omega \gg 1$, i.e. a strong confining potential. We now want to see whether this observation carries over to the entire distribution as well. Our starting point is Eq. , the parameterization of the phase-space density in terms of the energy $E$ and the phase-space angle $\alpha$. Dividing by $\Omega$, Eq.  can be written as $$\begin{aligned} \partial_\alpha P_P(\varepsilon,\alpha) = -\frac{1}{\Omega} \mathcal{L}_{\varepsilon,\alpha} P_P(\varepsilon,\alpha) \label{KFP-lattice-E2}.\end{aligned}$$ For $\Omega \gg 1$, we then have to leading order $\partial_\alpha P_P(\varepsilon,\alpha) = 0$. Thus the phase-space density is to leading order independent of the angle $\alpha$. Physically $\Omega \gg 1$ means that the oscillation frequency in the trap is much larger than the damping rate. A particle would thus complete many oscillations before its energy changes substantially due to dissipation and its motion can be approximately described by the Hamiltonian dynamics. In this sense, the expansion in large $\Omega$ is a weak damping limit, also referred to as the underdamped approximation [@str63]. This can be made more rigorous by comparing the change in energy due to damping $\Delta E_\text{diss}$ and diffusion $\Delta E_\text{diff}$ per period of the unperturbed oscillation, $$\begin{aligned} \Delta E_\text{diss} = \oint_\mathcal{T} \text{d}\tilde{x} \ F_\text{fric}(\tilde{p}(\tilde{x})) .\end{aligned}$$ Here we use the dimensionful units and denote by $\mathcal{T}$ the closed circular Hamiltonian orbit. Employing the form for the friction force and changing to dimensionless units, we find $$\begin{aligned} \Delta \varepsilon_\text{diss} = -\Omega^{-1} \oint_\mathcal{T} \text{d}x \frac{p(x)}{1+p^2(x)} .\end{aligned}$$ Since the energy $\varepsilon = (x^2+p^2)/2$ is conserved along Hamiltonian orbits, we may use this to express $p$ as a function of $x$ and obtain $$\begin{aligned} \Delta \varepsilon_\text{diss} &= 4 \Omega^{-1} \int_{0}^{\sqrt{2 \varepsilon}} \text{d}x \frac{\sqrt{2 \varepsilon - x^2}}{1 + 2 \varepsilon - x^2} \nonumber \\ &= 2 \pi \Omega^{-1} \bigg(1 - \frac{1}{\sqrt{1+2\varepsilon}}\bigg). \label{dissipation}\end{aligned}$$ The energy change due to diffusion is given by the momentum diffusion coefficient, $$\begin{aligned} \Delta E_\text{diff} = \frac{D_0 T}{m},\end{aligned}$$ where $T = 2 \pi/\omega$ is the period of the oscillation. In dimensionless units, this reads $$\begin{aligned} \Delta \varepsilon_\text{diff} = 2 \pi D \Omega^{-1} . \label{diffusion}\end{aligned}$$ From Eqs.  and , we conclude that the energy change due to dissipation and diffusion is proportional to $\Omega^{-1}$, which justifies the approach of using the Hamiltonian evolution with a slowly changing energy. Comparing the relative change in energy per oscillation, we have $$\begin{aligned} \frac{\Delta \varepsilon_\text{diss}}{\varepsilon} &= \frac{2 \pi \big(1 - \frac{1}{\sqrt{1+2\varepsilon}}\big)}{\Omega \ \varepsilon} \nonumber \\ \frac{\Delta \varepsilon_\text{diff}}{\varepsilon} &= \frac{2 \pi D }{\Omega \ \varepsilon} \label{dissipation-diffusion}\end{aligned}$$ We notice that both quantities are small not only for large frequency $\Omega$, but also for large energy $\varepsilon$. For the diffusive contribution this would be true even for linear friction. For the dissipative term, on the other hand, for linear friction, $\Delta \varepsilon_\text{diss}/\varepsilon$ is found to be independent of $\varepsilon$ and thus small only for large frequency. Thus for the nonlinear friction force discussed above, the system becomes underdamped both for large frequency and for large energy. The former limit will be discussed in the following, whereas we will exploit the latter property in Section \[sec:large-E\]. Having assured ourselves that the underdamped approximation works for our system, we proceed to discuss what results we can obtain from it. From the leading order behavior $\partial_\alpha P_P(\varepsilon,\alpha) = 0$, we conclude $P_P(\varepsilon,\alpha) \simeq P_\varepsilon(\varepsilon)/(2 \pi) + \mathcal{O}(\Omega^{-1})$. Plugging this into Eq.  and integrating over $\alpha$, we obtain an equation for the leading order energy density $P_\varepsilon(\varepsilon)$, $$\begin{aligned} \partial_\varepsilon \bigg[ 1 - \frac{1}{\sqrt{1+2\varepsilon}} + D \varepsilon \partial_\varepsilon \bigg] P_\varepsilon(\varepsilon) = 0 . \label{energy-density-FP}\end{aligned}$$ Solving this energy diffusion equation, we find the main result of this section, $$\begin{aligned} P_\varepsilon(\varepsilon) &= Z_\varepsilon^{-1} \big(1 + \sqrt{1+2\varepsilon}\big)^{-\frac{2}{D}} \label{energy-density-stationary} \\ \text{with} \quad Z_\varepsilon &= \frac{2^{-\frac{2}{D}} D}{(1-D)(2-D)} \nonumber .\end{aligned}$$ Clearly this density is very different both from the Boltzmann-Gibbs density Eq.  and the Tsallis-like form Eq.  obtained without the confining potential. ![Energy density as a function of energy for $D=0.3$ and different values of $\Omega$. The solid red line is Eq. , the dashed red line the corresponding asymptotic $\varepsilon^{-1/D}$-behavior. The circles are numerical simulations for $\Omega = 2$, the squares are for $\Omega = 0.1$. \[fig:edist\]](edist.png){width="47.00000%"} ![Ratio of the numerically obtained energy density and the analytical large-frequency result Eq. . Circles are for $\Omega = 2$, squares for $\Omega = 0.1$. The red dashed line indicating the value of $1$ has been added for clarity. \[fig:edist-rel\]](edist_rel.png){width="47.00000%"} We compare Eq.  to numerical simulations in Figs. \[fig:edist\] and \[fig:edist-rel\]. Good agreement is already obtained for moderate values of $\Omega \approx 2$, suggesting that the corrections to the leading-order behavior are generally small. This is substantiated by Fig. \[fig:edist-rel\], where we show the ratio between the numerically obtained energy density and Eq. . For $\Omega = 2$ this ratio is close to $1$, whereas for $\Omega = 0.1$, we see that the likelihood of large energies is increased, in agreement with the observation from Section \[sec:small-D\] that the total energy increases towards small $\Omega$. As for the case without the confining potential, Eq. , this stationary density has power-law tails in the energy, $P_\varepsilon(\varepsilon) \propto \varepsilon^{-1/D}$. Importantly, the exponent of the tails is universal for large energies, even when $\Omega$ is not large, see Fig. \[fig:edist\]. Compared to the free case $P_\varepsilon(\varepsilon) \propto \varepsilon^{-1/(2D)-1/2}$, however, the exponent of the tails is different, and the energy density for the confined system decays more rapidly at large energies. Note that in both cases $D=1$ marks a transition where the stationary density is no longer normalizable and thus the solution becomes time-dependent for $D>1$. A time-dependent solution is also needed to describe the moments of the density beyond certain values of $D$. For example, the average energy is given by $$\begin{aligned} \langle \varepsilon \rangle = \frac{D(2-D)}{(2-3D)(1-2D)} . \label{average-energy}\end{aligned}$$ This expression diverges at $D = 1/2$, and the average energy increases as a function of time for $D > 1/2$. This behavior is akin to the divergence of the kinetic energy at $D = 1/3$ for the free case [@kat97; @lut03; @kes10]. While we do not investigate the time-dependent solution at this point, we note that the divergence of the average energy occurs at larger values of $D$ when we include the confining potential as compared to the free case. Since $D$ is a measure of the ratio of heating due to spontaneous emission and cooling due to the lattice, this implies that the potential actually improves the dissipation mechanism. This is in agreement with the observation from the small-$D$ expansion in Section \[sec:small-D\], where we found that in the presence of the confining potential, the kinetic energy is reduced. The divergence of the average energy at $D = 1/2$ also explains the results for the equipartition ratio depicted in Fig. \[fig:equipart2\]. For $D$ slightly below $1/2$, the average of the energy is dominated by the large-energy tails of the phase-space density. Since, as we saw before, the underdamped limit is also valid for large energies, the average energy is well-described by a phase-space density that depends only on the Hamiltonian and consequently, we have equipartition of energy in this regime. For large $\Omega$ this holds for any $D$ and the kinetic and potential energy are equal to leading order in $\Omega$, $\langle \varepsilon_k \rangle = \langle \varepsilon_p \rangle = \langle \varepsilon \rangle/2$. ![Momentum distribution for $D = 0.3$. The black line is the result without the confining potential, the red line is obtained by numerically integrating the large-$\Omega$ expansion of the phase-space density over $x$. The circles are the result of Langevin simulations for $\Omega = 4$. Clearly, the confinement leads to a narrower momentum distribution, which suggests improved cooling (see Sec. \[sec:experimental\] for a more detailed discussion). \[fig:pdist\]](pdist.png){width="47.00000%"} From Eq. , we can immediately deduce the phase-space density for large $\Omega$, $$\begin{aligned} P(x,p) &= P_\varepsilon(H(x,p))/(2 \pi) \nonumber \\ & = (2 \pi Z_\varepsilon)^{-1} \big(1 + \sqrt{1+x^2+p^2}\big)^{-\frac{2}{D}} . \label{phase-space-large-omega}\end{aligned}$$ By integrating over $p$ or $x$, respectively, we can in principle obtain the marginal densities $P_x(x)$ and $P_p(p)$. These have no closed-form representation, however it is easy to see that their asymptotic tails behave as $$\begin{aligned} P_p(p) \simeq \frac{\sqrt{\pi} \Gamma\big(\frac{2-D}{2 D}\big)}{2 \pi Z_\varepsilon \Gamma\big(\frac{1}{D}\big)} |p|^{-\frac{2}{D}+1}, \label{momentum-PDF-confined}\end{aligned}$$ and likewise for $P_x(x)$. As with the energy density, the tails $P_p(p) \propto |p|^{-2/D+1}$ are markedly different from the free case $P_p(p) \propto |p|^{-1/D}$ Eq. . The resulting momentum distribution is shown in Fig. \[fig:pdist\]. So far, we have focused on the leading order term, which depends only on energy. We expand the full phase space density with respect to $\Omega$, we obtain $$\begin{aligned} P_P(\varepsilon,\alpha) = \frac{P_\varepsilon(\varepsilon)}{2 \pi} \Bigg[1 + \frac{f_1(\varepsilon,\alpha)}{\Omega} + \mathcal{O}(\Omega^{-2}) \Bigg]. \label{large-omega-expansion}\end{aligned}$$ The function $f_1(\varepsilon,\alpha)$ represents the (generally angle-dependent) first order correction to the angle-independent leading order result. Plugging this expansion into the Kramers-Fokker-Planck-equation , we find an equation for $f_1(\varepsilon,\alpha)$ by equating the coefficients of the terms of order $\Omega^{-1}$, $$\begin{aligned} \partial_\alpha f_1(\varepsilon,\alpha) &= - \frac{\mathcal{L}_{\varepsilon,\alpha} P_\varepsilon(\varepsilon)}{P_\varepsilon(\varepsilon)} \label{f_1-equation} \\ \Rightarrow f_1(\varepsilon,\alpha) &= - \frac{1}{P_\varepsilon(\varepsilon)} \int_{0}^{\alpha} \text{d}\alpha' \mathcal{L}_{\varepsilon,\alpha'} P_\varepsilon(\varepsilon) + \tilde{f}_1(\varepsilon) \nonumber .\end{aligned}$$ Here $\tilde{f}_1(\varepsilon)$ is some function that depends only on $\varepsilon$. Performing the integral over $\alpha$, we get $$\begin{aligned} f_1(\varepsilon,\alpha) &= -\frac{\sin(\alpha)\cos(\alpha)}{1+2\varepsilon \sin^2(\alpha)} - \frac{\phi(\varepsilon,\alpha)}{(1+2\varepsilon)^{\frac{3}{2}}} - \frac{\sin(2\alpha)}{2(1+2\varepsilon)(\varepsilon(\cos(2\alpha)-1)-1)} - \Bigg[\alpha - \frac{\phi(\varepsilon,\alpha)}{\sqrt{1+2\varepsilon}} + D \cos(\alpha) \sin(\alpha) \Bigg] \frac{\partial_\varepsilon P_\varepsilon(\varepsilon)}{P_\varepsilon(\varepsilon)} \nonumber \\ &\qquad \qquad - D \Bigg[\alpha-\cos(\alpha)\sin(\alpha)\Bigg] \frac{\partial_\varepsilon \varepsilon \partial_\varepsilon P_\varepsilon(\varepsilon)}{P_\varepsilon(\varepsilon)} + \tilde{f}_1(\varepsilon). \label{anglefunc1}\end{aligned}$$ Here, we introduced the function $$\begin{aligned} \phi(\varepsilon,\alpha) = \arctan\Big(\sqrt{1+2\varepsilon}\tan(\alpha)\Big) + \pi \Big\lfloor \frac{\alpha}{\pi} + \frac{1}{2} \Big\rfloor ,\end{aligned}$$ where $\lfloor x \rfloor$ denotes the floor function, i.e. the largest integer smaller than $x$. This function removes the discontinuities in $\arctan(\tan(\alpha))$ introduced by the divergence of $\tan(\alpha)$ at $\alpha = (2k+1)\pi/2$. Equation also places a condition on $P_\varepsilon(\varepsilon)$. In order to be consistent, we require $f_1(\varepsilon,\alpha) = f_1(\varepsilon,\alpha+2\pi)$. Using the above definition of the function $\phi(\varepsilon,\alpha)$, we find from Eq.  $$\begin{aligned} 0 &= f_1(\varepsilon,\alpha)-f_1(\varepsilon,\alpha + 2 \pi) \nonumber \\ & = \frac{2\pi}{P_\varepsilon(\varepsilon)}\bigg( \partial_\varepsilon \Big(1-\frac{1}{\sqrt{1+2\varepsilon}}\Big) P_\varepsilon(\varepsilon) + D \partial_\varepsilon \varepsilon \partial_\varepsilon P_\varepsilon(\varepsilon) \bigg). \label{pe-consistency}\end{aligned}$$ Taking the derivative outside the parentheses, we recover precisely Eq.  and thus the solution is consistent. Plugging in Eq. , we find for the corresponding terms in Eq.  $$\begin{aligned} \frac{\partial_\varepsilon P_\varepsilon(\varepsilon)}{P_\varepsilon(\varepsilon)} &= -\frac{2}{D(1+2\varepsilon + \sqrt{1+2\varepsilon})} \nonumber \\ \frac{\partial_\varepsilon \varepsilon \partial_\varepsilon P_\varepsilon(\varepsilon)}{P_\varepsilon(\varepsilon)} &= \frac{4\varepsilon \sqrt{1+2\varepsilon}-2D(1+\varepsilon+\sqrt{1+2\varepsilon})}{D^2(1+2\varepsilon)^{\frac{3}{2}}(1+\sqrt{1+2\varepsilon})^2}\end{aligned}$$ Before we proceed to discuss this result, let us determine the as of yet unknown function $\tilde{f}_1(\varepsilon)$. In order to find it, we require a consistency condition like , which we obtain from the second order of the large-$\Omega$ expansion. For the latter, we have the equation $$\begin{aligned} \partial_\alpha f_2(\varepsilon,\alpha) = - \frac{1}{P_\varepsilon(\varepsilon)} \mathcal{L}_{\varepsilon,\alpha} \Big[ f_1(\varepsilon,\alpha) P_\varepsilon(\varepsilon) \Big] \label{f2}.\end{aligned}$$ In particular, we want $f_2(\varepsilon,\alpha)$ to be periodic in $\alpha$, so we necessarily have $$\begin{aligned} \int_{0}^{2\pi} \text{d}\alpha' \mathcal{L}_{\varepsilon,\alpha'} \Big[f_1(\varepsilon,\alpha') P_\varepsilon(\varepsilon) \Big] = 0.\end{aligned}$$ Since the entire problem is $\pi$-periodic (corresponding to $(x,p) \rightarrow (-x,-p)$), we only need to consider the integral from $0$ to $\pi$. We split the integral at $\alpha' = \pi/2$, $$\begin{aligned} 0 &= \int_{0}^{\pi/2} \text{d}\beta \ \mathcal{L}'_{\varepsilon,\beta} \Big[f_1\Big(\varepsilon,\frac{\pi}{2}-\beta\Big) P_\varepsilon(\varepsilon) \Big] \nonumber \\ & \quad + \int_{0}^{\pi/2} \text{d}\gamma \ \mathcal{L}''_{\varepsilon,\gamma} \Big[f_1\Big(\varepsilon,\frac{\pi}{2}+\gamma\Big) P_\varepsilon(\varepsilon) \Big],\end{aligned}$$ where we introduced the new variables $\beta = \pi/2-\alpha$ and $\gamma = \alpha-\pi/2$. Here $\mathcal{L}'_{\varepsilon,\beta}$ and $\mathcal{L}''_{\varepsilon,\gamma}$ are the correspondingly transformed operators. From the symmetry of $\mathcal{L}_{\varepsilon,\alpha}$ (see Eq. ), it can easily be seen that the operator transforms in the same manner for $\beta$ and $\gamma$, so that $\mathcal{L}'_{\varepsilon,\alpha} = \mathcal{L}''_{\varepsilon,\alpha}$ and we can thus write $$\begin{aligned} 0 &= \int_{0}^{\pi/2} \text{d}\beta \mathcal{L}'_{\varepsilon,\beta} \Big[\Big(f_1\Big(\varepsilon,\frac{\pi}{2}-\beta\Big)+f_1\Big(\varepsilon,\frac{\pi}{2}+\beta\Big)\Big) P_\varepsilon(\varepsilon) \Big].\end{aligned}$$ Now, the angle-dependent part of Eq.  is antisymmetric around $\alpha = \pi/2$. While this is not immediately apparent, it can easily be shown after a little algebra. Thus this angle dependent part drops out of the integral and we have $$\begin{aligned} 0 &= \int_{0}^{\pi/2} \text{d}\alpha \ \mathcal{L}_{\varepsilon,\alpha} \Big[\tilde{f}_1(\varepsilon) P_\varepsilon(\varepsilon) \Big],\end{aligned}$$ where we reverted to the original angle $\alpha$. This gives us $$\begin{aligned} \partial_\varepsilon \bigg[ 1 - \frac{1}{\sqrt{1+2\varepsilon}} + D \varepsilon \partial_\varepsilon \bigg]\tilde{f}_1(\varepsilon) P_{\varepsilon}(\varepsilon) = 0, \label{f1-energy}\end{aligned}$$ which is exactly Eq.  but now for $\tilde{f}_1(\varepsilon) P_{\varepsilon}(\varepsilon)$. An obvious solution to this equation is $\tilde{f}_1(\varepsilon) = C$ where $C$ is an arbitrary constant. Demanding that the first order phase-space density is normalized then immediately yields $C=0$. In principle, there exists a second solution to Eq. , however, this can be shown to be non-normalizable. Thus the angle-independent part $\tilde{f}_1(\varepsilon)$ in Eq.  is indeed zero. Armed with this knowledge, we can proceed to examine the solution in more detail. ![First order angle-dependent correction to the phase-space density as a function of the angle for different energies and $D=0.3$. \[fig:anglefunc\]](anglefunc.png){width="47.00000%"} The function $f_1(\varepsilon,\alpha)$ is shown in Fig. \[fig:anglefunc\]. First of all, we note that the correction is small both for large and for small energies and thus the dependence on the angle $\alpha$ is strongest for intermediate energies. For small energies, the friction force is almost linear and we recover the Boltzmann-Gibbs result with small corrections, see Section \[sec:small-D\]. For large energies, we can expand the expression and find $$\begin{aligned} f_1(\varepsilon,\alpha) \simeq \frac{-D \cot(\alpha)+(1+D)\sin(2\alpha)}{2 D \varepsilon} + \mathcal{O}(\varepsilon^{-\frac{3}{2}}). \label{anglefunc-large-E}\end{aligned}$$ We see that indeed the function is of order $\varepsilon^{-1}$ for large energies. However, the resulting expression diverges at $\alpha = 0$ and all $\alpha = \pi$, which would lead to (nonintegrable) divergences in the probability density. To resolve this issue, let us expand around $\alpha = 0$ instead, $$\begin{aligned} f_1(\varepsilon,\alpha) \simeq -\frac{1+2\varepsilon+2\sqrt{1+2\varepsilon}}{(1+2\varepsilon)(1+\sqrt{1+2\varepsilon})^2} 2 \varepsilon \alpha + \mathcal{O}(\alpha^2). \label{anglefunc-small-alpha}\end{aligned}$$ This expression is of course regular at $\alpha = 0$, but tends to $-\alpha$ for large $\varepsilon$, in contrast to the behavior observed in Fig. \[fig:anglefunc\]. So, which of the two expansions is correct? The answer is of course, both, depending on the relative size of $\alpha$ and $\varepsilon$, as the limits $\varepsilon \rightarrow \infty$ and $\alpha \rightarrow 0$ do not commute. As long as $\alpha$ is farther away than $1/\sqrt{2 \varepsilon}$ from $0$ or $\pi$, the large-energy expansion $\eqref{anglefunc-large-E}$ yields the correct result. However, as we approach one of the points $\alpha = 0$ or $\alpha = \pi$, we have to use the small-angle expansion $\eqref{anglefunc-small-alpha}$. These non-commuting limits can be glimpsed from Fig. \[fig:anglefunc\], where we see sharp features emerging at $\alpha = 0$ and $\pi$ for large energies. Physically, these features stem from the nonlinearity of the friction force. As a high-energy particle oscillates in the confining potential, it is fast during most of its orbit, and the dissipation is weak, leading to an overall decrease of the effects with energy. Close to the turning points $\alpha = 0, \pi$, however, the particle slows down and the friction increases, up to the point where it becomes Stokes-like. Thus, even at high energies, there always remains a small section of phase space where the friction is relevant. \ ![Phase-space probability density for $D = 0.3$ and $\Omega = 2$, plotted using the position and momentum $(x,p)$ (top) respectively energy and phase-space angle $(E,\alpha)$ (bottom). The colored surfaces are the results of numerical Langevin simulations, the black lines correspond to the first-order large-$\Omega$ expansion Eq. . Both results agree well and exhibt an obvious asymmetry in phase-space. \[fig:phasedist\]](phasedist_angle.png "fig:"){width="47.00000%"} The phase-space density up to first order in $\Omega$ is shown in Fig. \[fig:phasedist\]. Even for moderately large values of $\Omega$, this first-order result captures the probability density very well, since the correction term $f_1(\varepsilon,\alpha)$ by itself is small. ![Asymmetric part of the phase-space probability density for $D = 0.3$ and $\Omega = 2$. The colored surfaces are the results of numerical Langevin simulations, the black lines correspond to the first-order large-$\Omega$ expansion Eq. . \[fig:phasedistas\]](phasedistas.png){width="47.00000%"} In Fig. \[fig:phasedistas\], we show the corresponding antisymmetric part, as defined in Eq. . The general features are very similar to the results from the small-$D$ expansion, which is no longer valid at these moderate values of $D$ (here $D=0.3$). This hints at the generality of the asymmetry in phase-space and the deviations from equipartition, even beyond the range of validity of the respective approximations. Note that, similarly to the small-$D$-expansion, the violation of energy equipartition is a second-order effect that is not included in Eq. . While it is in principle also possible to obtain higher order corrections, the calculations for obtaining them are much more involved. ![Second order angle-dependent correction to the phase-space density as a function of the angle for different energies and $D=0.3$. \[fig:anglefunc2\]](anglefunc2.png){width="47.00000%"} In Fig. \[fig:anglefunc2\], we show the result for the second order correction obtained by numerically solving Eq. . This does indeed exhibit the expected feature, an increased probability along the $x$-axis ($\alpha = 0$ and $\alpha = \pi$), which is responsible for the enhancement of the potential energy compared to the kinetic one. Besides the fact that violations of equipartition are of order $\Omega^{-2}$, we note that the coefficient $f_2(\varepsilon,\alpha)$ of the second order term is numerically much smaller than the first order one. This explains why we see good agreement with first order result even at moderate $\Omega$, see Figs. \[fig:phasedist\] and \[fig:phasedistas\]. Large energies: Asymptotic power laws \[sec:large-E\] ===================================================== In the previous Section, we discussed that the dynamics becomes underdamped not only for large frequencies (strong trapping) but also generically for large energies, see Eq.  and discussion following it. In this Section we want to formalize this and show that many of the features of the large-$\Omega$ expansion in fact carry over to the large-energy behavior at arbitrary $\Omega$. We adopt a similar formalism as before, this time, however, expanding in terms of energy, $$\begin{aligned} P_P(\varepsilon,\alpha) \simeq N \varepsilon^{\beta} \Bigg[1 + \frac{g_{1/2}(\alpha)}{\varepsilon^{\frac{1}{2}}} + \frac{g_1(\alpha)}{\varepsilon} + \mathcal{O}(\varepsilon^{\frac{3}{2}}) \Bigg] , \label{large-energy-expansion-outside}\end{aligned}$$ with some normalization constant $N$. This expansion warrants some explanation. Since we know that the system becomes underdamped for large energies, we anticipate that, similar to the result , the phase-space density is to leading order independent of the angle $\alpha$. Another way to see this is that the operator $\mathcal{L}_{\varepsilon,\alpha}$ in Eq.  is apparently of order $E^{-1}$ for large energies, so that we have $\Omega \partial_\alpha P(\varepsilon,\alpha) = \mathcal{O}(E^{-1})$. The power-law behavior of the leading order term and half-integer order terms in the expansion are motivated by the large-energy expansion of the stationary energy density , $$\begin{aligned} P_\varepsilon(\varepsilon) \simeq Z_\varepsilon^{-1} (2 \varepsilon)^{-\frac{1}{D}} \bigg( 1 - \frac{\sqrt{2}}{D \varepsilon^{\frac{1}{2}}} + \frac{1}{D^2 \varepsilon} + \mathcal{O}(\varepsilon^{-\frac{3}{2}}) \bigg) \label{large-omega-large-e},\end{aligned}$$ where similar terms occur. We plug the expansion into the Kramers-Fokker-Planck equation and expand for large energies, resulting in an equation containing powers of $\varepsilon$. Equating orders of $\varepsilon$, we find conditions on the functions $g_{1/2}(\alpha)$, $g_1(\alpha)$ and so on. Up to order $\varepsilon^{-1}$, these read $$\begin{aligned} \partial_\alpha &g_{1/2}(\alpha) = 0 \nonumber \\ \partial_\alpha &g_{1}(\alpha) = \frac{1}{2 \Omega} \Big( - (1+2\beta)(2 D \beta \sin^2(\alpha) - 1) \nonumber \\ &\qquad + D \beta \cos^2(\alpha) + \cot^2(\alpha) \Big). \end{aligned}$$ The first condition obviously demands that $g_{1/2}(\alpha) = \tilde{g}_{1/2}$ is a constant, while the second one gives us after integration $$\begin{aligned} g_1(\alpha) &= -\frac{1}{2 \Omega} \Big( \cot(\alpha) + \beta(2 \alpha(D \beta-1) \nonumber \\ &\qquad - D(1+\beta) \sin(2\alpha) ) \Big) + \tilde{g}_1 . \label{large-E-order1}\end{aligned}$$ Demanding that $g_1(\alpha)$ should be $2\pi$-periodic, we get a condition on $\beta$ that yields $\beta = 0$ or $\beta = -1/D$. Since the first solution is non-normalizable at large energies, $\beta = -1/D$ is the relevant solution. This shows that the power-law tails $P_\varepsilon(\varepsilon) \propto \varepsilon^{-1/D}$ are in fact the universal large energy behavior. Apart from the constant $\tilde{g}_1$, the above is then completely equivalent to the large-energy expansion of the large-$\Omega$ solution, Eq. , which is reassuring. The constants $\tilde{g}_{1/2}$ and $\tilde{g}_1$ have to be determined from next order in the expansion. However, Eq.  has the same issue that we already observed with Eq. : It diverges at $\alpha = 0, \pi$. The reason for this is that we expanded $\mathcal{L}_{\varepsilon,\alpha}$ for large $\varepsilon$ without paying attention to $\alpha$. But as we saw before, the limits $\varepsilon \rightarrow \infty$ and $\alpha \rightarrow 0$ do not commute. Because of this, we need to perform a similar expansion, but in the region where $\alpha$ is close to $0$ or $\pi$, i.e. where close to the turning points, where the energy is large but the momentum is not. In this region, however, the position $x$ is of order $\sqrt{\varepsilon}$, so we can expand Eq.  for large $x$. Recalling equation , $$\begin{aligned} \Bigg[ \Omega \bigg( - p \partial_x + x \partial_p \bigg) + \partial_p \bigg(\frac{p}{1+p^2} + D \partial_p \bigg) \Bigg] P_S(x,p) = 0 , \label{KFP-lattice-p}\end{aligned}$$ where we explicitly denote the solution inside the strip by $P_S$. we see that for $x$ large and $p$ of order $1$, the very first term is of order $1/x$, the second one of order $x$ and the final two terms are of order $1$. Note that $P_P(\varepsilon,\alpha) = P_S(\sqrt{2 \varepsilon} \cos(\alpha),\sqrt{2 \varepsilon} \sin(\alpha))$, i. e. the Jacobian of the variable transformation is unity. We write down an expansion for the phase-space density similar to Eq. , $$\begin{aligned} P_S(x,p) \simeq M |x|^{\gamma} \Bigg[1 + \frac{h_{1/2}(p)}{x} + \frac{h_1(p)}{x^2} + \mathcal{O}(x^{-3}) \Bigg] , \label{large-energy-expansion-inside}\end{aligned}$$ where $M$ is a normalization constant. We plug this into Eq. , expand for large $x$ and evaluate the coefficients, $$\begin{aligned} \Omega \partial_p &h_{1/2}(p) = -\partial_p \frac{p}{1+p^2} \nonumber \\ \Omega \partial_p &h_{1}(p) = -\partial_p \bigg(\frac{p}{1+p^2} + D \partial_p \bigg) h_{1/2}(p).\end{aligned}$$ Solving for $h_{1/2}(p)$ and $h_1(p)$, we have $$\begin{aligned} &h_{1/2}(p) = -\frac{1}{\Omega} \frac{p}{1+p^2} + \tilde{h}_{1/2} \nonumber \\ &h_1(p) = \frac{D + p^2(1-D)}{\Omega^2 (1+p^2)^2} - \frac{\tilde{h}_{1/2} \ p}{\Omega(1+p^2)} + \tilde{h}_1 .\end{aligned}$$ So far, the two expansions Eqs.  and are independent of each other. Whereas Eq.  is valid at most points in phase space, where $|p| \gg 1$ (or equivalently $|\alpha| \gg 1/\sqrt{2\varepsilon}$), Eq.  describes the strip where $|p| \lesssim 1$ (or equivalently $|\alpha| \lesssim \sqrt{2\varepsilon}$). However, they describe an expansion of the same function in these different areas of phase-space, and consequently should be related. In particular taking the $|p| \rightarrow \infty$ limit of Eq.  inside the strip should match onto the $\alpha \rightarrow 0$ limit of Eq.  outside the strip. From the zeroth-order term this immediately gives $\gamma = 2 \beta = -2/D$ and $M = 2^{1/D} N$. Further expanding $g_{1/2}$ and $g_1(\alpha)$ around $\alpha = 0$ gives $$\begin{aligned} &g_{1/2}(\alpha) \varepsilon^{-\frac{1}{2}} = \tilde{g}_{1/2} \varepsilon^{-\frac{1}{2}} \nonumber \\ &g_1(\alpha) \varepsilon^{-1} \simeq \Big[\tilde{g}_1 -\frac{1}{2 \Omega \alpha} + \mathcal{O}(\alpha) \Big] \varepsilon^{-1}. \label{outside-inside}\end{aligned}$$ We need to compare this to the large-$p$ expansion of $h_{1/2}(p)$ and $h_{1}(p)$, $$\begin{aligned} &h_{1/2}(p) x^{-1} \simeq \Big[\tilde{h}_{1/2} - \frac{1}{\Omega} \Big( \frac{1}{p} - \frac{1}{p^3} + \mathcal{O}(p^{-5}) \Big)\Big] x^{-1} \nonumber \\ &h_1(p) x^{-2} \simeq \Big[\tilde{h}_1 + \frac{1}{\Omega^2} \frac{1-D}{p^2} - \frac{\tilde{h}_{1/2}}{\Omega} \Big(\frac{1}{p} - \frac{1}{p^3} \Big) \nonumber \\ & \qquad \qquad \qquad \qquad + \mathcal{O}(p^{-4}) \Big] x^{-2}.\end{aligned}$$ Close to $\alpha = 0$, we further have $p \simeq \sqrt{2 \varepsilon} (\alpha + \mathcal{O}(\alpha^3))$ and $x \simeq \sqrt{2 \varepsilon} (1 + \mathcal{O}(\alpha^2))$ and thus $$\begin{aligned} &h_{1/2}(p) (2 \varepsilon)^{-\frac{1}{2}} \simeq \tilde{h}_{1/2} (2 \varepsilon)^{-\frac{1}{2}} - \frac{1}{2\Omega} \frac{1}{\alpha \varepsilon} + \mathcal{O}(\varepsilon^{-3/2}) \nonumber \\ &h_1(p) (2\varepsilon)^{-1} \simeq \tilde{h}_1 (2\varepsilon)^{-1} + \mathcal{O}(\varepsilon^{-\frac{3}{2}}). \label{inside-outside}\end{aligned}$$ Matching the coefficients of different orders in $\varepsilon$ between Eqs.  and then connects the integration constants of the two expansions: $\tilde{h}_{1/2} = \tilde{g}_{1/2}/\sqrt{2}$ and $\tilde{h}_1 = \tilde{g}_1/2$. As remarked before, we need to go to higher orders in $\varepsilon$ to find the integration constants $\tilde{g}_{1/2}$ and $\tilde{g}_1$. This is similar to the function $\tilde{f}_1(\varepsilon)$ we found for the large-$\Omega$ expansion in Eq. . This procedure is carried out up to order $\varepsilon^{-2}$ in Appendix \[app:large-e\]. The result, plotted in Fig. \[fig:phasedist-large-e\], for the expansion up to first order, in terms of $\varepsilon$ and $\alpha$ is $$\begin{aligned} P_{P}(\varepsilon,\alpha) \simeq N \varepsilon^{-\frac{1}{D}} \left\lbrace \begin{array}{ll} 1 - \frac{\sqrt{2}}{D} \varepsilon^{-1/2} + \Big[ \frac{1}{D^2} + \frac{1}{2 \Omega} \Big( \big(1 + \frac{1}{D}\big) \sin(2\alpha) - \cot(\alpha) \Big) \Big] \varepsilon^{-1},& \text{for} \; \sqrt{2 \varepsilon} \alpha \gg 1 \\[2 ex] 1 - \Big[\frac{\sqrt{2}}{D} + \frac{\sqrt{\varepsilon} \alpha}{ \Omega (1+2 \varepsilon \alpha^2)} \Big] \varepsilon^{-1/2} + \Big[\frac{1}{D^2} + \frac{1}{2 \Omega^2} \Big[ \frac{2 D -1}{(1+2 \varepsilon \alpha^2)^2} + \frac{D-D^2 + 2 \sqrt{2 \varepsilon} \alpha \Omega}{D(1+2 \varepsilon \alpha^2)} \Big] \Big] \varepsilon^{-1}, & \text{for} \; \sqrt{2 \varepsilon} \alpha \lesssim 1 . \label{large-e-final} \end{array} \right.\end{aligned}$$ ![Phase-space probability density for $D = 0.3$ and $\Omega = 0.5$ from the large-energy expansion Eq. , plotted using energy and phase-space angle $(E,\alpha)$. The thick black lines denote the approximate boundary between the strip and outside area. Note that for large energies, the matching at the boundary works increasingly well. \[fig:phasedist-large-e\]](large-e-plot.png){width="47.00000%"} The above discussion shows that the power-law tail found in the large-$\Omega$ result is in fact generic for large energies, where the system always is underdamped, independent of $\Omega$. To leading order the stationary state is thus characterized by the energy. Just like for small $D$ and large $\Omega$, the corrections lead to a complex structure of the phase-space density. We note that while the large-energy expansion agrees with the first order large-$\Omega$ expansion in the appropriate limits, the contribution from the strip actually also contains terms of order $\Omega^{-2}$. In that sense, just as the first order large-$\Omega$ expansion contains information that is not contained in the first order large-energy one, the converse is also true. Probability currents and detailed balance \[sec:detailed-balance\] ================================================================== In the previous sections we discussed the unusual stationary state behavior of confined atoms in Sisyphus cooling. We already saw that the stationary state is non-thermal in that there exists no well-defined temperature and that energy equipartition does not hold. In that sense, we may refer to this stationary state as a non-equilibrium stationary state. In order to differentiate this further, we want to see whether or not detailed balance holds. Detailed balance, which is often taken as the defining property of an equilibrium system, means that for every possible transition in the system, the forward and backward process are equally likely [@ris86; @gar96]. In terms of the transition probabilities, this can be expressed as [@gar96] $$\begin{aligned} P(x',p',t+\tau ; x,p,t) = P(x,-p,t+\tau ; x',-p',t), \label{detailed-balance}\end{aligned}$$ where $P(x',p',t' ; x, p, t)$ is the joint probability density for finding a particle at $(x,p)$ at time $t$ and at $(x',p')$ at time $t'$. Here the minus sign in front of the momentum is a consequence of the momentum being odd under time reversal. For the stationary state in a Markovian system, we can express this in terms of the conditional probability density $P(x',p',\tau \vert x, p, 0)$, $$\begin{aligned} P(x',p',\tau &\vert x,p,0) P(x,p) \nonumber \\ & = P(x,-p,\tau \vert x',-p',0) P(x',-p').\end{aligned}$$ For $\tau = 0$, the conditional probability density reduces to a delta-function, $P(x',p',0 \vert x,p,0) = \delta(x-x') \delta(p-p')$. Since this function is even under $(p,p') \rightarrow (-p,-p')$, detailed balance implies that $P(x,p) = P(x,-p)$, i.e. the stationary probability density has to be even under momentum reversal [@gar96]. As we saw in Sections \[sec:small-D\] and \[sec:large-freq\], the stationary state of our system of confined cold atoms does not satisfy this property and therefore does not respect detailed balance. For a Kramers-Fokker-Planck equation describing one-dimensional underdamped motion in a potential, $$\begin{aligned} \Bigg[ - p \partial_x + \partial_p \bigg[ U'(x) - F_\text{fric}(p) + D_p \partial_p \bigg] \Bigg] P(x,p) = 0, \label{KFP-general}\end{aligned}$$ detailed balance is equivalent to the conditions [@gra71; @ris86] $$\begin{aligned} D_{p} - D_{-p} &= 0 \label{db-conditions-1} \\ \partial_x J_x^\text{rev} + \partial_p J_p^\text{rev} &= 0 \label{db-conditions-2}\\ J_x^\text{ir} = J_p^\text{ir} &= 0 \label{db-conditions-3}\end{aligned}$$ \[detailed-balance-conditions\] Note that we use dimensionless variables here. For a confined particle, the reversible currents $J_i^\text{rev}$ correspond to the oscillatory motion of the particle in the potential. The irreversible currents, on the other hand are induced by the dissipative and fluctuating forces due to the bath. Equation  implies that the momentum diffusion coefficient $D_p$ should be an even function of $p$. This condition holds for the optical lattice system. For Eq.  the reversible and irreversible probability currents are given by $$\begin{aligned} J_x^\text{rev} = p P(x,p), \qquad &J_p^\text{rev} = -U'(x) P(x,p), \label{prob-current} \\ J_x^\text{ir} = 0, \qquad &J_p^\text{ir} = \Big( F_\text{fric}(p) - D_p \partial_p \Big) P(x,p) . \nonumber\end{aligned}$$ Plugging this into Eq. , we see that detailed balance is equivalent to stationary solution $P(x,p)$ solving both the Hamiltonian and “bath” part of the Kramers-Fokker-Planck equation individually, $$\begin{aligned} \bigg[ - p \partial_x + U'(x) \partial_p \bigg] P(x,p) &= 0 \label{db-hamilton} \\ \bigg[ - F_\text{fric}(p) + D_p \partial_p \bigg] P(x,p) &= 0 \label{db-bath} .\end{aligned}$$ Equation implies that $P(x,p) = f(p^2/2 + U(x))$. From Eq. , we further have $$\begin{aligned} P(x,p) = g(x) \exp \bigg[{\int_{0}^{p} \text{d}p' \ \frac{F_\text{fric}(p')}{D_p'}} \bigg] .\end{aligned}$$ The two expressions for $P(x,p)$ are compatible only if a fluctuation-dissipation relation holds, $$\begin{aligned} \frac{F_\text{fric}(p)}{D_p} = - \tilde{\beta} p \label{db-flucdiss},\end{aligned}$$ and we have the Boltzmann-Gibbs density $$\begin{aligned} P(x,p) = N \exp \bigg[ -\tilde{\beta} \Big( \frac{p^2}{2} + U(x) \Big) \bigg],\end{aligned}$$ where $\tilde{\beta} > 0$ plays the role of an inverse effective temperature. Thus for the general class of systems described by Eq. , an (effective) Boltzmann-Gibbs ensemble is the only possible solution obeying detailed balance. For Sisyphus cooling, we have from Eq.  in terms of dimensionless variables $F_\text{fric} = -p/(1+p^2)$, $D_p = D(1 + \mathfrak{D} (1+p^2))$ and thus $$\begin{aligned} \frac{F_\text{fric}(p)}{D_p} = \frac{-\frac{p}{1+p^2}}{D + \frac{D \mathfrak{D}}{1+p^2}},\end{aligned}$$ with $\mathfrak{D} = D_1/D_0$, see Eq. . This satisfies the condition only for $D \rightarrow 0$, where we obtain the Boltzmann-Gibbs density, see Section \[sec:small-D\] and Appendix \[app:d1\]. In this limit $D_1/\gamma p_c^2$ corresponds to the above defined effective temperature $\tilde{\beta}$. For $D_1 = 0$, as assumed in the previous sections, the effective temperature in the Boltzmann-Gibbs limit $D = D_0/\gamma p_c^2 \rightarrow 0$ vanishes. The violation of detailed balance and thus the non-equilibrium nature are due to the absence of a fluctuation-dissipation relation like Eq.  between the noise and the friction force. While the friction force and the momentum-dependent part $D \mathfrak{D}/(1+p^2)$ of the diffusion coefficient (see Eq.  and the following discussion) are both due to the motion of the atoms in the optical lattice and thus obey a fluctuation-dissipation relation, the momentum-independent part $D$ of the diffusion coefficient, which represents spontaneous emission of photons, has no dissipative counterpart. We now want to quantify the violation of the detailed balance conditions Eq. . We focus on the case where the diffusion coefficient is even in the momentum so that Eq.  is satisfied. The remaining two conditions are then equivalent to Eqs.  and . These two equations imply a geometric property for the total probability current $\vec{J} = (J_x, J_p)$, where $J_i = J_i^\text{rev} + J_i^\text{ir}$. As is easily verified by direct calculation, as long as detailed balance holds, the probability current is perpendicular to the gradient of the density $\vec{\nabla}P = (\partial_x P, \partial_p P)$, i. e. $\vec{J} \cdot \vec{\nabla} P = 0$ [@gra71b]. This means that for a system with detailed balance, the probability current always flows along equi-probability lines. For the optical lattice system with $D \neq 0$, on the other hand, detailed balance is broken and thus $\vec{J} \cdot \vec{\nabla} P \neq 0$. The current is related to the local mean phase-space velocity via $\vec{v}_P = \vec{J}/P$, where by definition $$\begin{aligned} \partial_t \left( \begin{array}{ll} \langle x \rangle \\ \langle p \rangle \end{array} \right) = \langle \vec{v}_P \rangle .\end{aligned}$$ In the steady state, the left hand side vanishes and thus the average phase-space velocity $\langle \vec{v}_P \rangle$ is zero, however, it is generally still non-zero locally. We use the scalar product between the phase-space velocity and the normalized gradient of the density $$\begin{aligned} \phi = \vec{v}_P \cdot \frac{1}{P}\vec{\nabla} P = \frac{\vec{J} \cdot \vec{\nabla} P}{P^2} \label{phidef}\end{aligned}$$ as a measure of the misalignment between current and equi-probability lines and hence detailed balance violation. In the large-$\Omega$ limit and for constant $D_p \equiv D_0$, we can write the probability density as $$\begin{aligned} P(x,p) \simeq P_0(x,p) \Big(1 + \frac{f(x,p)}{\Omega} \Big) + \mathcal{O}(\Omega^{-2})\end{aligned}$$ with $P_0(x,p) = P_\varepsilon[(x^2+p^2)/2]/(2 \pi)$ and $f(x,p) = f_1[(x^2+p^2)/2,\arctan(p/x)]$, see Eq. . From the definition of the probability current Eq. , we then find $$\begin{aligned} \phi &= \frac{1}{P_0(x,p)} \Bigg[\bigg[ P_0(x,p) \Big[ p \partial_x - x \partial_p \Big] f(x,p) \bigg] \\ & \qquad - \bigg[ \frac{p}{1+p^2} \partial_p + D \big\lbrace \partial_p \ln(P_0(x,p)) \big\rbrace \partial_p \bigg] P_0(x,p) \Bigg] \nonumber \\ & \qquad + \mathcal{O}(\Omega^{-1}). \nonumber\end{aligned}$$ Note that $\phi$ depends on the first order correction $f(x,p)$ even in the limit $\Omega \rightarrow \infty$. Further recognizing that $[p \partial_x - x \partial_p] f(x,p) = -\partial_\alpha f_1(\varepsilon,\alpha)$ and using Eqs.  and , we finally find, $$\begin{aligned} \phi &= \partial_p \frac{p}{1+p^2} + D \partial_p^2 \ln(P_0(x,p)) \label{phi-result} \\ &= \partial_p \frac{p}{1+p^2} - 2 \partial_p^2 \ln(1+\sqrt{1+p^2 + x^2}) + \mathcal{O}(\Omega^{-1}) \nonumber.\end{aligned}$$ Intriguingly, $\phi$ is to leading order independent of $D$. This seems counter-intuitive, since the detailed balance violations should vanish for $D \rightarrow 0$. As it turns out, this is an artifact of setting $D_1 = 0$, since then, as discussed above, the effective temperature for $D \rightarrow 0$ is zero and there is thus no well-defined detailed-balance preserving state in this limit. Repeating the calculation for a non-zero $D_1$ yields $\lim_{D \rightarrow 0} \phi = 0$ as it should, see Appendix \[app:d1-omega\]. The average $\langle \phi \rangle$ over all phase-space is zero, reflecting the fact that there is no global probability current in the steady state. We thus interpret $\phi$ as a measure of local flow due to the detailed balance violation of the system. The quantity $\phi$ for $\Omega \rightarrow \infty$ is shown in Fig. \[fig:current-angle\], where we see that the local phase-space velocity can be both parallel and antiparallel to the density gradient. The local mean phase-space velocity $\vec{v}_P = (\dot{x},\dot{p})$ represents the average change in position and momentum of particles located at a phase space point $(x,p)$. In areas where $\phi$ is positive, particles on average move from a low-density to a high-density region. This is mostly observed for small momenta, where the friction is effective and thus causes a flow towards the central, low energy region. In most areas of phase-space, the flow is directed towards larger energies due to the weak friction, balancing out the inward flow close to the $(p = 0)$-axis. We stipulate that probability currents that do not flow along equi-probability lines are generally connected to a non-zero entropy production as a measurable consequence of the non-equilibrium nature of the system. The breaking of detailed balance necessarily implies the presence of irreversible probability currents in the system [@cho11]. As was shown in Ref. [@esp10], a distinct contribution to the entropy production also arises from the breaking of detailed balance. Indeed, the entropy production can be directly expressed via the irreversible currents [@tom10], which in the case of momentum-dependent forces leads to an anomalous entropy production [@kwo16]. We suggest that the geometric properties of the probability current discussed above may yield a more detailed understanding of the precise way the system deviates from equilibrium. We leave this investigation to future work. ![The scalar product between the mean velocity and the gradient of the phase-space density $\phi = \vec{v}_P \cdot \vec{\nabla} P /P$, Eq. . The color corresponds to the value of $\phi$, the arrows denote the direction of the current parallel to the gradient; inward for positive $\phi$, outward for negative $\phi$. \[fig:current-angle\]](curangle.png){width="47.00000%"} Physical considerations for the cold atom system \[sec:experimental\] ===================================================================== So far, we treated the solution of our main equation from a mathematical point of view, with only occasional reference to the actual system of Sisyphus cooling of confined atoms. In the following, we first delineate the experimentally relevant parameter regime in terms of the dimensionless parameters $D$ and $\Omega$. We also discuss the effects of the momentum dependent diffusion coefficient and Stark shifts induced by introducing the confining potential. Finally, we present numerical simulations for two example sets of parameters to estimate the magnitude of the observed effects. In the following we will refer to the dimensionful coordinates as $\tilde{x}$ and $\tilde{p}$, while $x$ and $p$ denote their dimensionless counterparts, see Eq. . *Magnitude of the confinement.* As the semiclassical description of Sisyphus cooling relies on spatially averaging the motion of the atoms over a wavelength of the cooling lattice [@cas90], it will work reliably only when the atoms are able to move over several lattice periods on time scale of interest. This puts a natural constraint on the magnitude of the confining potential; the latter has to be weak enough so as to not localize the atoms on the length scale of the cooling lattice. We can estimate the magnitude of the confining field by demanding that the typical spread of the atomic cloud should be much larger than the lattice period, $\langle \tilde{x}^2 \rangle \gg (2 \pi/k)^2$, where $k$ is the wave vector of the cooling laser. As long as this holds and our theory is valid, we have $\langle \tilde{x}^2 \rangle = p_c^2/(m \omega)^2 \langle x^2 \rangle$, where $\langle x^2 \rangle$ is of order $1$. Thus we should have $\omega \ll p_c k/(2 \pi m)$, or $\Omega \ll p_c k/(2 \pi m \gamma)$. In terms of the lattice parameters, the parameters $p_c$, $\gamma$, $D_0$ and $D_1$ can be expressed as [@cas90] $$\begin{aligned} p_c &= \frac{m \Gamma s_0}{9 k}, \qquad \gamma = \frac{3 \hbar k^2 |\delta|}{m \Gamma}, \\ D_0 &= \frac{11 \hbar^2 k^2 \Gamma s_0}{18}, \qquad D_1 = \frac{\hbar^2 k^2 \delta^2 s_0}{\Gamma}, \nonumber\end{aligned}$$ where $\Gamma$ is the natural linewidth of the atomic transition, $\delta$ is the laser detuning and $s_0 = I/I_s/(1+4\delta^2/\Gamma^2)$ is the saturation parameter with the laser intensity $I$ and the saturation intensity $I_s$. Consequently, we get the condition $$\begin{aligned} \Omega \ll \frac{1}{108 \pi} \frac{\hbar \Gamma}{E_r} \frac{\Gamma s_0 }{\delta} ,\end{aligned}$$ where $E_r = \hbar^2 k^2/(2 m)$ is the photon recoil energy. For Cesium ${}^{133}\text{Cs}$, which is commonly used in Sisyphus cooling experiments $E_r = 5.4 \cdot 10^{-29} \text{J}$ and $\Gamma = 7.4 \cdot 10^{6} \text{s}^{-1}$ [@ste03], and thus $$\begin{aligned} \Omega \ll 1.19 \frac{\Gamma s_0}{\delta} .\end{aligned}$$ In the following, we consider two exemplary sets of parameters, $\delta = 10 \Gamma$ and $I = 20 I_s$ (large detuning) and $\delta = 1.5 \Gamma$ and $I = 3.3 I_s$ (moderate detuning). Both choices lead to a lattice depth $U_0 \approx 130 E_r$, which is the point where the minimal kinetic energy for free atoms is reached according to the semiclassical treatment [@cas90; @per00]. Since $D$ is related to $U_0$ by $D = 22 E_r/U_0$, this leads to a value of $D \approx 0.17$ [^1]. For these kinds of lattice parameters, we are thus at moderate values of $D$, which are far enough from the point where the average energy diverges ($D = 1/2$, see Eq. ) to be able to characterize the system by the stationary solution, yet still large enough for the deviations from Boltzmann-Gibbs to become important, see Section \[sec:small-D\]. In the large detuning limit the bound on $\Omega$ turns out to be $\Omega \ll 5.9 \cdot 10^{-3}$, whereas for moderate detuning we find $\Omega \ll 0.26$. In both cases, $\Omega$ has to be very small, which means that we cannot employ the underdamped description of Section \[sec:large-freq\]. However, we saw in Eq.  that the deviations from energy equipartition are most pronounced at small $\Omega$ and thus these effects are rather important in this regime. For moderate detuning, and not too small $\Omega$, also the asymmetry of the phase-space density, quantified in Eq. , might be observed. For the simulations, we take $\Omega$ to be one tenth of the above limit, i. e. $\Omega = 5.9 \cdot 10^{-4}$ for large detuning and $\Omega = 0.026$ for moderate detuning. *Momentum-dependent diffusion coefficient.* In Section \[sec:optical-lattice\], we saw that the diffusion coefficient $D_p$ actually depends on momentum, but we so far ignored this since it does not change the qualitative results of our analysis. The momentum-dependent diffusion coefficient is given by $$\begin{aligned} D_{\tilde{p}} = D_0 + \frac{D_1}{1+\frac{\tilde{p}^2}{p_c^2}} ,\end{aligned}$$ which reduces to the momentum-independent case for $D_1 = 0$. The momentum-dependent part is small for large momenta $|\tilde{p}| \ll p_c$, where the nonlinearity of the friction force becomes important. Precisely for this reason, the qualitative results, which hinge on this nonlinearity, are not changed by the introduction of a finite $D_1$, in particular the parameter $D = D_0/(\gamma p_c^2)$, which controls the behavior is unchanged. As long as the ratio $\mathfrak{D} = D_1/D_0$ is not too large, we can repeat the same procedure as in Section \[sec:small-D\] for the small-$D$ expansion including the momentum-dependent diffusion coefficient, see Appendix \[app:d1\]. As a result, we see that a nonzero $D_1$ essentially amplifies the deviations from the Boltzmann-Gibbs behavior. This is intuitively reasonable, as an enhanced diffusion coefficient at small momenta will push the particles to higher momenta, where they feel the effect of the nonlinear friction more strongly. In terms of the lattice parameters, the ratio $D_1/D_0$ can be expressed as [@cas90] $$\begin{aligned} \mathfrak{D} = \frac{D_1}{D_0} = \frac{18}{11} \frac{\delta^2}{\Gamma^2} .\end{aligned}$$ For moderate detuning, $D_1$ is thus comparable to $D_0$, $\mathfrak{D} \approx 3.7$; for large detuning, $D_1$ is much bigger than $D_0$, $\mathfrak{D} \approx 160$. In both cases the modified small-$D$ expansion presented in Appendix \[app:d1\] is not applicable and we have to rely on numerical simulations. We note, however, that the trend of increasing deviations from Boltzmann-Gibbs behavior with increasing $\mathfrak{D}$ persists. *Stark shifts.* In the previous discussion, we treated the harmonic confinement as an additional force that acts on the atoms but does not otherwise affect the friction or diffusion terms in the Kramers equation . Experimentally, the confinement could be realized by a static, or low-frequency, electric field. Then, the atomic states experience a position-dependent light shift due to the interaction with the trapping field. This light shift affects the detuning $\delta$ and thus the cooling mechanism. For a two-level atom, the DC Stark shift to the ground state is can be obtained from second order perturbation theory [@gri00], $$\begin{aligned} \delta \epsilon_{g} = \frac{|\langle e | H_{i} | g \rangle|^2}{\epsilon_{e}-\epsilon_{g}},\end{aligned}$$ where $g$ and $e$ denote the ground and excited state, $H_{I}$ is the interaction Hamiltonian and $(\delta)\epsilon$ the respective energy (shift). In a static electric field of amplitude $\mathcal{E}$, the interaction Hamiltonian reads $$\begin{aligned} H_i = - \vec{\mathcal{E}} \vec{\mu} ,\end{aligned}$$ with the atomic dipole operator $\mu$. This gives us $$\begin{aligned} \Delta \epsilon_{g} = - \frac{1}{2} \alpha_0 \mathcal{E}^2 \qquad \text{with} \qquad \alpha_0 = -2 \frac{|\langle e | \mu | g \rangle|^2}{\epsilon_{e}-\epsilon_{g}},\end{aligned}$$ $\alpha_0$ being the static polarizability of the atom. This quasistatic approximation is valid when the frequency of the electric field that constitutes the confinement is much lower than the frequency of the atomic transition $\omega_0$ [@gri00]. Introducing the intensity $\mathcal{I} = 2 c \varepsilon_0 \mathcal{E}^2$, this reads $$\begin{aligned} \Delta \epsilon_{g} = - \frac{\alpha_0}{4 c \epsilon_0} \mathcal{I},\end{aligned}$$ or in terms of the potential $U = -\alpha_0/(2 \epsilon_0 c) \mathcal{I}$, $$\begin{aligned} \Delta \epsilon_{g}(\tilde{x}) = \frac{1}{2} U(\tilde{x}).\end{aligned}$$ The excited state has the opposite shift $\Delta \epsilon_e = -\Delta \epsilon_g$. In our case, the trapping potential is harmonic, and we get the estimate estimate for the total light shift $\Delta \epsilon = \Delta \epsilon_g - \Delta \epsilon_e$ relative to the detuning, $$\begin{aligned} \frac{\Delta \epsilon}{\hbar \delta} = \frac{1}{\hbar \delta} U(\tilde{x}) = \frac{m \omega^2 }{2 \hbar \delta} \tilde{x}^2 = \frac{m v_c^2}{2 \hbar \delta} x^2,\end{aligned}$$ where in the last step, we replaced the physical position $\tilde{x}$ by the our dimensionless variable $x$. Using that $\langle x^2 \rangle$ is of order one, we find for the average relative light shift $$\begin{aligned} \frac{\Delta \epsilon}{\hbar \delta} \approx \frac{1}{324} \frac{\hbar \Gamma}{E_r} \frac{\Gamma s_0^2}{\delta} \approx 1.2 \frac{\Gamma s_0^2}{\delta} ,\end{aligned}$$ where the rightmost expression is again for Cesium. For large detuning, we find that the relative light shifts are approximately $3.1\cdot 10^{-4}$. In the large-detuning regime, we can thus treat the confining potential as a classical force without worrying about the additional induced light shifts. For moderate detuning we find relative shifts of around $0.09$, which are small but might still lead to a position-dependent detuning and thus cooling rate. From this point of view, it is thus advantageous to work in the large-detuning regime. *Numerical results.* Since our limiting expansions are not valid for the parameters stated above, let us discuss some results from numerical Langevin simulations with the above parameters. Let us first discuss the case of moderate detuning. Here, we find a stationary average potential energy $\langle E_p \rangle \approx 90 E_r$ and average kinetic energy $\langle E_k \rangle \approx 60 E_r$, which yields an equipartition ratio of $1.5$. The potential energy is thus significantly enhanced with respect to the kinetic one. This is mirrored in a discernible difference between the position and momentum distributions, see Fig. \[fig:exper-xpdist1\]. Interestingly, the average kinetic energy is smaller than the value for the same parameters without the confining potential, $\langle E_k \rangle_\text{free} \approx 83 E_r$, which was obtained in Refs. [@hod95; @per00]. We further find a small but discernible asymmetry in the phase-space density, resulting in an asymmetry parameter of $\eta \approx 1.02$. The power law tails of the energy distribution can only be observed at very large energies $E \gtrsim 10^3 E_r$, see Fig. \[fig:exper-edist\]. For large detuning, this asymmetry vanishes within the accuracy of the simulations, however, both the potential and kinetic energy are reduced considerably, to $\langle E_p \rangle \approx 68 E_r$ and $\langle E_k \rangle \approx 21 E_r$. The imbalance between potential and kinetic energy becomes even larger at an equipartition ratio of $3.2$ and the position and momentum distribution differ significantly, see Fig. \[fig:exper-xpdist2\]. Most strikingly, the average kinetic energy is well below the minimum value for Sisyphus cooling without confinement, $\langle E_k \rangle_\text{free} \approx 66 E_r$, within the semiclassical picture of course [@cas90; @per00]. At large detuning, the power law tails of the energy distribution cannot be observed for reasonable values of the total energy, see Fig. \[fig:exper-edist\]. We note that a larger potential energy compared to the kinetic one was found in terms of a diffusion approximation in Ref. [@per00]. However, this approximation predicts the divergence of potential and kinetic energy at different values of $D$, which is in contradiction to the result Eq.  that is confirmed by the numerical simulations. Further, within the diffusion approximation, the kinetic energy is unaffected by the confinement and always corresponds to the result for unconfined atoms, again in contradiction to our findings. ![Energy probability density for model Cesium as a function of energy in units of the recoil energy. The red line corresponds to medium detuning ($\Delta = 1.5 \Gamma$, $I = 3.3 I_s$), the blue line to large detuning ($\Delta = 10 \Gamma$, $I = 20 I_s$). The dashed line is the expected asymptotic power law $P_E(E) \propto E^{-1/D}$ for large energies. \[fig:exper-edist\]](exper1-edist.png){width="47.00000%"} ![Momentum (red) and position (blue) probability density versus positon/momentum in units of the recoil momentum $p_r = \hbar k/m$ and the lattice wavelength $\lambda = 2 \pi/k$ for medium detuning. The black line is the momentum probability density for the same parameters without the confining potential. \[fig:exper-xpdist1\]](exper1-xpdist.png){width="47.00000%"} ![Momentum (red) and position (blue) probability density versus positon/momentum in units of the recoil momentum $p_r = \hbar k/m$ and the lattice wavelength $\lambda = 2 \pi/k$ for large detuning. The black line is the momentum probability density for the same parameters without the confining potential. \[fig:exper-xpdist2\]](exper2-xpdist.png){width="47.00000%"} *Measurement protocol.* A possible protocol to measure both the position and momentum density of the atoms independently may be the following: Taking into account the above considerations, the atoms could trapped within a confining field, provided by a laser beam that is very far red-detuned from the atomic transitions, or by an electrostatic potential. For simplicity one could trap the atoms in all spatial directions but only have the optical lattice along one direction, similar to the setup employed in Ref. [@sag12] to observe the diffusion of the atoms. Then we let the atoms relax into a stationary state, the characteristic time scale for this is estimated from our simulations being around 10 ms for $\Delta = 1.5 \Gamma$ and around 100 ms for $\Delta = 10 \Gamma$. In the stationary state, we could then determine the position density of the cloud, by illuminating and imaging it. In order to obtain the momentum density, we turn off both the confining field and the optical lattice rapidly and take snapshots of the ensuing ballistic expansion of the cloud. Once the size of the cloud is significantly bigger than its stationary extension in the trapped state, the position of the atoms will essentially be proportional to their momentum at the time of release times the flight time and we can thus extract the momentum density in the trapped state. We note that also this kind of time-of-flight measurement was applied before in [@dou06] to determine the steady momentum density of Sisyphus cooled atoms without confinement. From the position- and momentum- density one can then obtain the average potential and kinetic energy by integration. Conclusion \[sec:conclusion\] ============================= In the preceding sections, we have discussed the nonequilibrium stationary state that results from confining cold atoms that are subject to Sisyphus cooling. We obtained analytical results in terms of expansions in three limiting regimes. The common theme of all these expansions is that the leading order phase-space density is a function of the Hamiltonian only. The advantage of this description is obvious: Instead of position and momentum, we only need the total energy to characterize the stationary state of the system. For small values of the parameter $D$, which corresponds to deep optical lattices in the physical system, we indeed recover a Boltzmann-Gibbs-like density describing the center, small-energy part of the phase-space density. However, this is only correct to leading order, as the correction terms for any finite $D$ break the equivalence between energy and probability and we need to describe the system in terms of position and momentum. A measurable consequence is that the average potential energy of the atoms will be larger than their average kinetic energy. This expansion does not, however, make any predictions about the tails of the probability density, which are always exponential within the approximation, while in reality, they are power laws. A complementary viewpoint is the underdamped limit, which is attained either when the confining potential is strong relative to the damping coefficient, or for very large energies. In this underdamped limit, the atoms perform oscillations in the confining potential, with an energy that changes slowly in time. A natural consequence is that the total energy describes the statistics to leading order. In this limit, we ascertain that the phase-space density and its marginals are indeed heavy tailed power-law densities. We stress that the exponent of the power-law is more negative than for the case without the confining potential, implying that confining the atoms does alter their statistics in a qualitative manner. Once we go beyond the leading order, we again find that the correction terms break the equivalence of energy and probability. Our analytical results are based on perturbation theory. In the non-perturbative regime, where we expect larger effects, we performed numerical simulations that confirm that the relevant features persist even beyond the validity of the expansions. In particular, the imbalance between potential and kinetic energy can potentially be very large, so that measuring the average potential energy in the trapped state does not allow one to infer their kinetic energy. Rephrasing this in terms of temperature, this means that the temperature extracted from a measurement of the potential energy will be larger than the actual kinetic temperature of the atoms, which often is the quantity of interest. The fact that the minimum kinetic temperature in the trapped state can actually be lower than without the confining potential is surprising and would not have been obtained from naively treating the system within the thermal Boltzmann-Gibbs approximation. Our results are thus interesting both from the viewpoint of statistical mechanics and cold atom physics. From a statistical mechanics perspective we have here a system that closely approximates the equivalence of energy and probability, that is a central hypothesis of equilibrium statistical mechanics, in certain limits, which, however, can also show strong deviations from this paradigm in other regimes. Predicting these intricate statistics for a well-controlled and existing experimental systems means that their experimental confirmation is within reach. As for the more practical application, by combining trapping and Sisyphus cooling, a continuously cooled atomic cloud with a well-defined stationary state can be realized. Our results suggest that in this situation, the kinetic temperature of the cloud can be even lower than without the trapping, as the confinement improves the efficiency of the cooling mechanism. **Acknowledgments.** This work was supported by the Israel Science Foundation. A. D. was partly employed as an International Research Fellow of the Japan Society for the Promotion of Science. Coefficients for the small-$D$-expansion \[app:small-d\] ======================================================== The coefficients in the expansion can be obtained by plugging the expansion into Eq. , $$\begin{aligned} \sum_{k=0}^{4} \sum_{l=0}^{4-k} a^{(1)}_{k l} &\mathcal{L}_0 z^{k} u^{l} - 3 u^2 + u^4 = 0 \\ \text{with} \qquad &\mathcal{L}_0 = \Omega(z \partial_{u} - u \partial_{z}) - u \partial_{u} + \partial_{u}^2 . \nonumber\end{aligned}$$ Acting with $\mathcal{L}_0$, $$\begin{aligned} \mathcal{L}_0 z^k u^l = \Omega (l z^{k+1} u^{l-1} - k z^{k-1} u^{l+1} ) - l u^l + l (l-1) u^{l-2},\end{aligned}$$ and demanding that the resulting equation should be valid independent of $u$ and $z$, we can equate the coefficient of any specific combination of powers $z^k u^l$ to zero. The resulting equations can be solved for the coefficients $a^{(1)}_{k l}$. We can drastically reduce the number of equations by noting the symmetry of our problem. The original equation for $g(z,u)$, is, just like the Fokker-Planck equation , invariant under an inversion $(z,u) \rightarrow (-z,-u)$. As the solution should obey the same symmetry, this immediately implies that all coefficients where $k+l$ is odd should be zero. The remaining equations then determine all of the remaining coefficients save $a^{(1)}_{0 0}$, $$\begin{aligned} a^{(1)}_{k l} &= 0 \qquad \text{for} \ k + l \ \text{odd}, \ \text{or} \ k + l > 4, \nonumber \\ a^{(1)}_{0 2} &= a^{(1)}_{3 1} = 0, \quad a^{(1)}_{0 4} = \frac{3 (1 + \Omega^2)}{4 (3 + 4 \Omega^2)}, \nonumber \\ a^{(1)}_{1 1} &= - \frac{3 \Omega}{4 (3 + 4 \Omega^2)}, \quad a^{(1)}_{1 3} = \frac{\Omega}{4 (3 + 4 \Omega^2)} \nonumber \\ a^{(1)}_{2 0} &= \frac{9}{2 (3 + 4 \Omega^2)}, \quad a^{(1)}_{2 2} = \frac{3 \Omega^2}{2 (3 + 4 \Omega^2)} \nonumber \\ a^{(1)}_{4 0} &= - \frac{3 \Omega)}{4 (3 + 4 \Omega^2)} .\end{aligned}$$ The final coefficient $a^{(1)}_{0 0}$ is fixed by demanding that the resulting probability density should be normalized, $$\begin{aligned} a^{(1)}_{0 0} = -\frac{3 (9 + 8\Omega^2)}{4 (3 + 4 \Omega^2)} .\end{aligned}$$ Asymptotic matching for the large-energy expansion \[app:large-e\] ================================================================== From the discussion in Section \[sec:large-E\], we know that the expansion outside the strip has singularities at $\alpha = 0, \pi$. This means that the expansion coefficients $g_k(\alpha)$ can have different values in the upper ($p>0$) and lower ($p < 0$) half-plane of phase space. To account for this, we modify Eq. , $$\begin{aligned} P^{\pm}_P(\varepsilon,\alpha) \simeq N \varepsilon^{\beta} \Bigg[1 + \sum_{k = 1}^{K} \frac{g^{\pm}_{k/2}(\alpha)}{\varepsilon^{\frac{k}{2}}} + \mathcal{O}(\varepsilon^{-\frac{K+1}{2}}) \Bigg], \label{appB-expansion-outside}\end{aligned}$$ where “+” denotes $p > 0$ and “–” denotes $p < 0$. Similarly, the approximate solutions to Eq.  inside the strip may be different depending on whether we are in the right $x > 0$ or left $x < 0$ half-plane of phase space. Instead of Eq. , we then write $$\begin{aligned} P^{\pm}_S(x,p) \simeq M |x|^{\gamma} \Bigg[1 + \sum_{k = 1}^{K} \frac{h^{\pm}_{k/2}(p)}{x^k} + \mathcal{O}(x^{-K-1}) \Bigg], \label{appB-expansion-inside}\end{aligned}$$ where in this case “+” stands for $x > 0$ and “–” for $x < 0$. We now plug the expansions respectively into the appropriate Equations respectively , keeping terms up to order $K = 2$. We find to lowest order in $\varepsilon$ $$\begin{aligned} g^{\pm}_{1/2}(\alpha) = \tilde{g}^{\pm}_{1/2}, \qquad h^{\pm}_{1/2}(p) = - \frac{1}{\Omega} \frac{p}{1+p^2} + \tilde{h}^{\pm}_{1/2}\label{appB-inside-order1/2}.\end{aligned}$$ Here $\tilde{g}$ and $\tilde{h}$ denote integration constants. The next order yields $$\begin{aligned} g^{\pm}_{1}(\alpha) &= \frac{(1+D)\sin(2\alpha)-D \cot(\alpha)}{2 D \Omega} - \frac{\alpha \beta}{\Omega} (1+ D\beta) + \tilde{g}^{\pm}_{1} \nonumber \\ h^{\pm}_{1}(p) &= \frac{D + (1-D)p^2}{\Omega^2 (1+p^2)^2} - \frac{p}{\Omega (1+p^2)} \tilde{h}^{\pm}_{1/2} + \tilde{h}^{\pm}_{1} \label{appB-inside-order1} .\end{aligned}$$ Next, we want to match the two solutions inside and outside the strip across the boundary. Replacing $x$ and $p$ by $\varepsilon$ and $\alpha$ we immediately find from the leading order term $$\begin{aligned} \gamma = \frac{\beta}{2}, \qquad M = \frac{N}{|2 \cos(\alpha)|^{\beta}} \simeq \frac{N}{2^{\beta}},\end{aligned}$$ since the expansion Eq.  is only valid for small $\alpha$. For the sub-leading orders, we match all terms of the same order in $\varepsilon$ and $\alpha$. In the upper half-plane, the matching corresponds to taking the limit $p \rightarrow +\infty$ for the solution inside the strip and again replacing $x$ and $p$ by $\varepsilon$ and $\alpha$, $$\begin{aligned} P^{\pm}_S(\sqrt{2 \varepsilon} \cos(\alpha), \sqrt{2 \varepsilon} \sin(\alpha)) \simeq N (2 \epsilon)^{\frac{\beta}{2}} |\cos(\alpha)|^{\beta} \Bigg[ 1 + \frac{\tilde{h}^{\pm}_{1/2}(p)}{\cos(\alpha)} (2 \varepsilon)^{-\frac{1}{2}} + \Big( \frac{\tilde{h}^{\pm}_{1}}{\cos^2(\alpha)} - \frac{1}{\Omega\sin(\alpha) \cos(\alpha)} \Big) (2\varepsilon)^{-1} + \mathcal{O}(\varepsilon^{-\frac{3}{2}}) \Bigg] . \end{aligned}$$ This should match the result from outside the strip for $\alpha \rightarrow 0, \pi$. To order $\varepsilon^{-\frac{1}{2}}$, we then find $$\begin{aligned} \tilde{g}^{+}_{1/2} = \frac{\tilde{h}^{+}_{1/2}}{\sqrt{2}\cos(\alpha)} = \frac{\tilde{h}^{+}_{1/2}}{\sqrt{2}}, \quad &\alpha \rightarrow 0+ \nonumber \\ \tilde{g}^{+}_{1/2} = \frac{\tilde{h}^{-}_{1/2}}{\sqrt{2}\cos(\alpha)} = -\frac{\tilde{h}^{-}_{1/2}}{\sqrt{2}}, \quad &\alpha \rightarrow \pi- .\end{aligned}$$ The terms of order $\varepsilon^{-1}$ give $$\begin{aligned} -\frac{\cot(\alpha)}{2\Omega} + \tilde{g}^{+}_{1} \simeq \frac{\tilde{h}^{+}_{1}}{2\cos^2(\alpha)} - \frac{1}{2\Omega\sin(\alpha)\cos(\alpha)}, \quad &\alpha \rightarrow 0+ \nonumber \\ -\frac{\cot(\alpha)}{2\Omega} - \frac{\pi}{\Omega} \beta (1+D\beta) + \tilde{g}^{+}_{1} \simeq \frac{\tilde{h}^{-}_{1}}{2\cos^2(\alpha)} - \frac{1}{2\Omega\sin(\alpha)\cos(\alpha)}, \quad &\alpha \rightarrow \pi-.\end{aligned}$$ Since the cotangent tends to $+\infty$ for $\alpha \rightarrow 0+$ and to $-\infty$ for $\alpha \rightarrow \pi-$, the singular terms cancel and the last two conditions simplify to $\tilde{g}^{+}_{1} = \tilde{h}^{+}_{1}/2 = -\tilde{h}^{-}_{1}+ \pi \beta (1+D\beta)/\Omega$. Similarly, we find in the lower half-plane, where we take the limit $p \rightarrow -\infty$ from inside the strip and the limits $\alpha \rightarrow 2\pi-$ respectively $\alpha \rightarrow \pi+$ from outside the strip, $\tilde{g}^{-}_{1/2} = \tilde{h}^{+}_{1/2}/\sqrt{2} = -\tilde{h}^{-}_{1/2}/\sqrt{2}$ and $\tilde{g}^{-}_{1} = \tilde{h}^{+}_{1}/2 = \tilde{h}^{-}_{1}/2 + \pi \beta (1+D\beta)/\Omega$. From the lowest two orders in $\varepsilon$, we thus find $$\begin{aligned} \tilde{g}^{+}_{1/2} &= \tilde{g}^{-}_{1/2} = \tilde{h}^{+}_{1/2}/\sqrt{2} = -\tilde{h}^{-}_{1/2}/\sqrt{2} \equiv a_{1/2} \nonumber \\ \tilde{g}^{+}_{1} &= \tilde{g}^{-}_{1} = \tilde{h}^{+}_{1}/2 = \tilde{h}^{-}_{1}/2 + \frac{\pi}{\Omega} \beta (1+D\beta) \equiv a_1 .\end{aligned}$$ We now use the fact that the phase-space probability density has to be invariant under $(x,p) \rightarrow (-x,-p)$, or equivalently $\alpha \rightarrow \alpha + \pi$, we must have for $0 < \alpha < \pi$ outside the strip $P_P^{+}(\epsilon,\alpha) = P_P^{-}(\epsilon,\alpha+\pi)$. To lowest order $\varepsilon^{-\frac{1}{2}}$ this condition is trivially fulfilled. However, at order $\varepsilon^{-1}$, we find a condition on the (as yet unknown) exponent $\beta$, $$\begin{aligned} g^{-}_{1}(\alpha+\pi) - g^{+}_{1}(\alpha) = - \frac{\pi}{\Omega} \beta (1+D\beta) = 0.\end{aligned}$$ Since the solution $\beta = 0$ is does not decay at large $\varepsilon$, we must have $\beta = -1/D$. The two lowest orders do not provide any constraint on the value of $a_{1/2}$ and $a_1$. To find the latter, we need to use the next order $\varepsilon^{-\frac{3}{2}}$, where we find outside the strip $$\begin{aligned} &g^{\pm}_{3/2}(\alpha) = \tilde{g}_{3/2}^{\pm} \nonumber \\ & \; + a_{1/2}\frac{(2 + D)((2+3D)\sin(2\alpha) - 2 \alpha D) - 4 D \cot(\alpha)}{8 D \Omega} .\end{aligned}$$ Inside the strip, we have for large $p$ $$\begin{aligned} h^{\pm}_{3/2}(p) &\simeq - \frac{(2+D)a_{1/2}}{\sqrt{2} D} p^2 + \frac{3(1+D)}{D \Omega} p + \tilde{h}^{\pm}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} + \frac{1 + D - 2 D a_1}{D \Omega} p^{-1} .\end{aligned}$$ Since this coefficient is multiplied by $x^{-3}$, the first two terms are of order $\varepsilon^{-\frac{1}{2}} \alpha^2$ (note that we focus on the expansion around $\alpha = 0$ for simplicity, the case $\alpha = \pi$ can be examined in a similar manner) and have no correspondence outside the strip. The next two terms are of order $\varepsilon^{-\frac{3}{2}}$ and should match the constant contribution from $g^{\pm}_{3/2}(\alpha)$. Note that we get an extra contribution of order $\varepsilon^{-\frac{3}{2}}$ from $h^{\pm}_{1}(p)$, Eq. . Summing up, we have the following conditions $$\begin{aligned} \tilde{g}_{3/2}^{+} - \frac{a_{1/2} \cot(\alpha)}{2 \Omega} &\simeq \frac{1}{(\sqrt{2} \cos(\alpha))^3} \Big( \tilde{h}^{+}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big) - \frac{a_{1/2}}{2 \Omega \cos^2(\alpha) \sin(\alpha)}, \qquad \alpha \rightarrow 0+ \nonumber \\ \tilde{g}_{3/2}^{+} - \frac{a_{1/2} \cot(\alpha)}{2 \Omega} - \frac{a_{1/2} \pi}{4 \Omega} &\simeq \frac{1}{(\sqrt{2} \cos(\alpha))^3} \Big( \tilde{h}^{-}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big) + \frac{a_{1/2}}{2 \Omega \cos^2(\alpha) \sin(\alpha)}, \qquad \alpha \rightarrow \pi- \nonumber \\ \tilde{g}_{3/2}^{-} - \frac{a_{1/2} \cot(\alpha)}{2 \Omega} - \frac{a_{1/2} \pi}{4 \Omega} &\simeq \frac{1}{(\sqrt{2} \cos(\alpha))^3} \Big( \tilde{h}^{-}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big) + \frac{a_{1/2}}{2 \Omega \cos^2(\alpha) \sin(\alpha)}, \qquad \alpha \rightarrow \pi+ \nonumber \\ \tilde{g}_{3/2}^{-} - \frac{a_{1/2} \cot(\alpha)}{2 \Omega} - \frac{a_{1/2} \pi}{2 \Omega} &\simeq \frac{1}{(\sqrt{2} \cos(\alpha))^3} \Big( \tilde{h}^{+}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big) - \frac{a_{1/2}}{2 \Omega \cos^2(\alpha) \sin(\alpha)}, \qquad \alpha \rightarrow 2 \pi - .\end{aligned}$$ The divergent terms once again cancel in the respective limits. The remaining system of equations, $$\begin{aligned} \tilde{g}_{3/2}^{+} &\simeq \frac{1}{2 \sqrt{2}} \Big( \tilde{h}^{+}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big), \qquad \alpha \rightarrow 0+ \nonumber \\ \tilde{g}_{3/2}^{+} - \frac{a_{1/2} \pi}{4 \Omega} &\simeq -\frac{1}{2 \sqrt{2}} \Big( \tilde{h}^{-}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big), \qquad \alpha \rightarrow \pi- \nonumber \\ \tilde{g}_{3/2}^{-} - \frac{a_{1/2} \pi}{4 \Omega} &\simeq -\frac{1}{2 \sqrt{2}} \Big( \tilde{h}^{-}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big), \qquad \alpha \rightarrow \pi+ \nonumber \\ \tilde{g}_{3/2}^{-} - \frac{a_{1/2} \pi}{2 \Omega} &\simeq \frac{1}{2 \sqrt{2}} \Big( \tilde{h}^{+}_{3/2} - \frac{ \pi (2+D)}{2 D \Omega} \Big), \qquad \alpha \rightarrow 2 \pi - ,\end{aligned}$$ is only solvable for $a_{1/2} = -\sqrt{2}/D$. Comparing this to the large-frequency expansion Eq. , this precisely recovers the first correction to the leading order power-law. The corresponding solution for the remaining coefficients reads $$\begin{aligned} \tilde{g}_{3/2}^{+} &\equiv a_{3/2} \nonumber \\ \tilde{g}_{3/2}^{-} &= a_{3/2} - \frac{\sqrt{2} \pi (2+D)}{4 D \Omega} \nonumber \\ \tilde{h}_{3/2}^{+} &= 2 \sqrt{2} a_{3/2} + \frac{\pi (2+D)}{2 D \Omega} \nonumber \\ \tilde{h}_{3/2}^{-} &= -2 \sqrt{2} a_{3/2} - \frac{\pi (2+D)}{2 D \Omega}.\end{aligned}$$ A similar but lengthy argument for the order $\varepsilon^{-2}$ terms fixes the value of $a_1 = 1/D^2$, in agreement with Eq. . Momentum-dependent diffusion coefficient - small $D$ \[app:d1\] =============================================================== In the presence of a the momentum-dependent diffusion coefficient $D_p = D_0 + D_1/(1+p^2/p_\text{c}^2)$, the rescaled Kramers-Fokker-Planck equation reads $$\begin{aligned} \Bigg[ \Omega \bigg( - u \partial_z + z \partial_u \bigg) + \partial_u \bigg(\frac{u}{1+D u^2} + \Big( 1 + \frac{\mathfrak{D}}{1 + D u^2} \Big) \partial_u \bigg) \Bigg] P_D(z,u) = 0, \label{KFP-lattice-D2}\end{aligned}$$ where we defined $\mathfrak{D} = D_0/D_1$ (see Section \[sec:experimental\]). In the limit $D \rightarrow 0$ with $\mathfrak{D}$ fixed, the solution is a slightly modified Boltzmann-Gibbs form, $$\begin{aligned} P_\text{BG}(z,u) = \frac{1}{2 \pi (1+\mathfrak{D})} e^{-\frac{z^2+u^2}{2 (1 + \mathfrak{D})}} .\end{aligned}$$ In complete analogy to Section \[sec:small-D\], we define an auxiliary function $g(z,u)$ via $$\begin{aligned} P_D(z,u) = P_\text{BG}(z,u) g(z,u),\end{aligned}$$ which yields an equation for $g(z,u)$ similar to Eq. , $$\begin{aligned} \Big[\mathcal{L}_0 &+ D \mathcal{L}_1 + D^2 \mathcal{L}_2\Big] g(z,u) = 0 \label{appC-auxiliary} \\ \mathcal{L}_0 &= \kappa^2 \big[ \Omega(z \partial_{u} - u \partial_{z}) - u \partial_{u} + \kappa \partial_{u}^2 \big] \nonumber \\ \mathcal{L}_1 &= 2 \Omega \kappa^2 (z u^2 \partial_{u} - u^3 \partial_{z}) - 3 \kappa u^2 + u^4 - \big(2 \kappa + \kappa^2 \big) u^3 \partial_{u} + \big(\kappa^2 + \kappa^3 \big) u^2 \partial_{u}^2 \nonumber \\ & \qquad + 2 \big(\kappa^2 - \kappa^3 \big) u \partial_u \nonumber \\ \mathcal{L}_2 &= \Omega \kappa^2 (z u^4 \partial_{u} - u^5 \partial_{z}) - \kappa u^4 + u^6 - 2 \kappa u^5 \partial_{u} + \kappa^2 u^4 \partial_{u}^2 \nonumber ,\end{aligned}$$ where we defined $\kappa = 1 + \mathfrak{D}$. The structure of the individual operators is the same as for $\mathfrak{D} = 0$, so that we can still expect an expansion of the form $$\begin{aligned} g(z,u) = 1 + \sum_{n = 1}^{M} D^{n} \sum_{k=0}^{4 n} \sum_{l=0}^{4 n - k} a^{(n)}_{k,l} z^k u^l \label{appC-expansion},\end{aligned}$$ for the solution. However, since the individual operators are now proportional to powers of $\kappa=1+\mathfrak{D}$, the expansion will only work for $\mathfrak{D}$ of at most order $1$. Plugging the expansion into Eq.  and solving for the coefficients, we find up to first order in $D$ $$\begin{aligned} P^{(1)}_D(z,u) = \frac{e^{-\frac{z^2+u^2}{2}}}{2 \pi \kappa} \Bigg[ 1 + \frac{D}{4 (3+4\Omega^2)} \bigg[ \frac{3 u^4}{\kappa^2} + \frac{18 z^2}{\kappa} - 27 + \bigg( \frac{4 u^3 z}{\kappa^2} - \frac{12 u z}{\kappa} \bigg) \Omega + \bigg(\frac{3(u^2+z^2)^2}{\kappa^2} - 24 \bigg) \Omega^2 \bigg] \Bigg],\end{aligned}$$ which for $\mathfrak{D} = 0$ ($\kappa = 1$) reduces to the previous result, Eq. . We can continue this expansion to higher orders, in particular, to second order in $D$, we find for the equipartition ratio Eq.  $$\begin{aligned} \chi^{(2)} = 1 + \frac{6 \kappa}{3 + 4 \Omega^2} D^2 .\end{aligned}$$ Since this expression increases linearly with $\kappa = 1 + \frac{D_1}{D_0}$, this shows that a nonzero $D_1$ enhances the deviations from equipartition. Momentum-dependent diffusion coefficient - large $\Omega$ \[app:d1-omega\] ========================================================================== Just like for the small-$D$ expansion, we can also repeat the large $\Omega$ expansion of Section \[sec:large-freq\] in the presence of a non-zero $D_1$. Employing the notation of Appendix \[app:d1\], the equivalent to the Kramers-Fokker-Planck equation including $D_1$ reads $$\begin{aligned} \Big[\Omega \partial_\alpha &+ \tilde{\mathcal{L}}_{\varepsilon,\alpha} \Big] P_P(\varepsilon,\alpha) = 0 \\ \text{with} \qquad \tilde{\mathcal{L}}_{\varepsilon,\alpha} &= \partial_\alpha \frac{\sin(\alpha)\cos(\alpha)}{1+2 \varepsilon \sin^2(\alpha)} + \partial_\varepsilon \frac{2 \varepsilon \sin^2(\alpha)}{1+2 \varepsilon \sin^2(\alpha)} \nonumber \\ & + D \bigg[ 2 \partial_\alpha \sin(\alpha) \cos(\alpha) \partial_\varepsilon + \frac{1}{2 \varepsilon} \partial_\alpha \cos^2(\alpha) \partial_\alpha + (\sin^2(\alpha) - \cos^2(\alpha)) \partial_\varepsilon + 2 \sin^2(\alpha) \partial_\varepsilon \varepsilon \partial_\varepsilon \bigg] \nonumber \\ & + D \mathfrak{D} \bigg[ \partial_\varepsilon \frac{2 \varepsilon \sin^2(\alpha)}{1 + 2 \varepsilon \sin^2(\alpha)} \partial_\varepsilon + \partial_\varepsilon \frac{\sin(\alpha) \cos(\alpha)}{1 + 2 \varepsilon \sin^2(\alpha)} \partial_\alpha + \partial_\alpha \frac{\sin(\alpha) \cos(\alpha)}{1 + 2 \varepsilon \sin^2(\alpha)} \partial_\varepsilon + \frac{1}{2 \varepsilon} \partial_\alpha \frac{2 \varepsilon \cos^2(\alpha)}{1 + 2 \varepsilon \sin^2(\alpha)} \partial_\alpha\bigg] . \nonumber\end{aligned}$$ As discussed in Section \[sec:large-freq\], for $\Omega \gg 1$ we have to leading order $\partial_\alpha P_P(\varepsilon,\alpha) = 0$. Then we can average the above equation over $\alpha$ to obtain an equation for the energy probability density, similar to Eq. , $$\begin{aligned} \partial_\varepsilon \Bigg[ 1 - \frac{1}{\sqrt{1+2 \varepsilon}} + D \bigg( \varepsilon + \mathfrak{D} \Big(1 - \frac{1}{\sqrt{1+2 \varepsilon}} \Big) \bigg) \partial_\varepsilon \Bigg] P_\varepsilon(\varepsilon) = 0 .\end{aligned}$$ The solution to this equation reads $$\begin{aligned} P_\varepsilon(\varepsilon) &= \frac{1}{Z_\varepsilon} \bigg(2 \mathfrak{D} + \Big(1 + 2 \varepsilon + \sqrt{1 + 2\varepsilon}\Big) \bigg)^{-\frac{1}{D}} \exp \Bigg[ \frac{2}{D} \frac{\text{artanh} \ \Big(\frac{1 + 2\sqrt{1+2\varepsilon}}{\sqrt{ 1 - 8 \mathfrak{D}}}\Big) - \text{artanh} \ \Big(\frac{3}{\sqrt{ 1 - 8 \mathfrak{D}}}\Big)}{\sqrt{1 - 8 \mathfrak{D}}} \Bigg]\end{aligned}$$ for $\mathfrak{D} \neq 1/8$ and $$\begin{aligned} P_\varepsilon(\varepsilon) &= \frac{1}{Z_\varepsilon} \Big( 1 + 2 \sqrt{1+2\varepsilon} \Big)^{-\frac{2}{D}} \exp \bigg[ - \frac{2}{1 + 2 \sqrt{1+2\varepsilon}} \bigg]\end{aligned}$$ for the special case $\mathfrak{D} = 1/8$. In particular, we see that the asymptotic behavior $P_\varepsilon(\epsilon) \sim \varepsilon^{-\frac{1}{D}}$ is not changed by the introduction of $D_1$ respectively $\mathfrak{D}$. Repeating the derivation of Section \[sec:detailed-balance\] for the quantity $\phi$ follows the same lines as outlined there; we find that Eq.  is modified slightly due to the momentum-dependent diffusion coefficient, $$\begin{aligned} \phi &= \partial_p \frac{p}{1+p^2} + \partial_p \Big( D \Big(1 + \frac{\mathfrak{D}}{1+p^2} \Big) \partial_p \ln(P_0(x,p)) \Big) \label{phi-result-d1} ,\end{aligned}$$ with $P_0(x,p) = P_\varepsilon((x^2+p^2)/2)/(2 \pi)$. In particular, this expression now depends explicitly on $D$ and vanishes in the limit $D \rightarrow 0$ with $D_1 = \gamma p_c^2 D \mathfrak{D}$ finite, $$\begin{aligned} \phi \simeq - \frac{1 + p^4 - x^2 - \frac{1+x^2}{\sqrt{1+p^2+x^2}} + p^2 \Big( 2 - \frac{1}{\sqrt{1+x^2+p^2}} + x^2 \Big( 1 + \frac{1}{\sqrt{1+x^2+p^2}}\Big) \Big) }{2(1+p^2)^2} \frac{\gamma p_c^2}{D_1} D + \mathcal{O}(D^2) .\end{aligned}$$ [^1]: The precise relation between $D$ and $U_0$ depends on the details of the atomic transition and the parametrization, however this only causes a slight change in the proportionality constant [@cas90; @mar96]

ArXiv

--- abstract: 'In the last few years one realized that if the state of a bipartite system can be written as $\sum_{i,j}p_{ij}|a_{i}\rangle \langle a_{i}|\otimes |b_{j}\rangle \langle b_{j}|$, where $\{|a_{i}\rangle \}$ and $\{|b_{j}\rangle \}$ form orthonormal basis for the subsystems and $\{p_{ij}\}$ is a probability distribution, then it possesses at most classical correlations. In this article we introduce a nonlinear witness providing a sufficient condition for classicality of correlations (absence of quantum discord) in a broad class of two-qubit systems. Such witness turns out to be necessary and sufficient condition in the case of Bell-diagonal states. We show that the witness introduced here can be readily experimentally implemented in nuclear magnetic resonance setups.' address: - | [Centro de Ciências Naturais e Humanas, Universidade Federal do ABC, R. Santa Adélia 166, 09210-170, Santo André, São Paulo, Brazil]{}\ [email protected] - | [Centro de Ciências Naturais e Humanas, Universidade Federal do ABC, R. Santa Adélia 166, 09210-170, Santo André, São Paulo, Brazil]{}\ [email protected] author: - JONAS MAZIERO - 'ROBERTO M. SERRA' title: 'CLASSICALITY WITNESS FOR TWO-QUBIT STATES' --- \#1[[\#1:]{}]{} Introduction ============ The characterization and quantification of quantum and classical correlations presented in quantum systems are among the principal and more interesting problems in quantum information science (QIS). The seeds for this program can be ascribed to the papers of Einstein, Podolsky and Rosen[@EPR] and Schrödinger[@Schrodinger], who somehow introduced us to the notions of nonlocal correlations and non-separability in composed quantum systems. In this context, early discussions about quantum and classical correlations attributed the difference between the two types of correlations to the nonlocal character of the former, which was associated with the violation of Bell’s inequalities[@Bell]. Subsequently, Werner gave an operational characterization of quantum correlations (at that time considered as synonymous of entanglement) as being the ones that cannot be generated by local operations and classical communication (LOCC)[@Werner]. The development of these ideas led to the today named theory of entanglement, that turned out to be a fruitful branch of research (see Ref. for a review). On the other side, recent studies have shown that entanglement is not the last word regarding the quantumness of correlations in composed quantum systems. Based on information-theoretical concepts, Ollivier and Zurek introduced the so called quantum discord as a measure of quantum correlations and showed that the correlations in a bipartite mixed state can have a quantum character even if it is separable[@Ollivier]. Oppenheim and co-workers came out with a similar conclusion from another quantum correlation measure, the quantum deficit, that was proposed based on a physical perspective[@Oppenheim]. After these early works on this subject, several measures of quantum correlation were proposed and analyzed (see Refs. and for a partial list of references). In particular the quantum discord has received a lot of attention. It has been subjected to experimental tests[@Lanyon; @Xu; @Diogo; @Auccaise], being recognized as a resource in several contexts[@Zurek; @Datta; @Lutz; @Shabani; @Cavalcanti], showing peculiar dynamic behavior under decoherence[@Maziero3; @Maziero4; @Mazzola; @Ferraro], and exhibiting an interesting link to quantum phase transitions[@Sarandy; @Chen; @Maziero5; @Maziero6]. Our goal in this work is not to make directly use of these more general measures of quantum correlations, but to introduce a witness for them. A Classification of Quantum States ================================== A brief summary concerning a classification of bipartite quantum states with relation to its correlations is in order. Any bipartite state that can be created via LOCC is said to be separable and its more general form reads $$\sum_{i}p_{i}\rho _{i}^{a}\otimes \rho _{i}^{b},$$ where $\{p_{i}\}$ is a probability distribution and $\rho _{i}^{a}$ and $\rho _{i}^{b}$ are quantum states for the two subsystems. By definition, a quantum system is entangled if its state is not separable. But contrary to our naive intuition, separability and classicality of correlations are not the same issue. In fact separable states can also possess quantum characteristics in its correlations. Actually, a system is at most classically correlated only if its state can be written as $$\sum_{i,j}p_{ij}|a_{i}\rangle \langle a_{i}|\otimes |b_{j}\rangle \langle b_{j}|, \label{CC}$$ with $\{|a_{i}\rangle \}$ and $\{|b_{j}\rangle \}$ forming orthonormal basis for the two subsystems and $\{p_{ij}\}$ being a probability distribution. Piani and colleagues gave an interesting characterization of this class of states as being the only ones whose correlations can be locally broadcast[@Piani]. The quantum correlations in bipartite states $\rho$ that cannot be cast as in Eq. (\[CC\]) are quantified, for instance, by the so called *quantum discord*, that can be defined as the difference between two classically-equivalent expressions for the mutual information: $$\mathcal{D}(\rho)=\mathcal{I}(\rho)-\max_{\hat{O}}\mathcal{J}(\rho). \label{discord}$$ The quantum mutual information, $$\mathcal{I}(\rho)=S(\rho^{a})+S(\rho^{b})-S(\rho),$$ is a quantifier for the total (quantum plus classical) correlation between the subsystems $a$ and $b$, where $S(\rho^{x})=-\mathrm{Tr}(\rho^{x}\log_{2}\rho^{x})$ is the von Neumann entropy, which measures the uncertainty about the system $x$, and $\rho^{a(b)}=\mathrm{Tr}_{b(a)}(\rho)$. The other version for mutual information reads $$\mathcal{J}(\rho)=S(\rho^{a})-\sum_{j}\mathrm{Pr}(o_{j})S(\rho_{j}^{a}),$$ and quantifies the information obtained about the subsystem $a$ when the observable $\hat{O}=\sum_{j}o_{j}|o_{j}\rangle\langle o_{j}|$ is measured on subsystem $b$. The maximization in Eq. (\[discord\]) is intended to use the observable $\hat{O}$ whose measurement yields the maximal amount of information about the subsystem $a$’s state. The state of the subsystem $a$ after the result $o_{j}$ is obtained, with probability $$\mathrm{Pr}(o_{j})=\mathrm{Tr}(\mathbf{I}^{a}\otimes|o_{j}\rangle\langle o_{j}|\rho),$$ in the measurement of $\hat{O}$ is given by $$\rho_{j}^{a}=\frac{1}{\mathrm{Pr}(o_{j})}\mathrm{Tr}_{b}[(\mathbf{I}^{a}\otimes|o_{j}\rangle\langle o_{j}|)\rho(\mathbf{I}^{a}\otimes|o_{j}\rangle\langle o_{j}|)],$$ where $\mathbf{I}^{ab}$ is the identity operator in the Hilbert’s space $\mathcal{H}_{ab}$. At last in our state classification comes the product states, $$\rho^{a}\otimes\rho^{b},$$ which possesses no correlations at all, where $\rho ^{a}$ and $\rho ^{b}$ are density operators of the subsystems. A typical problem in QIS is to quantify how far a given state and the aforementioned states are one from the another or simply to distinguish them. The former task is ordinarily performed using measures of correlation, that are ultimately obtained from experimental state tomography. But in some circumstances it is enough to know, for example, if the correlations in the system are classical or quantum. In these situations we would like to witness classicality, in analogy to what is done with entanglement witnesses[@Terhal], without doing the usually demanding quantum state tomography and also avoiding the generally hard numerical optimization procedures needed for the calculation of measures of quantum correlation. However, in contrast to the space of separable states, the set of classically correlated states is not convex. For that reason, as was proved by Rahimi and SaiToh[@Rahimi1], a linear witness cannot do the job in general. In Ref. , a nonlinear quantum correlation witness, whose calculation involves a maximization over the set of classical correlated states, was proposed and computed for some very specific cases. In this article we introduce a nonlinear witness providing a sufficient condition for the classicality of correlations in a wide class of two-qubit states. For Bell-diagonal states, such witness is necessary and sufficient condition for the absence of quantumness in the correlations of the system. As will be shown in the sequence, the classicality witness introduced here can be readily implemented in experimental contexts such as, for example, in nuclear magnetic resonance (NMR) setups. Witness for Quantum Correlations ================================ Here we are interested in systems whose state takes the following form: $$\rho =\frac{1}{4}\left(\mathbf{I}^{ab}+\vec{x}\ldotp\vec{\sigma}^{a}\otimes \mathbf{I}^{b}+\mathbf{I}^{a}\otimes \vec{y}\ldotp\vec{\sigma}^{b}+\sum_{i=1}^{3}c_{i}\sigma _{i}^{a}\otimes\sigma _{i}^{b}\right), \label{2qubit}$$ where $c_{i}\in \Re $, $\vec{x},\vec{y}\in \Re ^{3}$ are constrained such that the eigenvalues of $\rho $ are not negative. Besides $\mathbf{I}^{k}$ is the identity operator acting on state space of system $k=a,b,ab$ and $\vec{\sigma}^{j}=(\sigma _{1}^{j},\sigma _{2}^{j},\sigma _{3}^{j})$ with $j=a,b$, where $\sigma _{1}^{j}=|0\rangle \langle 1|+|1\rangle \langle 0|$, $\sigma _{2}^{j}=-i(|0\rangle \langle 1|-|1\rangle \langle 0|)$, and $\sigma_{3}^{j}=|0\rangle \langle 0|-|1\rangle \langle 1|$ are the Pauli operators acting in the state space of the subsystem $j$ and $\{|0\rangle ,|1\rangle \}$ is the usual computational basis. It is worthwhile to mention that this class of states is quite general and it appears routinely in several theoretical[@Maziero3; @Maziero4; @Sarandy] and experimental (as, for example, in optical and NMR setups[@Xu; @Diogo]) contexts. Let us regard observables represented by the following set of hermitian operators: $$\begin{aligned} \hat{O}_{i}&=&\sigma_{i}^{a}\otimes\sigma_{i}^{b}, \\ \hat{O}_{4}&=&\vec{z}\ldotp\vec{\sigma}^{a}\otimes\mathbf{I}^{b}+\mathbf{I} ^{a}\otimes\vec{w}\ldotp\vec{\sigma}^{b},\end{aligned}$$ where $i=1,2,3$ and $\vec{z},\vec{w}\in\Re^{3}$ with $||\vec{z}||=||\vec{w}||=1$. We observe that the directions $\vec{z}$ and $\vec{w}$ should be picked out randomly. Now we consider a relation among these observables as follows $$W_{\rho}=\sum_{i=1}^{3}\sum_{j=i+1}^{4} |\langle \hat{O}_{i}\rangle_{\rho}\langle \hat{O}_{j}\rangle_{\rho}|, \label{witness}$$ where $\langle \hat{O}_{i}\rangle_{\rho}=\mathrm{Tr}(\hat{O}_{i}\rho)$ and $|x|$ is the absolute value of $x$. We see that $W_{\rho}=0$ if and only if the average value of at least three of the four observables defined above is zero. Thus, if we note that $\langle\hat{O}_{i}\rangle_{\rho}=c_i$ for $i=1,2,3$ and $\langle\hat{O}_{4}\rangle_{\rho}=\vec{z}.\vec{x}+\vec{w}.\vec{y}$, it follows that the only way in which we warrant that $W_{\rho}=0$ (independently of the directions $\vec{z}$ and $\vec{w}$) is if the state $\rho$ assumes the form of one of the following states $$\begin{aligned} \chi_{i}&=&\frac{1}{4}\left(\mathbf{I}^{ab}+c_{i}\sigma_{i}^{a}\otimes\sigma_{i}^{b}\right), \\ \chi_{4}&=&\frac{1}{4}\left(\mathbf{I}^{ab}+ \vec{x}\ldotp\vec{\sigma}^{a}\otimes\mathbf{I}^{b}+\mathbf{I}^{a}\otimes\vec{y}\ldotp\vec{\sigma}^{b}\right),\end{aligned}$$ where $i=1,2,3$. It turns out that all these four states can be straightforwardly set in the form of Eq. (\[CC\]), and hence are at most classically correlated. Therefore $W_{\rho}=0$ is a sufficient condition for $\rho$ to be classically correlated. Moreover, for the so called Bell-diagonal class of states, $$\rho^{bd}=\frac{1}{4}\left(\mathbf{I}^{ab}+\sum_{i=1}^{3}c_{i}\sigma_{i}^{a}\otimes\sigma_{i}^{b}\right), \label{BD}$$ $W_{\rho^{bd}}=0$ is a necessary and sufficient condition for classicality. This result follows by noting that, in this case, $\rho^{bd}$ being classical correlated implies that it must take the form: $$\frac{1}{4}\left(\mathbf{I}^{ab}+c_{i}\sigma_{i}^{a}\otimes\sigma_{i}^{b}\right),$$ with $i=1$ or $i=2$ or $i=3$ (see e.g. Ref. ), and thus implies $W_{\rho^{bd}}=0$. Example: Correlations in the Werner’s State ------------------------------------------- As an example we apply the witness given in Eq. (\[witness\]) to the Werner’s state, $$\rho^{w}=(1-\alpha)\frac{\mathbf{I}^{ab}}{4}+\alpha|\Psi^{-}\rangle\langle\Psi^{-}|, \label{werner}$$ where $0\le\alpha\le1$ and $$|\Psi^{-}\rangle=\frac{1}{\sqrt{2}}(|01\rangle-|10\rangle).$$ By completeness we recall that the Werner’s state violates the CHSH inequality[@CHSH] for $\alpha\ge1/2$ and violates the Peres-Horodecki criterion[@Peres; @Horodeckis1] for $\alpha>1/3$. By a direct calculation one obtain that, for this state, $W_{\rho^{w}}=3\alpha^{2}$. As the Werner’s state belongs to the Bell-diagonal class, and in this case $W_{\rho^{w}}=0$ is necessary and sufficient condition for classicality, it follows that $\rho^{w}$ possesses quantumness in its correlations for all $\alpha\neq0$. It is worth mentioning that the same result is obtained when we use the quantum discord to study the character of correlations in the state (\[werner\])[@Ollivier]. Experimental Implementation =========================== In what follows we present some relations between correlation functions and magnetizations showing that the classicality witness introduced in this article can be readily implemented using the already developed tools of nuclear magnetic resonance (NMR). In these systems the qubits are encoded using nuclear spins and unitary transformations are obtained through suitable sequences of radio-frequency pulses. The natural observables in NMR experiments are the local transverse magnetizations, which are obtained directly from the NMR signal[@NMR]. Let us consider the following equalities $$\begin{aligned} &&\sigma_{1}^{a}\otimes\sigma_{1}^{b}=\mathrm{CNOT}_{a\rightarrow b}(\sigma_{1}^{a}\otimes\mathbf{I}^{b})\mathrm{CNOT}_{a\rightarrow b}, \\ &&\sigma _{2}^{a}\otimes \sigma _{2}^{b} =R_{3}^{\dagger }\left( \sigma _{1}^{a}\otimes\sigma _{1}^{b}\right) R_{3}, \\ &&\sigma _{3}^{a}\otimes \sigma _{3}^{b}=R_{2}^{\dagger }\left( \sigma _{1}^{a}\otimes\sigma _{1}^{b}\right) R_{2},\end{aligned}$$ where $$\begin{aligned} &&\mathrm{CNOT}_{a\rightarrow b} =|0\rangle \langle 0|\otimes \mathbf{I}^{b}+|1\rangle \langle 1|\otimes \sigma _{1}^{b}, \\ &&R_{k} =R_{k}^{a}(\pi /2)\otimes R_{k}^{b}(\pi /2), \\ &&R_{k}^{j}(\pi/2 )=\cos(\pi/4)\mathbf{I}^{j}-i\sin(\pi/4)\sigma_{k}^{j},\end{aligned}$$ with $j=a,b$ and $k=2,3$. Now, if we define the states $$\begin{aligned} \eta &=&\mathrm{CNOT}_{a\rightarrow b}(\rho )\mathrm{CNOT}_{a\rightarrow b}, \label{eta} \\ \zeta &=&\mathrm{CNOT}_{a\rightarrow b}(R_{3}\rho R_{3}^{\dagger})\mathrm{CNOT}_{a\rightarrow b}, \label{zeta} \\ \xi &=&\mathrm{CNOT}_{a\rightarrow b}(R_{2}\rho R_{2}^{\dagger})\mathrm{CNOT}_{a\rightarrow b}, \label{xi}\end{aligned}$$ then the following set of relations between correlation functions and magnetizations is obtained $$\begin{aligned} \langle \sigma _{1}^{a}\otimes \sigma _{1}^{b}\rangle _{\rho } &=&\langle \sigma _{1}^{a}\otimes \mathbf{I}^{b}\rangle _{\eta }, \\ \langle \sigma _{2}^{a}\otimes \sigma _{2}^{b}\rangle _{\rho } &=&\langle \sigma _{1}^{a}\otimes \mathbf{I}^{b}\rangle _{\zeta }, \\ \langle \sigma _{3}^{a}\otimes \sigma _{3}^{b}\rangle _{\rho } &=&\langle \sigma _{1}^{a}\otimes \mathbf{I}^{b}\rangle _{\xi}.\end{aligned}$$ Thus, by looking at these relations, one can note that the classicality witness defined in Eq. (\[witness\]) can be straightforwardly implemented in NMR setups[@NMR]. More specifically, the correlation functions $\langle\sigma_{i}^{a}\otimes\sigma_{i}^{b}\rangle_{\rho}$ are obtained by running the experiment three times. In each realization of the experiment, one must prepare the system in the state $\rho$ and, after doing the local-unitary and controlled-NOT operations to achieve the states $\eta$, $\zeta$, and $\xi$ (as shown in Eqs. (\[eta\])-(\[xi\])), measure the magnetization in the $x$-direction on qubit $a$. Concluding Remarks ================== It is important to stress that a crucial aspect that one should take into account when dealing with witnesses (or criteria) for the presence (or absence) of quantumness in the correlations of a composed system (for recent related works see Refs. , and ) is not only to reduce the number of required experimental settings in relation to those involved in quantum state tomography, but also to escape the hard numerical optimization procedures generally involved in the evaluation of quantum correlation measures. Here we have introduced a nonlinear witness providing a sufficient condition for classicality in the correlations of a broad class of two-qubit systems. Such a witness is necessary and sufficient condition in the case of Bell-diagonal states. The classicality witness introduced in this work has a straightforward experimental implementation, precluding any additional numerical optimization process[@Rahimi1] or ancillary qubits[@Zhang], as required by other proposals. Such feature is an important advantage for experimental bench tests of classicality. Furthermore, by regarding some useful relations between correlation functions and magnetizations we observed that the classicality witness given in Eq. (\[witness\]) can be readily experimentally implemented in, for instance, NMR setups. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful for the funding from UFABC, CAPES, FAPESP, and the Brazilian National Institute for Science and Technology of Quantum Information (INCT-IQ). We thank L. C. Céleri for discussions. [References]{} A. Einstein, B. Podolsky and N. Rosen, *Phys. Rev.* **47** (1935) 777. E. Schrödinger, *Proc. Cambridge Philos. Soc.* **31** (1935) 555. J. S. Bell, *Speakable and Unspeakable in Quantum Mechanics* (Cambridge University Press, Cambridge, 1988). R. F. Werner, *Phys. Rev. A* **40** (1989) 4277. R. Horodecki, P. Horodecki, M. Horodecki and K. Horodecki, *Rev. Mod. Phys.* **81** (2009) 865. H. Ollivier and W. H. Zurek, *Phys. Rev. Lett.* **88** (2001) 017901. J. Oppenheim, M. Horodecki, P. Horodecki and R. Horodecki, *Phys. Rev. Lett.* **89** (2002) 180402. J. Maziero, L. C. Céleri and R. M. Serra, Symmetry aspects of quantum discord, http://arxiv.org/abs/1004.2082 (2010). L. C. Céleri, J. Maziero and R. M. Serra, *Int. J. Quant. Inf.* **9** (2011) 1837. B. P. Lanyon, M. Barbieri, M. P. Almeida and A. G. White, *Phys. Rev. Lett.* **101** (2008) 200501. J.-S. Xu, X.-Y. Xu, C.-F. Li, C.-J. Zhang, X.-B. Zou and G.-C. Guo, *Nat. Commun.* **1** (2010) 7. D. O. Soares-Pinto, L. C. Celeri, R. Auccaise, F. F. Fanchini, E. R. deAzevedo, J. Maziero, T. J. Bonagamba and R. M. Serra, *Phys. Rev. A* **81** (2010) 062118. R. Auccaise, L. C. Celeri, D. O. Soares-Pinto, E. R. deAzevedo, J. Maziero, A. M. Souza, T. J. Bonagamba, R. S. Sarthour, I. S. Oliveira, R. M. Serra, *Phys. Rev. Lett.* **107** (2011) 140403. W. H. Zurek, *Phys. Rev. A* **67** (2003) 012320. A. Datta, A. Shaji and C. M. Caves, *Phys. Rev. Lett.* **100** (2008) 050502. R. Dillenschneider and E. Lutz, *EPL* **88** (2009) 50003. A. Shabani and D. A. Lidar, *Phys. Rev. Lett.* **102** (2009) 100402. D. Cavalcanti, L. Aolita, S. Boixo, K. Modi, M. Piani and A. Winter, *Phys. Rev. A* **83** (2011) 032324. J. Maziero J, L. C. Céleri, R. M. Serra and V. Vedral, *Phys. Rev. A* **80** (2009) 044102. J. Maziero, T. Werlang, F. F. Fanchini, L. C. Celeri and R. M. Serra, *Phys. Rev. A* **81** (2010) 022116. L. Mazzola, J. Piilo and S. Maniscalco, *Phys. Rev. Lett.* **104** (2010) 200401. A. Ferraro, L. Aolita, D. Cavalcanti, F. M. Cucchietti and A. Acin, *Phys. Rev. A* **81** (2010) 052318. M. S. Sarandy, *Phys. Rev. A* **80** (2009) 022108. Y.-X. Chen and S.-W. Li, *Phys. Rev. A* **81** (2010) 032120. J. Maziero, H. C. Guzman, L. C. Celeri, M. S. Sarandy and R. M. Serra, *Phys. Rev. A* **82** (2010) 012106. J. Maziero, L. C. Céleri, R. M. Serra and M. S. Sarandy, Long-range quantum discord in critical spin systems, http://arxiv.org/abs/1012.5926 (2010). M. Piani, P. Horodecki and R. Horodecki, *Phys. Rev. Lett.* **100** (2008) 090502. B. M. Terhal, *Theor. Comput. Sci.* **287** (2002) 313. R. Rahimi and A. SaiToh, *Phys. Rev. A* **82** (2010) 022314. M. D. Lang and C. M. Caves, *Phys. Rev. Lett.* **105** (2010) 150501. J. F. Clauser, M. A. Horne, A. Shimony and R. A. Holt, *Phys. Rev. Lett.* **23** (1969) 880. A. Peres, *Phys. Rev. Lett.* **77** (1996) 1413. M. Horodecki, P. Horodecki and R. Horodecki, *Phys. Lett. A* **223** (1996) 1. I. S. Oliveira, T. J. Bonagamba, R. S. Sarthour, J. C. C. Freitas and E. R. deAzevedo, *NMR Quantum Information Processing* (Elsevier, Amsterdam, 2007) C. Zhang, S. Yu, Q. Chen and C. H. Oh, *Phys. Rev. A* **84** (2011) 032122. B. Bylicka and D. Chruściński, *Phys. Rev. A* **81** (2010) 062102. B. Dakić, V. Vedral and Č Brukner, *Phys. Rev. Lett.* **105** (2010) 190502. L. Chen, E. Chitambar, K. Modi and G. Vacanti, *Phys. Rev. A* **83** (2011) 020101(R).

ArXiv

--- author: - 'ELIZEU SANTOS-NETO DAVID CONDON NAZARENO ANDRADE ADRIANA IAMNITCHI MATEI RIPEANU' bibliography: - 'all\_2012.bib' title: 'Reuse, Temporal Dynamics, Interest Sharing, and Collaboration in Social Tagging Systems ' ---

ArXiv

--- abstract: 'Identifying the sets of operations that can be executed simultaneously is an important problem appearing in many parallel applications. By modeling the operations and their interactions as a graph, one can identify the independent operations by solving a graph coloring problem. Many efficient sequential algorithms are known for this NP-Complete problem, but they are typically unsuitable when the operations and their interactions are distributed in the memory of large parallel computers. On top of an existing distributed-memory graph coloring algorithm, we investigate two compatible techniques in this paper for fast and scalable distributed-memory graph coloring. First, we introduce an improvement for the distributed post-processing operation, called recoloring, which drastically improves the number of colors. We propose a novel and efficient communication scheme for recoloring which enables it to scale gracefully. Recoloring must be seeded with an existing coloring of the graph. Our second contribution is to introduce a randomized color selection strategy for initial coloring which quickly produces solutions of modest quality. We extensively evaluate the impact of our new techniques on existing distributed algorithms and show the time-quality tradeoffs. We show that combining an initial randomized coloring with multiple recoloring iterations yields better quality solutions with the smaller runtime at large scale.' author: - 'Ahmet Erdem Sar[i]{}yüce[^1]' - Erik Saule - 'Ümit V. Çatalyürek' bibliography: - 'paper.bib' title: On Distributed Graph Coloring with Iterative Recoloring --- Introduction ============ In parallel computing, the problem of organizing computations so that no two concurrent procedures access shared resources simultaneously appears often. This problem might be solved by ordering them explicitly so that concurrent accesses can not happen, using concurrency control mechanisms such as locks, lock-free data structures or transactional memory. Even when using concurrency controls mechanisms, it is important to minimize the access to the lock or page of the transactional memory. The problem can be modeled as a graph coloring problem where the vertices of the graphs are operations of the problem and edges represent concurrent accesses to a resource. A coloring of the graph is a partition of the vertices of the graph in a number of independent sets. Minimizing the number of colors (i.e., the number of independent sets) reduces the number of synchronization points in the computation and enhances the efficiency of the parallel computers. Graph coloring appears in many other applications, including, but not limited to, optimization [@CM83], efficient computation of sparse Jacobian and Hessian matrices [@GMP05], preconditioners [@Saad99], iterative solution of sparse linear systems [@Jones94], sparse tiling [@Strout02], printed circuit testing [@Garey_circuit], eigenvalue computation [@Manne98aparallel], frequency assignment [@gamst_freq], parallel numerical computation [@ABC94] and register allocation [@Chaitin] areas. There exist different types of the graph coloring problem. For instance, the distance-$k$ coloring problem requires that vertices separated by less than or equal to $k$ edges to have different colors. In this paper, we are only interested in distance-1 coloring; however, we believe that all the techniques and results presented in this paper can be extended to the other variants of the graph coloring problem. Let $G=(V, E)$ be a graph with ${\ensuremath{|V|}}$ vertices and ${\ensuremath{|E|}}$ edges. Set of neighbors of a vertex $v$ is $adj(v)$; its cardinality, also called the [*degree*]{} of $v$, is [$\delta_{v}$]{}. The degree of the vertex having the most neighbors is ${\ensuremath{\Delta}}= \max_v {\ensuremath{\delta_{v}}}$. A distance-1 coloring ${\ensuremath{C}}:V \rightarrow {\ensuremath{\mathbb{N}}}$ is a function that maps each vertex of the graph to a color (represented by an integer), such that two adjacent vertices get different colors, i.e., $\forall (u, v)\in E, {\ensuremath{C}}(u) \neq {\ensuremath{C}}(v)$. Without loss of generality, the number of colors used is $\max_{u\in V} {\ensuremath{C}}(u)$. Finding a coloring with as few colors as possible is an optimization problem. The problem of deciding whether a graph can be colored with less than $k$ colors is known to be NP-Complete for arbitrary graphs [@matula_SL]. Therefore, finding the minimal number of colors a graph can be colored with (also called the chromatic number of the graph) is NP-Hard. Recently, it has been shown that, for all $\epsilon > 0$, it is NP-Hard to [*approximate*]{} the graph coloring problem within ${\ensuremath{|V|}}^{1-\epsilon}$ [@Zuckerman]. Hence, we are focusing on heuristics which work well in practice on many graphs. We are studying the distributed-memory graph coloring problem in this paper. We focus on obtaining high quality solutions for large-scale scientific parallel applications. In those applications, the computational model (the graph) is already distributed to the nodes of the parallel machine. If the graph is sufficiently small, with a naive approach, one can aggregate it in the memory of a single node and color it there. It would actually be better to take advantage of partitioning in this case, i.e., first coloring the [*interior*]{} vertices (vertices for which all their neighbors are local) in parallel and then color the other, [*boundary vertices*]{} (vertices that have at least one non-local neighbor), sequentially, by aggregating the graph induced by them in the memory of a single processor. However, even if one implements such optimizations, one can still achieve significantly faster solutions by coloring in parallel [@BGMBC-jpdc]. If the graph is too large to fit in the memory of a single computer, coloring in distributed memory is inevitable. Also, people usually use coloring as a tool in contexts where good and quick solutions are desired. Therefore, one cannot afford repartitioning the graph for the sole purpose of coloring, since repartitioning has a higher computational complexity than coloring itself. In this work, we aim to achieve good quality coloring with good runtime in distributed-memory settings. As a starting point, we use the framework presented in [@BGMBC-jpdc]; it mainly relies on processor-local greedy coloring techniques with multiple iterations to converge to a valid solution. Traditionally, the problem of improving the quality of coloring obtained from greedy coloring methods is addressed by considering the vertices in an order that have good properties for coloring. We investigated such techniques in [@HiPC11] and found that processor-local ordering techniques do not yield quality improvement at large scale. Good global ordering techniques exhibit little parallelism and will not yield good runtimes. In this paper, we investigate recoloring, a post-processing operation which refines an existing coloring. We showed in [@HiPC11] that the quality of the solution the recoloring procedure generates is not affected by the scale of the distributed-memory machine. However, the communications required by the algorithm make it non-scalable in terms of the runtime. In this paper, we show how recoloring can be made scalable in terms of runtime by improving its communication scheme. In particular, we use piggybacking to reduce the number of communications. With a fast recoloring method, it becomes efficient to run recoloring multiple times. This allows to potentially start with a faster initial coloring algorithm of lower quality. Pursuing this idea, we investigate the Random-X Fit color selection strategy [@Gebremedhin02paralleldistance-k] for generating a first coloring. Random-X Fit leads to a solution of modest quality but with a balanced color distribution which makes it very suitable for recoloring. We show that combining these different techniques allows to reach better time-quality trade-offs than previously existing algorithms. We extensively evaluate all the techniques we propose at different scales in terms of both quality and runtime. This allow us to identify two sets of parameters “speed” and “quality” which enables the user to achieve the tradeoff she is interested in without having to understand the inner working of our coloring framework. The remaining of the document is organized as follows. Section \[sec:rel\_work\] presents the different coloring algorithms existing for sequential and parallel architectures, including the previous distributed-memory coloring algorithm which is the reference algorithm we use and ordering solutions. Section \[sec:improv\] discusses the techniques we use to improve coloring in distributed-memory architecture. The proposed techniques are experimentally evaluated on real-world graphs and on random graphs in Section \[sec:expe\]. Final conclusions and ideas to improve further are given in Section \[sec:ccl\]. Graph Coloring Algorithms {#sec:rel_work} ========================= Graph coloring is one of the well studied problems in the literature [@Ellis:1989:LVG:74142.74153; @dubr81; @GMP05; @kosow]. Literature is abundant with many different techniques, such as the one that utilizes greedy coloring [@matula_72; @kosow], cliques [@Turner:1988:AKC:48880.48884] and Zykov trees [@dubr81]. In the following sections, we will first present a simple greedy sequential coloring algorithm and how it can be improved with vertex visit orderings and recoloring. We then briefly discuss other shared- and distributed-memory parallel coloring algorithms. Sequential Coloring ------------------- In spite of the existing pessimistic theoretical results and the existence of more complicated algorithms, for many graphs that shows up in practice, solutions that are provably optimal or near optimal can be obtained using a simple [*greedy*]{} algorithm [@CM83]. In this algorithm, the vertices are visited in some [*order*]{} and the smallest permissible color at each iteration is assigned to the vertex. Algorithm \[a:greedy\] gives the pseudocode of this technique. Different ways for improving the resulting number of colors are presented in the literature. Changing the color selection strategy has an impact in the quality of coloring. Choosing the smallest permissible color, as stated in Algorithm \[a:greedy\], is known as the [*First Fit*]{} strategy. Selecting a color based on an initial estimate of the number of colors in a distributed-memory setting is proposed in [@BGMBC-jpdc] and called [*Staggered First Fit*]{}. The [*Least Used*]{} strategy picks (locally) the least used color so far so that a more even color distribution is achieved. Apart from that, Gebremedhin et al. [@Gebremedhin02paralleldistance-k] proposed [*randomized*]{} color selection strategies. Algorithm \[a:greedy\] has two nice properties. First, for any vertex-visit ordering, it produces a coloring with at most $1+{\ensuremath{\Delta}}$ colors. Second, for some vertex-visit orderings it will produce an optimal coloring [@GMP05]. Many heuristics for ordering the vertices have been proposed in the literature [@GMP05]. These heuristics can be grouped into two categories: [*static*]{} and [ *dynamic*]{} orderings. [*Largest First*]{} (LF) and [*Smallest Last*]{} (SL) [@Matula1983] orderings are [*static*]{} orderings, in the sense that the coloring order is obtained before the coloring starts. [*Saturation Degree*]{} [@Brelaz1979] and [*Incidence Degree*]{} orderings are [*dynamic*]{} orderings in which the coloring order of the vertices is obtained while the coloring is done. We refer the reader to [@GMP05] for a summary of these ordering techniques. The LF ordering, introduced by Welsh and Powell [@Welsh01011967], visits the vertices in non-increasing order of their degree. SL obtains the ordering from backwards via selecting a vertex with the minimum degree to be ordered last and removing it from the graph for the rest of the ordering phase. Then, the next vertex with the minimum degree within the remaining graph is selected to be ordered second-to-last. This procedure is repeated until all vertices are ordered. Apart from the different pre-ordering techniques, Culberson [@Culberson92iteratedgreedy] introduces a coloring algorithm, called [*Iterated Greedy*]{} (IG), where starting with an initial greedy coloring, vertices are iteratively **recolored** based on the coloring of the previous iteration. Culberson shows that, if the vertices belonging to the same color class (i.e., the vertices of the same color) at the end of initial coloring are colored consecutively, then the number of colors will either decrease or stay the same. Several different permutations of color classes are considered based on the colors of the vertices in the previous iteration, the number of vertices, degree of the vertices, randomness and combination of them. Culberson suggests that a hybrid approach, which changes the permutation strategy at each recoloring iteration is more effective than applying the same permutation strategy at each round. He proposed using random permutations of the colors to break cycles where number of colors stays the same after some number of iterations that appear when deterministic methods are used exclusively. The first work evaluating the effects of the recoloring scheme on parallel graph coloring is [@Gebremedhin_parallelgraph] where the focus is on shared-memory computers. Also, Goldberg et al. suggest to use recoloring in multigraph edge-coloring [@JGT:JGT3190080115]. Distributed-Memory Parallel Coloring {#sec:jpdc} ------------------------------------ In the literature, Bozdağ et al. [@BGMBC-jpdc] introduced the first scalable distributed-memory parallel graph coloring framework. We believe that, their work is the only distributed-memory coloring algorithm that provides a parallel speedup, therefore our work is based on their algorithm, which we present here briefly. It is assumed that the graph is distributed onto the distributed memory of computers. Each vertex belongs to a single processing unit. At each processing unit, information of the edges connected to any owned vertex is kept. In other words, for an edge $(u, v)$, if both $u$ and $v$ belong to a single processing unit, then only this processor has the information of that edge. Given the edge $(u, v)$, if $u$ and $v$ belong to two different processing units, then $(u, v)$ is only know by these two processing units. Each processing unit colors the vertices it has. If a vertex $u$ has a neighbor vertex $v$, owned by another processing unit, then $u$ and $v$ are [ *boundary*]{} vertices. On the other hand, if all neighbors of a vertex $u$ belong to same processing unit with $u$, then $u$ is called as [*internal*]{} vertex. The coloring procedure is performed in multiple rounds. All the colorless vertices are tentatively colored by the greedy coloring algorithm at each round. Two neighbor vertices, belonging to two different processing units, may end up having the same color at the end of this phase, i.e., a conflict can occur. Such conflicts are independently detected by each processing unit. To resolve a conflict, one of the vertices will keep its color whereas the other one is scheduled to be colored again in the next round. Ties are broken based on a random total ordering, which is obtained beforehand. The algorithm continues to iterate in multiple rounds until there is no conflict. Each round of the coloring procedure is executed in supersteps to reduce the number of conflicts. Each processing unit colors a specified number of owned vertices at each superstep. After that, colors of the boundary vertices (colored in that super step) are exchanged among [*neighbor*]{} processing units. There are two types of coloring procedures based on the communication mechanism between processing units. Coloring is called as [*synchronous*]{}, if a processing unit waits for its neighbors to complete their super step before beginning the next one. [*Synchronous*]{} coloring guarantees that two vertices can be in conflict only if they are colored in the same superstep. The other option, in which no waiting mechanism is enforced, is named as [*asynchronous*]{}. The superstep size matters for the coloring procedure, since a larger size decreases the number of exchanged messages with likely high number of conflicts, whereas the smaller size increases the exchanged message number with expectedly low number of conflicts. ### Vertex-visit Ordering The orderings considered in [@BGMBC-jpdc] only consider the partitioning of the graph and not the properties of the graph itself. Three orderings were investigated in their work, coloring internal vertices first, boundary vertices first and coloring the vertices in the order they are stored in the memory (which was called [*unordered*]{} in [@BGMBC-jpdc], here we will call it [*Natural*]{} ordering). In [@HiPC11], we investigated two more successful ordering heuristics, namely [*Largest First*]{} (LF) and [*Smallest Last*]{} (SL) techniques. The LF ordering can be computed in $O({\ensuremath{|V|}})$ time [@Welsh01011967]. Using a carefully designed bucket data structure allows to implement SL with a complexity of $O({\ensuremath{|E|}})$ [@Matula1983]. In distributed memory, we let each processor compute an ordering of the graph based on the knowledge it has and therefore it is not guaranteed that the coloring computed on a single processor and on multiple processors will be the same. We also verified this fact by scalability experiments in [@HiPC11]. We showed that LF ordering provides less number of colors than sequential Natural ordering on less than 8 processors. SL obtains much fewer number of colors. However, when the number of processors increases, the advantages of LF and SL orderings disappear. On 512 processors, the choice of vertex-visit ordering does not yield a significant difference in number of colors. So, vertex-visit orderings are not beneficial for quality at large scale since the number of internal vertices decreases when the scale increases and the ordering is done locally at each processor. Distributed-Memory Iterative Recoloring {#sec:improv} ======================================= Culberson [@Culberson92iteratedgreedy] presented the use of iterative recoloring for improving the number of colors in a sequential algorithm. In [@HiPC11], we extended this work to distributed-memory graph coloring. The recoloring idea naturally fits distributed-memory graph coloring. Independents sets obtained at the end of the initial coloring are used to efficiently color the graph. Our [*recoloring algorithm*]{} (RC) proceeds in as many steps as the number of colors in the initial solution. All the vertices in the same color class (i.e., having the same color in the previous coloring round) are colored at the same step in the recoloring process. Then the processors exchange the color information with their neighboring processors at the end of the step. Notice that, although a processor does not have any vertices with that color, it will wait for other processors to finish coloring at that step. This procedure guarantees that no conflict is created by the end of recoloring. Therefore, recoloring in sequential and in distributed memory lead to the same solution, if the initial coloring is same, making the recoloring scalable in terms of number of colors. However, the algorithm is fairly synchronous since a processor can not start the $i$-th step before its neighbors finish their $(i-1)$-th step. Moreover, there is no guarantee that two processors will have a similar number of vertices in each color, thus potentially leading to a load imbalance. To prevent some potential load imbalance due to synchronous execution of RC, we also propose a second recoloring approach, named [*asynchronous recoloring*]{} (aRC). In this approach, each processor computes their vertex-visit orderings independently using the initial coloring, and apply a second parallel coloring with this new vertex-visit ordering using the algorithm recalled in Section \[sec:jpdc\]. Note that, conflicts can happen at the end of this procedure, so this second parallel coloring step proceeds in conflict resolution rounds. We expect that this approach will not be as good as synchronous recoloring in terms of number of colors, but it might be beneficial for trading the quality for runtime. In the recoloring process, all the vertices in the same color class must be colored in a consecutive manner, but one can choose any permutation of the color classes and this permutation affects the number of colors well. Therefore different permutations of the color classes should be taken into account for sequential and parallel recoloring. We considered four permutations of color classes: [ *Reverse*]{} order (RV) of colors, [*Non-Increasing*]{} number of vertices (NI), where the color classes are ordered in the non-increasing order of their vertex counts, [*Non-Decreasing number*]{} of vertices (ND), which is similarly derived, and [ *Random*]{} order of color classes where the color classes are ordered randomly using the Knuth shuffling procedure in linear time [@Culberson92iteratedgreedy]. We compute global NI and ND permutations by communicating between processors before coloring to exchange the number of vertices at each color class. Improvements on the Communication Scheme of Synchronous Recoloring {#sec:improved_comm} ------------------------------------------------------------------ The communication scheme in [*synchronous recoloring*]{} algorithm results in fine-grained communications where the number of messages are high but each of these messages is small. This behavior lowers network performance and results in slower execution. To improve this fine-grained communication scheme, we apply piggybacking techniques. In [*synchronous recoloring*]{}, vertices belonging to same color class are colored in the same step and colors are communicated at the end of that step immediately. However, immediate communication is not necessary for two boundary vertices that are not colored in consecutive steps. For example, for two neighbor boundary vertices $a$ and $b$, belonging to different processors, if $color\_order(a) \textgreater color\_order(b)$ (in other words if $a$ will be colored later than $b$), then the coloring of $b$ does not depend on the coloring of $a$. Of course $b$ will need the new color of $a$ during the next recoloring iteration, so sending the color of $a$ to $b$’s processor can be deferred to the end of the recoloring iteration. For the reverse case, a similar technique can be used. Assume again that $color\_order(a) \textgreater color\_order(b)$, then $a$ needs to know the new color of $b$ before being colored. So, sending the color of $b$ can happen at any time between step $color\_order(a)$ and step $color\_order(b)$. A processor $P_1$ accumulates the color information in a buffer to send to a processor $P_2$. $P_1$ only sends the whole buffer at the color step before the step where $P_2$ needs any of the information contained in the whole buffer. This way, the color information of vertices are piggybacked and sent in a minimum number of messages which improves the performance of the communication subsystem. In order to be able to apply the piggybacking technique, each processor needs to have the knowledge of when and from whom to receive messages for each iteration. For this purpose, there is a need for pre-communication at the beginning of each recoloring iteration among processors so that each get the information of from whom to receive at which step. The only information processors should send to each other is the list of color classes when they will communicate with each other. If $P_1$ is to send something to $P_2$ at color class $c$, then $P_2$ will know this at the end of the pre-communication and it will wait an incoming message from $P_1$ at color class $c$. As expected, there will be an overhead for this process. We investigate the gained time by piggybacking as well as the overhead time incurred by pre-communication in the experiments section. Figure \[fig:piggy\] presents the situation between two processors with a total of six boundary vertices after initial coloring. Vertices that are not on the boundary between processors $P_1$ and $P_2$ are not depicted. Without the use of piggybacking, $P_1$ sends three messages to $P_2$ containing the information of a single vertex at the end of steps 1, 3 and 12. It also needs to send empty messages during the other steps. $P_2$ needs to send similar three messages. Using piggybacking, $P_1$ sends a message to $P_2$ at the end of step 4; the message contains both the color information of vertices $v_B$ and $v_C$. Then $P_2$ sends the color information of vertex $v_D$. Finally, $P_1$ will send the color information of vertex $v_A$ at the end of the recoloring iteration and $P_2$ will send the color information of both $v_E$ and $v_F$. Notice that in this setting no empty messages are exchanged. Piggybacking is able to remove all empty messages and furthermore reduces the total number of non-empty messages from 6 to 4 between $P_1$ and $P_2$. Random X-Fit ------------ We expect that synchronous recoloring will be scalable in terms of number of colors. Piggybacking will make recoloring scalable in terms of runtime, especially if the number of vertices per color is roughly balanced. So, we investigate a fast coloring alternative for the initial coloring. In [@BGMBC-jpdc], Bozdağ et al. proposed two different color selection strategies; First Fit and Staggered First Fit. We propose another color selection strategy, Random-X Fit, which uniformly selects a random color from the first X available colors [@Gebremedhin02paralleldistance-k]. We expect that Random-X Fit will significantly reduce the number of conflicts in the coloring procedure leading to a fast coloring. Moreover, it should balance the number of vertices in each color which should help the recoloring procedure being fast. Random-X Fit is expected to give solutions with high number of colors. But, we believe that recoloring is strong enough to fix any bad number of colors obtained in the initial coloring. And the coordinated use of Random-X Fit and recoloring should make the procedure fast. Experiments {#sec:expe} =========== In the experiments we study the effectiveness of the recoloring procedure in Section \[subsec:recoloring\] and prove three points: - recoloring can significantly improved solution of low quality in a few iterations (Section \[subsubsec:seq\]) - our piggybacking techniques makes recoloring scale much better (Section \[sec:improved\_comm\_exp\]) - and in a distributed memory setting, recoloring improves significantly the quality of the solution and incurs a low runtime cost allowing to use multiple recoloring iterations (Section \[sec:dist\]). Then we show in Section \[sec:usingrand\] that using a Random X-Fit coloring leads to a low quality solution but, when coupled with recoloring, it leads to better quality solution in a smaller amount of time than using a first fit coloring. This allows us to study the time-quality trade-off of the coloring problem and we exhibit two sets of parameters that can be used to obtain either a fast solution “speed” or a good solution “quality”. Before showing experimental results, we present how the experiments are conducted in Section \[sec:expesetting\]. Experimental Setting {#sec:expesetting} -------------------- We implemented all the algorithms in Zoltan [@zoltanug_v3], an MPI-based C library for parallel partitioning, load balancing, coloring and data management services for distributed-memory systems. For the graphs, we partitioned them by ParMETIS [@parmetis31] version 3.1.1 or with a simple block partitioning onto the parallel platform. All of our experiments are performed on an in-house cluster with 64 computing nodes. There are two Intel Xeon E5520 (quad-core clocked at 2.27GHz) processors at each node along with 48GB of main memory, and 500 GB of local hard disk. Interconnection between the nodes are done through 20Gbps DDR oversubscribed InfiniBand. Nodes run CentOS with the Linux kernel 2.6.32. The C code is compiled with GCC 4.1.2 using the -O2 optimization flag. We tried different MPI implementations and MVAPICH2 in version 1.6 is chosen to utilize the InfiniBand interconnect efficiently. Experiments are performed on number of processors which are powers of 2 from 1 to 512 processors. Each physical machine has 8 cores. When we do the allocation of processors on the cluster, we first use processors on different nodes to highlight distributed memory issue. That means, when 64 processors are used, each allocated processor (core) is on a different machine. Using 128 processors allocates 2 cores per machine and an allocation of 512 processors uses the 8 cores of the 64 machines. We have shown in [@HiPC11] that this setting is the fastest configuration. We run the experiments on six real-world application graphs and three synthetically generated graphs. The real-world graphs are from different application areas including linear car analysis, finite element, structural engineering and automotive industry [@GM00; @SH04]. We obtained them from the University of Florida Sparse Matrix Collection[^2] and the Parasol project. Table \[tab:prop\_real\] gives the list of the graphs and their main properties. We also listed the number of colors obtained with a sequential run of the three vertex-visit orderings in the table. Lastly, runtime of the [*Natural*]{} coloring in sequential is given in table. One important thing to note is that the biggest real-world graph takes less than half a second to color sequentially using a [*Natural*]{} ordering, which means that the distributed coloring of all the graphs very challenging. --------- --------- ------------ ----- ---- ---- ---- --------- auto 448,695 3,314,611 37 13 12 10 0.1103s bmw3\_2 227,362 5,530,634 335 48 48 37 0.0836s hood 220,542 4,837,440 76 40 39 34 0.0752s ldoor 952,203 20,770,807 76 42 42 34 0.3307s msdoor 415,863 9,378,650 76 42 42 35 0.1458s pwtk 217,918 5,653,257 179 48 42 33 0.0820s --------- --------- ------------ ----- ---- ---- ---- --------- : Properties of real-world graphs.[]{data-label="tab:prop_real"} ------ ------------ ------------- -------- ----- ---- ---- ER 16,777,216 134,217,624 42 12 10 10 Good 16,777,216 134,181,065 1,278 28 15 14 Bad 16,777,216 133,658,199 38,143 146 89 88 ------ ------------ ------------- -------- ----- ---- ---- : Properties of synthetic graphs.[]{data-label="tab:prop_rmat"} The synthetically generated graphs are RMAT graphs, introduced by Chakrabarti et al. [@RMAT]. When generating the RMAT graphs, the adjacency matrix of the graph is subdivided in 4 equal parts and edges are distributed to these parts with a given probabilities. Specifying different probabilities allows the generation of different classes of random graphs. We generated three graphs, RMAT-ER, RMAT-Good and RMAT-Bad. Their degree distributions for the four parts are $(0.25, 0.25, 0.25, 0.25)$, $(0.45, 0.15, 0.15, 0.25)$ and $(0.55, 0.15, 0.15, 0.15)$, respectively. We generate these three graphs to create variety of challenges for our distributed memory algorithms. For instance, RMAT-ER is in the class of Erdős-Rényi random graphs which are known to be hard to partition. Other two graphs are scale-free graphs with small-world property and power-law degree distribution. This class of graphs are also difficult to partition and easily create very unbalanced workloads if a vertex-partitioning scheme is used. Properties of those graphs are given in Table \[tab:prop\_rmat\]. In distributed experiments, we partition these graphs using a block partitioning. For all the experiments we will present both the number of colors obtained and the runtime of the method when the number of processors varies. The real-world graphs all show the same trends and their results are aggregated in the following manner. Each value (number of colors and runtime) is first normalized with respect to the value obtained by the [*Natural*]{} ordering of the same graph on one processor. Then the normalized value for different graphs are aggregated using a geometric mean. The three randomly generated graphs are presented independently for number of colors and they will be aggregated for runtime results, which are normalized with respect to [*Natural*]{} ordering on 4 processors, since they show the same behavior in terms of runtime. Recoloring {#subsec:recoloring} ---------- ### Sequential Setting {#subsubsec:seq} In Figure \[fig:seq\_study\], we present the effect of vertex-visit orderings and multiple iterations of recoloring on the number of colors in sequential settings, which was also presented in [@HiPC11]. Different vertex-visit orderings are multiplexed with different color permutation strategies in each chart. For example. NAT+RC-ND gives the [*Natural*]{} ordering coupled with recoloring with the Non-Decreasing number of vertex order of the color classes. Note that the charts begin with 0 iteration which shows the quality of the vertex-visit ordering only. The first important result in those charts is that the LF vertex ordering results in lower number of colors than the [*Natural*]{} ordering with $0.96$ normalized number of colors without recoloring and SL is the best vertex ordering strategy with $0.78$ normalized number of colors. The second result is that three tested permutations of the color classes result in a decrease of the number of colors. Among them, the NI permutation leads to smallest improvement whereas the ND permutation gives the smallest number of colors by obtaining $0.8$ normalized number of colors after $20$ iterations for all three orderings. Lastly, the best number of colors is obtained by combining the SL vertex ordering with ND color permutation (SL+RC-ND). ND permutation leads to lower number of colors due to the selection of color classes. Success of a permutation can be measured by its ability to remove as many color classes as possible. In ND permutation strategy, color classes with fewer vertices are selected first so that the classes with larger number of vertices can merge with them. In this paper, we investigate the effect of randomness of the color class permutations in the sequential setting to get out of the local plateaus. For this purpose, we run experiments on real-world graphs with the three vertex-visit orderings, namely Natural, Largest First and Smallest Last. We choose the ND permutation as a reference point since it gives the best number of colors. The color permutation that picks a random permutation uniformly is denoted RAND. Using the RAND permutation at every $x$ iterations and using ND for the other iterations is denoted (ND-RAND%$x$). For instance using RAND every five iterations is denoted (ND-RAND%5) while using it every 10 iterations is denoted (ND-RAND%10). We also investigate the variant we call (ND-RAND%$2^i$) which uses random permutation only at iterations which are power of two; in other words the ND permutation is used at each iteration except at iterations 2, 4, 8, 16, ... where the permutation is RAND. Figure \[fig:random\_perm\] presents the results for this experiment. Results are obtained by running all the tests 10 times and taking the average over all runs. For the NAT vertex-visit ordering, randomness helps reducing the number of colors. The RAND permutation is better than the ND permutation for all number of iterations. However, rarefied randomness outperforms pure randomness. As the frequency of randomness decreases, the number of colors decreases as well. For example, ND-RAND%10 is better than ND-RAND%5. Two main conclusions can be drawn from this experiment. First, randomization helps getting ND out of local plateau as is strongly suggested by the sudden decrease in number of colors at the 10th iteration of ND-RAND%10. Second, ND might need numerous iterations to refine a coloring: the normalized number of colors does not change between iteration 32 and 60 of ND. The ND-RAND%$2^i$ manages to successfully use these two properties by trying to get out of local plateau aggressively during the first iterations and then, by letting ND to slowly refine its coloring. On the other hand, the LF and SL orderings show different behaviors than the NAT ordering. Randomness helps reducing the number of colors during the first iterations (up to 20 iterations in our experiments). But the ND permutation outperforms all of the other permutations on large number of iterations. This picture suggests that insisting on the ND permutation for the LF and SL vertex-visit orderings gives the best results if one is interested on high number of iterations. Overall, it can be said that randomness provides a sharp decrease in number of colors at the beginning of the process, but then the number of colors converges rapidly. Since we start from a lower number of colors in LF and SL orderings, it is expected to converge quicker and this is the case in LF and SL orderings with high number iterations of random permutations. Figure \[fig:seq\_study\] showed that the SL+RC-ND combination outperformed all other vertex-visit orderings and color class permutations presented in the sequential case. We will then focus on comparing the distributed synchronous recoloring (RC) and the asynchronous one (aRC) using the non recolored [*Smallest Last*]{} ordering as a reference. However, before diving into more comparison, we need to investigate the effect of the piggybacking technique on the communication scheme of synchronous recoloring procedure. ### Improvements on Communication Scheme of Recoloring {#sec:improved_comm_exp} We show here the runtime improvement obtained on recoloring by applying the piggybackingtechniques discussed in Section \[sec:improved\_comm\]. We run the experiments with 8 processors per node configuration to see the impact easily. Figure \[fig:reduced\_comm\] presents a comparison between a base implementation of recoloring (without piggybacking presented in [@HiPC11]) and the improved implementation using piggybacking with detailed timings. Using piggybacking needs a preparation phase, but it takes at most 12% of the improved total coloring time in the worst case. On the other hand, piggybacking provides a huge improvement in recoloring time. Our experiments show that piggybacking provides 80% less number of messages in average, when compared to base implementation. Improved coloring with piggybacking gives 20% to 70% improvement over base total coloring time. These results indicate that accumulating messages into bigger chunks provides a better communication for recoloring and we will then use the improved recoloring technique when we are making runtime comparisons. ![Comparison of one recoloring iteration of base recoloring in [@HiPC11] and improved scheme on the real-world graphs.[]{data-label="fig:reduced_comm"}](charts/reduce_comm.pdf){width=".491\linewidth"} ### Distributed Setting {#sec:dist} Figure \[fig:comp\_recolor\_real\] presents the performance of recoloring on the real-world graphs. We compare the combination of First-Fit color selection, [*Smallest Last*]{} ordering and synchronous communication (denoted FSS as in,  [@HiPC11]), with additional synchronous and asynchronous recoloring. As expected, the synchronous recoloring reduces the normalized number of colors significantly providing to keep the normalized number of colors below the one obtained using the sequential [*Largest First*]{} even on 256 processors, bringing a 18% improvement in the number of colors to FSS. However the synchronous recoloring takes more time than the FSS coloring, reaching a normalized runtime of $2.01$ while FSS has a normalized runtime of $0.60$ on 256 processors. Asynchronous recoloring provides a middle ground by allowing to obtain a better quality coloring than FSS is able to achieve. However, asynchronous recoloring gives almost the same runtime with synchronous recoloring. It shows a different picture than the one in [@HiPC11] since the piggybacking technique in synchronous recoloring brings significant improvement. Synchronous recoloring is as fast as the asynchronous recoloring, and leads to higher quality solutions. The impact of recoloring for number of colors is presented on the randomly generated graphs in Figures \[fig:comp\_rc\_ER\], \[fig:comp\_rc\_Good\], and  \[fig:comp\_rc\_Bad\]. Figure \[fig:comp\_rc\_time\] shows the aggregated runtime results for RMAT graphs. The number of colors of FSS was mainly given by the number of conflicts it generated. The vertex-visit ordering computed by the asynchronous recoloring does not avoid the majority of these conflicts and the improvement in number of colors compared to FSS is less than 10%. Synchronous recoloring shows a different picture. While FSS used many colors because of a high number of conflicts on RMAT-Good and RMAT-Bad, the synchronous recoloring does not yield any conflict. Therefore, it obtains a much better number of colors, close to the sequential [*Largest First*]{} and [*Smallest Last*]{} orderings with at most 50% improvement compared to FSS. In terms of runtime, the absence of conflict and the size of the graphs make the synchronous recoloring procedure very scalable, inducing a very low overhead compared to the initial coloring when the number of processors is high. ![Impact of the number of recoloring iterations on real-world graphs in distributed memory [@HiPC11].[]{data-label="fig:comp_recolor_iter_dist"}](charts/multi_iters.pdf){width=".60\linewidth"} Given the benefit of the recoloring procedure, it is of the interest to catch the sequential improvement on the number of colors obtained by performing the recoloring procedure multiple times. The effect of multiple recoloring iterations in distributed-memory settings on the real-world graphs is given in Figure \[fig:comp\_recolor\_iter\_dist\]. Single iteration of recoloring provides significant gain over no recoloring by staying below sequential [*Largest First*]{} when number of processors is 512. On the other hand, running 10 iterations of the recoloring procedure on 512 processors provides very good number of colors, which is close to the sequential [*Smallest Last*]{} vertex-visit ordering. Using Random-X Fit with Recoloring {#sec:usingrand} ---------------------------------- We investigate the effect of Random-5 Fit, Random-10 Fit and Random-50 Fit within our coloring framework with different number of recoloring iterations. We believe that Random-X Fit can provide an initial solution that is fast and that balances the number of vertex per color which is very suitable for recoloring. Various number of processors were tried, but they lead to similar results, so we only present results using 32 processes on 32 nodes. Firstly, we investigate the original coloring with Random-X Fit color selection strategy. We run experiments with varying and combining all of the parameters we have: superstep size (500, 1000, 5000 and 10000), vertex-visit ordering (Internal First and Smallest Last), synchronous and asynchronous communication patterns for original coloring and color selection strategies (First Fit, Random-5 Fit, Random-10 Fit and Random-50 Fit). Quality and runtime results are presented in Figure \[fig:random1\]. Note that, superstep size and communication pattern does not have a significant impact on quality-runtime trade-off, so we clustered the nearer points and tagged them with respect to their color selection strategy and vertex-visit ordering. For example, R5Ixx stands for Random-5 Fit with Internal First vertex-visit ordering where superstep size and communication pattern vary. As expected, Internal First ordering gives better runtime results than SL ordering while the situation is the other way around for the number of colors. As the X factor in Random-X Fit strategy increases, the number of colors degrades since the selection is done from a larger set. Next, we investigate the use of Random-X color selection strategy in the context of recoloring. Results for one recoloring iteration are shown in Figure \[fig:random2\]. All Random-X color selection strategies give better number of colors than First Fit with one recoloring iteration. It means that although Random-X color selection strategies start recoloring with higher number of colors than First Fit, they result in less colors at the end of the recoloring. In our opinion, this result is very important and shows that random color selection strategies provide perfect ground for recoloring. However, Random-X color selection strategies are not as good as First Fit in terms of runtime. Since the runtime of recoloring is correlated with the number of colors at the beginning of recoloring procedure, the recoloring procedure after Random-X Fit strategy takes longer. Despite this fact, R(5|10)Ixx combinations give better number of colors and better runtime than FSxx combinations, which were suggested as the best combination for recoloring in [@HiPC11]. We also investigate the impact of Random-X color selection strategies with two recoloring iterations. Figure \[fig:random3\] presents the results for this experiment. Results are similar to the ones presented in Figure \[fig:random2\]. The trade-off between number of colors and runtime is presented in Figure \[fig:bigpicture\] where results with 0,1 and 2 recoloring iterations are combined. Notice that, in Figure \[fig:bigpicture\], R(5|10)IxxND1 is better than FIxxND2 and FSxxND2 both in terms of number of colors and runtime. In summary, the random color selection strategy compensates the impact of vertex-visit ordering and the best combination to be suggested becomes R(5|10)IxxND1, which is using the Random-5 or Random-10 color selection strategy with Internal First Ordering and one Non-Decreasing recoloring iteration. This allows us to identify two interesting sets of parameters. This makes R(5|10)IxxND1 a good candidate for having a solution of good quality and we call this set of parameters “quality”. For a obtaining a solution quickly, it is better not to use recoloring. The random strategies tend to significantly hinders the quality of the solution in this case and using First Fit is leading to a higher quality solution at no runtime overhead. The local ordering of the vertices makes little difference in quality because of the distributed execution of algorithm; therefore the fastest ordering (Internal first) is preferred. This makes FIxxND0 a good candidate for obtaining quickly a solution of moderate quality and we call this set of parameters “speed”. Conclusion {#sec:ccl} ========== In this paper we investigated different ways of improving the number of colors in distributed-memory graph coloring algorithms. We investigated recoloring by utilizing the independent sets of vertices resulted by an existing solution to color large number of vertices independently with little synchronization or with little conflicts. We showed that graphs that result in a larger number of conflicts benefit significantly from synchronous recoloring in terms of the number of colors. The runtime overhead is being small in such cases, therefore multiple iterations of recoloring can be used to obtain even fewer colors. We also showed that synchronous recoloring can be implemented using piggybacking to improve network access patterns; making synchronous recoloring as fast as the previously proposed asynchronous recoloring. Overall, recoloring proved to be a scalable concept for the number of colors when number of processors increases, thanks to the equivalence to a sequential procedure. Last but may be most interestingly, we showed that the Random-X Fit color selection strategy works quite well with recoloring and provides better number of colors and runtime results than the vertex-visit ordering solutions. Overall, we suggest two combinations of the parameters; if the user is interested in a quick coloring with decent quality, then the “speed” policy FIxxND0 (First Fit, Internal first, no recoloring) is suitable; if the user needs better coloring, then the “quality” R(5|10)IxxND1 combination of Random-X Fit, with Internal First ordering and some Non-Decreasing recoloring iterations is best for large number of processors. As a future work, more thorough study of Random-X Fit strategy could lead to a better understanding of the coloring problem and could suggest ways to derandomize this color selection strategy. Acknowledgments {#acknowledgments .unnumbered} =============== This work was partially supported by the U.S. Department of Energy SciDAC Grant DE-FC02-06ER2775 and NSF grants CNS-0643969, OCI-0904809 and OCI-0904802. [^1]: Corresponding author: [email protected]. Tel: 614-688-9637, Fax: 614-688-6600 [^2]: <http://www.cise.ufl.edu/research/sparse/matrices/>

ArXiv

--- abstract: 'If $f:[a,b]\to \mathbb{R}$, with $a<b$, is continuous and such that $a$ and $b$ are mapped in opposite directions by $f$, then $f$ has a fixed point in $I$. Suppose that $f:\mathbb{C}\to\mathbb{C}$ is map and $X$ is a continuum. We extend the above for certain continuous maps of dendrites $X\to D, X\subset D$ and for positively oriented maps $f:X\to \mathbb{C}, X\subset \mathbb{C}$ with the continuum $X$ not necessarily invariant. Then we show that in certain cases a holomorphic map $f:\mathbb{C}\to\mathbb{C}$ must have a fixed point $a$ in a continuum $X$ so that either $a\in \mathrm{Int}(X)$ or $f$ exhibits rotation at $a$.' address: | Department of Mathematics\ University of Alabama at Birmingham\ Birmingham, AL 35294-1170 author: - Alexander Blokh - Lex Oversteegen bibliography: - '/lex/references/refshort.bib' date: 'September 21, 2008' title: 'Fixed points in non-invariant plane continua' --- [^1] [^2] Introduction ============ By ${\mathbb{C}}$ we denote the plane and by ${{\mathbb{C}}^\infty}$ the Riemann sphere. The fixed point problem, attributed to [@ster35], is one of the central problems in plane topology. It can be formulated as follows. \[fpt\] Suppose that $f:{\mathbb{C}}\to {\mathbb{C}}$ is continuous and $f(X)\subset X$ for a non-separating continuum $X$. Does there always exist a fixed point in $X$? As is, Problem \[fpt\] is not yet solved. The most well-known particular case for which it is solved is that of a map of a closed interval $I=[a, b]$, $a<b$ into itself in which case there must exist a fixed point in $I$. In fact, in this case a more general result can be proven of which the existence of a fixed point in an invariant interval is a consequence. Namely, instead of considering a map $f:I\to I$ consider a map $f:I\to {\mathbb{R}}$ such that either (a) $f(a)\ge a$ and $f(b)\le b$, or (b) $f(a)\le a$ and $f(b)\ge b$. Then still there must exist a fixed point in $I$ which is an easy corollary of the Intermediate Value Theorem applied to the function $f(x)-x$. Observe that in this case $I$ need not be invariant under $f$. Observe also that without the assumptions on the endpoints, the conclusion on the existence of a fixed point inside $I$ cannot be made because, e.g., a shift map on $I$ does not have fixed points at all. The conditions (a) and (b) above can be thought of as boundary conditions imposing restrictions on where $f$ maps the boundary points of $I$ in ${\mathbb{R}}$. Our main aim in this paper is to consider some other cases for which Problem \[fpt\] is solved in affirmative (i.e., the existence of a fixed point in an invariant continuum is established) and replace for them the invariantness of the continuum by boundary conditions in the spirit of the above “interval version” of Problem \[fpt\]. More precisely, instead of assuming that $f(X)\subset X$ we will make some assumptions on the way that $f$ maps points of ${\overline}{f(X){\setminus}X}\cap X$. First though let us discuss particular cases for which Problem \[fpt\] is solved. They can be divided into two categories: either $X$ has specific properties, or $f$ has specific properties. The most direct extension of the “interval particular case” of Problem \[fpt\] is, perhaps, the following well known theorem (see for example[@nadl92]). \[dendr\] If $f:D\to D$ is a continuous map of a dendrite into itself then it has a fixed point. Here $f$ is just a continuous map but the continuum $D$ is very nice. Theorem \[dendr\] can be generalized to the case when $f$ maps $D$ into a dendrite $X\supset D$ and certain conditions onto the behavior of the points of the boundary of $D$ in $X$ are fulfilled. This presents a “non-invariant” version of Problem \[fpt\] for dendrites and can be done in the spirit of the interval case described above. Moreover, with some additional conditions it has consequences related to the number of periodic points of $f$. The details and exact statements can be found to Section 2 devoted to the dendrites. Another direction is to consider specific maps of the plane on arbitrary non-separating continua. Let us go over known results here. Cartwright and Littlewood [@cartlitt51] have solved Problem \[fpt\] in affirmative for orientation preserving homeomorphisms of the plane. This result was generalized to all homeomorphisms by Bell [@bell78]. The existence of fixed points for orientation preserving homeomorphisms of the *entire plane* under various conditions was considered in [@brou12a; @brow84a; @fath87; @fran92; @guil94; @fmot07], and of a point of period two for orientation reversing homeomorphisms in [@boni04]. The result by Cartwright and Littlewood deals with the case when $X$ is an invariant continuum. In parallel with the interval case, we want to extend this to a larger class of maps of the plane (i.e., not necessarily one-to-one) such that certain “boundary” conditions are satisfied. Our tools are mainly based on [@fmot07] and apply to *positively oriented maps* which generalize the notion of an orientation preserving homeomorphism (see the precise definitions in Section \[prel\]). Our main topological results are Theorems \[fixpt\] and  \[locrot\]. The precise conditions in them are somewhat technical - after all, we need to describe “boundary conditions” of maps of arbitrary non-separating continua. However it turns out that these conditions are satisfied by holomorphic maps (in particular, polynomials), allowing us to obtain a few corollaries in this case, essentially all dealing with the existence of periodic points in certain parts of the Julia set of a polynomial and degeneracy of certain impressions. Preliminaries {#prel} ============= A map $f:X\to Y$ is *perfect* if for each compact set $K\subset Y$, $f^{-1}(K)$ is also compact. *All maps considered in this paper are prefect.* Given a continuum $K\subset {\mathbb{C}}$, denote by $U_{\infty}(K)$ the component of ${\mathbb{C}}{\setminus}K$ containing infinity, and by $T(K)$ the *topological hull of $K$*, i.e. the set ${\mathbb{C}}{\setminus}U_{\infty}(K)$. By ${\mathbb{S}^1}$ we denote the unit circle which we identify with ${\mathbb{R}}\mod{\mathbb{Z}}$. In this section we will introduce a new class of maps which are the proper generalization of an orientation preserving homeomorphism. For completeness we will recall some related results from [@fmot07] where these maps were first introduced. Let $f:U \to {\mathbb{C}}$ be a map from a simply connected domain $U$ into the plane. Let $S$ be a positively oriented simple closed curve in $U$, and $w{\not\in}f(S)$ be a point. Define $f_w:S\to{\mathbb{S}^1}$ by $$f_p(x)=\frac{f(x)-w}{|f(x)-w|}.$$ Then $f_w$ has a well-defined [*degree*]{} (also known as the [*winding number*]{} of $f|_S$ about $w$), denoted ${\text{degree}}(f_w)={\mathrm{win}}(f,S,w)$. A map $f:U \to {\mathbb{C}}$ from a simply connected domain $U$ is *positively oriented* (respectively, [*negatively oriented*]{}) provided for each simple closed curve $S$ in $U$ and each point $w\in f(T(S))\setminus f(S)$, ${\text{degree}}(f_w)>0$ (${\text{degree}}(f_w)<0$, respectively). A holomorphic map $f:{\mathbb{C}}\to{\mathbb{C}}$ is a prototype of a positively oriented map. Hence the results obtained in this paper apply to them. However, in general positively oriented maps do not have to be light (i.e., a positively oriented map can map a subcontinuum of ${\mathbb{C}}$ to a point). Observe also, that for points $w{\not\in}T(f(S))$ we have ${\text{degree}}(f_w)=0$. A map $f:{\mathbb{C}}\to{\mathbb{C}}$ is *oriented* provided for each simple closed curve $S$ and each $x\in T(S)$, $f(x)\in T(f(S))$. Every positively or negatively oriented map is oriented (indeed, otherwise there exists $x\in T(S)$ with $f(x){\not\in}T(f(S))$ which implies that ${\mathrm{win}}(f,S,f(x))=0$, a contradiction). A map $f:{\mathbb{C}}\to {\mathbb{C}}$ is *confluent* provided for each subcontinuum $K\subset{\mathbb{C}}$ and every component $C$ of $f^{-1}(K)$, $f(C)=K$. It is well known that both open and monotone maps (and hence compositions of such maps) of continua are confluent. It follows from a result of Lelek and Read [@leleread74] that each confluent mapping of the plane is the composition of a monotone map and a light open map. Theorem \[orient\] is obtained in [@fmot07]. \[orient\] Suppose that $f:{\mathbb{C}}\to{\mathbb{C}}$ is a surjective map. Then the following are equivalent: 1. \[pnorient\] $f$ is either positively or negatively oriented; 2. \[iorient\]$f$ is oriented; 3. \[conf\] $f$ is confluent. Moreover, if $f$ satisfies these properties then for any non-separating continuum $X$ we have $f({\mbox{Bd}}(X))\supset {\mbox{Bd}}(f(X))$. Let $X$ be an non-separating plane continuum. A [*crosscut*]{} of $U={\mathbb{C}}{\setminus}X$ is an open arc $A\subset U$ such that ${\mbox{Cl}}(A)$ is an arc with exactly two endpoints in ${\mbox{Bd}}(U)$. Evidently, a crosscut of $U$ separates $U$ into two disjoint domains, exactly one unbounded. Let $S$ be a simple closed curve in ${\mathbb{C}}$ and suppose $g:S\to{\mathbb{C}}$ has no fixed points on $S$. Since $g$ has no fixed points on $S$, the point $z-g(z)$ is never $0$. Hence the unit vector $v(z)=\frac{g(z)-z}{|g(z)-z|}$ always exists. Let $z(t)$ be a convenient counterclockwise parameterization of $S$ by $t\in {\mathbb{S}^1}$ and define the map ${\overline}{v}=v\circ z:{\mathbb{S}^1}\to {\mathbb{S}^1}$ by $${\overline}{v}(t)=v(z(t))=\frac{g(z(t))-z(t)}{|g(z(t))-z(t)|}.$$ Then ${\mathrm{Ind(g,S)}}$, the *index of $g$ on $S$*, is the [*degree*]{} of ${\overline}v$. The following theorem (see, e.g., [@fmot07]) is a major tool in finding fixed points of continuous maps of the plane. \[basic\] Suppose that $S\subset {\mathbb{C}}$ is a simple closed curve and $f:T(S)\to {\mathbb{C}}$ is a continuous map such that ${\mathrm{Ind(f,S)}}\not=0$. Then $f$ has a fixed point in $T(S)$. Theorem \[basic\] applies to Problem \[fpt\] as follows. Given a non-separating continuum $X\subset {\mathbb{C}}$ one constructs a simple closed curve $S$ approximating $X$ so that the index of $f$ on $S$ can be computed. If it is not equal to zero, it implies the existence of a fixed point in $T(S)$, and if $S$ is tight enough, in $X$. Hence our efforts should be aimed at constructing $S$ and computing ${\mathrm{Ind(f,S)}}$. One way of doing so is to use Bell’s notion of variation which we will now introduce. Suppose that $X$ is a non-separating plane continuum and $S$ is a simple closed curve such that $X\subset T(S)$ and $S\cap X$ consists of more than one point. Then we will call $S$ a *bumping simple closed curve of $X$*. Any subarc of $S$, both of whose endpoints are in $X$, is called a *bumping arc of $X$* or a *link of $S$*. Note that any bumping arc $A$ of $X$ can be extended to a bumping simple closed curve $S$ of $X$. Hence, every bumping arc has a natural order $<$ inherited from the positive circular order of a bumping simple closed curve $S$ containing $A$. If $a<b$ are the endpoints of $A$, then we will often write $A=[a,b]$. Also, by the *shadow $Sh(A)$ of $A$*, we mean the union of all bounded components of ${\mathbb{C}}{\setminus}(X\cup A)$. The [*standard junction*]{} $J_0$ is the union of the three rays $R_i=\{z\in{\mathbb{C}}\mid z=re^{i\pi/2}, r\in[0,\infty)\}$, $R_+=\{z\in{\mathbb{C}}\mid z=re^{0}, r\in[0,\infty)\}$, $R_-=\{z\in{\mathbb{C}}\mid z=re^{i\pi}, r\in[0,\infty)\}$, having the origin $0$ in common. By $U$ we denote the lower half-plane $\{z\in{\mathbb{C}}\mid z=x+iy, y<0\}$. A [*junction*]{} $J_v$ is the image of $J_0$ under any orientation-preserving homeomorphism $h:{\mathbb{C}}\to{\mathbb{C}}$ where $v=h(0)$. We will often suppress $h$ and refer to $h(R_i)$ as $R_i$, and similarly for the remaining rays and the region $h(U)$. Let $f:{\mathbb{C}}\to{\mathbb{C}}$ be a map, $X$ be a non-separating plane continuum, $A=[a,b]$ be a bumping subarc of $X$ with $a<b$, $\{f(a),f(b)\}\subset X$ and $f(A)\cap A={\emptyset}$. We define the [*variation of $f$ on $A$ with respect to $X$*]{}, denoted ${\mathrm{Var(f,A)}}$, by the following algorithm: 1. Choose an [orientation preserving]{} homeomorphism $h$ of ${\mathbb{C}}$ such that $h(0)=v\in A$ and $X\subset h(U)\cup\{v\}$. 2. \[crossings\] [*Crossings:*]{} Consider the set $K=[a,b]\cap f^{-1}(J_v)$. Move along $A$ from $a$ to $b$. Each time a point of $[a,b]\cap f^{-1}(R_+)$ is followed immediately by a point of $[a,b]\cap f^{-1}(R_i)$ in $K$, count $+1$. Each time a point of $[a,b]\cap f^{-1}(R_i)$ is followed immediately by a point of $[a,b]\cap f^{-1}(R_+)$ in $K$, count $-1$. Count no other crossings. 3. The sum of the crossings found above is the variation, denoted ${\mathrm{Var(f,A)}}$. It is shown in [@fmot07] that the variation does not depend on the choice of a junction satisfying the above listed properties. Informally, one can understand the notion of variation as follows. Since $f(A)\cap A={\emptyset}$, we can always complete $A$ with another arc $B$ (now connecting $b$ to $a$) to a simple closed curve $S$ disjoint from $J_v$ so that $v{\not\in}f(S)$. Then it is easy to see that ${\mathrm{win}}(f, S, v)$ cab be obtained by summing up ${\mathrm{Var(f,A)}}$ and the similar count for the arc $B$ (observe that the latter is *not* the variation of $B$ because to compute that we will need to use another junction “based” at a point of $B$). Any partition $A=\{a_0<a_1<\dots<a_n<a_{n+1}=a_0\}\subset X\cap S$ of a bumping simple closed curve $S$ of a non-separating continuum $X$ such that for all $i$, $f(a_i)\in X$ and $f([a_i,a_{i+1}])\cap [a_i,a_{i+1}]={\emptyset}$ is called an *allowable partition of $S$.* We will call the bumping arcs $[a_i,a_{i+1}]$ *links (of $S$)*. It is shown in [@fmot07] that the variation of a bumping arc is well-defined. Moreover, it follows from that paper (see Theorem 2.12 and Remark 2.19) that: \[FMOT\] Let $S$ be a simple closed curve, $X=T(S)$ and let $a_0<a_1<\dots<a_n<a_0=a_{n+1}$ be points in $S$ (in the positive circular order around $S$) such that for each $i$, $f(a_i)\in T(S)$ and, if $Q_i=[a_i,a_{i+1}]$, then $f(Q_i)\cap Q_i={\emptyset}$.\ Then $${\mathrm{Ind(f,S)}}=\sum_{i} {\mathrm{Var(f,Q_i)}}+1.$$ Observe that if the points $a_i, i=1, \dots, n$ satisfying the properties of Theorem \[FMOT\] can be chosen then there are no fixed points of $f$ in $S$ and ${\mathrm{Ind(f,S)}}$ is well-defined. Theorem \[FMOT\] shows that if we define ${\mathrm{Var(f,S)}}=\sum_i{\mathrm{Var(f,Q_i)}}$, then ${\mathrm{Var(f,S)}}$ is well defined and independent of the choice of the allowable partition of $S$ and of the choice of the junctions used to compute ${\mathrm{Var(f,Q_i)}}$. \[closed\] Let $f:{\mathbb{C}}\to{\mathbb{C}}$ be a map, $X$ be a non-separating continuum and $C=[a,b]$ be a bumping arc of $X$ with $a<b$. Let $v\in [a,b]$ be a point and let $J_v$ be a junction such that $J_v\cap (X\cup C)=\{v\}$. Suppose that $J_v\cap f(X)={\emptyset}$. Then there exists an arc $I$ such that $S=I\cup C$ is a bumping simple closed curve of $X$ and $f(I)\cap J_v={\emptyset}$. Since $f(X)\cap J_v={\emptyset}$, it is clear that there exists an arc $I$ with endpoints $a$ and $b$ near X such that $I\cup C$ is a simple closed curve, $X\subset T(I\cup C)$ and $f(I)\cap J_v={\emptyset}$. This completes the proof. The next corollary gives a sufficient condition for the non-negativity of the variation of an arc. \[posvar\] Suppose $f:{\mathbb{C}}\to{\mathbb{C}}$ is a positively oriented map, $C=[a,b]$ is a bumping arc of $X$, $f(C)\cap C={\emptyset}$ and $J_v\subset[{\mathbb{C}}{\setminus}X]\cup\{v\}$ is a junction with $J_v\cap C=\{v\}$. Suppose that $f(\{a,b\})\subset X$ and there exists a continuum $K\subset X$ such that $f(K)\cap J_v\subset \{v\}$. Then ${\mathrm{Var(f,C)}}\geq 0$. Observe that since $f(a), f(b)\in X$ then ${\mathrm{Var(f,C)}}$ is well-defined. Consider a few cases. Suppose first that $f(K)\cap J_v={\emptyset}$. Then, by Lemma \[closed\], there exists an arc $I$ such that $S=I\cup C$ is a bumping simple closed curve around $K$ and $f(I)\cap J_v={\emptyset}$ (it suffices to choose $I$ very close to $K$). Then $v\in{\mathbb{C}}\setminus f(S)$. Hence ${\mathrm{Var(f,C)}}=\text{win}(f,S,v)\geq 0$. Suppose next that $f(K)\cap J_v=\{v\}$. Then we can perturb the junction $J_v$ slightly in a small neighborhood of $v$, obtaining a new junction $J_d$ such that intersections of $f(C)$ with $J_v$ and $J_d$ are the same (and, hence, yield the same variation) and $f(K)\cap J_d={\emptyset}$. Now proceed as in the first case. Main results ============ Dendrites --------- In this subsection we generalize Theorem \[dendr\] to non-invariant dendrites. We will also show that in certain cases the dendrite must contain infinitely many periodic cutpoints (recall that if $Y$ is a continuum and $x\in Y$ then ${\mbox{val}}_Y(x)$ is the number of connected components of $Y{\setminus}\{x\}$, and $x$ is said to be an *endpoint (of $Y$)* if ${\mbox{val}}_Y(x)=1$, a *cutpoint (of $Y$)* if ${\mbox{val}}_Y(x)>1$ and a *vertex/branchpoint (of $Y$)* if ${\mbox{val}}_Y(x)>2$). These results have applications in complex dynamics [@bco08]. In this subsection given two points $a, b$ of a dendrite we denote by $[a, b], (a, b], [a, b), (a, b)$ the unique closed, semi-open and open arcs connecting $a$ and $b$ in the dendrite. More specifically, unless otherwise specified, the situation considered in this subsection is as follows: $D_1\subset D_2$ are dendrites and $f:D_1\to D_2$ is a continuous map. Set $E={\overline}{D_2{\setminus}D_1}\cap D_1$. In other words, $E$ consists of points at which $D_2$ “grows” out of $D_1$. A point $e\in E$ may be an endpoint of $D_1$ (then there is a unique component of $D_2{\setminus}\{e\}$ which meets $D_1$) or a cutpoint of $D_1$ (then there are several components of $D_2{\setminus}\{e\}$ which meet $D_1$). The following theorem is a simple extension of the real result claiming that if there are points $a<b$ in ${\mathbb{R}}$ such that $f(a)<a, f(b)>b$ then there exists a fixed point $c\in (a, b)$ (case (b) described in Introduction). \[fixpt-1\] Suppose that $a, b\in D_1$ are such that $a$ separates $f(a)$ from $b$ and $b$ separates $f(b)$ from $a$. Then there exists a fixed point $c\in (a, b)$ which is a cutpoint of $D_1$ (and hence $D_2$). In particular if there are two points $e_1\ne e_2\in E$ such that $f(e_i)$ belongs to a component of $D_2{\setminus}\{e_i\}$ disjoint from $D_1$ then there exists a fixed point $c\in (a, b)$ which is a cutpoint of $D_1$ (and hence $D_2$). It follows that we can find a sequence of points $a_{-1}, \dots$ in $(a, b)$ such that $f(a_{-n-1})=a_{-n}$ and $a_{-n-1}$ separates $a_{-n}$ from $b$. Clearly, $\lim_{n\to {\infty}} a_{-n}=c\in [a, b]$ is a fixed point as desired. If there are two points $e_1\ne e_2\in E$ such that $f(e_i)$ belongs to a component of $D_2{\setminus}\{e_i\}$ disjoint from $D_1$ then the above applies to them. The other real case (case (a)) described in Introduction is somewhat more difficult to generalize. Definition \[bouscr\] extends it (i.e. the real case when $a<b$ are points of ${\mathbb{R}}$ such that $f(a)>a$ and $f(b)<b$) onto dendrites. \[bouscr\] Suppose that in the above situation the map $f$ is such that for each non-fixed point $e\in E$, $f(e)$ is contained in a component of $D_2{\setminus}\{e\}$ which meets $D_1$. Then we say that $f$ has the *boundary scrambling property* or that it *scrambles the boundary*. Observe that if $D_1$ *is* invariant then $f$ automatically scrambles the boundary. The next definition presents a useful topological version of repelling at a fixed point. \[wkrep\] Suppose that $a\in D_1$ is a fixed point and that there exists a component $B$ of $D_1{\setminus}\{a\}$ such that arbitrarily close to $a$ in $B$ there exist fixed cutpoints of $f$ or points $x$ separating $p$ from $f(x)$. Then say that $a$ is a *weakly repelling fixed point (of $f$ in $B$)*. A periodic point $a$ is said to be *weakly repelling* if there exists $n$ and a component $B$ of $D_1{\setminus}\{a\}$ such that $a$ is a weakly repelling fixed point of $f^n$ in $B$. It is easy to see that a fixed point $a$ is weakly repelling in $B$ if and only if either $a$ is a limit of fixed cutpoints of $f$ in $B$, or there exist a neighborhood $U$ of $a$ in ${\overline}{B}$ and a point $x\in U{\setminus}\{a\}$ such that $U$ contains no fixed points but $a$ and $x$ separates $p$ from $f(x)$. Indeed, in the latter case by continuity there exists a point $x_1\in (a, x)$ such that $f(x_{-1})=x$ and this sequence of preimages can be extended towards $a$ inside $(a, x)$ so that it converges to $a$ (otherwise it would converge to a fixed point inside $U$ distinct from $a$, a contradiction). In particular, *if $a$ is a weakly repelling fixed point of $f$ then $a$ is a weakly repelling fixed point of $f^n$ for any $n$*. Moreover, since there are only countably many vertices of $D_2$ and their images under $f$ and its powers, we can choose $x$ and its backward orbit converging to $a$ so that all its points are cutpoints of $D_2$ of valence $2$. From now on we assume that to each weakly repelling fixed point $a$ of $f$ in $B$ which is not a limit point of fixed cutpoints in $B$ we associate a point $x_a=x\in B$ of valence $2$ in $B$ separating $a$ from $f(x)$ and a neighborhood $U_a=U\subset {\overline}{B}$ which is the component of ${\overline}{B}{\setminus}\{x\}$ containing $a$. As an important tool we will need the following retraction closely related to the described above situation. \[retr\] For each $x\in D_2$ there exists a unique arc (possibly a point) $[x,d_x]$ such that $[x,d_x]\cap D_1=\{d_x\}$. Hence there exists a natural monotone retraction $r:D_2\to D_1$ defined by $r(x)=d_x$, and the map $g=g_f=r\circ f:D_1\to D_1$ which is a continuous map of $D_1$ into itself. We call the map $r$ the *natural retraction (of $D_2$ onto $D_1$)* and the map $g$ the *retracted (version of) $f$*. The map $g$ is designed to make $D_1$ invariant so that Theorem \[dendr\] applies to $g$ and allows us to conclude that there are $g$-fixed points. However these points are not necessarily fixed points of $f$. Indeed, $g(x)=x$ means that $r\circ f(x)=x$. Hence it really means that $f$ maps $x$ to a point belonging to a component of $D_2{\setminus}D_1$ which “grows” out of $D_1$ at $x$. In particular, it means that $x\in E$. Thus, if $g(x)=x$ and $x\notin E$ then $f(x)=x$. In general, it follows from the construction that if $f(x)\ne g(x)$, then $g(x)\in E$ because points of $E$ are exactly those points of $D_1$ to which points of $D_2{\setminus}D_1$ map under $r$. We are ready to prove our first lemma in this direction. \[fixpt0\] Suppose that $f$ scrambles the boundary. Then $f$ has a fixed point. We may assume that there are no $f$-fixed points $e\in E$. By Theorem \[dendr\] the map $g_f=g$ has a fixed point $p\in D_1$. It follows from the fact that $f$ scrambles the boundary that points of $E$ are not $g$-fixed. Hence $p\notin E$, and by the argument right before the lemma $f(p)=p$ as desired. It follows from Proposition \[fixpt-1\] and Proposition \[fixpt0\] that the only behavior of points in $E$ which does not force the existence of a fixed point in $D_1$ is when exactly one point $e\in E$ maps into a component of $D_2{\setminus}\{e\}$ which is disjoint from $D_1$ whereas any other point $e'\in E$ maps into a component of $D_2{\setminus}\{e'\}$ which is not disjoint from $D_1$. The next lemma shows that in some cases $p$ can be chosen to be a cutpoint of $D_1$. \[fxctpt\] Suppose that $f$ scrambles the boundary and all $f$-fixed endpoints of $D_1$ are weakly repelling. Then there is a fixed cutpoint of $f$. Suppose that $f$ has no fixed cutpoints. By Proposition \[fixpt0\], the set of fixed points of $f$ is not empty. Hence we may assume that *all* fixed points of $f$ are endpoints of $D_1$. Let $a, b$ be distinct fixed points of $f$. Then it follows that either $U_a\subset U_b$, or $U_b\subset U_a$, or $U_a\cap U_b={\emptyset}$. Indeed, suppose that $x_a\in U_b$. Let us show that then $U_a\subset U_b$. Indeed, otherwise by Lemma \[fixpt0\].(1) there exists a fixed point $c\in (x_a, x_b)$, a contradiction. Now, suppose that neither $x_a\in U_b$ nor $x_b\in U_a$. Then clearly $U_a\cap U_b={\emptyset}$. Consider an open covering of the set of all fixed points $a\in D_1$ by their neighborhoods $U_a$ and choose a finite subcover. By the above we may assume that its consists of pairwise disjoint sets $U_{a_1}, \dots, U_{a_k}$. Consider now the component $Q$ of $D_1$ whose endpoints are the points $a_1, \dots, a_k$. It is easy to see that $f|_Q$, with $Q$ considered as a subdendrite of $D_2$, scrambles the boundary and has no fixed endpoints. Hence an $f$-fixed point $p\in Q$, existing by Lemma \[fixpt0\], must be a cutpoint of $D_1$. Lemma \[fxctpt\] is helpful in the next theorem which shows that under some rather weak assumptions on periodic points the map has infinitely many periodic cutpoints. \[infprpt\] Suppose that $f:D\to D$ is continuous and all its periodic points are weakly repelling. Then $f$ has infinitely many periodic cutpoints. By way of contradiction suppose that there are finitely many periodic cutpoints of $f$. Without loss of generality we may assume that these are points $a^1, \dots, a^k$ each of which is *fixed* under $f$. Let $A=\cup^k_{i=1} a^i$ and let $B$ be component of $D{\setminus}A$. Then ${\overline}{B}$ is a subdendrite of $D$ to which the above tools apply: $D$ plays the role of $D_2$, ${\overline}{B}$ plays the role of $D_1$, and $E$ is exactly the boundary ${\mbox{Bd}}(B)$ of $B$ (by the construction ${\mbox{Bd}}(B)\subset A$). Suppose that each point $a\in {\mbox{Bd}}(B)$ is weakly repelling in $B$. Then by the assumptions of the theorem Lemma \[fxctpt\] applies to this situation. It follows that there exists a fixed cutpoint of $B$, a contradiction. Hence for some $a\in {\mbox{Bd}}{B}$ we have that $a$ is *not* weakly repelling in $B$. Since by the assumptions $a$ *is* weakly repelling, there exists another component, say, $C$, of $D{\setminus}A$ such that $a\in {\mbox{Bd}}(C)$ and $a$ *is* weakly repelling in $C$. We can now apply the same argument to $C$. If all boundary points of $C$ are weakly repelling in $C$ then by Lemma \[fxctpt\] $C$ will contain a fixed cutpoint, a contradiction. Hence there exists a point $d\in A$ such that $d$ is *not* weakly repelling in $C$ and a component $F$ of $D{\setminus}A$ whose closure meets ${\overline}{C}$ at $d$, and $d$ *is* weakly repelling in $F$. Clearly, after finitely many steps this process will have to end ultimately leading to a component $Z$ of $D{\setminus}A$ such that all points of ${\mbox{Bd}}(Z)$ are weakly repelling in $Z$. Since here the set ${\mbox{Bd}}(Z)$ plays the role of the set $E$ from above and by the assumptions of the theorem we see that Lemma \[fxctpt\] applies to $Z$ and there exists a fixed cutpoint of $Z$, a contradiction. An important application of Theorem \[infprpt\] is to the dendritic *topological Julia sets*. They can be defined as follows. Consider an equivalence relation $\sim$ on the unit circle ${\mathbb{S}^1}$. Equivalence classes of $\sim$ will be called *($\sim$-)classes* and will be denoted by boldface letters. A $\sim$-class consisting of two points is called a *leaf*; a class consisting of at least three points is called a *gap* (this is more restrictive than Thurston’s definition in [@thur85]; we follow [@bl02] in our presentation). Fix an integer $d>1$ and denote the map $z^d:{\mathbb{S}^1}\to{\mathbb{S}^1}$ by ${\sigma}_d$. Then $\sim$ is said to be a *($d$-)invariant lamination* if: (E1) $\sim$ is *closed*: the graph of $\sim$ is a closed set in ${\mathbb{S}^1}\times {\mathbb{S}^1}$; (E2) $\sim$ defines a *lamination*, i.e., it is *unlinked*: if ${\mathbf{g}}_1$ and ${\mathbf{g}}_2$ are distinct $\sim$-classes, then their convex hulls ${\mbox{Ch}}({\mathbf{g}}_1), {\mbox{Ch}}({\mathbf{g}}_2)$ in the unit disk ${\mathbb{D}}$ are disjoint, (D1) $\sim$ is *forward invariant*: for a class ${\mathbf{g}}$, the set ${\sigma}_d({\mathbf{g}})$ is a class too which implies that (D2) $\sim$ is *backward invariant*: for a class ${\mathbf{g}}$, its preimage ${\sigma}_d^{-1}({\mathbf{g}})=\{x\in {\mathbb{S}^1}: {\sigma}_d(x)\in {\mathbf{g}}\}$ is a union of classes; (D3) for any gap ${\mathbf{g}}$, the map ${\sigma}_d|_{{\mathbf{g}}}: {\mathbf{g}}\to {\sigma}_d({\mathbf{g}})$ is a *covering map with positive orientation*, i.e., for every connected component $(s, t)$ of ${\mathbb{S}^1}\setminus {\mathbf{g}}$ the arc $({\sigma}_d(s), {\sigma}_d(t))$ is a connected component of ${\mathbb{S}^1}\setminus {\sigma}_d({\mathbf{g}})$. The lamination in which all points of ${\mathbb{S}^1}$ are equivalent is said to *degenerate*. It is easy to see that if a forward invariant lamination $\sim$ has a class with non-empty interior then $\sim$ is degenerate. Hence equivalence classes of any non-degenerate forward invariant lamination are totally disconnected. Call a class ${\mathbf{g}}$ *critical* if ${\sigma}_d|_{{\mathbf{g}}}: {\mathbf{g}}\to {\sigma}({\mathbf{g}})$ is not one-to-one, *(pre)critical* if ${\sigma}_d^j({\mathbf{g}})$ is critical for some $j\ge 0$, and *(pre)periodic* if ${\sigma}^i_d({\mathbf{g}})={\sigma}^j_d({\mathbf{g}})$ for some $0\le i<j$. A gap ${\mathbf{g}}$ is *wandering* if ${\mathbf{g}}$ is neither (pre)periodic nor (pre)critical. Let $p: {\mathbb{S}^1}\to J_\sim={\mathbb{S}^1}/\sim$ be the quotient map of ${\mathbb{S}^1}$ onto its quotient space $J_\sim$, let $f_\sim:J_\sim \to J_\sim$ be the map induced by $\sigma_d$. We call $J_\sim$ a *topological Julia set* and the induced map $f_\sim$ a *topological polynomial*. It is easy to see that if ${\mathbf{g}}$ is a $\sim$-class then ${\mbox{val}}_{J_\sim}(p({\mathbf{g}}))=|{\mathbf{g}}|$ where by $|A|$ we denote the cardinality of a set $A$. \[lamwkrp\] Suppose that the topological Julia set $J_\sim$ is a dendrite and $f_\sim:J_\sim\to J_\sim$ is a topological polynomial. Then all periodic points of $f_\sim$ are weakly repelling and $f_\sim$ has infinitely many periodic cutpoints. Suppose that $x$ is an $f_\sim$-fixed point and set ${\mathbf{g}}=p^{-1}(x)$. If $x$ is an endpoint then ${\mathbf{g}}$ is a singleton. Connect $x$ with a point $y\ne x$. Then the arc $[x, y]\subset J_\sim$ contains points $y_k\to x$ of valence $2$ because, as is well known, there are no more than countably many vertices of $J_\sim$. It follows that $\sim$-classes $p^{_1}(y_k)$ are leaves separating ${\mathbf{g}}$ from the rest of the circle and repelled from ${\mathbf{g}}$ under the action of ${\sigma}$. Hence $f_\sim(y_i)$ is separated from $x$ by $y_i$ and so $x$ is weakly repelling. Suppose that $x$ is not an endpoint. Choose a very small connected neighborhood $U$ of $x$. It is easy to see that each component $A$ of $U{\setminus}\{x\}$ corresponds to a unique chord $\ell_A\in {\mbox{Bd}}({\mbox{Ch}}({\mathbf{g}}))$. Moreover, for each component $A$ of $U{\setminus}\{x\}$ there exists a unique component $B$ of $U{\setminus}\{x\}$ such that $f_\sim(A)\cap B\ne {\emptyset}$. Hence there is a map $h$ from the set ${\mathcal{A}}$ of all components of $U{\setminus}\{x\}$ to itself. Suppose that there exists $E\in {\mathcal{A}}$ and $n>0$ such that $h^n(E)=E$. Then it follows that the endpoints of $\ell_E$ are fixed under ${\sigma}^n$. Connect $x$ with a point $y\in E$ and choose, as in the previous paragraph, a sequence of points $y_k\in [x, y], y_k\to x$ of valence $2$. Then again by the repelling properties of ${\sigma}^n$ it follows that $f_\sim(y_i)$ is separated from $x$ by $y_i$ and so $x$ is weakly repelling (for $f^m_\sim$ in $E$). It remains to show that there $E\in {\mathcal{A}}$ with $h^n(E)=E$ for some $n>0$ must exist. Suppose otherwise. To each component $C$ of $U{\setminus}\{x\}$ we associate the corresponding component $J_C$ of $J_\sim{\setminus}\{a\}$ containing $C$. Then there are only finitely many such components $C$ of $U{\setminus}\{x\}$ that $J_C$ contains a critical point; denote their collection by $\mathcal C$. Let us show that an eventual $h$-image of every $E\in {\mathcal{A}}$ must coincide with an element of $\mathcal C$. Indeed, otherwise there is a component $E\in {\mathcal{A}}$ such that all $h^k(E)$ are distinct and the map $f_\sim|_{J_{h^k(E)}}$ is a homeomorphism. Clearly, this implies the existence of a wandering subcontinuum $K$ of $J_\sim$. However by Theorem C [@bl02] this is impossible. Hence all periodic points of $f_\sim$ are weakly repelling and by Theorem \[infprpt\] $f_\sim$ has infinitely many periodic cutpoints. Positively oriented maps of the plane ------------------------------------- In this subsection we will first obtain a general fixed point theorem which shows that if a non-separating plane continuum, not necessarily invariant, maps in an appropriate way, then it must contain a fixed point. This extends Theorem \[fmot1\]. Let us denote the family of all positively oriented maps of the plane by ${\mathcal{P}}$. \[fmot1\] Suppose that $f\in {\mathcal{P}}$ and that $X\subset {\mathbb{C}}$ is a non-separating continuum such that $f(X)\subset X$. Then there exists a fixed point $p\in X$. To proceed we will need to generalize Corollary \[posvar\] to a more general situation. To this end we introduce a definition similar to the one given for dendrites in the previous section. \[scracon\] Suppose that $f\in {\mathcal{P}}$ and $X$ is non-separating continuum. Suppose that there exist $n\ge 0$ disjoint non-separating continua $Z_i$ such that the following properties hold: 1. $f(X){\setminus}X\subset \cup_i Z_i$; 2. for all $i$, $Z_i\cap X=K_i$ is a non-separating continuum; 3. for all $i$, $f(K_i)\cap [Z_i{\setminus}K_i]={\emptyset}$. Then the map $f$ is said to *scramble the boundary (of $X$*). If instead of (3) we have 1. for all $i$, either $f(K_i)\subset K_i$, or $f(K_i)\cap Z_i={\emptyset}$ then we say that $f$ *strongly scrambles the boundary (of $X$)*. In either case, $K_i$’s are called *exit continua (of $X$)*. Note that since $Z_i\cap X$ is a continuum, $X\cup (\bigcup Z_i)$ is a non-separating continuum. Speaking of maps which (strongly) scramble the boundary, we always use the notation from this definition unless explicitly stated otherwise. Observe that in the situation of Definition \[scracon\] if $X$ is invariant then $f$ automatically strongly scrambles the boundary because we can simply take the set of exit continua to be empty. Also, if $f$ strongly scrambles the boundary of $X$ and $f(K_i)\not\subset K_i$ for any $i$, then it is easy to see that there exists ${\varepsilon}>0$ such that for every point $x\in X$ either $d(x, Z_i)>{\varepsilon}$, or $d(f(x),Z_i)>{\varepsilon}$. Let us now prove the following technical lemma. \[posvar+\] Suppose that $f:{\mathbb{C}}\to{\mathbb{C}}$ scrambles the boundary of $X$. Let $Q$ be a bumping arc of $X$ with endpoints $a<b\in X$ such that $f(\{a,b\})\subset X$ and $f(Q)\cap Q={\emptyset}$. Then ${\mathrm{Var(f,Q)}}\ge 0$. We will use the notation as specified in the lemma. Suppose first that $Q{\setminus}\bigcup Z_i\ne {\emptyset}$ and choose $v\in Q{\setminus}\bigcup Z_i$. Since $v\in Q{\setminus}\bigcup Z_i$ and $X\cup (\bigcup Z_i)$ is non-separating, there exists a junction $J_v$, with $v\in Q$, such that $J_v\cap[X\cup Q\cup \bigcup Z_i]= \{v\}$ and, hence, $J_v\cap f(X)\subset\{v\}$. Now the desired result follows from Corollary \[posvar\]. Observe that if $Q{\setminus}\cup Z_i={\emptyset}$ then $Q\subset Z_i$ for some $i$ and so $Q\cap X\subset K_i$. In particular, both endpoints $a,b$ of $Q$ are contained in $K_i$. Choose a point $v\in Q$. Then again there is a junction connecting $v$ and infinity outside $X$ (except possibly for $v$). Since all sets $Z_j, j\ne i$ are positively distant from $v$ and $X\cup (\bigcup_{i\ne j} Z_i)$ is non-separating, the junction $J_v$ can be chosen to avoid all sets $Z_j, j\ne i$. Now, by (3) $f(K_i)\cap J_v\subset \{v\}$, hence by Lemma \[posvar\] ${\mathrm{Var(f,Q)}}\ge 0$. Lemma \[posvar+\] is applied in Theorem \[fixpt\] in which we show that a map which strongly scrambles the boundary has fixed points. In fact, it is a major technical tool in our other results too. Indeed, if we can construct a bumping simple closed curve $S$ around $X$ such that the endpoints of its links map back into $X$ while these links move completely off themselves, the lemma would imply that the variation of $S$ is non-negative. By Theorem \[FMOT\] this would imply that the index of $S$ is positive. Hence by Theorem \[basic\] there are fixed points in $T(S)$. Choosing $S$ to be sufficiently tight we see that there are fixed points in $X$. For the sake of convenience we now sketch the proof of Theorem \[fixpt\] which allows us to emphasize the main ideas rather than details. The main steps in constructing $S$ are as follows. First we assume by way of contradiction that a map $f:{\mathbb{C}}\to {\mathbb{C}}$ has no fixed points in $X$. Then by Theorem \[fmot1\] it implies that $f(X)\not\subset X$ and that $f(K_i)\not\subset K_i$ for any $i$. By the definition of strong scrambling then $f(K_i)$ is far away from $Z_i$ for any $i$. Now, since there are no fixed points in $X$ we can choose the links in $S$ to be very small so that they will all move off themselves. However some of them will have endpoints mapping outside $X$ which prevents us from directly applying Lemma \[posvar+\] to them. These links will be enlarged by concatenating them so that the images of the endpoints of these concatenations are inside $X$ *and* these concatenations still map off themselves. The bumping simple closed curve $S$ then remains as before, however the representation of $S$ as the union of links changes because we enlarge some of them. Still, the construction shows that Lemma \[posvar+\] applies to the new “bigger” links and as before this implies the existence of a fixed point in $X$. To achieve the goal of replacing some links in $S$ by their concatenations we consider the links which are mapped outside $X$ in detail using the fact that $f$ strongly scrambles the boundary (indeed, all other links are such that Lemma \[posvar+\] already applies to them). The idea is to consider the links of $S$ whose concatenation is a connected piece of $S$ mapping into one $Z_i$. Then if we begin the concatenation right before the images of links enter $Z_i$ and stop it right after the images of the links exit $Z_i$ we will have one condition of Lemma \[posvar+\] satisfied because the endpoints of the thus constructed new “big” concatenation link $T$ of $S$ map into $X$. We now need to verify that $T$ moves off itself under $f$. Indeed, this is easy to see for the end-links of $T$: each end-link has the image “crossing” into $Z_i$ from $X{\setminus}Z_i$, hence the images of end-links are close to $K_i$. However the sets $K_i$ are mapped far away from $Z_i$ by the definition of strong scrambling and because none of $K_j$’s is invariant by the assumption. This implies that the end-links themselves must be far away from $Z_i$. If now we move from link to link inside $T$ we see that those links cannot approach $Z_i$ too closely because if they do they will have to “cross over $K_i$” into $Z_i$, and then their images will have to be close to the image of $K_i$ which is far away from $Z_i$, a contradiction with the fact that all links in $T$ have endpoints which map into $Z_i$. In other words, the dynamics of $K_i$ prevents the new bigger links from getting even close to $Z_i$ under $f$ which shows that they move off themselves as desired. As before, we now apply Theorem \[FMOT\] to see that ${\mathrm{Ind(f,S)}}={\mathrm{Var(f,S)}}+1$ and then Theorem \[basic\] to see that this implies the existence of a fixed point in $X$. Given a compact set $K$ denote by $B(K, {\varepsilon})$ the open set of all points whose distance to $K$ is less than ${\varepsilon}$. \[fixpt\] Suppose that $f:{\mathbb{C}}\to{\mathbb{C}}$ strongly scrambles the boundary of $X$. Then $f$ has a fixed point in $X$. If $f(X)\subset X$ then the result follows from [@fmot07]. Similarly, if there exists $i$ such that $f(K_i)\subset K_i$, then $f$ has a fixed point in $T(K_i)\subset X$ and we are also done. Hence we may assume $f(X){\setminus}X\ne{\emptyset}$ and $f(K_i)\cap Z_i={\emptyset}$ for all $i$. Suppose that $f$ is fixed point free. Then there exists ${\varepsilon}>0$ such that for all $x\in X$, $d(x,f(x))>{\varepsilon}$. We may assume that ${\varepsilon}<\min\{d(Z_i,Z_j)\mid i\ne j\}$. We now choose constants $\eta', \eta, {\delta}$ and a bumping simple closed curve $S$ of $X$ so that the following holds. 1. $0<\eta'<\eta<{\delta}<{\varepsilon}/3$. 2. For each $x\in X\cap B(K_i,3{\delta})$, $d(f(x), Z_i)>3{\delta}$. 3. For each $x\in X{\setminus}B(K_i,3{\delta})$, $d(x,Z_i)>3\eta$. 4. For each $i$ there exists a point $x_i\in X$ such that $f(x_i)=z_i\in Z_i$ and $d(z_i, X)>3\eta$. 5. $X\subset T(S)$ and $A=X\cap S=\{a_0<a_1<\dots<a_n<a_{n+1}=a_0\}$ in the positive circular order around $S$. 6. $f|_{T(S)}$ is fixed point free. 7. For the closure $Q_i=[a_i, a_{i+1}]$ of a component of $S{\setminus}X$, ${\mbox{diam}}(Q)+{\mbox{diam}}(f(Q))<\eta$. 8. For any two points $x, y\in X$ with $d(x, y)<\eta'$ we have $d(f(x), f(y))<\eta$. 9. $A$ is an $\eta'$-net in ${\mbox{Bd}}(X)$. Observe that by the triangle inequality, $Q_i\cap f(Q_i)={\emptyset}$ for every $i$. **Claim 1.** *There exists a point $a_j$ such that $f(a_j)\in X{\setminus}\cup {\overline}{B(Z_i, \eta)}$.* *Proof of Claim 1*. Set ${\overline}{B(Z_i, 3\eta)}=T_i$ and show that there exists a point $x\in {\mbox{Bd}}(X)$ with $f(x)\in X{\setminus}\cup T_i$. Indeed, suppose first that $n=1$. Then $f(K_1)\subset X{\setminus}T_1$ and we can choose any point of $K_1\cap {\mbox{Bd}}(X)$ as $x$. Now, suppose that $n\ge 2$. Observe that the sets $T_i$ are pairwise disjoint compacta. By Theorem \[orient\] $f({\mbox{Bd}}(X))\supset {\mbox{Bd}}(f(X))$. Hence there are points $x_1\ne x_2$ in ${\mbox{Bd}}(X)$ such that $f(x_1)\in Z_1\subset T_1, f(x_2)\in Z_2\subset T_2$. Since the sets $f^{-1}(T_i)\cap X$ are pairwise disjoint non-empty compacta we see that the set $V={\mbox{Bd}}(X){\setminus}\cup f^{-1}(T_i)$ is non-empty (because ${\mbox{Bd}}(X)$ is a continuum). Now we can choose any point of $V$ as $x$. It remains to notice that by the choice of $A$ we can find a point $a_j$ such that $d(a_j, x)<\eta'$ which implies that $d(f(a_j), f(x))<\eta$ and hence $f(a_j)\in X{\setminus}\cup {\overline}{B(Z_i, \eta)}$ as desired. There exists a point $x_1$ such that $f(x_1)=z_1$ is more than $3\eta$-distant from $X$. We can find $a\in A$ such that $d(a, x_1)<\eta'$ and hence $f(a){\not\in}X$ is at least $2\eta$-distant from $X$. On the other hand, by Claim 1 there are points of $A$ mapped into $X$. Let $<$ denote the circular order on the set $\{0,1,\dots,n+1\}$ defined by $i<j$ if $a_i<a_j$ in the positive circular order around $S$. Then we can find $n(1)<m(1)$ such that the following claims hold. 1. $f(a_{n(1)-1})\in X{\setminus}\cup Z_i$. 2. $f(a_r)\in f(X){\setminus}X$ for all $r$ with $n(1)\le r\le m(1)-1$ (and so, since ${\mbox{diam}}(f(Q_u))<{\varepsilon}/3$ for any $u$ and $d(Z_s,Z_t)>{\varepsilon}$ for all $s\ne t$, there exists $i(1)$ with $f(a_r)\in Z_{i(1)}$ for all $n(1)\le r<m(1)$). 3. $f(a_{m(1)}\in X{\setminus}\cup Z_i$. Consider the arc $Q'_1=[a_{n(1)-1},a_{m(1)}]\subset S$ and show that $f(Q'_1)\cap Q'_1={\emptyset}$. As we walk along $Q'_1$, we begin outside $Z_{i(1)}$ at $f(a_{n(1)-1})$, then enter $Z_{i(1)}$ and walk inside it, and then exit $Z_{i(1)}$ at $f(a_{m(1)})$. Since every step in this walk is rather short (${\mbox{diam}}(Q_i)+{\mbox{diam}}(f(Q_i))<\eta$), we see that $d(f(a_{n(1)-1}), Z_{i(1)})<\eta$ and $d(f(a_{m(1)}), Z_{i(1)})<\eta$. On the other hand for each $r, n(1)\le r<m(1)$ we have $f(a_r)\in Z_{i(1)}$, hence we see that $d(f(a_r), Z_{i(1)})<\eta$ for each $r, n(1)\le r<m(1)$. This implies that $d(a_r, K_{i(1)})>3\delta$ (because otherwise $f(a_r)$ would be farther away from $Z_{i(1)}$) and so $d(a_r), Z_{i(1)})>3\eta$ (because $a_r\in X{\setminus}B(K_{i(1)}, 3\delta)$). Since ${\mbox{diam}}(Q)+{\mbox{diam}}(f(Q))<\eta$, then $d(Q'_1, Z_{i(1)})>2{\delta}>2\eta$. On the other hand, $d(f(a_r), Z_{i(1)})<\eta$ similarly implies that $d(f(Q'_1), Z_{i(1)})<2\eta$. Thus indeed $f(Q'_1)\cap Q'_1={\emptyset}$. This allows us to replace the original division of $S$ into its prime links $Q_1, \dots, Q_n$ by a new one in which $Q'_1$ plays the role of a new prime link; in other words, we simply delete the points $\{a_{n(1)},\dots, a_{m(1)-1}\}$ from $A$. Continuing in the same manner and moving along $S$, in the end we obtain a finite set $A'=\{a_0=a'_0<a'_1<\dots<a'_k\}$ such that for each $i$ we have $f(a'_i)\in X\subset T(S)$ and for each arc $Q'_i=[a'_i,a'_{i+1}]$ we have $f(Q'_i)\cap Q'_i={\emptyset}$. Hence, by Theorem \[FMOT\], ${\mathrm{Ind(f,S)}}=\sum_{Q'_{i}} {\mathrm{Var(f,Q'_i)}}+1$. Since by Lemma \[posvar+\], ${\mathrm{Var(f,Q'_i)}}\ge 0$ for all $i$, ${\mathrm{Ind(f,S)}}\ge 1$ contradicting the fact that $f$ is fixed point free in $T(S)$. Maps with isolated fixed points ------------------------------- Now we consider maps $f\in {\mathcal{P}}$ with isolated fixed points; denote the set of such maps by ${\mathcal{P}}_i$. We need a few definitions. \[crit\] Given a map $f:X\to Y$ we say that $c\in X$ is a *critical point of $f$* if for any neighborhood $U$ of $c$, there exist $x_1\ne x_2\in U$ such that $f(x_1)=f(x_2)$. Hence, if $x$ is not a critical point of $f$, then $f$ is locally one-to-one near $x$. If a point $x$ belongs to a continuum collapsed under $f$ then $x$ is critical; also any point which is an accumulation point of collapsing continua is critical. However in this case the map around $x$ may be monotone. A more interesting case is when the map around $x$ is not monotone; then $x$ is a *branchpoint* of $f$ and it is critical even if there are no collapsing continua close by. One can define the *local degree* $\deg_f(a)$ as the number of components of $f^{-1}(y)$, for a point $y$ close to $f(a)$, which are non-disjoint from a small neighborhood of $a$. Then branchpoints are exactly the points at which the local degree is more than $1$. Notice that since we do not assume any smoothness, a critical point may well be both fixed (periodic) and be such that small neighborhoods of $c=f(c)$ map over themselves by $f$. The next definition is closely related to that of the index of the map on a simple closed curve. \[indpt\] Suppose that $x$ is a fixed point of a map $f\in {\mathcal{P}}_i$. Then the *local index of $f$ at $x$*, denoted by ${\mathrm{Ind(f,x)}}$, is defined as ${\mathrm{Ind(f,S)}}$ where $S$ is a small simple closed curve around $x$. It is easy to see that if $f\in {\mathcal{P}}_i$, then the local index is well-defined, i.e. does not depend on the choice of $S$. By modifying a translation map one can give an example of a homeomorphism of the plane which has exactly one fixed point $x$ with local index $0$. Still in some cases the local index at a fixed point must be positive. \[toprepat\] A fixed point $x$ is said to be *topologically repelling* if there exists a sequence of simple closed curves $S_i\to \{x\}$ such that $x\in {\mbox{int}}(T(S_i))\subset T(S_i)\subset {\mbox{int}}(T(f(S_i))$. A fixed point $x$ is said to be *topologically attracting* if there exists a sequence of simple closed curves $S_i\to \{x\}$ not containing $x$ and such that $x\in {\mbox{int}}(T(f(S_i))\subset T(f(S_i))\subset {\mbox{int}}(T(S_i))$. \[ind1\] If $a$ is a topologically repelling fixed point then ${\mathrm{Ind(f,a)}}=\deg_f(a)\ge 1$ where $d$ is the local degree. If however $a$ is a topologically attracting fixed point then ${\mathrm{Ind(f,a)}}=1$. Consider the case of the repelling fixed point $a$. Then it follows that, as $x$ runs along a small simple closed curve $S$ with $a\in T(S)$, the vector from $x$ to $f(x)$ produces the same winding number as the vector from $a$ to $f(x)$, and it is easy to see that the latter equals $\deg_f(a)$. The argument with attracting fixed point is similar. If however $x$ is neither topologically repelling nor topologically attracting, then ${\mathrm{Ind(f,x)}}$ could be greater than $1$ even in the non-critical case. Indeed, take a neutral fixed point of a rational function. Then it follows that if $f'(x)\ne 1$ then ${\mathrm{Ind(f,x)}}=1$ while if $f'(x)=1$ then ${\mathrm{Ind(f,x)}}$ is the multiplicity at $x$ (i.e., the local degree of the map $f(z)-z$ at $x$). This is related the following useful theorem. It is a version a more general, topological *argument principle* stated in the convenient for us form. \[argupr\] Suppose that $f\in {\mathcal{P}}_i$. Then for any simple closed curve $S\subset{\mathbb{C}}$ which contains no fixed points of $f$ its index equals the sum of local indices taken over all fixed points in $T(S)$. Theorem \[argupr\] implies Theorem \[basic\] but provides more information. In particular if $S$ were a simple closed curve and if we knew that the local index at any fixed point $a\in T(S)$ is $1$, it would imply that ${\mathrm{Ind(f,S)}}$ equals the number $n(f, S)$ of fixed points of $f$ in $T(S)$. By the above analysis this holds if all $f$-fixed points in $T(S)$ are either repelling, or attracting, or neutral and such that $f$ has a complex derivative $f'$ in a small neighborhood of $x$, and $f'(x)\ne 1$. In the spirit of the previous parts of the paper, we are still concerned with finding $f$-fixed points inside non-invariant continua of which $f$ (strongly) scrambles the boundary. However we now specify the types of fixed points we are looking for. Thus, the main result of this subsection proves the existence of specific fixed points in non-degenerate continua satisfying the appropriate boundary conditions and shows that in some cases such continua must be degenerate. It is in this latter form that we apply the result later on in Section 4. Given a non-separating continuum $X$, a ray $R\subset {\mathbb{C}}{\setminus}X$ from $\infty$ which lands on $x\in X$ (i.e., ${\overline}{R}{\setminus}R=\{x\}$) and a crosscut $Q$ of $X$ we say that $Q$ and $R$ *cross essentially* provided there exists $r\in R$ such that the subarc $[x,r]\subset {\overline}{R}$ is contained in ${\mbox{Sh}}(Q)$. The next definition complements the previous one. \[repout\] If $f(p)=p$ and $p\in {\mbox{Bd}}(X)$ then we say that $f$ *repels outside $X$ at $p$* provided there exists a ray $R\subset {\mathbb{C}}{\setminus}X$ from $\infty$ which lands on $p$ and a sequence of simple closed curves $S^j$ bounding closed disks $D^j$ such that $D^1\supset D^2\supset\dots, $ $\cap D^j=\{p\}$, $f(D^1\cap X)\subset X$, $f({\overline}{S^j{\setminus}X})\cap D^j={\emptyset}$ and for each $j$ there exists a component $Q^j$ of $S^j{\setminus}X$ such that $Q^j\cap R\ne{\emptyset}$ and ${\mathrm{Var(f,Q^j)}}\ne 0$. If $f\in {\mathcal{P}}$ and $f$ scrambles the boundary of $X$, then by Lemma \[posvar+\], for any component $Q$ of $S^j{\setminus}X$ we have ${\mathrm{Var(f,Q)}}\ge 0$ so that in this case ${\mathrm{Var(f,Q^j)}}>0$. The next theorem is the main result of this subsection. \[locrot\] Suppose that $f\in {\mathcal{P}}_i$, and $X\subset{\mathbb{C}}$ is a non-separating continuum or a point. Moreover, the following conditions hold. 1. \[fV\] For each fixed point $p\in X$ we have that ${\mathrm{Ind(f,x)}}=1$ and $f$ repels outside $X$ at $p$. 2. The map $f$ scrambles the boundary of $X$. Moreover, either $f(K_i)\cap Z_i={\emptyset}$, or there exists a neighborhood $W_i$ of $K_i$ with $f(W_i\cap X)\subset X$. Then $X$ is a point. Suppose that $X$ is not a point. Since $f\in {\mathcal{P}}_i$, there exists a simply connected neighborhood $V$ of $X$ such that all fixed points $\{p_1,\dots,p_m\}$ of $f|_{{\overline}{V}}$ are contained in $X$. We will show that then $f$ must have at least $m+1$ fixed points in $V$, a contradiction. The proof will proceed like the proof of Theorem \[fixpt\]: we construct a tight bumping simple closed curve $S$ such that $X\subset T(S)\subset V$. We will show that for an appropriate $S$, ${\mathrm{Var(f,S)}}\ge m$. Hence ${\mathrm{Ind(f,S)}}={\mathrm{Var(f,S)}}+1\ge m+1$ and by Theorem \[argupr\] $f$ must have at least $m+1$ fixed points in $V$. Let us choose neighborhoods $U_i$ of exit continua $K_i$ satisfying conditions listed below. 1. For $n_1<i\le n$ by assumption (2) of the theorem we may assume that $f(U_i\cap X)\subset X$. 2. For $1\le i\le n_1$ we may assume that $d(U_i\cup Z_i,f(U_i))>0$. 3. We may assume that $T(X\cup \bigcup {\overline}{U_i})\subset V$ and ${\overline}{U_i}\cap{\overline}{U_k}={\emptyset}$ for all $i\ne k$. 4. We may assume that every fixed point of $f$ contained in ${\overline}{U_i}$ is contained in $K_i$. Let $\{p_1,\dots,p_t\}$ be all fixed points of $f$ in $X{\setminus}\bigcup_i K_i$ and let $\{p_{t+1},\dots, p_m\}$ be all the fixed points contained in $\bigcup K_i$. Observe that then by the choice of neighborhoods $U_i$ we have $p_i\in X{\setminus}{\overline}{\cup U_s}$ if $1\le i\le t$. Also, it follows that for each $j, t+1\le j\le n$ there exists a unique $r_j, n_1<r_j\le n$ such that $p_j\in K_{r_j}$. For each fixed point $p_j\in X$ choose a ray $R_j\subset {\mathbb{C}}{\setminus}X$ landing on $p_j$, as specified in Definition \[repout\], and a small simple closed curve $S_j$ bounding a closed disk $D_j$ such that the following claims hold. 1. $D_i\cap R_j={\emptyset}$ for all $i\ne j$, 2. $f({\overline}{S_j{\setminus}X})\cap D_j={\emptyset}$. 3. $T(X\cup \bigcup_j D_j)\subset V$. 4. $[D_j\cup f(D_j)]\cap [D_k\cup f(D_k)]={\emptyset}$ for all $j\ne k$. 5. $f(D_j\cap X)\subset X$. 6. Denote by $Q(j,s)$ the components of $S_j{\setminus}X$; then there exists $Q(j,s(j))$, a component of $S_j{\setminus}X$, with ${\mathrm{Var(f,Q(j,s(j)))}}>0$ and $Q(j,s(j))\cap R_j\ne{\emptyset}$. 7. $[D_j\cup f(D_j)]\cap \bigcup {\overline}{U_i}={\emptyset}$ for all $1\le j\le t$. 8. If $t<j\le n$ then $[D_j\cup f(D_j)]\subset U_{r_j}$. Note that by (1) for all $i\ne j$, ${\mbox{Sh}}(Q(j,s(j))\cap Q(i,s(i))={\emptyset}$. We need to choose a few constants. First choose ${\varepsilon}>0$ such that for all $x\in X{\setminus}\bigcup D_j$, $d(x,f(x))>3{\varepsilon}$. Then by continuity we can choose $\eta>0$ such that for each set $H\subset V$ of diameter less than $\eta$ we have ${\mbox{diam}}(H)+{\mbox{diam}}(f(H))<{\varepsilon}$ and for each crosscut $C$ of $X$ disjoint from $\cup D_j$ we have that $f(C)$ is disjoint from $C$ (observe that outside $\cup D_j$ all points of $X$ move by a bounded away from zero distance). Finally we choose ${\delta}>0$ so that the following inequalities hold: 1. $3{\delta}<{\varepsilon}$, 2. $3{\delta}<d(Z_i,Z_j)$ for all $i\ne j$, 3. \[eU\] $3{\delta}<d(Z_i, [X\cup f(X)]{\setminus}[Z_i\cup U_i])$, 4. $3{\delta}<d(K_i,{\mathbb{C}}{\setminus}U_i)$, 5. if $f(K_i)\cap Z_i={\emptyset}$, then $3{\delta}<d(f(U_i),Z_i\cup U_i)$. Also, given a crosscut $C$ we can associate to its endpoints external angles ${\alpha}, {\beta}$ whose rays land at these endpoints from the appropriate side of $X$ determined by the location of $C$ (so that the open region of the plane enclosed by a tight equipotential between $R_{\alpha}$ and $R_{\beta}$, the segments of the rays from the equipotential to the endpoints of $C$, and $C$ itself, is disjoint from $X$). Thus we can talk about the angular measure of $Q(j, s(j))$; denote by ${\beta}$ the minimum of all such angular measures taken over all crosscuts $Q(j, s(j))$. Now, choose a bumping simple closed curve $S'$ of $X$ which satisfies two conditions: all its links are (a) less than ${\delta}$ in diameter, and (b) are of angular measures less than ${\beta}$. Clearly this is possible. Then we amend $S'$ as follows. Let us follow $S'$ in the positive direction starting at a link outside $\cup D_j$. Then at some moment for the first time we will be walking along a link of $S'$ which enters some $D_j$. As it happens, the link $L'$ of $S'$ intersects some $Q(j, s)$ with endpoints $a, b$ and enters the shadow ${\mbox{Sh}}(Q(j,s))$. Later on we will be walking outside ${\mbox{Sh}}(Q(j,s))$ moving along some link $L''$. In this case we replace the entire segment of $S'$ from $L'$ to $L''$ by three links: the first one is a deformation of $L'$ which has the same initial endpoint as $L'$ and the terminal point as $a$, then $Q(j, s)$, and then a deformation of $L''$ which begins at $b$ and ends at the same terminal point as $L''$. In this way we make sure that for all crosscuts $Q(j, s)$ either they are links of $S'$ or they are contained in the shadow of a link of $S'$. Moreover, by the choice of ${\beta}$ crosscuts $Q(j, s(j))$ will have to become links of our new bumping simple closed curve $S$. By the choice of $\eta$ and by the properties of crosscuts $Q(j, s)$ it follows that any link of $S$ is disjoint from its image, and for each $j$, $Q(j,s(j))\subset S$ and ${\mathrm{Var(f,Q(j,s(j)))}}>0$. We want to compute the variation of $S$. Each link $Q(j,s(j))$ contributes at least $1$ towards ${\mathrm{Var(f,S)}}$, and we want to show that all other links have non-negative variation. To do so we want to apply Lemma \[posvar+\]. Hence we need to verify two conditions on a crosscut listed in Lemma \[posvar+\]. One of them follows from the previous paragraph: all links of $S$ move off themselves. However the other condition of Lemma \[posvar+\] may not be satisfied by some links of $S$ because some of their endpoints may map off $X$. To ensure that for our bumping simple closed curve endpoints $e$ of its links map back into $X$ we have to enlarge links of $S$ and replace some of them by their concatenations (this is similar to what was done in Theorem \[fixpt\]). Then we will have to check if the new “bigger” links still have images disjoint from themselves. Suppose that $ X\cap S=A=\{a_0<a_1<\dots<a_n\}$ and $a_0\in A$ is such that $f(a_0)\in X$ (the arguments similar to those in Theorem \[fixpt\] show that we can may this assumption without loss of generality). Let $t'$ be minimal such that $f(a_{t'})\not\in X$ and $t''>t'$ be minimal such that $f(a_{t''})\in X$. Then $f(a_{t'})\in Z_i$ for some $i$. Denote by $[a_l, a_r]$ a subarc of $S$ with the endpoints $a_r$ and $a_l$ and moving from $a_l$ to $a_r$ is the positive direction. Since every component of $[a_{t'},a_{t''}]{\setminus}X$ has diameter less than ${\delta}$, $f(a_t)\in Z_i{\setminus}X$ for all $t'\le t<t''$. Moreover, for $t'\le t<t''$, $a_t\not \in U_i$. To see this note that if $f(K_i)\cap Z_i={\emptyset}$, then by the above made choices $f(U_i)\cap Z_i={\emptyset}$, and if $f(K_i)\cap Z_i\ne{\emptyset}$, then $f(U_i\cap X)\subset X$ by the assumption. Hence it follows from the property (\[eU\]) of the constant ${\delta}$ that $f([a_{t'-1},a_{t''}])\cap [a_{t'-1},a_{t''}]={\emptyset}$ and we can remove the points $a_t$, for $t'\le t<t''$ from the partition $A$ of $S$. By continuing in the same fashion we obtain a subset $A'\subset A$ such that for the closure of each component $C$ of $S{\setminus}A'$, $f(C)\cap C={\emptyset}$ and for both endpoints $a$ and $a'$ of $C$, $\{f(a),f(a')\}\subset X$. Moreover, for each $j$, $Q(j,j(s))$ is a component of $S{\setminus}A'$. Now we can apply a variation of the standard argument sketched in Section 2 after Theorem \[basic\] and applied in the proof of Theorem \[fixpt\]; in this variation instead of Theorem \[basic\] we use the fact that $f$ satisfies the argument principle. Indeed, by Theorem \[FMOT\] and Lemma \[posvar+\], ${\mathrm{Ind(f,S)}}\ge \sum_j {\mathrm{Var(f,Q(j,j(s)))}}+1\ge m+1$ and by the Theorem \[argupr\] $f$ has at least $m+1$ fixed points in $T(S)\subset V$, a contradiction. Theorem \[locrot\] implies the following \[degenerate\] Suppose that $f$ and a non-degenerate $X$ satisfy all the conditions stated in Theorem \[locrot\]. Then either $f$ does not repel outside $X$ at one of its fixed points, or the local index at one of its fixed points is not equal to $1$. The last lemma of this section gives a sufficient and verifiable condition for a fixed point $a$ belonging to a locally invariant continuum $X$ to be such that the map $f$ repels outside $X$; we apply the lemma in the next section. \[repel\] Suppose that $f:{\mathbb{C}}\to{\mathbb{C}}$ is positively oriented, $X\subset{\mathbb{C}}$ is a continuum and $p$ is a fixed point of $f$ such that: 1. there exists a neighborhood $U$ of $p$ such that $f|_U$ is one-to-one and $f(U\cap X)\subset X$, 2. there exists a closed disk $D\subset U$ containing $p$ in its interior such that $f(\partial D)\cap D={\emptyset}$ and $\partial D{\setminus}X$ has at least two components, 3. there exists a ray $R\subset{\mathbb{S}}{\setminus}X$ from infinity such that ${\overline}{R}=R\cup \{p\}$, $f|_R:R\to R$ is a homeomorphism and for each $x\in R$, $f(x)$ separates $x$ from $\infty$ in $R$. Then there exists a component $C$ of $\partial D{\setminus}X$ so that $C\cap R\ne{\emptyset}$, ${\mathrm{Var(f,C)}}=+1$ and $f$ repels outside $X$ at $p$. We may assume that $X{\setminus}U$ contains a continuum. Let ${\mathbb{D}}^{\infty}={\mathbb{S}}{\setminus}{\overline}{{\mathbb{D}}}$ be the open disk at infinity and let ${\varphi}:{\mathbb{D}}^{\infty}\to {\mathbb{S}}{\setminus}X$ be a conformal map such that ${\varphi}({\infty})={\infty}$. Then $T={\varphi}^{-1}(R)$ is a ray in ${\mathbb{D}}^\infty$ which compactifies on a point ${\widehat{p}}\in S^1$. Let $Q_j$ be all components of ${\varphi}^{-1}(\partial D{\setminus}X)$. Then each $Q_j$ is a crosscut of ${\mathbb{D}}^{\infty}$. Let $O=\{z\in{\mathbb{D}}^{\infty}\mid f\circ {\varphi}(z)\in{\mathbb{S}}{\setminus}X$ and define $F:O\to{\mathbb{D}}^{\infty}$ by $F(z)={\varphi}^{-1}\circ f\circ {\varphi}(z)$. Note that $T\cup \bigcup Q_j\subset O$. We may assume that ${\overline}{Q_1}$ separates ${\widehat{p}}$ from ${\infty}$ in ${\overline}{{\mathbb{D}}^{\infty}}$ and that no other $Q_j$ separates $Q_1$ from ${\infty}$ in ${\mathbb{D}}^{\infty}$. **Claim.** $F(Q_1)$ separates $Q_1$ from ${\infty}$ in ${\mathbb{D}}^{\infty}$. *Proof of Claim.* Let $T_{\infty}$ be the component of $T{\setminus}Q_1$ which contains ${\infty}$ and let $T_p$ be the component of ${\overline}{T}{\setminus}Q_1$ which contains ${\widehat{p}}$. Let $a={\overline}{T_{\infty}}\cap Q_1$ and $b={\overline}{T_p}\cap Q_1$. Choose a point $b'\in T_p$ very close to $b$ so that the subarc $[b,b']\subset T_p$ is contained in ${\varphi}^{-1}(D)$. Let $T'_p\subset T_p$ be the closed subarc from ${\widehat{p}}$ to $b'$. Choose an open arc $A$ in the bounded component of ${\mathbb{D}}^{\infty}{\setminus}Q_1$, very close to $Q_1$ from $a$ to the point $b'\in T_p$ so that $f|_{T'_p\cap A\cup T_{\infty}}$ is one to-one. Put $Z=T'_p\cap A\cup T_{\infty}$, then $Q_1\cap Z=\{a\}$ and $F(Q_1)\cap F(Z)=\{F(a)\}$. Since $F$ is a local orientation preserving homeomorphism near $a$, $F(Z)$ enters the bounded component of ${\mathbb{D}}^{\infty}{\setminus}F(Q_1)$ at $F(a)$ and never exits this component after entering it. Moreover, if $q$ is an endpoint of $Q_1$, then points very close to $q$ on $Q_1$ and their images are on the same side of $T$. Since $F(a)$ separates $a$ from ${\infty}$ on $Z$ and an initial segment of $T_p$ (with endpoint ${\widehat{p}}$) is contained in $F(Z)$, ${\widehat{p}}\in{\mbox{Sh}}(F(Q_1))$. This completes the proof of the claim. Let us compute the variation ${\mathrm{Var(F,Q_1)}}$ of the crosscut $Q_1$ with respect to the continuum $S^1$. Since the computation is independent of the choice of the Junction [@fmot07], we can choose a junction $J_v$ with junction point $v\in Q_1$ so that each of the three rays $R_+$, $R_i$ and $R_-$ intersect $F(Q_1)$ in exactly one point. Hence ${\mathrm{Var(F,Q_1)}}=+1$. Since ${\varphi}$ is an orientation preserving homeomorphism, ${\mathrm{Var(f,{\varphi}(Q_1))}}=+1$ and we are done. Applications ============ The results in the previous section can be used to obtain results in complex dynamics (see for example [@bco08]). We will show that in certain cases continua (e.g., impressions of external rays) must be degenerate. Suppose that $P:{\mathbb{C}}\to{\mathbb{C}}$ is a complex polynomial of degree $d$ with a connected Julia set $J$. Let the filled-in Julia set be denoted by $K=T(J)$. We denote the external rays of $K$ by $R_{\alpha}$. It is well known [@douahubb85] that if the degree of $P$ is $d$ and $\sigma:{\mathbb{C}}\to{\mathbb{C}}$ is defined by $\sigma(z)=z^d$, then $P(R_{\alpha})=R_{\sigma({\alpha})}$. Let for ${\lambda}\in{\mathbb{C}}$, $L_{\lambda}:{\mathbb{C}}\to{\mathbb{C}}$ be defined by $L_{\lambda}(z)={\lambda}z$. Suppose that $p$ is a fixed point in $J$ and ${\lambda}=f'(p)$ with $|{\lambda}|>1$ (i.e., $p$ is a *repelling* fixed point). Then there exists neighborhoods $U\subset V$ of $p$ and a conformal isomorphism ${\varphi}:V\to{\mathbb{D}}$ such that for all $z\in U$, $P(z)={\varphi}^{-1}\circ L_{\lambda}\circ {\varphi}(z)$. Now, a fixed point $p$ is *parabolic* if $P'(p)=e^{2\pi i r}$ for some rational number $r\in\mathbb{Q}$. A nice description of the local dynamics at a parabolic fixed point can be found in [@miln00]. If $p$ is a repelling or parabolic fixed point then [@douahubb85] there exist $k\ge 1$ external rays $R_{{\alpha}(i)}$ such that ${\sigma}|_{\{{\alpha}(1),\dots,{\alpha}(k)\}}: \{{\alpha}(1),\dots,{\alpha}(k)\}\to \{{\alpha}(1),\dots,{\alpha}(k)\}$ is a permutation, $P(R_{{\alpha}(i)})=R_{{\sigma}({\alpha}(i))}$, for each $j$, $R_{{\alpha}(j)}$ lands on $p$ and no other external rays land on $p$. Also, if $P(R_{{\alpha}(i)})=R_{{\alpha}(i)}$ for some $i$, then ${\sigma}({\alpha}(j))={\alpha}(j)$ for all $j$. It is known that two distinct external rays are not homotopic in the complement of $K$. Given an external ray $R_{\alpha}$ of $K$ we denote by $\Pi({\alpha})={\overline}{R_{\alpha}}{\setminus}R_{\alpha}$ the *principle set of ${\alpha}$*, and by ${\mbox{Imp}}({\alpha})$ the *impression of ${\alpha}$* (see [@miln00]). Given a set $A\subset S^1$, we extend the above notation by $\Pi(A)=\bigcup_{{\alpha}\in A} \Pi({\alpha})$ and ${\mbox{Imp}}(A)=\bigcup_{{\alpha}\in A} {\mbox{Imp}}({\alpha})$. Let $X\subset K$ be a non-separating continuum or a point such that: 1. \[P1\] Pairwise disjoint non-separating continua/points $E_1\subset X, \dots, E_m\subset X$ and finite sets of angles $A_1=\{{\alpha}^1_1, \dots, {\alpha}^1_{i_1}\}, \dots, A_m=\{{\alpha}^m_1, \dots, {\alpha}^m_{i_m}\}$ are given with $i_k\ge 2, 1\le k\le m$. 2. \[P2\] We have $\Pi(A_j)\subset E_j$ (so the set $E_j\cup (\cup^{i_j}_{k=1} R_{{\alpha}^j_k})=E'_j$ is closed and connected). 3. \[P3\] $X$ intersects a unique component $C$ of ${\mathbb{C}}{\setminus}\cup E'_j$, such that $X{\setminus}\bigcup E_j=C\cap K$. We call such $X$ a *general puzzle-piece* and call the continua $E_i$ the *exit continua* of $X$. Observe that if $U$ is a Fatou domain then either a general puzzle-piece $X$ contains $U$, or it is disjoint from $U$. For each $j$, the set $E'_j$ divides the plane into $i_j$ open sets which we will call *wedges (at $E_j$)*; denote by $W_j$ the wedge which contains $X{\setminus}E_j$. Let us now consider the condition (1) of Theorem \[locrot\]. It is easy to see that applied “as is” to the polynomial $P$ at parabolic points it is actually not true. Indeed, as explained above the local index at parabolic fixed points at which the derivative equals $1$ is greater than $1$. And indeed, in our case there are fixed rays landing at all fixed points, therefore [@miln00] the derivatives at all the parabolic points in $X$ are equal to $1$. The idea which allows us to solve this problem is that we can change our map $P$ inside the parabolic domains in question without compromising the rest of the arguments and making these parabolic points topologically repelling. The thus constructed new map $g$ will satisfy conditions of Theorem \[locrot\]. \[pararepel\] Suppose that $X$ is a continuum and $p\in X$ is a parabolic point of a polynomial $f$ and $R$ is a fixed external which lands at $p$. Then $f$ repels outside $X$ at $p$. Let $p\in X$ be a parabolic fixed point and let $F_i$ be the parabolic domains containing $p$ in their boundaries $B_i$. Since there are fixed rays landing at $p$, all $F_i$’s are forward invariant. By a nice recent result of Yin and Roesch [@roesyin08], the boundary $B_i$ of each $F_i$ is a simple closed curve and $f|_{B_i}$ is conjugate to the map $z\to z^{d(i)}$ for some $d(i)\ge 2$. Let $\psi:F_i\to{\mathbb{D}}$ be a conformal isomorphism. Since ${\mbox{Bd}}(F_i)$ is a simple closed curve, $\psi$ extends to a homeomorphism. Since $f|_{B_i}$ is conjugate to the map $z\to z^{d(i)}$, it now follows that the map $P|_{{\overline}{F_i}}$ can be replaced by a map topologically conjugate by $\psi$ to the map $g_i(z)=z^{d(i)}$ on the closed unit disk. Let $g$ be the map defined by $g(z)=P(z)$ for each $z\in{\mathbb{C}}{\setminus}\bigcup F_i$ and $g(z)=g_i(z)$ when $z\in F_i$. Then $g$ is clearly a positively oriented map. The well-known analysis of the dynamics of $P$ around parabolic points [@miln00] implies that $P$ repels points away from $p$ outside parabolic domains $F_i$. In other words, we can find a sequence of simple closed curves $S_i$ which satisfy conditions of Definition \[toprepat\] and show that $p$ is a topologically repelling point of $g$. Hence the local index ${\mathrm{Ind(g,p)}}$ at $p$ equals $1$. On the other hand, by Lemma \[repel\] and properties of $X$ it follows that $g$ repels outside $X$ at $p$. Since $f$ and $g$ coincide outside $X$, $f$ also repels at $p$. The following corollary follows from Theorem \[locrot\]. \[pointdyn\] Suppose that $X\subset K$ is a non-separating continuum or a point. Then the following claims hold for $X$. 1. \[1\] Suppose that $X$ is a general puzzle-piece with exit continua $E_1, \dots, E_m$ such that either $P(E_i)\subset W_i$, or $E_i$ is a fixed point. If $X$ does not contain an invariant parabolic domain, all fixed points which belong to $X$ are repelling or parabolic, and all rays landing at them are fixed, then $X$ is a repelling or parabolic fixed point. 2. \[2\] Suppose that $X\subset J$ is an invariant continuum, all fixed points which belong to $X$ are repelling or parabolic, and all rays landing at them are fixed. Then $X$ is a repelling or parabolic fixed point. By way of contradiction we can assume that $X$ is not a point. Let us show that no parabolic domain with a fixed point on its boundary can intersect $X$. Indeed, in the case (2) $X\subset J$ and no Fatou domain intersects $X$, so there is nothing to prove. In the case (1) observe that since $X$ is a general puzzle-piece, it has to contain the closure of the entire parabolic domain with a fixed point, say, $p$ on its boundary. Then the fact that all external rays landing at $p$ are fixed implies that all parabolic domains containing $p$ in their boundaries are invariant. Since by the assumptions $X$ contains no invariant parabolic domain, it does not contain any of them. So, $X$ is disjoint from all parabolic domains containing a fixed point in their boundaries. To apply Theorem \[locrot\] we need to verify that its conditions apply. It is easier to check the condition (2) first. To do so, observe first that $f(X)\cap X\ne {\emptyset}$. Indeed, otherwise no set $E_i$ is a fixed point and $f(X)$ must be contained in one of the wedges formed by some $E'_l$ but not in the wedge $W_l$. This implies that $E_l$ neither is a fixed point, nor is mapped in $W_l$, a contradiction. Thus, $f(X)\cap X\ne {\emptyset}$ and we can think of $f(X)$ as a continuum which “grows” out of $X$. Now, any component of $f(X){\setminus}X$ which intersects $E_k$ for some $k$ must be contained in one of the wedges at $E_k$, but not in $W_k$. Take the closure of their union and then its topological hull union $E_i$ and denote it by $Z_i$. It is easy to check now that with these sets $Z_i$ the map $P$ scrambles the boundary of $X$. Moreover, if $E_i$ is mapped into $W_i$ then clearly $P(E_i)\cap Z_i={\emptyset}$ (because $Z_i$ is contained in the other wedges at $E_i$ but is disjoint from $W_i$). On the other hand, if $E_i$ is a fixed point then it is a repelling or parabolic fixed point with a few external fixed rays landing at it. Hence in a small neighborhood $U_i$ of $E_i$ the intersection $U_i\cap X$ maps into $X$ as desired in the condition (2) of Theorem  \[locrot\]. By Lemmas \[repel\] and \[pararepel\] $P$ repels outside $X$ at any fixed point in $X$. Moreover, using the map $g$ constructed in the proof of Lemma \[pararepel\], we see that $g$ is topologically repelling at $p$ and, hence ${\mathrm{Ind(g,p)}}=+1$. Hence the conditions of Theorem \[locrot\] are satisfied for the map $g$. Thus, by Theorem \[locrot\] we conclude that $X$ is a point as desired. The following is an immediate corollary of Theorem \[pointdyn\]. \[rot-neutr\] Suppose that for a non-separating *non-degenerate* continuum $X\subset K$ one of the following facts hold. 1. $X$ is a general puzzle-piece with exit continua $E_1, \dots, E_m$ such that either $P(E_i)\subset W_i$, or $E_i$ is a fixed point. 2. $X\subset J$ is an invariant continuum. Then either $X$ contains a non-repelling and non-parabolic fixed point, or $X$ contains an invariant parabolic domain, or $X$ contains a repelling or parabolic fixed point at which a non-fixed ray lands. Finally, the following corollary is useful in proving the degeneracy of certain impressions and establishing local continuity of the Julia set at some points. Let $P:{\mathbb{C}}\to{\mathbb{C}}$ be a complex polynomial and $R_{\alpha}$ is a fixed external ray landing on a repelling or parabolic fixed point $p\in J$. Suppose that $T({\mbox{Imp}}({\alpha}))$ contains only repelling or parabolic periodic points. Then ${\mbox{Imp}}({\alpha})$ is degenerate. Let $X={\mbox{Imp}}({\alpha})$. Since $R_{\alpha}$ is a fixed external ray, $P(X)\subset X$. Clearly $P$ does not rotate at $p$. Suppose that $p'$ is another fixed point of $P$ in $X$ and $R_{\beta}$ is an external ray landing at $p'$. Then $P(R_{\beta})$ also lands on $p'$. If $P$ rotates at $p'$, then $p'$ is a cut point of $X$. This would contradict the fact that $X={\mbox{Imp}}({\alpha})$. Hence $P$ does not rotate at any fixed point in $X$ and the result follows from Corollary \[pointdyn\]. [FMOT07]{} V. Akis, *On the plane fixed point problem*, Topology Proc. **24** (1999), 15–31. H. Bell, *On fixed point properties of plane continua*, Trans. A. M. S. **128** (1967), 539–548. [to3em]{}, *A fixed point theorem for plane homeomorphisms*, Fund. Math. **100** (1978), 119–128, See also: Bull. A. M. S. 82(1976), 778-780. A. Blokh and G. Levin, *An inequality for laminations, Julia sets and ‘growing trees’*, Erg. Th. and Dyn. Sys., **22** (2002), pp. 63–97. M. Bonino, *A [B]{}rouwer like theorem for orientation reversing homeomorphisms of the sphere*, Fund. Math. **182** (2004), 1–40. A. Blokh, C. Clinton, L. Oversteegen, *Locally connected models for Julia sets*, preprint (2008). L. E. J. Brouwer, *Beweis des ebenen [T]{}ranslationessatzes*, Math. Ann. **72** (1912), 35–41. Morton Brown, *A new proof of [B]{}rouwer’s lemma on translation arcs*, Houston J. of Math. **10** (1984), 35–41. L. Carleson and T. W. Gamelin, *Complex dynamics*, Universitext: Tracts in Mathematics, Springer-Verlag, 1993. M. L. Cartwright and J. E. Littlewood, *Some fixed point theorems*, Annals of Math. **54** (1951), 1–37. A. Douady and J. H. Hubbard, *Étude dynamique des polynômes complexes I, II* Publications Mathématiques d’Orsay **84-02** (1984), **85-04** (1985). A. Douady and J. H. Hubbard, *On the dynamics of polynomial-like mappings*, Ann. Sci. École Norm. Sup. (4) **18** (1985), no. 2, 287–343. Albert Fathi, *An orbit closing proof of [B]{}rouwer’s lemma on translation arcs*, L’enseignement Mathématique **33** (1987), 315–322. R. J. Fokkink, J.C. Mayer, L. G. Oversteegen, and E.D. Tymchatyn, *[H]{}arold [B]{}ell and the plane fixed point problem*, preprint arXiv:0805.1184 (2008). J. Franks, *A new proof of the [B]{}rouwer plane translation theorem*, Ergodic Theory and Dynamical Systems **12** (1992), 217–226. L. Guillou, *Téotème de translation plane de [B]{}rouwer et généralisations du téotème de [P]{}oincaré-[B]{}irkhoff*, Topology **33** (1994), 331–351. S. D. Iliadis, *Location of continua on a plane and fixed points*, Vestnik Moskovskogo Univ. Matematika **25** (1970), no. 4, 66–70, Series [I]{}. A. Lelek and D. Read, *Compositions of confluent mappings and some other classes of functions*, Colloq. Math., **29** (1974), 101–112. J. Milnor, *Dynamics in one complex variable*, second ed., Vieweg, Wiesbaden, 2000. S. B. Nadler, Jr., [*Continuum theory*]{}, Marcel Dekker Inc., New York, 1992. P. Roesch, Y. Yin, *The boundary of bounded polynomial Fatou components*, Comptes Rendus Mathematique **346** (2008), 877-880. K. Sieklucki, *On a class of plane acyclic continua with the fixed point property*, Fund. Math. **63** (1968), 257–278. Sternbach, Problem [**107**]{} (1935), in: [*The [S]{}cottish [B]{}ook: [M]{}athematics from the [S]{}cottish [C]{}afé*]{}, Birkhauser, Boston, 1981, 1935. W. P. Thurston, *On the geometry and dynamics of iterated rational maps*, Preprint, 1985. G. T. Whyburn, *Analytic topology*, vol. 28, AMS Coll. Publications, Providence, RI, 1942. [^1]: The first author was partially supported by NSF grant DMS-0456748 [^2]: The second author was partially supported by NSF grant DMS-0405774

ArXiv

--- abstract: 'We model the growth of a cell population by a piecewise deterministic Markov branching tree. Each cell splits into two offsprings at a division rate $B(x)$ that depends on its size $x$. The size of each cell grows exponentially in time, at a rate that varies for each individual. We show that the mean empirical measure of the model satisfies a growth-fragmentation type equation if structured in both size and growth rate as state variables. We construct a nonparametric estimator of the division rate $B(x)$ based on the observation of the population over different sampling schemes of size $n$ on the genealogical tree. Our estimator nearly achieves the rate $n^{-s/(2s+1)}$ in squared-loss error asymptotically. When the growth rate is assumed to be identical for every cell, we retrieve the classical growth-fragmentation model and our estimator improves on the rate $n^{-s/(2s+3)}$ obtained in [@DHRR; @DPZ] through indirect observation schemes. Our method is consistently tested numerically and implemented on [*Escherichia coli*]{} data.' author: - 'Marie Doumic[^1]' - 'Marc Hoffmann[^2]' - 'Nathalie Krell[^3]' - 'Lydia Robert[^4]' bibliography: - 'biblio.bib' title: 'Statistical estimation of a growth-fragmentation model observed on a genealogical tree' --- [ [*Keywords:*]{} Growth-fragmentation, cell division equation, Nonparametric estimation, Markov chain on a tree.\ [*Mathematical Subject Classification:*]{} 35A05, 35B40, 45C05, 45K05, 82D60, 92D25, 62G05, 62G20.\ ]{} Introduction ============ Size-structured models and their inference {#size-strutured models} ------------------------------------------ In mathematical biology, physiologically structured equations [@Metz] allow to describe the temporal evolution of a population characterised by state variables such as age, size, growth, maturity, protein content and so on (see for instance [@Metz; @Pe] and the references therein). A paradigmatic example is given by the growth-fragmentation or size-structured cell division equation. For the evolution of a bacterial population it reads $$\label{transport-fragmentation eq} \left\{ \begin{array}{l} \partial_t n(t,x) + \tau \partial_x\big(x\, n(t,x)\big) + B(x)n(t,x) = 4B(2x)n(t,2x) \\ \\ n(0,x)= n^{(0)}(x), x \geq 0, \end{array} \right.$$ and it quantifies the density $n(t,x)$ of cells having size $x$ (the state variable) at time $t$. A common stochastic mechanism for every single cell is attached to : 1. The size $x = x(t)$ of a cell at time $t$ evolves exponentially according to the deterministic evolution $dx(t)=\tau x(t)dt$, where $\tau >0$ is the growth rate of each cell, that quantifies their ability to ingest a common nutrient. 2. Each cell splits into two offsprings according to a division rate $B(x)$ that depends on its current size $x$. 3. At division, a cell of size $x$ gives birth to two offsprings of size $x/2$ each, what is called *binary fission*. Model is thus entirely determined by the parameters $\big(\tau, B(x), x \in [0,\infty)\big)$. Typically, $\tau$ is assumed to be known or guessed [@DMZ], and thus inference about mainly concerns the estimation of the division rate $B(x)$ that has to be taken from a nonparametric perspective.\ By use of the general relative entropy principle, Michel, Mischler and Perthame showed that the approximation $n(t,x) e^{-\lambda_0 t} \approx N(x)$ is valid [@MMP], with $\lambda_0 >0$, and where $(\lambda_0,N)$ is the dominant eigenpair related to the corresponding eigenvalue problem, see [@M1; @Pe; @DG; @LP; @CCM; @BCG]. The “stationary" density $N(x)$ of typical cells after some time has elapsed enables to recover $(B(x), x \in {\mathcal D}\big)$ for a compact ${\mathcal D} \subset (0,\infty)$ by means of the regularisation of an inverse problem of ill-posedness degree 1. From a deterministic perspective, this is carried out in [@PZ; @DPZ; @DT]. From a statistical inference perspective, if an $n$-sample of the distribution $N(x)$ is observed and if $B(x)$ has smoothness $s>0$ in a Sobolev sense, it is proved in [@DHRR] that $B(x)$ can be recovered in squared-error loss over compact sets with a rate of convergence $n^{-s/(2s+3)}$. Both deterministic and stochastic methodology of [@DPZ] and [@DHRR] are motivated by experimental designs and data such as in [@Kubitschek; @DMZ]. However, they do not take into account the following two important aspects: - Bacterial growth exhibits variations in the individual growth rate $\tau$ as demonstrated for instance in [@Sturm]. One would like to incorporate variability in the growth rate within the system at the level of a single cell. This requires to modify Model . - Recent evolution of experimental technology enables to track the whole genealogy of cell populations (along prescribed lines of descendants for instance), affording the observation of other state variables such as size at division, lifetime of a single individual and so on [@Robert]. Making the best possible use of such measures is of great potential impact, and needs a complementary approach. The availability of observation schemes at the level of cell individuals suggests an enhancement of the statistical inference of $\big(B(x), x \in {\mathcal D}\big)$, possibly enabling to improve on the rates of convergence obtained by indirect measurements such as in [@DHRR; @DPZ]. This is the purpose of the present paper. Results of the paper -------------------- ### Statistical setting {#statistical-setting .unnumbered} Let $${\mathcal U} = \bigcup_{k=0}^\infty \{0,1\}^k$$ denote the binary genealogical tree (with $\{0,1\}^0:=\{\emptyset\}$). We identify each node $u \in {\mathcal U}$ with a cell that has a size at birth $\xi_u$ and a lifetime $\zeta_u$. In the paper, we consider the problem of estimating $\big(B(x), x \in [0,\infty)\big)$ over compact sets of $(0,\infty)$. Our inference procedure is based on the observation of $$\label{micro dataset} \big((\xi_u, \zeta_u), u \in {\mathcal U}_n\big).$$ where ${\mathcal U}_n \subset {\mathcal U}$ denotes a connected subset of size $n$ containing the root $u = \emptyset$. Asymptotics are taken as $n\rightarrow \infty$. Two important observation schemes are considered: the sparse tree case, when we follow the system along a given branch with $n$ individuals, and the full tree case, where we follow the evolution of the whole binary tree up to the $N_n$-th generation, with $N_n \approx \log_2 n$. In this setting, we are able to generalise Model and allow the growth rate $\tau$ to vary with each cell $u \in {\mathcal U}$. We assume that a given cell $u$ has a random growth rate $\tau_u=v \in {\mathcal E} \subset (0,\infty)$ (later constrained to live on a compact set). Moreover, this value $v$ is inherited from the growth rate $v'$ of its parent according to a distribution $\rho(v',dv)$. Since a cell splits into two offsprings of the same size, letting $u^-$ denote the parent of $u$, we have the fundamenal relationship $$\label{fundamental} 2\,\xi_u = \xi_{u^-}\exp\big(\tau_{u^-} \zeta_{u^-}\big)$$ that enables to recover the growth rate $\tau_u$ of each individual in ${\mathcal U}_n$ since ${\mathcal U}_n$ is connected by assumption, possibly leaving out the last generation of observed individuals, but this has asymptotically no effect on a large sample size approach. ### Variability in the growth rate {#variability-in-the-growth-rate .unnumbered} In the case where the growth rate can vary for each cell, the density $n(t,x)$ of cells of size $x$ at time $t$ does not follow Eq.  anymore and an extended framework needs to be considered. To that end, we structure the system with an additional variable $\tau_u=v,$ which represents the growth rate and depends on each individual cell $u$. We construct in Section \[microscopic model\] a branching Markov chain $\big((\xi_u, \tau_u), u \in {\mathcal U}\big)$ that incorporates variability for the growth rate in the mechanism described in Section \[size-strutured models\]. Equivalently to the genealogical tree, the system may be described in continuous time by a piecewise deterministic Markov process $$\big(X(t),V(t)\big) = \Big(\big(X_1(t),V_1(t)\big),\big(X_2(t),V_2(t)\big),\ldots\Big),$$ which models the process of sizes and growth rates of the living particles in the system at time $t$, with value in $\bigcup_{k=0}^\infty {\mathcal S}^k$, where ${\mathcal S} = [0,\infty)\times {\mathcal E}$ is the state space of size times growth rate. Stochastic systems of this kind that correspond to branching Markov chains are fairly well known, both from a theoretical angle and in applications; a selected list of contributions is [@BDMV; @Cloez; @MT] and the references therein. By fragmentation techniques inspired by Bertoin [@bertoin], see also Haas [@Haas], we relate the process $(X,V)$ to a growth-fragmentation equation as follows. Define $$\langle n(t,\cdot), \varphi\rangle = {{\mathbb{E}}}\Big[\sum_{i = 1}^\infty \varphi\big(X_i(t), V_i(t)\big)\Big]$$ as the expectation of the empirical measure of the process $(X,V)$ over smooth test functions defined on ${\mathcal S}$. We prove in Theorem \[sol transport general\] that, under appropriate regularity conditions, the measure $n(t,\cdot)$ that we identify with the temporal evolution of the density $n(t,x,v)$ of cells having size $x$ and growth rate $v$ at time $t$ is governed (in a weak sense[^5]) by $$\label{transport variabilite} \left\{ \begin{array}{rl} & \partial_t n(t,x,v) + v\, \partial_x\big(x\, n(t,x, v)\big)+B(x)n(t,x,v) \\ \\ & =\; 4B(2x)\int_{{\mathcal E}} \rho(v',v)n(t,2x,dv'), \\ \\ & n(0,x,v)= n^{(0)}(x,v), x \geq 0. \end{array} \right.$$ This result somehow legitimates our methodology: by enabling each cell to have its own growth rate and by building-up new statistical estimators in this context, we still have a translation in terms of the approach in [@DPZ]. In particular, if we assume a constant growth rate $\tau>0$, we then take $\rho(v', dv) = \delta_{\tau}(dv)$ (where $\delta$ denotes the Dirac mass) and we retrieve the standard growth-fragmentation equation . The proof of Theorem \[sol transport general\] is obtained via a so-called many-to-one formula, established in Proposition \[many-to-one\] in Section \[a many-to-one formula via a tagged branch\]. Indeed, thanks to the branching property of the system, it is possible to relate the behaviour of additive functionals like the mean empirical measure to the behaviour of a so-called tagged cell (like a tagged fragment in fragmentation process), that consists in following the behaviour of a single line of descendants along a branch where each node is picked at random, according to a uniform distribution. This approach, inspired by fragmentation techniques, is quite specific to our model and enables to obtain a relatively direct proof of Theorem \[transport variabilite\]. ### Nonparametric estimation of the growth rate {#nonparametric-estimation-of-the-growth-rate .unnumbered} In Section \[statistical analysis\] we take over the problem of estimating $(B(x), x \in {\mathcal D})$ for some compact ${\mathcal D} \subset (0,\infty)$. We assume we have data of the form , and that the mean evolution of the system is governed by . The growth rate kernel $\rho$ is unknown and treated as a nuisance parameter. A fundamental object is the transition kernel $${\mathcal P}_{B}(\boldsymbol{x},d\boldsymbol{x'}) = {{\mathbb{P}}}\big((\xi_u,\tau_u)\in d\boldsymbol{x'}\big|\,(\xi_{u^-},\tau_{u^-})=\boldsymbol{x}\big)$$ of the size and growth rate distribution $(\xi_u,\tau_u)$ at the birth of a descendant $u \in {\mathcal U}$, given the size of birth and growth rate of its parent $(\xi_{u^-},\tau_{u^-})$. We define in Section \[upper rates\] a class of division rates and growth rate kernels such that if $(B,\rho)$ belongs to this class, then the transition ${\mathcal P}_{B}$ is geometrically ergodic and has a unique invariant measure $\nu_B(d\boldsymbol{x}) = \nu_B(x,dv)dx$. From the invariant measure equation $$\nu_B{\mathcal P}_B=\nu_B$$ we obtain in Proposition \[rep B via inv\] the explicit representation $$\label{rep B anticipate} B(x)=\frac{x}{2}\frac{\nu_B(x/2)}{{{\mathbb{E}}}_{\nu_B}\Big[\frac{1}{\tau_{u^-}}{\bf 1}_{\{\xi_{u^-} \leq x,\;\xi_u \geq x/2\}}\Big]}.$$ where $\nu_B(x) = \int_{{\mathcal E}}\nu_B(x,dv)$ denotes the first marginal of the invariant distribution $\nu_B$. A strategy for constructing and estimator $B$ consists in replacing the right-hand size of by its empirical counterpart, the numerator being estimated via a kernel estimator of the first maginal of $\nu_B(d\boldsymbol{x})$. Under local H" older smoothness assumption on $B$ of order $s>0$, we prove in Theorem \[upper bound\] that for a suitable choice of bandwidth in the estimation of the invariant density, our estimator achieves the rate $n^{-s/(2s+1)}$ in squared-error loss over appropriate compact sets ${\mathcal D} \subset (0,\infty)$, up to an inessential logarithmic term when the full tree observation scheme is considered. We see in particular that we improve on the rate obtained in [@DHRR]. Our result quantifies the improvement obtained when estimating $B(x)$ from data $\big((\xi_u, \zeta_u), u \in {\mathcal U}_n\big)$, as opposed to overall measurements of the system after some time has elapsed as in [@DHRR]. We provide a quantitative argument based on the analysis of a PDE that explains the reduction of ill-posedness achieved by our method over [@DHRR] in Section \[num\].\ In order to obtain the upper bound of Theorem \[upper bound\], a major technical difficulty is that we need to establish uniform rates of convergence of the empirical counterparts to their limits in the numerator and denominator of when the data are spread along a binary tree. This can be done via covariance inequalities that exploit the fact that the transition ${\mathcal P}_{B}$ is geometrically ergodic (Proposition \[prop transition\]) using standard Markov techniques, see [@MT; @B]. The associated chain is however not reversible, and this yields an extraneous difficulty: the decay of the correlations between $\varphi(\xi_u, \tau_u)$ and $\varphi(\xi_v, \tau_v)$ for two nodes $u,v\in {\mathcal U}_n$ are expressed in terms of the sup-norm of $\varphi$, whenever $|\varphi({\boldsymbol{x}})| \leq {\mathbb V}(\boldsymbol{x})$ is dominated by a certain Lyapunov function ${\mathbb V}$ for the transition ${\mathcal P}_{B}$. However, the typical functions $\varphi$ we use are kernels that depend on $n$ and that are not uniformly bounded in sup-norm as $n \rightarrow \infty$. This partly explains the relative length of the technical Sections \[covariance inequalities\] and \[rate of convergence for the empirical measure\]. Organisation of the paper ------------------------- In Section \[microscopic model\], we construct the model $\big((\xi_u, \tau_u), u \in {\mathcal U}\big)$ of sizes and growth rates of the cells as a Markov chain along the genealogical tree. The discrete model can be embedded into a continuous time piecewise deterministic Markov process $(X,V)$ of sizes and growth rates of the cells present at any time within the system. In Theorem \[sol transport general\] we explicit the relation between the mean empirical measure of $(X,V)$ and the growth-fragmentation type equation \[transport variabilite\]. In Section \[statistical analysis\], we explicitly construct an estimator $\widehat B_n$ of $B$ by means of the representation given by in Section \[estimation of the division rate\]. Two observation schemes are considered and discussed in Section \[two observation schemes\], whether we consider data along a single branch (the sparse tree case) or along the whole genealogy (the full tree case). The specific assumptions and the class of admissible division rates $B$ and growth rate kernels $\rho$ are discussed in Section \[upper rates\], and an upper bound for $\widehat B_n$ in squared-error loss is given in our main Theorem \[upper bound\]. Section \[numerical implementation\] shows and discusses the numerical implementation of our method on simulated data. In particular, ignoring the variability in the reconstruction dramatically deterioriates the accuracy of estimation of $B$. We also explain from a deterministic point perspective the rate improvement of our method compared with [@DHRR] by means of a PDE analysis argument in Section \[num\]. The parameters are inspired from real data experiments on [*Escherichia coli*]{} cell cultures. Section \[proofs\] is devoted to the proofs. A Markov model on a tree {#microscopic model} ======================== The genealogical construction ----------------------------- Recall that ${\mathcal U} := \bigcup_{n=0}^\infty \{0,1\}^n$ (with $\{0,1\}^0:=\{\emptyset\}$) denotes the infinite binary genealogical tree. Each node $u \in {\mathcal U}$ is identified with a cell of the population and has a mark $$(\xi_u, b_u, \zeta_u, \tau_u),$$ where $\xi_u$ is the size at birth, $\tau_u$ the growth rate, $b_u$ the birthtime and $\zeta_u$ the lifetime of $u$. The evolution $\big(\xi_t^{u}, t \in [b_u,b_u+\zeta_u)\big)$ of the size of $u$ during its lifetime is governed by $$\label{piecewise deterministic} \xi_t^u=\xi_u \exp\big(\tau_u (t-b_u)\big)\;\;\text{for}\;\;t\in [b_u,b_u+\zeta_u).$$ Each cell splits into two offsprings of the same size according to a division rate $B(x)$ for $x\in (0,\infty)$. Equivalently $$\label{def div rate} {{\mathbb{P}}}\big(\zeta_{u}\in [t,t+dt]\,\big|\,\zeta_{u}\geq t,\xi_{u}=x, \tau_u=v\big) = B\big(x\exp(v t)\big)dt.$$ At division, a cell splits into two offsprings of the same size. If $u^-$ denotes the parent of $u$, we thus have $$\label{fundamental} 2\,\xi_u = \xi_{u^-}\exp\big(\tau_{u^-} \zeta_{u^-}\big)$$ Finally, the growth rate $\tau_u$ of $u$ is inherited from its parent $\tau_{u^-}$ according to a Markov kernel $$\label{markov heritage} \rho(v,dv')={{\mathbb{P}}}(\tau_u\in dv'\,|\,\tau_{u^-}=v),$$ where $v >0$ and $\rho(v,dv')$ is a probability measure on $(0,\infty)$ for each $v>0$. Eq. , , and completely determine the dynamics of the model $\big((\xi_u, \tau_u), u \in {\mathcal U}\big)$, as a Markov chain on a tree, given an additional initial condition $(\xi_\emptyset, \tau_\emptyset)$ on the root. The chain is embedded into a piecewise deterministic continuous Markov process thanks to by setting $$(\xi^u_t, \tau_t^u) =\big(\xi_u \exp\big(\tau_u( t-b_u)\big), \tau_u\big)\;\;\text{for}\;\;t \in [b_u, b_u+\zeta_u)$$ and $(0,0)$ otherwise. Define $$\big(X(t),V(t)\big) = \Big(\big(X_1(t),V_1(t)\big),\big(X_2(t),V_2(t)\big),\ldots\Big)$$ as the process of sizes and growth rates of the living particles in the system at time $t$. We have an identity between point measures $$\label{identity point measures} \sum_{i = 1}^\infty {\bf 1}_{\{X_i(t)>0\}}\delta_{(X_i(t),V_i(t))} = \sum_{u \in {\mathcal U}}{\bf 1}_{\{b_u \leq t < b_u +\zeta_u\}}\delta_{(\xi_t^u,\tau_t^u)}$$ where $\delta$ denotes the Dirac mass. In the sequel, the following basic assumption is in force. \[basic assumption\] The division rate $x \leadsto B(x)$ is continuous. We have $B(0)=0$ and $\int^\infty x^{-1}B(x)dx=\infty$. The Markov kernel $\rho(v,dv')$ is defined on a compact set ${\mathcal E} \subset (0,\infty)$. \[existence processus\] Work under Assumption \[basic assumption\]. The law of $$\big((X(t),V(t)), t \geq 0\big)\;\;\text{or}\;\;\big((\xi_u, \tau_u), u \in {\mathcal U}\big)\;\;\text{or}\;\;\big((\xi_t^u, \zeta_t^u), t \geq 0, u \in {\mathcal U}\big)$$ is well-defined on an appropriate probability space. If $\mu$ is a probability measure on the state space ${\mathcal S} = [0,\infty)\times {\mathcal E}$, we shall denote indifferently by ${{\mathbb{P}}}_\mu$ the law of any of the three processes above where the root $(\xi_{\emptyset}, \tau_{\emptyset})$ has distribution $\mu$. The construction is classical (see for instance [@bertoin] and the references therein) and is outlined in Appendix \[construction of the chain\]. The behaviour of the mean empirical measure ------------------------------------------- Denote by ${\mathcal C}_0^1({\mathcal S})$ the set of real-valued test functions with compact support in the interior of ${\mathcal S}$. \[sol transport general\] Work under Assumption \[basic assumption\]. Let $\mu$ be a probability distribution on ${\mathcal S}$. Define the distribution $n(t,dx,dv)$ by $$\langle n(t,\cdot),\varphi \rangle = {{\mathbb{E}}}_{\mu}\Big[\sum_{i = 1}^\infty\varphi\big(X_i(t), V_i(t)\big)\Big]\;\;\text{for every}\;\;\varphi \in {\mathcal C}^1_0({\mathcal S}).$$ Then $n(t,\cdot)$ solves (in a weak sense) $$\left\{ \begin{array}{rl} & \partial_t n(t,x,v) + v\, \partial_x\big(x n(t,x, v)\big)+B(x)n(t,x,v) \\ \\ & =\; 4B(2x)\int_{{\mathcal E}} \rho(v',v)n(t,2x,dv'), \\ \\ & n(0,x,v)= n^{(0)}(x,v), x \geq 0. \end{array} \right.$$ with initial condition $n^{(0)}(dx,dv) = \mu(dx,dv)$. Theorem \[sol transport general\] somehow legitimates our methodology: by enabling each cell to have its own growth rate and by building-up new statistical estimators in this context, we still have a translation in terms of the approach in [@DPZ]. In particular, we will be able to compare our estimation results with [@DHRR]. Our proof is based on fragmentation techniques, inspired by Bertoin [@bertoin] and Haas [@Haas]. Alternative approaches to the same kind of questions include the probabilistic studies of Chauvin [*et al.*]{} [@CRW], Bansaye [*et al.*]{} [@BDMV] or Harris and Roberts [@HR] and the references therein. Statistical estimation of the division rate {#statistical analysis} =========================================== Two observation schemes {#two observation schemes} ----------------------- Let ${\mathcal U}_n \subset {\mathcal U}$ denote a subset of size $n$ of connected nodes: if $u$ belongs to ${\mathcal U}_n$, so does its parent $u^-$. We look for a nonparametric estimator of the division rate $$y\leadsto B(y) = B\big(y,(\xi_u, \tau_u), u \in {\mathcal U}_n)\big)\;\;\text{for}\;\;y \in (0,\infty)$$ Statistical inference is based on the observation scheme $$\big((\xi_u, \tau_u), u \in {\mathcal U}_n \big)$$ and asymptotic study is undertaken as the population size of the sample $n \rightarrow \infty$. We are interested in two specific observation schemes. We observe every pair $(\xi_u,\tau_u)$ over the first $N_n$ generations of the tree: $${\mathcal U}_n = \{u \in {\mathcal U},\;\;|u| \leq N_n\}$$ with the notation $|u|=n$ if $u = (u_0, u_1,\ldots, u_n) \in {\mathcal U}$, and $N_n$ is chosen such that that $2^{N_n}$ has order $n$. We follow the first $n$ offsprings of a single cell, along a fixed line of descendants. This means that for some $u\in {\mathcal U}$ with $|u|=n$, we observe every size $\xi_u$ and growth rate $\tau_u$ of each node $(u_0)$, $(u_0,u_1)$, $(u_0,u_1,u_2)$ and so on up to a final node $u=(u_0,u_1,\ldots, u_{n})$. \[premiere remarque\] For every $n \geq 1$, we tacitly assume that there exists a (random) time $T_n<\infty$ almost surely, such that for $t \geq T_n$, the observation scheme ${\mathcal U}_n$ is well-defined. This is a consequence of the behaviour of $B$ near infinity that we impose later on in below. Estimation of the division rate {#estimation of the division rate} ------------------------------- ### Identification of the division rate {#identification-of-the-division-rate .unnumbered} We denote by $\boldsymbol{x} = (x,v)$ an element of the state space ${\mathcal S} = [0,\infty)\times {\mathcal E}$. Introduce the transition kernel $${\mathcal P}_{B}(\boldsymbol{x},d\boldsymbol{x'}) = {{\mathbb{P}}}\big((\xi_u,\tau_u)\in d\boldsymbol{x'}\big|\,(\xi_{u^-},\tau_{u^-})=\boldsymbol{x}\big)$$ of the size and growth rate distribution $(\xi_u,\tau_u)$ at the birth of a descendant $u \in {\mathcal U}$, given the size at birth and growth rate of its parent $(\xi_{u^-},\tau_{u^-})$. From , we infer that ${{\mathbb{P}}}(\zeta_{u^-}\in dt\,\big|\,\xi_{u^-}=x, \tau_{u^-}=v)$ is equal to $$B\big(x\exp(v t)\big)\exp\Big(-\int_0^tB\big(x\exp(v s)\big)ds\Big)dt.$$ Using formula , by a simple change of variables $${{\mathbb{P}}}\big(\xi_u\in dx'\big|\,\xi_{u^-}=x, \tau_{u^-}=v\big) =\frac{B(2x')}{v x'}{\bf 1}_{\{x' \geq x/2\}}\exp\big(-\int_{x/2}^{x'} \tfrac{B(2s)}{vs}ds\big)dx'.$$ Incorporating , we obtain an explicit formula for $${\mathcal P}_{B}(\boldsymbol{x},d\boldsymbol{x'}) ={\mathcal P}_B\big((x,v), x', dv')dx',$$ with $${\mathcal P}_B\big((x,v), x', dv') = \frac{B(2x')}{v x'}{\bf 1}_{\{x' \geq x/2\}}\exp\big(-\int_{x/2}^{x'} \tfrac{B(2s)}{v s}ds\big) \rho(v,dv'). \label{density explicit}$$ Assume further that ${\mathcal P}_B$ admits an invariant probability measure $\nu_B(d\boldsymbol{x})$, [*i.e.*]{} a solution to $$\label{def mesure invariante} \nu_B{\mathcal P}_B=\nu_B,$$ where $$\mu{\mathcal P}_B(d\boldsymbol{y})=\int_{{\mathcal S}}\mu(d\boldsymbol{x}){\mathcal P}_B(\boldsymbol{x},d\boldsymbol{y})$$ denotes the left action of positive measures $\mu(d\boldsymbol{x})$ on $\mathcal{S}$ for the transition ${\mathcal P}_B$. \[rep B via inv\] Work under Assumption \[basic assumption\]. Then ${\mathcal P}_B$ admits an invariant probability measure $\nu_B$ of the form $\nu_B(d\boldsymbol{x}) = \nu_B(x,dv)dx$ and we have $$\label{representation cle} \nu_B(y) = \frac{B(2y)}{y} {{\mathbb{E}}}_{\nu_B}\Big[\frac{1}{\tau_{u^-}}{\bf 1}_{\{\xi_{u^-} \leq 2y,\;\xi_u \geq y\}}\Big]$$ where ${{\mathbb{E}}}_{\nu_B}[\cdot]$ denotes expectation when the initial condition $(\xi_\emptyset, \tau_\emptyset)$ has distribution $\nu_B$ and where we have set $\nu_B(y) = \int_{{\mathcal E}}\nu_B(y,dv')$ in for the marginal density of the invariant probability measure $\nu_B$ with respect to $y$. We exhibit below a class of division rates $B$ and growth rate kernels $\rho$ that guarantees the existence of such an invariant probability measure. ### Construction of a nonparametric estimator {#construction-of-a-nonparametric-estimator .unnumbered} Inverting and applying an appropriate change of variables, we obtain $$\label{first rep} B(y)=\frac{y}{2}\frac{\nu_B(y/2)}{{{\mathbb{E}}}_{\nu_B}\Big[\frac{1}{\tau_{u^{-}}}{\bf 1}_{\{\xi_{u^-} \leq y,\;\xi_u \geq y/2\}}\Big]},$$ provided the denominator is positive. Representation suggests an estimation procedure, replacing the marginal density $\nu_B(y/2)$ and the expectation in the denominator by their empirical counterparts. To that end, pick a kernel function $$K: [0,\infty)\rightarrow {{\mathbb{R}}},\;\;\int_{[0,\infty)}K(y)dy=1,$$ and set $K_h(y)=h^{-1}K\big(h^{-1}y\big)$ for $y\in [0,\infty)$ and $h>0$. Our estimator is defined as $$\begin{aligned} \widehat B_n(y) & = \frac{y}{2}\,\frac{n^{-1}\sum_{u\in {\mathcal U}_n}K_h(\xi_u-y/2)}{n^{-1}\sum_{u \in {\mathcal U}_n} \frac{1}{\tau_{u^-}}{\bf 1}_{\displaystyle \{\xi_{u^-}\leq y, \xi_u \geq y/2\}} \bigvee \varpi}, \label{def estimator}\end{aligned}$$ where $\varpi >0$ is a threshold that ensures that the estimator is well defined in all cases and $x \bigvee y = \max\{x,y\}$. Thus $(\widehat B_n(y), y\in {\mathcal D})$ is specified by the choice of the kernel $K$, the bandwidth $h>0$ and the threshold $\varpi>0$. \[prop K\] The function $K$ has compact support, and for some integer $n_0 \geq 1$, we have $\int_{[0,\infty)}x^kK(x)dx={\bf 1}_{\{k=0\}}\;\;\text{for}\;\;0 \leq k\leq n_0.$ Error estimates {#upper rates} --------------- We assess the quality of $\widehat B_n$ in squared-loss error over compact intervals ${\mathcal D}$. We need to specify local smoothness properties of $B$ over ${\mathcal D}$, together with general properties that ensure that the empirical measurements in converge with an appropriate speed of convergence. This amounts to impose an appropriate behaviour of $B$ near the origin and infinity. ### Model constraints {#model-constraints .unnumbered} For $\lambda >0$ and a vector of positive constants $\mathfrak{c}=(r,m, \ell, L)$, introduce the class ${\mathcal F}^\lambda(\mathfrak{c})$ of continuous functions $B:[0,\infty)\rightarrow [0,\infty)$ such that $$\label{loc control} \int_{0}^{r/2}x^{-1}B(2x)dx \leq L,\;\;\;\int_{r/2}^{r}x^{-1}B(2x)dx \geq \ell,$$ and $$\label{poly control} B(x) \geq m\, x^\lambda\;\;\text{for}\;\;\;x\geq r.$$ Similar conditions on the behaviour of $B$ can also be found in [@DG], in a deterministic setting. Assumption \[basic assumption\] is satisfied as soon as $B \in {\mathcal F}^\lambda(\mathfrak{c})$. As mentioned in Remark \[premiere remarque\], there are arbitrarily many divisions for sufficiently large $n$, thanks to and our observation scheme ${\mathcal U}_n$ is thus well-defined under . Let $\rho_{\min}, \rho_{\max}$ be two probability measures on ${\mathcal E}$. We define ${\mathcal M}(\rho_{\min},\rho_{\max})$ as the class of Markov transitions $\rho(v,dv')$ on ${\mathcal E}$ such that $$\label{geo ergo variabilite} \rho_{\min}(A) \leq \rho(v,A) \leq \rho_{\max}(A),\;\;A \subset {\mathcal E},v \in {\mathcal E}.$$ Control ensure the geometric ergodicity of the process of variability in the growth rate. Let us be given in the sequel a vector of positive constants $\mathfrak{c}=(r,m, \ell, L)$ and $0<e_{\min} \leq e_{\max}$ such that ${\mathcal E} \subset [e_{\min}, e_{\max}]$. We introduce the Lyapunov function $$\label{first def Lyapu} {\mathbb V}(x,v)={\mathbb V}(x)=\exp\big(\tfrac{m}{e_{\min} \lambda}x^\lambda\big)\;\;\text{for}\;\;(x,v)\in {\mathcal S}.$$ The function ${\mathbb V}$ controls the rate of the geometric ergodicity of the chain with transition ${\mathcal P}_B$ and will appear in the proof of Proposition \[prop transition\] below. Define $$\delta = \delta(\mathfrak{c}):=\frac{1}{1-2^{-\lambda}} \exp\big(- (1-2^{-\lambda})\tfrac{m}{e_{\max} \lambda} r^\lambda\big).$$ \[contrainte constante\] Let $\lambda >0$. We have $\delta(\mathfrak{c})<1$. In the case of the full tree observation scheme, we will need more stringent (and technical) conditions on $\mathfrak{c}$. Let $\gamma_{B,{\mathbb V}}$ denote the spectral radius of the operator ${\mathcal P}_B - 1 \otimes \nu_B$ acting on the Banach space of functions $g:{\mathcal S}\rightarrow {{\mathbb{R}}}$ such that $$\sup\{|g(\boldsymbol{x})|/{\mathbb V}(\boldsymbol{x}), \boldsymbol{x}\in {\mathcal S}\}<\infty,$$ where ${\mathbb V}$ is defined in above. \[the full tree assumption\] We have $\delta(\mathfrak{c})<\tfrac{1}{2}$ and moreover $$\label{condition norme operateur} \sup_{B \in {\mathcal F}^\lambda(\mathfrak{c})}\gamma_{B,{\mathbb V}}<\tfrac{1}{2}.$$ It is possible to obtain bounds on $\mathfrak{c}$ so that holds, by using explicit (yet intricate) bounds on $\gamma_{B,{\mathbb V}}$ following Fort [*et al.*]{} or [@FMP], Douc [*et al.*]{} [@DMR], see also Baxendale [@B]. ### Rate of convergence {#rate-of-convergence .unnumbered} We are ready to state our main result. For $s>0$, with $s=\lfloor s\rfloor + \{s\}$, $0< \{s\} \leq 1$ and $\lfloor s\rfloor$ an integer, introduce the Hölder space ${\mathcal H}^s({\mathcal D})$ of functions $f:{\mathcal D}\rightarrow {{\mathbb{R}}}$ possessing a derivative of order $\lfloor s \rfloor$ that satisfies $$\label{def sob} |f^{\lfloor s \rfloor}(y)-f^{\lfloor s \rfloor}(x)| \leq c(f)|x-y|^{\{s\}}.$$ The minimal constant $c(f)$ such that holds defines a semi-norm $|f|_{{\mathcal H}^s(\mathcal{D})}$. We equip the space ${\mathcal H}^s(\mathcal D)$ with the norm $$\|f\|_{{\mathcal H}^s(\mathcal D)} = \|f\|_{L^\infty({\mathcal D})} + |f|_{{\mathcal H}^s({\mathcal D})}$$ and the Hölder balls $${\mathcal H}^s({\mathcal D}, M) = \{B,\;\|B\|_{{\mathcal H}^s({\mathcal D})} \leq M\},\;M>0.$$ \[upper bound\] Work under Assumption \[contrainte constante\] in the sparse tree case and Assumption \[the full tree assumption\] in the full tree case. Specify $\widehat B$ with a kernel $K$ satisfying Assumption \[prop K\] for some $n_0>0$ and $$h=c_0n^{-1/(2s+1)},\;\;\varpi_n = (\log n)^{-1}.$$ For every $M>0$ there exist $c_0=c_0(\mathfrak{c}, M)$ and $d(\mathfrak{c})\geq0$ such that for every $0<s<n_0$ and every compact interval ${\mathcal D}\subset (d(\mathfrak{c}),\infty)$ such that $\inf {\mathcal D} \geq r/2$, we have $$\sup_{\rho, B}{{\mathbb{E}}}_{\mu}\big[\|\widehat B_n-B\|_{L^2({\mathcal D})}^2\big]^{1/2} \lesssim (\log n)n^{-s/(2s+1)},$$ where the supremum is taken over $$\rho \in {\mathcal M}(\rho_{\min},\rho_{\max})\;\;\text{and}\;\;B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s({\mathcal D}, M),$$ and ${{\mathbb{E}}}_{\mu}[\cdot]$ denotes expectation with respect to any initial distribution $\mu(d\boldsymbol{x})$ for $(\xi_\emptyset,\tau_\emptyset)$ on ${\mathcal S}$ such that $\int_{\mathcal S}{\mathbb V}(\boldsymbol{x})^2\mu(d\boldsymbol{x})<\infty$. Several remarks are in order: [**1)**]{} We obtain the classical rate $n^{-s/(2s+1)}$ (up to a log term) which is optimal in a minimax sense for density estimation. It is presumably optimal in our context, using for instance classical techniques for nonparametric estimation lower bounds on functions of transition densities of Markov chains, see for instance [@GHR]. [**2)**]{} The extra logarithmic term is due to technical reasons: we need it in order to control the decay of correlations of the observations over the full tree structure. [**3)**]{} The knowledge of the smoothness $s$ that is needed for the construction of $\widehat B_n$ is not realistic in practice. An adaptive estimator could be obtained by using a data-driven bandwidth in the estimation of the invariant density $\nu_B(y/2)$ in . The Goldenschluger-Lepski bandwidth selection method [@GL2], see also [@DHRR] would presumably yield adaptation, but checking the assumptions still requires a proof in our setting. We implement data-driven bandwidth in the numerical Section \[numerical implementation\] below. Numerical implementation {#numerical implementation} ======================== Protocol and results -------------------- ### Generating simulated data {#generating-simulated-data .unnumbered} Given a division rate $B(x)$, a growth rate kernel $\rho$, an initial distribution $\mu({d{\boldsymbol{x}}})$ for the node $(\xi_{\emptyset},\tau_{\emptyset})$ (as in Theorem \[upper bound\]) and a dataset size $n=2^{N_n}$, we simulate the full tree and the sparse tree schemes recursively: 1. Given $(\xi_{u^-},\tau_{u^-}),$ we select at random its lifetime $\zeta_{u^-}$ (by a rejection sampling algorithm) with probability density $$t \leadsto B\big(\xi_{u^-}\exp(\tau_{u^-} t)\big)\exp\Big(-\int_0^tB\big(\xi_{u^-} \exp(\tau_{u^-} s)\big)ds\Big).$$ following the computations of Section \[estimation of the division rate\]. 2. We derive the two sizes at birth $\xi_u$ (with $u = (u^-, 0)$ and $(u^-,1)$) by Formula . 3. We simulate at random the growth rates $\tau_{u}$ (by the rejection sampling algorithm) according to the distribution $\rho(\tau_{u^-}, dv).$ 4. For the sparse tree case, we select only one offspring (either $(u^-,0)$ of $(u^-,1)$), whereas we keep both for the full tree case. In order to stay in line with previous simulations of [@DHRR] we pick $B(x)=x^2$. We fix $\mu(d{\boldsymbol x})$ as the uniform distribution over $[1/3,3]\times {\mathcal E}$, with ${\mathcal E} = [0.2, 3]$. As for the growth rate kernel, we implement $$\rho(v,dv')= g(v'-v)dv'$$ where $g$ is a uniform distribution over $[1-\alpha, 1+\alpha]$ for some $\alpha >0$, and dilated by a scaling factor so that $\big(\int (v'-v)^2\rho(v,dv')\big)^{1/2}=1/2$. We also condition the values of $\tau_u$ to stay in ${\mathcal E}$ (by rejection sampling). ### Implementing $\widehat B_n$ {#implementing-widehat-b_n .unnumbered} We implement $\widehat{B}_n$ using Formula . We pick a standard Gaussian kernel $K(x)=(2\pi)^{-1/2}\exp(-x^2/2)$, for which $n_0=1$ in Assumption ; henceforth we expect a rate of convergence of order $n^{-1/3}$ at best. We evaluate $\widehat{B}_n$ on a regular grid $x_1=\Delta x,\cdots x_m, =m \Delta x$ with $\Delta x = n^{-1/2}$ and $x_m=5$. Thus $x_m$ is large enough so that $\nu_B(x/2)$ becomes negligible for $x\geq x_m$ and $\Delta x$ is smaller than $n^{-1/3}$ to avoid numerical discrepancies. For tractability purposes, we wish to avoid the use of any relationship between the nodes $u \in {\mathcal U}_n$. Indeed, whereas it is quite easy to label $u^-$ and $u$ in the sparse tree case, it is a bit more difficult to track the parent of each individual in the full tree case if we do not want to double the memory. As a consequence, we simply reformulate  into $$\begin{aligned} \widehat B_n(y) & = \frac{y}{2}\,\frac{n^{-1}\sum_{u\in {\mathcal U}_n}K_h(\xi_u-y/2)}{n^{-1}\sum_{u \in {\mathcal U}_n} \frac{1}{\tau_{u}}{\bf 1}_{\displaystyle \{\xi_{u}\leq y \leq \xi_{u}e^{\tau_{u}\zeta_{u}} \}} \bigvee \varpi}. \label{def estimator2}\end{aligned}$$ We take $h_n=n^{-1/3}$ for the bandwidth according to Theorem \[upper bound\] to serve as a proof of concept. Data-driven choices could of course be made, such as the Goldenschluger and Lepski’s method [@GL2; @DHRR], and improve the already fairly good results shown in Figure \[fig:nvariab2\]. Finally, we also test whether taking into account variability in the growth rate improves significantly or not the estimate of $B,$ replacing $\tau_u$ by its mean value $n^{-1}\sum_{u\in {\cal U}_n} \tau_u$ everywhere in Formula , thus ignoring growth variability in that case. ### Numerical results {#numerical-results .unnumbered} We display our numerical results as specified above in Figures \[fig:nvariab\], \[fig:nvariab2\] and \[fig:n2puis10\]. Figure \[fig:nvariab\] displays the reconstruction of $B$ on the full tree scheme for a simulated sample of size $n=2^{17}$. At a visual level, we see that the estimation deteriorates dramatically when the variability is ignored in the region where $\nu_B$ is small, while our estimator still shows good performances. ![\[fig:nvariab\] [*Reconstruction for $n=2^{17}$ and $\varpi=n^{-1/2}.$ When the variability in the growth rate is ignored, the estimate reveals unsatisfactory. The parameter values are the reference ones.*]{}](Figure1.pdf){width="11cm" height="7cm"} ![\[fig:nvariab2\] [*Error v.s. $n$ for the full tree and the sparse tree case on a log-log scale. The error actually proves better than the upper rate of convergence announced in Theorem \[upper bound\], and $\varpi$ may be taken smaller than $\log(n)$. Estimates are comparable for both schemes. The parameter values are the reference ones.*]{}](Figure2.pdf){width="11cm" height="7cm"} In Figure \[fig:nvariab2\], we plot on a log-log scale the empirical mean error of our estimation procedure for both full tree and sparse tree schemes. The numerical results agree with the theory. The empirical error is computed as follows: we compute $$\label{error def} e_i = \frac{\|\widehat B - B\|_{\Delta x, m}}{\|B\|_{\Delta x, m, \varpi}},\;\;i=1,\ldots, M,$$ where $\|\cdot\|_{\Delta x,m, \varpi}$ denotes the discrete norm over the numerical sampling described above, conditioned on the fact that the denominator in is larger than $\varpi1/\log(n)$. We end up with a mean-empirical error $\overline{e}=M^{-1}\sum_{i=1}^Me_i$. The number of Monte-Carlo samples is chosen as $M=100$. In Figure \[fig:n2puis10\], we explore further the degradation of the estimation process on the region where $\nu_B$ is small, plotting $95\%$ confidence intervals of the empirical distribution of the estimates, based on $M=100$ Monte-Carlo samples. Finally, Table \[a table!\] displays the relative error for the reconstruction of $B$ according to . The standard deviation is computed as $(M^{-1}\sum_{i=1}^M\big(e_i-\overline{e}\big)^2)^{1/2}$. \[a table!\] \[tab:num\] $\log_2(n)$ $5$ $6$ $7$ $8$ $9$ $10$ ------------------------- -------- -------- -------- -------- -------- -------- $\overline{e}$ 0.2927 0.1904 0.1460 0.1024 0.0835 0.0614 std. dev. 0.1775 0.0893 0.0627 0.0417 0.0364 0.0241 : [*Relative error $\overline{e}$ for $B$ and its standard deviation, with respect to $n$ (on a log scale). The error is computed using with $\varpi=1/\log(n).$*]{} ![\[fig:n2puis10\] [*Reconstruction for $n=2^{10},$ error band for $95\%$, full tree case, over $M=100$ simulations, with $\varpi=1/{n}$ in order to emphasise that the larger $x,$ the smaller $\nu_B$ and the larger the error estimate*]{}.](Figure4.pdf){width="12cm" height="7cm"} We also carried out control experiments for other choices of variability kernel $\rho(v,dv')$ for the growth rate. These include $\rho(v,dv')=g(v')dv'$, so that the variability of an individual is not inherited from its parent, a Gaussian density for $g$ with the same prescription for the mean and the variance as in the uniform case, conditioned to live on $[e_{\min}, e_{\max}]$. We also tested the absence of variability, with $\rho(v,dv')=\delta_{\tau}(dv')$, with $\tau=1$. None of these control experiments show any significant difference from the case displayed in Figures \[fig:nvariab\], \[fig:nvariab2\] and \[fig:n2puis10\]. ### Analysis on [*E. coli*]{} data {#analysis-on-e.-coli-data .unnumbered} Finally, we analyse a dataset obtained through microscopic time-lapse imaging of single bacterial cells growing in rich medium, by Wang, Robert et al. [@Robert]. Thanks to a microfluidic set-up, the experimental conditions are well controlled and stable, so that the cells are in a steady state of growth (so-called balanced growth). The observation scheme corresponds to the sparse tree case: at each generation, only one offspring is followed. The growth and division of the cells is followed by microscopy, and image analysis allows to determine the time evolution of the size of each cell, from birth to division. We picked up the quantities of interest for our implementation: for each cell, its size at birth, growth rate and lifetime. We consider that cells divide equally into two daughter cells, neglecting the small differences of size at birth between daughter cells. Each cell grows exponentially fast, but growth rates exhibit variability. Our data is formed by the concatenation of several lineages, each of them composed with a line of offsprings coming from a first single cell picked at random in a culture. Some of the first and last generations were not considered in order to avoid any experimental disturbance linked either to non stationary conditions or to aging of the cells. We proceed as in the above protocol. Figure \[fig:exper\] shows the reconstructed $B$ and $\nu_B$ for a sample of $n=2335$ cells. Though much more precise and reliable, thanks both to the experimental device and the reconstruction method, our results are qualitatively in accordance with previous indirect reconstructions carried out in [@DMZ] on old datasets published in [@Kubitschek] back in 1969. The reconstruction of the division rate is prominent here since it appears to be the last component needed for a full calibration of the model. Thus, our method provides the biologists with a complete understanding of the size dependence of the biological system. Phenotypic variability between genetically identical cells has recently received growing attention with the recognition that it can be genetically controlled and subject to selection pressures [@Kaern]. Our mathematical framework allows the incorporation of this variability at the level of individual growth rates. It should allow the study of the impact of variability on the population fitness and should be of particular importance to describe the growth of populations of cells exhibiting high variability of growth rates. Several examples of high variability have been described, both in genetically engineered or natural bacterial populations [@Sturm; @Tan_Marguet_You_2009]. ![\[fig:exper\] [*Estimation of $B$ (dotted line) and $\nu_B$ (solid line) on experimental data of E. coli dividing cells, $n=2335$. In abscissae, the bacterial length is in arbitrary unit.*]{}](Figure3.pdf){width="12cm" height="7cm"} Link with the deterministic viewpoint {#num} ------------------------------------- Considering the reconstruction formula , let us give here some insight from a deterministic analysis perspective. For the sake of clarity, let us focus on the simpler case when there is no variability, so that for all $u\in {\cal U}_n$ we have $\tau_u=\tau>0$ a fixed constant. Formula  comes from , which in the case $\tau_u=\tau $ simplifies further into $$\label{formule B} B(y)=\frac{\tau y}{2}\frac{\nu_B(y/2)}{{{\mathbb{E}}}_{\nu_B}\Big[{\bf 1}_{\{\xi_{u^-} \leq y,\;\xi_u \geq y/2\}}\Big]}=\frac{\tau y}{2}\frac{\nu_B(y/2)}{\int_{y/2}^y \nu_B(z)dz}.$$ We also notice that, in this particular case, we do not need to measure the lifetime of each cell in order to implement . Define $N(y)=\tfrac{1}{2}\frac{\nu_B(y/2)}{B(y)},$ or equivalently $\nu_B(x)=2B(2x)N(2x).$ Differentiating , we obtain $$\partial_x(\tau x N)=2B(2x)N(2x)-B(x)N(x)$$ which corresponds to the stationary state linked to the equation $$\left\{\begin{array}{l} {\partial_t}n(t,x)+ \tau \partial_x \big(x\, n(t,x)\big)=2B(2x)n(t,2x)-B(x)n(t,x), \\ \\ n(0,x)= n^{(0)}(x), x \geq 0. \end{array} \right.\label{eq:transport-fragmentation conservative}$$ Eq. exactly corresponds to the behaviour of the tagged cell of Section \[a many-to-one formula via a tagged branch\] below, in a (weak) sense: $$n(t,dx) = {{\mathbb{P}}}(\chi(t)\in dx)$$ where $\chi(t)$ denotes the size at time $t$ along a branch picked at random, see Section \[a many-to-one formula via a tagged branch\]. Existence and uniqueness of an invariant measure $\nu_B$ has an analogy to the existence of a steady state solution for the PDE , and the convergence of the empirical measure to the invariant rejoins the stability of the steady state [@PPS]. The equality $\nu_B(x)=2B(2x)N(2x)$ may be interpreted as follows: $N(x)$ is the steady solution of Eq. , and represents the probability density of a cell population dividing at a rate $B$ and growing at a rate $x \tau$, but when only one offspring remains alive at each division so that the total quantity of cells remains constant. The fraction of dividing cells is represented by the term $B(x)N(x)$ in the equation, with distribution given by $\tfrac{1}{2}\nu_B(x/2),$ whereas the fraction of newborn cells is $2B(2x)N(2x)$. Eq.  can be written in terms of $BN$ as $$\label{surprise} B(y)=\frac{\tau y BN(y)}{\int_y^{2y} B(z)N(z) dz}.$$ This also highlights why we obtain a rate of convergence of order $n^{-s/(2s+1)}$ rather than the rate $n^{-s/(2s+3)}$ obtained with indirect measurements as in [@DHRR]. In that latter case, we observe a $n$-sample with distribution $N$. As shown in [@DHRR], one differentiation is necessary to estimate $B$ therefore we have a degree of ill-posedness of order 1. In the setting of the present paper, we rather observe a sample with distribution $BN$, and $B$ can be recovered directly from and we have here a degree of ill-posedness of order $0$. Proofs ====== The notation $\lesssim$ means inequality up to a constant that does not depend on $n$. We set $a_n \sim b_n$ when $a_n \lesssim b_n$ and $b_n \lesssim a_n$ simultaneously. A mapping $f:{\mathcal E}\rightarrow {{\mathbb{R}}}$ or $g:[0,\infty)\rightarrow {{\mathbb{R}}}$ is implicitly identified as a function on ${\mathcal S}$ via $f(x,v)=f(x)$ and $g(x,v)=g(v)$. A many-to-one formula via a tagged cell {#a many-to-one formula via a tagged branch} --------------------------------------- For $u \in {\mathcal U}$, we set $m^i u$ for the $i$-th parent along the genealogy of $u$. Define $$\overline{\tau_t^u} = \sum_{i = 1}^{|u|}\tau_{m^i u}\zeta_{m^i u}+\tau_t^u(t-b_u)\;\;\text{for}\;\;t\in [b_u, b_u + \zeta_u)$$ and $0$ otherwise for the cumulated growth rate along its ancestors up to time $t$. In the same spirit as tagged fragments in fragmentation processes (see the book by Bertoin [@bertoin] for instance) we pick a branch at random along the genealogical tree at random: for every $k \geq 1$, if $\vartheta_k$ denotes the node of the tagged cell at the $k$-th generation, we have $${{\mathbb{P}}}(\vartheta_k = u)=2^{-k}\;\;\text{for every}\;\;u\in {\mathcal U}\;\;\text{such that}\;\;|u|=k,$$ and $0$ otherwise. For $t\geq 0$, the relationship $$b_{\vartheta_{C_t}} \leq t < b_{\vartheta_{C_t}}+\zeta_{\vartheta_{C_t}}$$ uniquely defines a counting process $(C_t, t \geq 0)$ with $C_0=0$. The process $C_t$ enables in turn to define a tagged process of size, growth rate and cumulated growth rate via $$\big(\chi(t), {\mathcal V}(t), \overline{{\mathcal V}}(t)\big) = \Big(\xi_t^{\vartheta_{C_t}}, \tau_{t}^{\vartheta_{C_t}}, \overline{\tau_t^{\vartheta_{C_t}}}\Big)\;\;\text{for}\;\;t \in [b_{\vartheta_{C_t}},b_{\vartheta_{C_t}}+\zeta_{\vartheta_{C_t}})$$ and $0$ otherwise. We have the representation $$\label{rep chi} \chi(t) = \frac{xe^{\overline{{\mathcal V}}(t)}}{2^{C_t}}$$ and since ${\mathcal V}(t) \in [e_{\min},e_{\max}]$, we note that $$\label{controle croissance cumulee} e_{\min}t \leq \overline{{\mathcal V}}(t) \leq e_{\max}t.$$ The behaviour of $\big(\chi(t), {\mathcal V}(t), \overline{{\mathcal V}}(t)\big)$ can be related to certain functionals of the whole particle system via a so-called many-to-one formula. This is the key tool to obtain Theorem \[sol transport general\]. \[many-to-one\] Work under Assumption \[basic assumption\]. For $x \in (0,\infty)$, let ${{\mathbb{P}}}_x$ be defined as in Lemma \[counting property\]. For every $t\geq 0$, we have $${{\mathbb{E}}}_x\big[\phi\big(\chi(t), {\mathcal V}(t), \overline{{\mathcal V}}(t)\big)\big] = {{\mathbb{E}}}_x\Big[\sum_{u \in {\mathcal U}}\xi_t^u\frac{e^{-\overline{\tau_t^u}}}{x}\phi\big(\xi_t^u, \tau_t^u, \overline{\tau_t^u}\big)\Big]$$ for every $\phi:{\mathcal S}\times [0,\infty) \rightarrow [0,\infty)$. For $v \in {\mathcal U}$, set $I_v = [b_{v}, b_v+\zeta_v)$. By representation , we have $$\begin{aligned} {{\mathbb{E}}}_{x}\big[\phi\big(\chi(t), {\mathcal V}(t), \overline{{\mathcal V}}(t)\big)\big] & = {{\mathbb{E}}}_{x}\big[\phi\big(\tfrac{xe^{\overline{{\mathcal V}}(t)}}{2^{C_t}}, {\mathcal V}(t), \overline{{\mathcal V}}(t)\big)\big] \\ & = {{\mathbb{E}}}_{x}\Big[\sum_{v \in {\mathcal U}}\phi\big(\tfrac{xe^{\overline{\tau_t^v}}}{2^{|v|}}, \tau_t^v, \overline{\tau_t^v}\big){\bf 1}_{\{t\in I_v, v=\vartheta_{C_t}\}}\Big].\end{aligned}$$ Introduce the discrete filtration ${\mathcal H}_n$ generated by $(\xi_u, \zeta_u, \tau_u)$ for every $u$ such that $|u|\leq n$. Conditioning with respect to ${\mathcal H}_{|v|}$ and noting that on $\{t \in I_v\}$, we have $${{\mathbb{P}}}\big(\vartheta_{C_t} = v\,|\,{\mathcal H}_{|v|}\big)=\frac{1}{2^{|v|}} = \frac{\xi_v e^{-\overline{\tau_{b_v}^v}}}{x},$$ we derive $$\begin{aligned} {{\mathbb{E}}}_{x}\Big[\sum_{v \in {\mathcal U}}\phi\big(\tfrac{xe^{\overline{\tau_t^v}}}{2^{|v|}}, \tau_t^v, \overline{\tau_t^v}\big){\bf 1}_{\{t\in I_v, v=\vartheta_{C_t}\}}\Big] = \,&{{\mathbb{E}}}_{x}\Big[\sum_{v \in {\mathcal U}}\xi_v \frac{e^{-\overline{\tau_{b_v}^v}}}{x}\phi\big(\tfrac{xe^{\overline{\tau_t^v}}}{2^{|v|}}, \tau_t^v, \overline{\tau_t^v}\big){\bf 1}_{\{t\in I_v\}}\Big] \\ =\,& {{\mathbb{E}}}_x\Big[\sum_{u \in {\mathcal U}}\xi_t^u\frac{e^{-\overline{\tau_t^u}}}{x}\phi\big(\xi_t^u, \tau_t^u, \overline{\tau_t^u}\big)\Big].\end{aligned}$$ Proof of Theorem \[sol transport general\] ------------------------------------------ We fix $x \in (0,\infty)$ and first prove the result for an initial measure $\mu_x$ as in Proposition \[many-to-one\]. Let $\varphi \in {\mathcal C}^1_0({\mathcal S})$ be nonnegative. By we have $$\langle n(t,\cdot), \varphi \rangle = {{\mathbb{E}}}_x\big[\sum_{i=1}^\infty\varphi\big(X_i(t),Z_i(t)\big)\big] = {{\mathbb{E}}}_x\big[\sum_{u \in {\mathcal U}}\varphi(\xi_t^u,\tau_t^u)\big]$$ and applying Proposition \[many-to-one\], we derive $$\label{n via many-to-one} \langle n(t,\cdot), \varphi \rangle =x\, {{\mathbb{E}}}_x\Big[\varphi\big(\chi(t), {\mathcal V}(t)\big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big].$$ For $h>0$, introduce the difference operator $$\Delta_h f(t) = h^{-1}\big(f(t+h)-f(t)\big).$$ We plan to study the convergence of $ \Delta_h\langle n(t,\cdot),\varphi \rangle$ as $h\rightarrow 0$ using representation in restriction to the events $\{C_{t+h}-C_t=i\}$, for $i=0,1$ and $\{C_{t+h}-C_t \geq 2\}$. Denote by ${\mathcal F}_t$ the filtration generated by the tagged cell $\big(\chi(s), {\mathcal V}(s),s \leq t\big)$. The following standard estimate proved in Appendix \[A1\] will be later useful. \[counting property\] Assume that $B$ is continuous. Let $x\in (0,\infty)$ and let $\mu_x$ be a probability measure on ${\mathcal S}$ such that $\mu_x(\{x\}\times {\mathcal E})=1$. Abbreviate ${{\mathbb{P}}}_{\mu_x}$ by ${{\mathbb{P}}}_x$. For small $h>0$, we have $${{\mathbb{P}}}_x(C_{t+h}-C_t = 1\,|\,{\mathcal F}_t)=B\big(\chi(t)\big)h+h\,\varepsilon(h),$$ with the property $|\varepsilon(h)| \leq \epsilon(h) \rightarrow 0$ as $h \rightarrow 0$, for some deterministic $\epsilon(h)$, and $${{\mathbb{P}}}_x(C_{t+h}-C_t \geq 2) \lesssim h^2.$$ Since $\varphi \in {\mathcal C}^1_0({\mathcal S})$, there exists $c(\varphi)>0$ such that $\varphi(y,v)=0$ if $y \geq c(\varphi)$. By , we infer $$\label{integrabilite reste} \Big|\varphi\big(\chi(t), {\mathcal V}(t)\big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big| \leq \sup_{y,v}\varphi(y,v)\frac{\exp(e_{\max}t)}{c(\varphi)}$$ By Lemma \[counting property\] and , we derive $$\label{first estimate poisson} {{\mathbb{E}}}_x\Big[\Delta_h\Big(\varphi\big(\chi(t), {\mathcal V}(t)\big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big){\bf 1}_{\{C_{t+h}-C_t \geq 2\}}\Big] \lesssim h.$$ On the event $\{C_{t+h}-C_t=0\}$, the process ${\mathcal V}(s)$ is constant for $s \in [t,t+h)$ and so is $\frac{e^{\overline{{\mathcal V}}(s)}}{\chi(s)}$ thanks to . It follows that $$\Delta_h\Big(\varphi\big(\chi(t), {\mathcal V}(t)\big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big)= \Delta_h\varphi\big(\chi(t), {\mathcal V}(s)\big)_{\big|_{s=t}}\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}$$ on $\{C_{t+h}-C_t=0\} $ and also $$\Big|\Delta_h\varphi\big(\chi(t), {\mathcal V}(s)\big)_{\big|_{s=t}}\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big| \leq \sup_{y,v}|\partial_y\varphi(y,v)|xe_{\max} \frac{\exp(2e_{\max}t)}{c(\varphi)}$$ on $\{C_{t+h}-C_t=0\}$ likewise. Since ${{\mathbb{P}}}_x(C_{t+h}-C_t=0)\rightarrow 1$ as $h\rightarrow 0$, by dominated convergence $$\begin{aligned} & \;x\,{{\mathbb{E}}}_x\Big[\Delta_h\Big(\varphi\big(\chi(t), {\mathcal V}(t)\big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big){\bf 1}_{\{C_{t+h}-C_t =0\}}\Big] \nonumber \\ \rightarrow &\; x\,{{\mathbb{E}}}_x\big[\partial_1\varphi\big(\chi(t), {\mathcal V}(t)\big){\mathcal V}(t)e^{\overline{{\mathcal V}}(t)}\big]\;\;\text{as}\;\;h\rightarrow 0. \label{convergence poisson 2}\end{aligned}$$ By Proposition \[many-to-one\] again, this last quantity is equal to $\langle n(t,dx,dv), xv\,\partial_x \varphi\rangle$. On $\{C_{t+h}-C_t=1\}$, we successively have $$\chi(t+h) = \frac{1}{2}\chi(t)+\varepsilon_1(h),$$ $$\varphi\big(\chi(t+h), {\mathcal V}(t+h)\big) = \varphi\big(\chi(t)/2, {\mathcal V}(t+h)\big)+\varepsilon_2(h)$$ and $$\exp\big(\overline{{\mathcal V}}(t+h)\big) = \exp\big(\overline{{\mathcal V}}(t)\big)+\varepsilon_3(h)$$ with the property $|\varepsilon_i(h)|\leq \epsilon_1(h) \rightarrow 0$ as $h\rightarrow 0$, where $\epsilon_1(h)$ is deterministic, thanks to and . Moreover, $${\mathcal V}(t+h) = \tau_{\vartheta_{C_t+1}}\;\;\text{on}\;\;\{C_{t+h}-C_t=1\}.$$ It follows that $$\begin{aligned} & {{\mathbb{E}}}_x\Big[\varphi\big(\chi(t+h), {\mathcal V}(t+h)\big)\frac{e^{\overline{{\mathcal V}}(t+h)}}{\chi(t+h)}{\bf 1}_{\{C_{t+h}-C_t =1\}}\Big] \\ = &\; {{\mathbb{E}}}_x\Big[\varphi\big(\chi(t)/2, \tau_{\vartheta_{C_t+1}}\big)\frac{2e^{\overline{{\mathcal V}}(t)}}{\chi(t)}{\bf 1}_{\{C_{t+h}-C_t =1\}}\Big]+\epsilon_2(h) \\ = &\; {{\mathbb{E}}}_x\Big[\varphi\big(\chi(t)/2, \tau_{\vartheta_{C_t+1}}\big)\frac{2e^{\overline{{\mathcal V}}(t)}}{\chi(t)}{\bf 1}_{\{C_{t+h}-C_t \geq 1\}}\Big]+\epsilon_3(h)\end{aligned}$$ where $\epsilon_2(h), \epsilon_3(h)\rightarrow 0$ as $h\rightarrow 0$, and where we used the second part of Lemma \[counting property\] in order to obtain the last equality. Conditioning with respect to ${\mathcal F}_t \bigvee \tau_{\vartheta_{C_t+1}}$ and using that $\{C_{t+h}-C_t \geq 1\}$ and $\tau_{\vartheta_{C_t+1}}$ are independent, applying the first part of Lemma \[counting property\], this last term is equal to $$\begin{aligned} & {{\mathbb{E}}}_x\Big[\varphi\big(\chi(t)/2, \tau_{\vartheta_{C_t+1}}\big)\frac{2e^{\overline{{\mathcal V}}(t)}}{\chi(t)}B\big(\chi(t)\big)h\Big] + \epsilon_3(h)\\ =\;& {{\mathbb{E}}}_x\Big[\int_{{\mathcal E}}\varphi\big(\chi(t)/2, v'\big)\rho\big({\mathcal V}(t), dv'\big)\frac{2e^{\overline{{\mathcal V}}(t)}}{\chi(t)}B\big(\chi(t)\big)h\Big] + \epsilon_4(h)\end{aligned}$$ where $\epsilon_4(h)\rightarrow 0$ as $h \rightarrow 0$. Finally, using Lemma \[counting property\] again, we derive $$\begin{aligned} & {{\mathbb{E}}}_x\Big[\Delta_h\Big(\varphi\big(\chi(t), {\mathcal V}(t)\big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}\Big){\bf 1}_{\{C_{t+h}-C_t =1\}}\Big] \nonumber \\ \rightarrow &\;{{\mathbb{E}}}_x\Big[\Big(\int_{{\mathcal E}}2\varphi\big(\chi(t)/2,v'\big)\rho\big({\mathcal V}(t), dv'\big)-\varphi\big(\chi(t),{\mathcal V}(t)\big)\Big)\frac{e^{\overline{{\mathcal V}}(t)}}{\chi(t)}B\big(\chi(t)\big)\Big] \label{convergence poisson 1}\end{aligned}$$ as $h \rightarrow 0$. By Proposition \[many-to-one\], this last quantity is equal to $$\big\langle n(t,dx,dv), \big(\int_{{\mathcal E}}2\varphi(x/2,v')\rho(v,dv')-\varphi(x,v)\big)B(x)\big\rangle$$ which, in turn, is equal to $$\big\langle n(t,2dx,dv),\int_{{\mathcal E}}4\varphi(x,v')\rho(v,dv')B(2x)\big\rangle-\big\langle n(t,dx,dv),\varphi(x,v)B(x)\big\rangle$$ by a simple change of variables. Putting together the estimates , and , we conclude $$\begin{aligned} &\partial_t \langle n(t,dx,dv), \varphi\rangle - \langle n(t,dx,dv), xv\partial_x \varphi\rangle + \langle n(t,dx,dv)B(x),\varphi\rangle \\ = & \; \big\langle n(t,2dx,dv),\int_{{\mathcal E}}4\varphi(x,v')\rho(v,dv')B(2x)\big\rangle,\end{aligned}$$ which is the dual formulation of . The proof is complete. Geometric ergodicity of the discrete model ------------------------------------------ We keep up with the notations of Sections \[microscopic model\] and \[statistical analysis\]. We first prove Proposition \[rep B via inv\]. ### Proof of Proposition \[rep B via inv\] {#proof-of-proposition-rep-b-via-inv .unnumbered} The fact that $\nu_B(d{\boldsymbol x}) = \nu_B(x,dv)dx$ readily follows from the representation ${\mathcal P}_{B}(\boldsymbol{x},d\boldsymbol{x'}) ={\mathcal P}_B\big((x,v), x', dv')dx'$ together with the invariant measure equation . It follows that for every $y \in (0,\infty)$, $$\begin{aligned} \nu_B(y, dv') & = \int_{{\mathcal S}} \nu_B(x,dv)dx\,{\mathcal P}_B\big((x,v),y,dv'\big) \\ & = \frac{B(2y)}{y} \int_{{\mathcal E}}\int_{0}^{2y}\nu_B(x, dv)\exp\big(-\int_{x/2}^y\tfrac{B(2s)}{v s}ds\big)\frac{\rho(v,dv')}{v}dx.\end{aligned}$$ By Assumption \[basic assumption\], we have $\int_{x/2}^\infty \tfrac{B(2s)}{s}ds=\infty$ hence $$\exp\big(-\int_{x/2}^y \tfrac{B(2s)}{v s}ds\big)= \int_{y}^\infty\tfrac{B(2s)}{v s}\exp\big(-\int_{x/2}^s\tfrac{B(2s')}{v s'}ds'\big)ds.$$ It follows that $\nu_B(y,dv')$ is equal to $$\begin{aligned} \,&\frac{B(2y)}{y} \int_{{\mathcal E}}\int_{0}^{2y}\nu_B(x, dv)dx \int_{y}^\infty\tfrac{B(2s)}{v s}\exp\big(-\int_{x/2}^s\tfrac{B(2s')}{v s'}ds'\big)ds\frac{\rho(v,dv')}{v} \\ =\;& \frac{B(2y)}{y} \int_{{\mathcal S}}\int_{[0,\infty)}{\bf 1}_{\displaystyle \{x\leq 2y, s \geq y\}}v^{-1} \nu_B(x,dv)dx\,{\mathcal P}_B\big((x,v),s,dv'\big)ds. \end{aligned}$$ Integrating with respect to $dv'$, we obtain the result. ### Geometric ergodicity {#geometric-ergodicity .unnumbered} We extend ${\mathcal P}_B$ as an operator acting on functions $f:{\mathcal S}\rightarrow [0,\infty)$ via $${\mathcal P}_Bf(\boldsymbol{x}) = \int_{{\mathcal S}}f(\boldsymbol{y}){\mathcal P}_B(\boldsymbol{x},d\boldsymbol{y})$$ If $k\geq 1$ is an integer, define ${\mathcal P}_B^k={\mathcal P}_B^{k-1}\circ {\mathcal P}_B$. \[prop transition\] Let $\mathfrak{c}$ satisfy Assumption \[contrainte constante\]. Then, for every $B \in {\mathcal F}^\lambda(\mathfrak{c})$ and $\rho \in {\mathcal M}(\rho_{\min})$, there exists a unique invariant probability measure of the form $\nu_B(d\boldsymbol{x}) = \nu_B(x,dv)dx$ on ${\mathcal S}$. Moreover, there exist $0<\gamma<1$, a function ${\mathbb V}: {\mathcal S} \rightarrow [1,\infty)$ and a constant $R$ such that $$\label{contraction} \sup_{B \in {\mathcal F}^\lambda(\mathfrak{c}), \rho \in {\mathcal M}(\rho_{\min})}\sup_{|g| \leq V}\big|{\mathcal P}^k_Bg(\boldsymbol{x})-\int_{{\mathcal S}}g(\boldsymbol{z})\nu_B(d\boldsymbol{z})\big| \leq R{\mathbb V}(\boldsymbol{x})\gamma^k$$ for every $\boldsymbol{x}\in {\mathcal S}$, $k\geq 0$, and where the supremum is taken over all functions $g: {\mathcal S} \rightarrow {{\mathbb{R}}}$ satisfying $|g(\boldsymbol{x})| \leq {\mathbb V}(\boldsymbol{x})$ for all $\boldsymbol{x}\in {\mathcal S}$. Moreover, under Assumption \[the full tree assumption\], we can take $\gamma < \frac{1}{2}$. Finally, the function ${\mathbb V}$ is $\nu_B$-integrable for every $B \in {\mathcal F}^{\lambda}({\mathfrak{c}})$ and is well defined. We will show in the proof that the function $\mathbb{V}$ defined in satisfies the properties announced in Proposition \[prop transition\]. ### Proof of Proposition \[prop transition\] {#proof-of-proposition-prop-transition .unnumbered} We follow the classical line of establishing successively a condition of minorisation, strong aperiodicity and drift for the transition operator ${\mathcal P}_B$ (see for instance [@MT; @B; @FMP]. We keep in with the notation of Baxendale [@B]). Recall that $0<e_{\min}\leq e_{\max}$ is such that ${\mathcal E} \subset [e_{\min},e_{\max}]$. Let $B \in {\mathcal F}^\lambda(\mathfrak{c})$. Define $$\label{def varphi} \varphi_B(y) = \frac{B(2y)}{e_{\max} y}\exp\big(-\int_0^y \tfrac{B(2s)}{e_{\min} s}ds\big).$$ Set ${\mathcal C} = (0, r)\times {\mathcal E}$, where $r$ is specified by $\mathfrak{c}$. For any measurable $\mathcal{X} \times A \subset {\mathcal S}$ and $(x,v) \in {\mathcal C}$, we have $$\begin{aligned} {\mathcal P}_B\big((x,v),\mathcal{X}\times A\big) & = \int_{A}\rho(v,dv')\int_{{\mathcal X} \cap [x/2,\infty]} \frac{B(2y)}{v y}\exp\big(-\int_{x/2}^y \tfrac{B(2s)}{v s}ds\big)dy \\ & \geq \rho_{\min}(A)\int_{{\mathcal X} \cap [r/2,\infty]} \varphi_B(y)dy.\end{aligned}$$ Define $$\Gamma_B(dy,dv)=c_B^{-1}{\bf 1}_{[r/2,\infty)}(y)\varphi_B(y)dy\,\rho_{\min}(dv),$$ where $$\begin{aligned} c_B = \int_{r/2}^\infty \varphi_B(y)dy \geq \frac{e_{\min}}{e_{\max}}\exp\big(-\tfrac{L}{e_{\min}}) =: \widetilde \beta >0\end{aligned}$$ by since $B \in {\mathcal F}^\lambda(\mathfrak{c})$. We have thus exhibited a small set ${\mathcal C}$, a probability measure $\Gamma_B$ and a constant $\widetilde \beta >0$ so that the minorisation condition $$\label{minorisation} {\mathcal P}_B\big((x,v),{\mathcal X}\times A\big) \geq \tilde \beta\, \Gamma_B({\mathcal X} \times A)$$ holds for every $(x,v)\in {\mathcal C}$ and ${\mathcal X}\times A \subset {\mathcal S}$, uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})$. We have $$\begin{aligned} \tilde \beta\, \Gamma_B({\mathcal C})& = \widetilde \beta\, c_B^{-1}\int_{{\mathcal E}}\rho_{\min}(dv)\int_{r/2}^r \frac{B(2y)}{e_{\max} y}\exp\big(-\int_{x/2}^y \tfrac{B(2s)}{e_{\min} s}ds\big)dy \nonumber\\ & \geq \widetilde \beta\, c_B^{-1}\int_{r/2}^r \varphi_B(y)dy \nonumber\\ & \geq \widetilde \beta \big(1-\exp\big(-\int_{r/2}^{r}\tfrac{B(2y)}{e_{\min} y}dy\big)\big) \nonumber\\ & \geq \widetilde \beta (1-\exp(-\tfrac{\ell}{e_{\min}})\big) =: \beta >0 \label{aperiodicity}\end{aligned}$$ where we applied for the last inequality. Let $B \in {\mathcal F}^\lambda(\mathfrak{c})$. Let ${\mathbb V}:{\mathcal S}\rightarrow [1,\infty)$ be continuously differentiable and such that for every $v\in {\mathcal E}$, $$\label{control V alinfini} \lim_{y \rightarrow \infty} {\mathbb V}(y,v)\exp\big(-2^\lambda \tfrac{m}{v \lambda}y^\lambda\big)=0.$$ For $x \geq r$, by and integrating by part with the boundary condition , we have, for every $v \in {\mathcal E}$, $$\begin{aligned} {\mathcal P}_B{\mathbb V}(x,v) & = \int_{{\mathcal E}}\rho(v,dv') \int_{x/2}^\infty {\mathbb V}(y,v')\frac{B(2y)}{v y} \exp\big(-\int_{x/2}^y \tfrac{B(2s)}{v s}ds\big)dy \nonumber \\ & \leq \int_{\mathcal {E}}\rho(v,dv') \int_{x/2}^\infty \partial_y {\mathbb V}(y,v') \exp\big(-\tfrac{m2^\lambda}{v}\int_{x/2}^y s^{\lambda-1}ds\big)dy \nonumber \\ & \leq \exp\big(\tfrac{m}{v\lambda}x^\lambda\big) \int_{\mathcal {E}}\rho(v,dv')\int_{\tfrac{m2^\lambda}{v \lambda}(x/2)^\lambda}^\infty {\mathbb V}\Big(\big(y\tfrac{v \lambda}{m 2^\lambda}\big)^{1/\lambda},v'\Big)e^{-y}dy. \nonumber\end{aligned}$$ Pick ${\mathbb V}(x,v)={\mathbb V}(x)=\exp\big(\tfrac{m}{e_{\min} \lambda}x^\lambda\big)$ defined in and note that is satisfied. With this choice, we further infer $$\begin{aligned} {\mathcal P}_B{\mathbb V}(x,v) & \leq {\mathbb V}(x,v) \int_{\mathcal {E}}\rho(v,dv') \int_{\tfrac{m2^\lambda}{v \lambda}(x/2)^\lambda}^\infty \exp\big(-(1-2^{-\lambda})y\big)dy \\ & \leq {\mathbb V}(x,v) \frac{1}{1-2^{-\lambda}} \exp\big(- (1-2^{-\lambda})\tfrac{m}{v \lambda} r^\lambda\big)\end{aligned}$$ since $x \geq r$. Recall that $$\delta(\mathfrak{c})= \frac{1}{1-2^{-\lambda}} \exp\big(- (1-2^{-\lambda})\tfrac{m}{e_{\max} \lambda} r^\lambda\big).$$ We obtain, for $x\geq r$ and $v \in {\mathcal E}$ $$\label{first drift} {\mathcal P}_B{\mathbb V}(x,v) \leq \delta(\mathfrak{c}) {\mathbb V}(x,v)$$ and we have $\delta(\mathfrak{c}) < 1$ by Assumption \[contrainte constante\]. We next need to control ${\mathcal P}_B{\mathbb V}$ outside $x \in [r,\infty)$, that is on the small set ${\mathcal C}$. For every $(x,v) \in {\mathcal C}$, we have $$\begin{aligned} {\mathcal P}_B{\mathbb V}(x,v) \leq &\int_{{\mathcal E}}\rho(v,dv')\Big(\int_{x/2}^{r/2} {\mathbb V}(y,v') \frac{B(2y)}{vy}dy \nonumber \\ &+ \int_{r/2}^\infty {\mathbb V}(y,v') \frac{B(2y)}{v y}\exp\big(-\int_{r/2}^y \tfrac{B(2s)}{v s}ds\big)dy\Big)\nonumber \\ \leq &\,e_{\min}^{-1}\sup_{y \in [0,r]}{\mathbb V}(y) L+\mu(\mathfrak{c}) {\mathbb V}(r/2) =: Q <\infty \label{second drift} \end{aligned}$$ where we used and the fact that $B \in {\mathcal F}^\lambda(\mathfrak{c})$. Combining together and , we conclude $$\label{drift condition} {\mathcal P}_B{\mathbb V}(\boldsymbol{x}) \leq \delta(\mathfrak{c}) {\mathbb V}(\boldsymbol{x}) {\bf 1}_{\{\boldsymbol{x} \notin {\mathcal C}\}}+ Q{\bf 1}_{\{\boldsymbol{x}\in {\mathcal C}\}}.$$ The minorisation condition together with the strong aperiodicity condition and the drift condition imply inequality by Theorem 1.1 in Baxendale [@B], with $R$ and $\gamma $ that explicitly depend on $\delta(\mathfrak{c})$, $\beta$, $\tilde \beta$, $V$ and $Q$. By construction, this bound is uniform in $B \in {\mathcal F}^\lambda(\mathfrak{c})$ and $\rho \in {\mathcal M}(\rho_{\min})$. More specifically, we have $$\gamma < \min\{\max\{\delta({c}), \rho_{{\mathbb V},B}\}, 1\}$$ therefore under Assumption \[contrainte constante\] we have $\gamma <1$ and under Assumption \[the full tree assumption\], we obtain the improvement $\gamma < \frac{1}{2}$. Further estimates on the invariant probability ---------------------------------------------- \[borne sup nu\] For any $\mathfrak{c}$ such that Assumption \[contrainte constante\] is satisfied and any compact interval ${\mathcal D} \subset (0,\infty)$, we have $$\sup_{B \in {\mathcal F}^\lambda(\mathfrak{c})\, \cap\, {\mathcal H}^s({\mathcal D}, M)}\sup_{x \in {2^{-1}\mathcal D}}\nu_B(x)<\infty,$$ with $\nu_B(x) = \int_{{\mathcal E}}\nu_B(x,dv)$. Since $B \in {\mathcal F}^\lambda(\mathfrak{c})$, $\nu_B$ is well-defined and satisfies $$\begin{aligned} \nu_B(x,dv) & = \frac{B(2x)}{x} \int_{{\mathcal E}}\int_{0}^{2x}\nu_B(y, dv')dy \exp\big(-\int_{y/2}^x\tfrac{B(2s)}{v' s}ds\big)\frac{\rho(v',dv)}{v'}.\end{aligned}$$ Hence $\nu_B(x,dv) \leq B(2x)(e_{\min}x)^{-1}\rho_{\max}(dv)$ and also $\nu_B(x) \leq B(2x)(e_{\min}x)^{-1}$. Since $B \in {\mathcal H}^s({\mathcal D},M)$ implies $\sup_{x\in 2^{-1}{\mathcal D}}B(2x) = \|B\|_{L^{\infty}({\mathcal D})} \leq M$, the conclusion follows. \[minoration mes inv\] For any $\mathfrak{c}$ such that Assumption \[contrainte constante\] is satisfied, there exists a constant $d(\mathfrak{c}) \geq 0$ such that for any compact interval ${\mathcal D}\subset (d(\mathfrak{c}),\infty)$, we have $$\inf_{B \in {\mathcal F}^\lambda(\mathfrak{c})}\inf_{x \in {\mathcal D}}\varphi_B(x)^{-1}\nu_B(x) >0,$$ where $\varphi_B(x)$ is defined in . Let $g:[0,\infty)\rightarrow [0,\infty)$ satisfy $g(x) \leq {\mathbb V}(x)=\exp(\tfrac{m}{e_{\min} \lambda}x^{\lambda})$ for every $x \in [0,\infty)$. By Proposition \[prop transition\], we have $$\label{integ unif mes inv} \sup_{B\in \mathcal{F}_\lambda^\nu(\mathfrak{c})}\int_{[0,\infty)}g(x)\nu_B(x)dx<\infty,$$ as a consequence of with $n=1$ together with the property that $\sup_{B\in \mathcal{F}_\lambda^\nu(\mathfrak{c})}{\mathcal P}_B{\mathbb V}(\boldsymbol{x})<\infty$ for every $\boldsymbol{x} \in \mathcal{S}$, as follows from in the proof of Proposition \[prop transition\]. Next, for every $x\in (0,\infty)$, we have $$\begin{aligned} \int_{2x}^\infty \nu_B(y)dy \; &\leq \exp(-\tfrac{m}{e_{\min} \lambda}(2x)^{\lambda})\int_{[0,\infty)}{\mathbb V}(y)\nu_B(y)dy\end{aligned}$$ and this bound is uniform in $B\in \mathcal{F}^\lambda(\mathfrak{c})$ by . Therefore, for every $x\in (0,\infty)$, we have $$\label{ineg qui tue} \sup_{B\in \mathcal{F}_\lambda^\nu(\mathfrak{c})} \int_{2x}^\infty \nu_B(y)dy \leq c(\mathfrak{c}) \exp(-\tfrac{m}{e_{\min} \lambda}(2x)^{\lambda})$$ for some $c(\mathfrak{c})>0$. Let $$\label{def d} d(\mathfrak{c})>\big(\frac{e_{\min} \lambda 2^{-\lambda}}{m} \log c(\mathfrak{c})\big)^{1/\lambda} {\bf 1}_{\{c(\mathfrak{c}) \geq 1\}}.$$ By definition of $\nu_B$, for every $x\in (0,\infty)$, we now have $$\begin{aligned} \nu_B(x,dv) & = \frac{B(2x)}{x}\int_{\mathcal{E}}\int_{0}^{2x} \nu_B(y,dv')\exp\big(-\int_{y/2}^x\tfrac{B(2s)}{v' s}ds\big)dy\frac{\rho(v', dv)}{v'} \\ & \geq \frac{B(2x)}{e_{\max} x}\exp\big(-\int_{0}^x\tfrac{B(2s)}{e_{\min} s}ds\big)\int_0^{2x} \nu_B(y)dy\, \rho_{\min}(dv)\\ & \geq \frac{B(2x)}{e_{\max} x}\exp\big(-\int_{0}^x\tfrac{B(2s)}{e_{\min} s}ds\big)\Big(1-c(\mathfrak{c}) \exp(-\tfrac{m}{e_{\min} \lambda}(2x)^{\lambda})\Big)\rho_{\min}(dv) \end{aligned}$$ where we used for the last inequality. By , for $x \geq d(\mathfrak{c})$ we have $$\big(1-c(\mathfrak{c}) \exp(-\tfrac{m}{e_{\min} \lambda}(2x)^{\lambda})\big)>0$$ and the conclusion readily follows by integration. Covariance inequalities {#covariance inequalities} ----------------------- If $u, w \in {\mathcal U}$, we define $a(u,w)$ as the node of the most recent common ancestor between $u$ and $w$. Introduce the distance $$\mathbb{D}(u,w) = |u|+|w|-2|a(u,w)|.$$ \[moment bound\] Work under Assumption \[contrainte constante\]. Let $\mu$ be a probability distribution on ${\mathcal S}$ such that $\int_{\mathcal {S}}{\mathbb V}(\boldsymbol{x})^2\mu(d\boldsymbol{x})<\infty$. Let $G: {\mathcal S}\rightarrow {{\mathbb{R}}}$ and $H:[0,\infty)\rightarrow {{\mathbb{R}}}$ be two bounded functions. Define $$Z(\xi_{u^-},\tau_{u^-}, \xi_u) = G(\xi_{u^-},\tau_{u^-})H(\xi_u) - {{\mathbb{E}}}_{\nu_B}[G(\xi_{u^-},\tau_{u^-})H(\xi_u)].$$ For any $u,w\in {\mathcal U}$ with $|u|,|w| \geq 1$, we have $$\begin{aligned} \big| {{\mathbb{E}}}_{\mu}\big[Z(\xi_{u^-},\tau_{u^-}, \xi_u) Z(\xi_{w^-},\tau_{w^-}, \xi_w) \big] \big| \lesssim \gamma^{\,\mathbb{D}(u,w)} \label{covariance GH}\end{aligned}$$ uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})$, with $\gamma$ and $\nu_B$ defined in of Proposition \[prop transition\]. In view of , with no loss of generality, we may (and will) assume that for every $(x,v)\in {\mathcal S}$ $$\label{controle lyapou} |G(x,v)|\leq {\mathbb V}(x)\;\;\text{and}\;\;|H(x)| \leq {\mathbb V}(x)$$ Applying repeatedly the Markov property along the branch that joins the nodes $a^-(u,w):=a(u^-,w^-)$ and $w$, we have $$\begin{aligned} &{{\mathbb{E}}}_{\mu}\big[G(\xi_{u^-}, \tau_{u^-})H(\xi_u)|\,\xi_{a^-(u,w)}, \tau_{a^-(u,w)}\big]\\ = &\;{\mathcal P}_B^{|u^-|-|a^-(u,w)|}(G\,{\mathcal P}_B H)\big(\xi_{a^-(u,w)},\tau_{a^-(u,w)}\big) \\ =&\; {\mathcal P}_B^{|u|-|a(u,w)|}(G\,{\mathcal P}_B H)\big(\xi_{a^-(u,w)},\tau_{a^-(u,w)}\big)\end{aligned}$$ with an analogous formula for $G(\xi_{w^-}, \tau_{w^-})H(\xi_w)$. Conditioning with respect to $\xi_{a^-(u,w)}, \tau_{a^-(u,w)}$, it follows that $$\begin{aligned} &{{\mathbb{E}}}_{\mu}\big[Z(\xi_{u^-},\tau_{u^-}, \xi_u) Z(\xi_{w^-},\tau_{w^-}, \xi_w) \big] \\ =\;& {{\mathbb{E}}}_{\mu}\Big[ \Big({\mathcal P}_B^{|u|-|a(u,w)|}(G\,{\mathcal P}_B H)(\xi_{a^-(u,w)}, \tau_{a^-(u,w)})-{{\mathbb{E}}}_{\nu_B}[G\,{\mathcal P}_B H(\xi_{\emptyset}, \tau_{\emptyset})]\Big)\\ &\;\hspace{1cm}\Big({\mathcal P}_B^{|w|-|a(u,w)|}(G\,{\mathcal P}_B H)(\xi_{a^-(u,w)}, \tau_{a^-(u,w)})-{{\mathbb{E}}}_{\nu_B}[G\,{\mathcal P}_B H(\xi_{\emptyset}, \tau_{\emptyset})]\Big)\Big].\end{aligned}$$ Applying Proposition \[prop transition\] thanks to Assumption \[contrainte constante\] and , we further infer $$\begin{aligned} & {{\mathbb{E}}}_{\mu}\big[Z(\xi_{u^-},\tau_{u^-}, \xi_u) Z(\xi_{w^-},\tau_{w^-}, \xi_w)\big] \\ \leq \;& R^2\sup_{(x,v)}G(x,v)^2H(x)^2{{\mathbb{E}}}_{\mu}\big[{\mathbb V}(\xi_{a^-(u,w)})^2\big]\gamma^{\,\mathbb{D}(u,w)}\\ \lesssim \; & \int_{\mathcal{S}}{\mathcal P}_B^{|a^-(u,w)|}\big({\mathbb V}^2\big)(\boldsymbol{x})\mu(d\boldsymbol{x})\, \gamma^{\,\mathbb{D}(u,w)}.\end{aligned}$$ We leave to the reader the straightfoward task to check that the choice of ${\mathbb V}$ in implies that ${\mathbb V}^2$ satisfies . It follows that Proposition \[prop transition\] applies, replacing ${\mathbb V}$ by ${\mathbb V}^2$ in . In particular, $$\label{majoration V deux} \sup_{B \in {\mathcal F}^\lambda(\mathfrak{c})}{\mathcal P}_B^{|a^-(u,w)|}\big({\mathbb V}^2\big)(\boldsymbol{x}) \lesssim 1+{\mathbb V}(\boldsymbol{x})^2.$$ Since ${\mathbb V}^2$ is $\mu$-integrable by assumption, inequality follows. \[moment bound bis\] Work under Assumption \[contrainte constante\]. Let $\mu$ be a probability on ${\mathcal S}$ such that $\int_{\mathcal{S}}{\mathbb V}(\boldsymbol{x})^2\mu(d\boldsymbol{x})<\infty$. Let $x_0$ be in the interior of $\frac{1}{2}{\mathcal D}$. Let $H:{{\mathbb{R}}}\rightarrow {{\mathbb{R}}}$ be bounded with compact support. Set $$\widetilde H\big(\tfrac{\xi_u-x_0}{h}\big) = H\big(\tfrac{\xi_u-x_0}{h}\big)-{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_\emptyset-x_0}{h}\big)\big].$$ For any $u,w\in {\mathcal U}$ with $|u|, |w| \geq 1$, we have $$\begin{aligned} \big| {{\mathbb{E}}}_{\mu}\big[\widetilde H\big(\tfrac{\xi_u-x_0}{h}\big)\widetilde H\big(\tfrac{\xi_w-x_0}{h}\big)\big] \big| \lesssim \; & \gamma^{\,\mathbb{D}(u,w)} \bigwedge h \gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))} \label{first covariance}\end{aligned}$$ uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s({\mathcal D}, M)$ for sufficiently small $h>0$. The first part of the estimate in the right-hand side of is obtained by letting $G=1$ in . We turn to the second part. Repeating the same argument as for and conditioning with respect to $\xi_{a(u,w)}$, we obtain $$\begin{aligned} & {{\mathbb{E}}}_{\mu}\big[\widetilde H\big(\tfrac{\xi_u-x_0}{h}\big)\widetilde H\big(\tfrac{\xi_w-x_0}{h}\big)\big]\nonumber\\ = \;& {{\mathbb{E}}}_{\mu}\Big[\big({\mathcal P}_B^{|u|-|a(u,w)|}H\big(\tfrac{\xi_{a(u,w)}-x_0}{h}\big) -{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_\emptyset-x_0}{h}\big)\big]\big) \nonumber\\ & \hspace{3cm} \big({\mathcal P}_B^{|w|-|a(u,w)|}H\big(\tfrac{\xi_{a(u,w)}-x_0}{h}\big) -{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_\emptyset-x_0}{h}\big)\big]\big) \label{decomp markov} \Big].\end{aligned}$$ Assume with no loss of generality that $|u| \leq |w|$ (otherwise, the same subsequent arguments apply exchanging the roles of $u$ and $w$). On the one hand, applying of Proposition \[prop transition\], we have $$\label{premiere maj h} \big|{\mathcal P}_B^{|w|-|a(u,w)|}H\big(\tfrac{\xi_{a(u,w)}-x_0}{h}\big)-{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_\emptyset-x_0}{h}\big)\big] \big| \leq R{\mathbb V}\big(\xi_{a(u,w)}\big)\gamma^{|w|-|a(u,w)|}.$$ On the other hand, identifying $H$ as a function defined on $\mathcal{S}$, for every $(x,v)\in {\mathcal S}$, we have $$\begin{aligned} \big|{\mathcal P}_BH\big(\tfrac{x-x_0}{h}\big)\big| & = \Big|\int_{x/2}^\infty H\big(h^{-1}(y-x_0)\big) \frac{B(2y)}{vy}\exp\big(-\int_{x/2}^y \tfrac{B(2s)}{vs}ds\big) dy\Big| \nonumber \\ & \leq \int_{[0,\infty)} \big|H\big(h^{-1}(y-x_0)\big)\big| \frac{B(2y)}{e_{\min} y} dy \nonumber\\ & \leq e_{\min}^{-1}\sup_{y\in\{x_0+h\,\mathrm{supp}(H)\}}\frac{B(2y)}{y} h \int_{[0,\infty)}|H(x)|dx \lesssim h. \label{obtention h}\end{aligned}$$ Indeed, since $x_0$ is in the interior of $\frac{1}{2}{\mathcal D}$ we have $\{x_0+h\,\mathrm{supp}(H)\} \subset \tfrac{1}{2}{\mathcal D}$ for small enough $h$ hence $\sup_{y\in\{x_0+h\,\mathrm{supp}(H)\}}B(2y) \leq M$. Now, since ${\mathcal P}_B$ is a positive operator and ${\mathcal P}_B {\bf 1} = {\bf 1}$, we derive $$\label{avant esperance h} \big|{\mathcal P}_B^{|u|-|a(u,w)|}H\big(\tfrac{\xi_{a(u,w)}-x_0}{h}\big)\big| \lesssim h$$ as soon as $|u|-|a(u,w)| \geq 1$, uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s({\mathcal D}, M)$. If $|u|=|a(u,w)|$, since $\int_{\mathcal{E}}\nu_B(dx,dv) = \nu_B(x)dx$, we obtain in the same way $$\begin{aligned} \big|{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_{a(u,w)}-x_0}{h}\big)\big] \big|& \leq \int_{[0,\infty)}\big|H\big(\tfrac{x-x_0}{h}\big)\big|\nu_B(x)dx \lesssim h \label{esperance h}\end{aligned}$$ using Lemma \[borne sup nu\]. We have $\big|{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_u-x_0}{h}\big)\big]\big| \lesssim h$ likewise. Putting together and we derive $$\big|{\mathcal P}_B^{|u|-|a(u,w)|}H\big(\tfrac{\xi_{a(u,w)}-x_0}{h}\big) -{{\mathbb{E}}}_{\nu_B}\big[H\big(\tfrac{\xi_u-x_0}{h}\big)\big]\big| \lesssim h$$ and this estimate is uniform in $B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s({\mathcal D}, M)$. In view of and , we obtain $${{\mathbb{E}}}_{\mu}\big[\widetilde H\big(\tfrac{\xi_u-x_0}{h}\big)\widetilde H\big(\tfrac{\xi_w-x_0}{h}\big)\big] \lesssim \; h\gamma^{|w|-|a(u,w)|}{{\mathbb{E}}}_{\mu}\big[{\mathbb V}\big(\xi_{a(u,w)}\big)\big].$$ We conclude in the same way as in Proposition \[moment bound\]. Rate of convergence for the empirical measure {#rate of convergence for the empirical measure} --------------------------------------------- For every $y\in (0,\infty)$ and $u\in {\mathcal U}$ with $|u|\geq 1$, define $$\label{def D} D(y) = {{\mathbb{E}}}_{\nu_B}\big[\tfrac{1}{\tau_{u^-}}{\bf 1}_{\{\xi_{u^-}\leq 2y,\,\xi_u \geq y\}}\big]$$ and $$\label{def D_n} D_n(y)_{\varpi} = n^{-1}\sum_{u \in {\mathcal U}_n} \tfrac{1}{\tau_{u^-}}{\bf 1}_{\displaystyle \{\xi_{u^-}\leq 2y, \xi_u \geq y\}} \bigvee \varpi.$$ \[moment lemma\] Work under Assumption \[contrainte constante\] in the sparse tree case and Assumption \[the full tree assumption\] in the full tree case. Let $\mu$ be a probability on ${\mathcal S}$ such that $\int_{\mathcal{S}}{\mathbb V}(\boldsymbol{x})^2\mu(d\boldsymbol{x})<\infty$. If $1 \geq \varpi = \varpi_n\rightarrow 0$ as $n \rightarrow \infty$, we have $$\label{convergence D_n} \sup_{y\in {\mathcal D}}{{\mathbb{E}}}_{\mu}\big[\big(D_n(y)_{\varpi_n}-D(y)\big)^2\big] \lesssim n^{-1}$$ uniformy in $B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s(2^{-1}{\mathcal D}, M)$ and $\rho \in {\mathcal M}(\rho_{\min}, \rho_{\max})$. We first need the following estimate \[minoration D\] Work under Assumption \[contrainte constante\]. Let $d(\mathfrak{c})$ be defined as in Lemma \[minoration mes inv\]. For every compact interval ${\mathcal D} \subset (d(\mathfrak{c}),\infty)$ such that $\inf {\mathcal D} \leq r/2$, we have $$\inf_{B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s(2^{-1}{\mathcal D}, M)}\inf_{y\in {\mathcal D}}D(y) > 0.$$ By and the definition of $\varphi_B$ in , we readily have $$D(y) = \frac{1}{e_{\max}}\varphi_B(y)^{-1}\nu_B(y)\exp\big(-\int_0^y \tfrac{B(2s)}{e_{\min} s}ds\big).$$ Since $B\in {\mathcal F}^\lambda(\mathfrak{c})\cap {\mathcal H}^s(2^{-1}{\mathcal D}, M)$, by applying and successively, we obtain $$\begin{aligned} \int_{0}^{\sup {\mathcal D}} \frac{B(2s)}{e_{\min} s}ds & \leq e_{\min}^{-1}L+\int_{r/2}^{\sup \mathcal{D}} \frac{B(2s)}{e_{\min} s}ds \\ & \leq e_{\min}^{-1}(L + M \log \tfrac{\sup {\mathcal D}}{r/2}) < \infty\end{aligned}$$ where we used that $\inf {\mathcal D} \leq r/2$. It follows that $$\inf_{y\in {\mathcal D}}\exp\big(-\int_0^y \tfrac{B(2s)}{e_{\min} s}ds\big) \geq \exp\big(-e_{\min}^{-1}(L + M \log \tfrac{\sup {\mathcal D}}{r/2}) \big)>0$$ and Lemma \[minoration D\] follows by applying Lemma \[minoration mes inv\]. Since $D_n(y)$ is bounded, we have $$\label{take off varpi} \big(D_n(y)_{\varpi_n}-D(y)\big)^2\lesssim \big(D_n(y)-D(y)\big)^2+{\bf 1}_{\{D_n(y) < \varpi_n\}}.$$ Next, take $n$ sufficiently large, so that $$0 < \varpi_n \leq q=\tfrac{1}{2}\inf_{B \in {\mathcal F}^\lambda(\mathfrak{c}) \cap {\mathcal H}^s(2^{-1}{\mathcal D}, M)}\inf_{y \in {\mathcal D}}D(y)$$ a choice which is possible thanks to Lemma \[minoration D\]. Since $$\{D_n(y) < \varpi_n\}\subset \{D_n(y)-D(y)< -q\},$$ integrating , we have that ${{\mathbb{E}}}_{\mu}\big[\big(D_n(y)_{\varpi_n}-D(y)\big)^2\big]$ is less than a constant times $$\begin{aligned} {{\mathbb{E}}}_{\mu}\big[\big(D_n(y)-D(y)\big)^2\big] + {{\mathbb{P}}}_{\mu}\big(|D_n(y)-D(y)| \geq q\big),\end{aligned}$$ which in turn is less than a constant times ${{\mathbb{E}}}_{\mu}\big[\big(D_n(y)-D(y)\big)^2\big]$. Set $G(x, v) = \frac{1}{v}{\bf 1}_{\{x \leq 2y\}}$ and $H(x)={\bf 1}_{\{x \geq y\}}$ and note that $G$ and $H$ are bounded on ${\mathcal S}$ (and also uniformly in $y \in {\mathcal D}$). It follows that $$\begin{aligned} D_n(y) -D(y) & = n^{-1}\sum_{u \in {\mathcal U}_n}\Big( G(\xi_{u^-}, \tau_{u^-})H(\xi_u)-{{\mathbb{E}}}_{\nu_B}\big[G(\xi_{u^-}, \tau_{u^-})H(\xi_u)\big]\Big). \end{aligned}$$ We then apply of Proposition \[moment bound\] to infer, with the same notation that $$\begin{aligned} & {{\mathbb{E}}}_{\mu}\big[\big(D_n(y)-D(y)\big)^2\big] \\ =&\;n^{-2}\sum_{u,w \in {\mathcal U}_n} {{\mathbb{E}}}_{\mu}\big[Z(\xi_{u^-},\tau_{u^-},\xi_u)Z(\xi_{w^-},\tau_{w^-},\xi_w)\big] \\ \lesssim &\; n^{-2}\sum_{u,w \in {\mathcal U}_n}\gamma^{\mathbb{D}(u,w)}\end{aligned}$$ uniformly in $y\in {\mathcal D}$ and $B \in {\mathcal F}^\lambda(\mathfrak{c})$. We further separate the sparse and full tree cases. We have $ \sum_{u,w \in {\mathcal U}_n}\gamma^{\mathbb{D}(u,w)} = \sum_{1 \leq |u|,|w| \leq n}\gamma^{||u|-|w||}$ by Proposition \[prop transition\], and this last quantity is of order $n$. We have $n \sim 2^{N_n}$, where $N_n$ is the number of generations used to expand ${\mathcal U}_n$. We evaluate $$\label{the big sum} \mathcal {N}_k=\sum_{|u|=k}\sum_{w \in\, {\mathcal U}_n} \gamma^{\mathbb{D}(u,w)}\;\;\text{for}\;\;k=0,\ldots, N.$$ For $k=0$, we have $${\mathcal N}_0=1+2\gamma + 4\gamma^2 +\ldots + 2^{N_n}\gamma^{N_n} = \tfrac{1-(2\gamma)^{N_n+1}}{1-2\gamma}=: \phi_\gamma(N_n).$$ Under Assumption \[the full tree assumption\], by Proposition \[prop transition\], we have $\gamma < \frac{1}{2}$ therefore $\phi_\gamma(N_n)$ is bounded as $n \rightarrow \infty$. For $k=1$, if we start with the node $u=(\emptyset, 0)$, then the contribution of its descendants in is given by $\phi_\gamma(N_n-1)$, to which we must add $\gamma$ for its ancestor corresponding to the node $u=\emptyset$ and also $\gamma\phi_\gamma(N_n)$ for the contribution of the second lineage of the node $u=\emptyset$. Finally, we must repeat the argument for the node $u=(\emptyset, 1)$. We obtain $$\mathcal{N}_1 = 2\big(\phi_\gamma(N_n-1)+\gamma+\gamma^2\phi_\gamma(N_n-1)\big).$$ More generally, proceeding in the same manner, we derive $$\begin{aligned} \mathcal{N}_{k} & =2^k\Big(\phi_\gamma(N_n-k)+\big(\gamma+\gamma^2\phi_\gamma(N_n-k)\big)+ \ldots \nonumber \\ &+ \gamma^i+\gamma^{i+1}\phi_\gamma\big(N_n-k+(i-1)\big)+ \ldots+\big(\gamma^k+\gamma^{k+1}\phi_\gamma(N_n-1)\big)\Big) \label{def Vk} \end{aligned}$$ for $k=1,\ldots, N_n$, and this last quantity is of order $2^k$. It follows that $$\sum_{u,w \in \,{\mathcal U}_n}\gamma^{\mathbb{D}(u,w)} = \sum_{k=0}^{N_n} {\mathcal N}_k \lesssim \sum_{k=1}^{N_n} 2^k \lesssim 2^{N_n} \lesssim n$$ and the conclusion follows likewise. Putting together the sparse and full tree case, we obtain the proposition. \[moment lemma bis\] Work under Assumption \[contrainte constante\] in the sparse tree case and Assumption \[the full tree assumption\] in the full tree case. Let $\mu$ be a probability on ${\mathcal S}$ such that $\int_{\mathcal{S}}{\mathbb V}(\boldsymbol{x})^2\mu(d\boldsymbol{x})<\infty$. We have $$\label{moment mes inv} \sup_{y \in {\mathcal D}}{{\mathbb{E}}}_{\mu}\big[\big(K_{h_n}\star \widehat \nu_n(y)-K_{h_n} \star \nu_B(y)\big)^2\big] \lesssim |\log h_n| {(nh_n)}^{-1}$$ uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})$. We have, with the notation of Proposition \[moment bound bis\] $$\begin{aligned} & {{\mathbb{E}}}_{\mu}\big[\big(K_{h_n}\star \widehat \nu_n(y)-K_{h_n} \star \nu_B(y)\big)^2\big] \nonumber \\ =\, &(nh_n)^{-2}{{\mathbb{E}}}_{\mu}\Big[\Big(\sum_{u\in\, {\mathcal U}_n}K\big(\tfrac{\xi_u-y}{h_n}\big)-{{\mathbb{E}}}_{\nu_B}\big[K\big(\tfrac{\xi_u-y}{h_n}\big)\big]\Big)^2\Big] \nonumber\\ =\, &(nh_n)^{-2}\sum_{u,w\,\in {\mathcal U}_n}\mathbb{E}_{\mu}\big[\widetilde K\big(\tfrac{\xi_u-y}{h_n}\big) \widetilde K\big(\tfrac{\xi_w-y}{h_n}\big)\big] \nonumber\\ \lesssim\, &(nh_n)^{-2} \sum_{u,w \in\, {\mathcal U}_n} \gamma^{\,\mathbb{D}(u,w)} \bigwedge h_n \label{estimation dist arbre} \gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))} \end{aligned}$$ by applying of Proposition \[moment bound bis\]. It remains to estimate . We have $a(u,w)=u$ if $|u|\leq |w|$ and $a(u,w)=w$ otherwise. It follows that $${{\mathbb{E}}}_{\mu}\big[\big(K_{h_n}\star \widehat \nu_n(y)-K_{h_n} \star \nu_B(y)\big)^2\big] \lesssim n^{-2}h_n^{-1}\sum_{u,w \in \,{\mathcal U}_n} \gamma^{\,\mathbb{D}(u,w)},$$ and since $\sum_{u,w \in \,{\mathcal U}_n} \gamma^{\,\mathbb{D}(u,w)}=\sum_{1 \leq |u|,|w| \leq n}\gamma^{||u|-|w||}$ is of order $n$ as soon as $\gamma <1$, we obtain the result. The computations are a bit more involved. Let us evaluate $$\sum_{|u|=k}\sum_{w \in {\mathcal U}_n}\gamma^{\mathbb{D}(u,w)}\bigwedge h_n\gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))}.$$ We may repeat the argument displayed in in order to evaluate the contribution of the term involving $\gamma^{\,\mathbb{D}(u,a(u,w))}$. However, in the estimate ${\mathcal N}_k$, each term $\gamma^i+\gamma^{i+1}\phi_\gamma\big(N_n-k+(i-1)\big)$ in formula may be replaced by $h_n\big(\gamma^i+\gamma\phi_\gamma\big(N_n-k+(i-1)\big)\big)$ up to constants. This corresponds to the correction given by $h_n\gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))}$. As a consequence, we obtain $$\begin{aligned} & \sum_{|u|=k}\sum_{w \in {\mathcal U}_n}\gamma^{\mathbb{D}(u,w)}\bigwedge h_n\gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))} \\ \lesssim &\; 2^k \sum_{i=1}^k h_n\big(\gamma^i+\gamma\phi_\gamma(N_n-k+i-1)\big) \bigwedge \gamma^i\big(1+\gamma \phi_\gamma(N_n-k+i-1)\big) \\ \lesssim &\; 2^k \sum_{i=1}^k h_n \wedge \gamma^{i}.\end{aligned}$$ Define $k_n^\star= \lfloor \tfrac{|\log h_n|}{|\log \gamma|}\rfloor$. We readily derive $$2^k\sum_{i=1}^k h_n \wedge \gamma^{i} = 2^k\big(\sum_{i=1}^{k_n^\star}h_n+\sum_{i=k_n^\star+1}^k \gamma^i\big) \lesssim 2^k h_n|\log h_n|,$$ ignoring the second term if $k_n^\star+1 \geq k$. Going back to , it follows that $$\begin{aligned} &(nh_n)^{-2} \sum_{u,w \in\, {\mathcal U}_n} \gamma^{\,\mathbb{D}(u,w)} \bigwedge h_n \gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))} \\ =\, &(nh_n)^{-2} \sum_{k=0}^{N_n}\sum_{|u|=k}\sum_{v \in {\mathcal U}_n} \gamma^{\,\mathbb{D}(u,w)} \bigwedge h_n \gamma^{\,\mathbb{D}(u,a(u,w)) \vee \,\mathbb{D}(w,a(u,w))} \\ \lesssim\, & (nh_n)^{-2} \sum_{k=0}^{N_n} 2^k h_n|\log h_n| \lesssim |\log h_n| (nh_n)^{-1}\end{aligned}$$ and the conclusion follows in the full case. Putting together the sparse and full tree cases, we obtain the proposition. Proof of Theorem \[upper bound\] -------------------------------- From $$\widehat B_n(2y)= y\frac{n^{-1}\sum_{u \in {\mathcal U}_n}K_{h_n}(\xi_u-y)}{n^{-1}\sum_{u \in {\mathcal U}_n} \frac{1}{\tau_{u^-}}{\bf 1}_{\displaystyle \{\xi_{u^-}\leq 2y, \xi_u \geq y\}} \bigvee \varpi_n}$$ and $$B(2y) =y \frac{\nu_B(y)}{ {{\mathbb{E}}}_{\nu_B}\big[\tfrac{1}{\tau_{u^-}}{\bf 1}_{\{\xi_{u^-}\leq 2y,\,\xi_u \geq y\}}\big]},$$ we plan to use the following decomposition $$\widehat B_n(2y)-B(2y) = y(I+II+III),$$ with $$\begin{aligned} I & =\frac{K_{h_n}\star \nu_B(y)-\nu_B(y)}{D(y)}, \\ II & =\frac{K_{h_n}\star \widehat \nu_n(y)-K_{h_n}\star \nu_B(y)}{D_n(y)_{\varpi_n}},\\ III & =\frac{K_{h_n}\star \nu_B(y)}{D_n(y)_{\varpi_n} D(y)}\big(D(y)-D_n(y)_{\varpi_n}\big), \end{aligned}$$ where $D(y)$ and $D_n(y)_{\varpi}$ are defined in and respectively. It follows that $$\begin{aligned} \|\widehat B_n-B\|_{L^2({\mathcal D})}^2 & = 2\int_{\tfrac{1}{2}{\mathcal D}}\big(\widehat B_n(2y)-B(2y)\big)^2dy \lesssim IV+V+VI,\end{aligned}$$ with $$\begin{aligned} IV & = \int_{\tfrac{1}{2}{\mathcal D}}\big(K_{h_n}\star \nu_B(y)-\nu_B(y)\big)^2\tfrac{y^2}{D(y)^2}dy\\ V & = \int_{\tfrac{1}{2}{\mathcal D}}\big(K_{h_n}\star \widehat \nu_n(y)-K_{h_n}\star \nu_B(y)\big)^2D_n(y)_{\varpi}^{-2}y^2dy\\ VI & = \int_{\tfrac{1}{2}{\mathcal D}} \big(D_n(y)_\varpi-D(y)\big)^2\big(K_{h_n}\star \nu_B(y)\big)^2\big(D_n(y)_\varpi D(y)\big)^{-2}y^2dy.\end{aligned}$$ We get rid of the term $\tfrac{y^2}{D(y)^2}$ by Lemma \[minoration D\] and the fact that ${\mathcal D}$ is bounded. By Assumption \[prop K\] and classical kernel approximation, we have for every $0 < s \leq n_0$ $$\label{biais mesure invariante} IV \lesssim \|K_{h_n}\star \nu_B-\nu_B\big\|_{L^2(2^{-1}{\mathcal D})}^2 \lesssim |\nu_B|_{{\mathcal H}^s(2^{-1}{\mathcal D})}^2h_n^{2s}.$$ \[reg transfer\] Let ${\mathcal D}\subset (0,\infty)$ be a compact interval. Let $B\in {\mathcal F}^\lambda(\mathfrak{c})$ for some ${\mathfrak{c}}$ satisfying Assumption \[contrainte constante\]. We have $$\|\nu_B\|_{{\mathcal H}^s(2^{-1}\mathcal{D})} \leq \psi\big(e_{\min},e_{\max}, {\mathcal D},\|B\|_{{\mathcal H}^s({\mathcal D})}\big)$$ for some continuous function $\psi$. Define $$\Lambda_B(x,y) = \int_{\mathcal E}\frac{\nu_B(y,dv')}{v'}\exp\big(-\int_{y/2}^x\tfrac{B(2s)}{v's}ds\big).$$ If $B\in{\mathcal H}^s({\mathcal D})$, then $x\leadsto \Lambda_B(x,y) \in {\mathcal H}^s(2^{-1}{\mathcal D})$ for every $y \in [0,\infty)$, and we have $$\|\Lambda_B(\cdot,y)\|_{{\mathcal H}^s(2^{-1}{\mathcal D})} \leq \psi_1(y,\|B\|_{{\mathcal H}^s({\mathcal D})}, e_{\min}, e_{\max})$$ for some continous function $\psi_1$. The result is then a consequence of the representation $\nu_B(x) = \frac{B(2x)}{x}\int_0^{2x} \Lambda_B(x,y)dy$. Going back to we infer from Lemma \[reg transfer\] that $|\nu_B|_{{\mathcal H}^s(2^{-1}\mathcal{D})}$ is bounded above by a constant that depends on $e_{\min}$, $e_{\max}$, ${\mathcal D}$ and $\|B\|_{{\mathcal H}^s({\mathcal D})}$ only. It follows that $$\label{controle IV} IV \lesssim h_n^{2s}$$ uniformly in $B \in {\mathcal H}^s({\mathcal D},M)$. We have $${{\mathbb{E}}}_{\mu}[V] \leq \varpi_n^{-2}|{\mathcal D}| \sup_{y \in 2^{-1}{\mathcal D}} y^2{{\mathbb{E}}}_{\mu}\big[\big(K_{h_n}\star \widehat \nu_n(y)-K_{h_n} \star \nu_B(y)\big)^2\big].$$ By of Proposition \[moment lemma bis\] we derive $$\label{controle V} {{\mathbb{E}}}_{\mu}[V] \lesssim \varpi_n^{-2}|\log h_n|(nh_n)^{-1}$$ uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})$. First, by Lemma \[minoration D\], the estimate $$\inf_{B \in {\mathcal F}^\lambda(\mathfrak{c})}\inf_{y\in 2^{-1}{\mathcal D}}D_n(y)_\varpi D(y) \gtrsim \varpi_n$$ holds. Next, $$\begin{aligned} \sup_{y \in 2^{-1}{\mathcal D}} |K_{h_n}\star \nu_B(y)| & = \sup_{y \in 2^{-1}{\mathcal D}}\big|\int_{[0,\infty)}K_{h_n}(z-y)\nu_B(z)dz\big| \nonumber \\ & \leq \sup_{y \in 2^{-1}{\mathcal D}_{h_n}}\nu_B(y)\|K\|_{L^1([0,\infty))} \label{borne phi l1}\end{aligned}$$ where $2^{-1}{\mathcal D}_{h_n} = \{y+z,\;y\in 2^{-1}{\mathcal D},\;z\in \text{supp}(K_{h_n})\} \subset \widetilde {\mathcal D}$, for some compact interval $\widetilde {\mathcal D}$ since $K$ has compact support by Assumption \[prop K\]. By Lemma \[borne sup nu\], we infer that holds uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})$. We derive $${{\mathbb{E}}}_{\mu}\big[VI\big] \lesssim \varpi_n^{-2}\sup_{y \in 2^{-1}{\mathcal D}}{{\mathbb{E}}}_{\mu}\big[\big(D_n(y)_{\varpi_n}-D(y)\big)^2\big].$$ Applying of Proposition \[moment lemma\], we conclude $$\label{controle VI} {{\mathbb{E}}}_{\mu}\big[VI\big] \lesssim \varpi_n^{-2}n^{-1}$$ uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})$. We put together the three estimates , and . We obtain $$\begin{aligned} {{\mathbb{E}}}_{\mu}\big[\|\widehat B_n-B\|_{L^2({\mathcal D})}^2\big] \lesssim h_n^{2s}+\varpi_n^{-2}|\log h_n|(nh_n)^{-1}+ \varpi_n^{-2}n^{-1}\end{aligned}$$ uniformly in $B \in {\mathcal F}^\lambda(\mathfrak{c})\,\cap \,{\mathcal H}^s({\mathcal D},M)$. The choice $h_n \sim n^{-1/(2s+1)}$ and the fact that $\varpi_n^{-2}$ grows logarithmically in $n$ yields the rate $n^{-s/(2s+1)}$ up to log terms. The proof is complete. Appendix ======== Construction of the discrete model {#construction of the chain} ---------------------------------- Fix an initial condition $\boldsymbol{x} = (x,v)\in {\mathcal S}$. On a rich enough probability space, we consider a Markov chain on the binary tree $(\tau_u, u \in {\mathcal U})$ with transition $\rho(v,dv')$ and initial condition $v$: if $u = (u_1,\ldots, u_k) \in {\mathcal U}$, we write $ui = (u_1,\ldots, u_k, i)$, $i=0,1$ for the two offsprings of $u$; we set $\tau_\emptyset=v$ and $$\tau_{u0}\sim \rho(\tau_{u}, dv')\;\;\text{and}\;\;\tau_{u1}\sim \rho(\tau_{u}, dv')$$ so that conditional on $\tau_{u}$, the two random variables $\tau_{u0}$ and $\tau_{u1}$ are independent. We also pick a sequence of independent standard exponential random variables $\big({\bf e}_u, u \in {\mathcal U}\big)$, independent of $(\tau_u, u \in {\mathcal U})$. The model $\big((\xi_u, \tau_u), u \in {\mathcal U}\big)$ is then constructed recursively. We set $$\xi_\emptyset=x,\;\;b_\emptyset=0,\;\;\tau_\emptyset = v\;\;\text{and}\;\;\zeta_\emptyset=F_{x,v}^{-1}({\bf e}_\emptyset)$$ where $F_{x,v}(t) = \int_0^t B\big(x\exp(vs)\big)ds$. For $u \in {\mathcal U}$ and $i=0,1$, we put $$\xi_{u0}=\xi_{u1}=e^{\tau_u \zeta_u}\frac{\xi_u}{2},\;\;b_{u0}=b_{u1}=b_u+\zeta_u,\;\;\zeta_{ui}=F^{-1}_{\xi_{ui}, \tau_{ui}}({\bf e}_{ui}).$$ To each node $u \in {\mathcal U}$, we then associate the mark $(\xi_i, b_u, \zeta_u, \tau_u)$ of the size, date of birth, lifetime and growth rate respectively of the individual labeled by $u$. One easily checks that Assumption \[basic assumption\] guarantees that the model is well defined. Proof of Lemma \[counting property\] {#A1} ------------------------------------ Note first that $$\{C_{t+h}-C_t \geq 1\} = \{t < b_{\vartheta_{C_t}}+\zeta_{\vartheta_{C_t}}\leq t+h\}.$$ Since moreover $\xi_{\vartheta_{C_t}} = x\exp\big(\overline{{\mathcal V}}(b_{\vartheta_{C_t}})\big)2^{-C_t}$, it follows by that $$\begin{aligned} &{{\mathbb{P}}}(C_{t+h}-C_t \geq 1\,|\,{\mathcal F}_t) \\ =& \int_{t-b_{\vartheta_{C_t}}}^{t+h-b_{\vartheta_{C_t}}}B\Big(\tfrac{xe^{\overline{{\mathcal V}}(b_{\vartheta_{C_t}})+s{\mathcal V}(s)}}{2^{-C_t}}\Big)\exp\Big(-\int_0^s B\Big(\tfrac{xe^{\overline{{\mathcal V}}(b_{\vartheta_{C_t}})+s'{\mathcal V}(s')}}{2^{-C_t}} \Big)ds'\Big)ds.\end{aligned}$$ Introduce the quantity $B\big(xe^{\overline{{\mathcal V}}(b_{\vartheta_{C_t}})+{\mathcal V}(t)(t-b_{\vartheta_{C_t}})}2^{-C_t}\big)$ within the integral. Noting that $\overline{{\mathcal V}}(b_{\vartheta_{C_t}})+{\mathcal V}(t)(t-b_{\vartheta_{C_t}}) = \overline{{\mathcal V}}(t)$ we obtain the first part of the lemma thanks to the representation and the uniform continuity of $B$ over compact sets. For the second part, introduce the $({\mathcal F}_t)$-stopping time $$\Upsilon_t = \inf\{s>t, C_s - C_t \geq 1\}$$ and note that $\{C_{t+h}-C_t \geq 1\}=\{\Upsilon_{t} \leq t+h\} \in {\mathcal F}_{\Upsilon_t}$. Writing $$\{C_{t+h}-C_t \geq 2\} = \{\Upsilon_t < t+h,\;\Upsilon_{\Upsilon_{t}} \leq t+h\}$$ and conditioning with respect to ${\mathcal F}_{\Upsilon_t}$, we first have $$\begin{aligned} & {{\mathbb{P}}}(C_{t+h}-C_t \geq 2) \\ = &\; {{\mathbb{E}}}\Big[\int_t^{t+h-\Upsilon_t}B\Big(\tfrac{xe^{\overline{{\mathcal V}}(b_{\vartheta_{C_t}})+s{\mathcal V}(s)}}{2^{C_t}}\Big)e^{-\int_0^s B\big(\frac{xe^{\overline{{\mathcal V}}(b_{\vartheta_{C_t}})+s'{\mathcal V}(s')}}{2^{C_t}} \big)ds'}ds\,{\bf 1}_{\{\Upsilon_t < t+h\}} \Big] \\ \leq &\; h \sup_{y \leq x \exp (2e_{\max} t)} B(y)\,{{\mathbb{P}}}(\Upsilon_t < t+h).\end{aligned}$$ In the same way, ${{\mathbb{P}}}(\Upsilon_t < t+h) \lesssim h$ and the conclusion follows. Acknowledgements {#acknowledgements .unnumbered} ---------------- [The research of M. Doumic is partly supported by the Agence Nationale de la Recherche, Grant No. ANR-09-BLAN-0218 TOPPAZ. The research of M. Hoffmann is partly supported by the Agence Nationale de la Recherche, (Blanc SIMI 1 2011 project CALIBRATION). We are grateful to V. Bansaye, A. Dalalyan, E. Moulines, A. Olivier, B. Perthame, V. Rivoirard and V.C. Tran for helpful discussion and comments.]{} [^1]: INRIA Rocquencourt, projet BANG, Domaine de Voluceau, BP 105, 781153 Rocquencourt, France and Laboratoire J.L. Lions, CNRS-UMR 7598, 4 place Jussieu 75005 Paris, France [email]{}: [email protected] [^2]: Université Paris-Dauphine, CNRS-UMR 7534, Place du maréchal De Lattre de Tassigny 75775 Paris Cedex 16, France. [email]{}: [email protected] [^3]: Université de Rennes 1, CNRS-UMR 6625, Campus de Beaulieu, 35042 Rennes Cedex, France. [email]{}: [email protected] [^4]: INRA and AgroParisTech, Micalis CNRS-UMR 1319, 78350 Jouy-en-Josas, France. [email]{}: [email protected] [^5]: For every $t\geq 0$, we actually have a Radon measure $n(t,dx ,dv)$ on ${\mathcal S}=[0,\infty)\times {\mathcal E}$: If $\varphi(x,v)$ is a function defined on ${\mathcal S}$, we define $\langle n(t,\cdot), \varphi\rangle = \int_{{\mathcal S}}\varphi(x,v)n(t,dx, dv)$ whenever the integral is meaningful. Thus has the following sense: for every sufficiently smooth test function $\varphi$ with compact support in ${\mathcal E}$, we have $$\begin{aligned} &\int_{{\mathcal S}}\partial_t n(t,dx,dv)\varphi(x,v)-\,vx n(t,dx, dv)\partial_x\varphi(x,v)+B(x)n(t,dx,dv)\varphi(x,v) \\ = & \;4\int_{{\mathcal S}}\big(B(2x)\int_{{\mathcal E}} \rho(v',dv)n(t,2dx,dv')\varphi(x,v)\big). \end{aligned}$$

ArXiv

--- abstract: 'Spatially coupled turbo-like codes (SC-TCs) have been shown to have excellent decoding thresholds due to the threshold saturation effect. Furthermore, even for moderate block lengths, simulation results demonstrate very good bit error rate performance (BER) in the waterfall region. In this paper, we discuss the effect of spatial coupling on the performance of TCs in the finite block-length regime. We investigate the effect of coupling on the error-floor performance of SC-TCs by establishing conditions under which spatial coupling either preserves or improves the minimum distance of TCs. This allows us to investigate the error-floor performance of SC-TCs by performing a weight enumerator function (WEF) analysis of the corresponding uncoupled ensembles. While uncoupled TC ensembles with close-to-capacity performance exhibit a high error floor, our results show that SC-TCs can simultaneously approach capacity and achieve very low error floor.' author: - | Saeedeh Moloudi, Michael Lentmaier, ,\ and Alexandre Graell i Amat,  [^1][^2] [^3] [^4] title: | Spatially coupled turbo-like codes:\ a new trade-off between waterfall and error floor --- Bound on minimum distance, expurgated bounds,spatially coupled turbo-like codes, union bound, weight enumerator analysis Introduction ============ Turbo-like codes (TCs) [@BerrouTC] and low-density parity check (LDPC) codes [@GallagerLDPCBook] are adopted in many communication standards because they can practically approach the Shannon limit. Recently, it has been proved that LDPC convolutional codes [@JimenezLDPCCC; @MichaelCLDPC] —also known as spatially coupled LDPC (SC-LDPC) codes—exhibit the remarkable threshold saturation phenomenon [@Kudekar_ThresholdSaturation; @Yedla2012; @Yedla2012Vector; @Yedla2014], i.e., for an SC-LDPC ensemble, the belief propagation (BP) decoder can achieve the threshold of the optimal maximum-a-posteriori (MAP) decoder of the underlying uncoupled ensemble. It then follows that regular SC-LDPC codes achieve capacity as their variable node degrees tend to infinity. Spatially coupled TCs (SC-TCs) were introduced in [@Moloudi_SCTurbo; @Moloudi_SCTC_Journal; @MoloudiHCC], and it was proved that threshold saturation also occurs for this class of codes. A density evolution analysis shows that, by having stronger component codes, SC-TCs can achieve excellent decoding thresholds with variable nodes of degree one and two only. In this paper, motivated by the excellent asymptotic behavior of SC-TCs, we investigate the performance of these codes in the finite block-length regime. We consider the same TC ensembles as those in [@Moloudi_SCTurbo; @Moloudi_SCTC_Journal; @MoloudiHCC], namely parallel concatenated codes (PCCs) [@BerrouTC], serially concatenated codes (SCCs) [@Benedetto98Serial; @AGiAa], braided convolutional codes (BCCs) [@ZhangBCC; @WindowBCC], and hybrid concatenated codes (HCCs) [@3DTurbo; @KollerHCC]. As the first step of our investigation, using the decoding thresholds of the binary erasure channel (BEC) obtained in [@Moloudi_SCTC_Journal; @MoloudiHCC] and the method described in [@LDPCPuncture; @UmarISIT2018], we predict the decoding thresholds over the additive white Gaussian noise (AWGN) channel. Using these thresholds together with the provided simulation results, we discuss the effect of spatial coupling on the performance of TCs in the waterfall region over the AWGN channel. Then, we investigate the effect of coupling on the error-floor performance of TCs. We prove that under certain conditions the minimum distance of a coupled SC-TC ensemble cannot get smaller than that of the corresponding TC ensemble. This means that the error-floor performance of the TCs is not degraded by spatial coupling. These conditions can be seen as a guideline for unwrapping the TC ensembles. This connection between the minimum distance of TC and SC-TC ensembles allows us to avoid the complexity of computing the weight enumerator functions (WEFs) of the coupled ensembles. Instead, we simply perform a WEF analysis for the uncoupled TC ensemble to investigate and discuss the distance properties of SC-TCs. Thus, we compute the WEFs of TC ensembles [@UnveilingTC; @Benedetto98Serial; @AbuSurrahGlobeCom07; @AbuSurra2011] to obtain bounds on their bit error rate (BER) performance and a bound on the minimum distance. Finally, in the last step of our investigation, we use the obtained results to discuss the overall performance of SC-TCs for the finite block-length regime. The remainder of the paper is organized as follows. In Section \[Turbo-like Codes\], we briefly describe several TC and SC-TC ensembles by use of the compact graph representation introduced in [@Moloudi_SCTC_Journal]. We discuss the decoding thresholds of these ensembles in Section \[Waterfall\]. In the same section, we provide some simulation results to discuss the waterfall region performance of SC-TCs. In Section \[Proof\], we prove that the minimum distance of SC-TC ensembles is either better or equal than that of the corresponding uncoupled ensemble. In Section \[WE\], we compute the average WEF of TC ensembles to obtain bounds on their BER performance and minimum distance. Finally, in Section \[con\], we discuss the trade-off between waterfall and error floor performance of SC-TCs, and we conclude the paper in the same section. Spatially Coupled Turbo-like Codes {#Turbo-like Codes} ================================== In this section, we briefly describe four major classes of TCs— namely, PCCs, SCCs, BCCs, and HCCs— and their coupled counterparts. In particular, we discuss PCCs and SC-PCCs with coupling memory $m=1$, and refer the interested reader to [@Moloudi_SCTC_Journal] for details on the other SC-TC ensembles and higher coupling memories, $m>1$. Fig. \[CGPCC\](a) shows the block diagram of a rate $R=1/3$ PCC encoder built of two recursive systematic convolutional encoders, referred to as upper and lower encoder. As shown in the figure, the information sequence [$\boldsymbol{u}$]{} is encoded by the upper encoder $\mathcal{C}^{\text{U}}$ to produce the upper parity sequence ${\ensuremath{\boldsymbol{v}}}^{\text{U}}$. Likewise, a reordered copy of ${\ensuremath{\boldsymbol{u}}}$ is encoded by the lower encoder $\mathcal{C}^{\text{L}}$ to produce the lower parity sequence ${\ensuremath{\boldsymbol{v}}}^{\text{L}}$. The corresponding permutation is denoted by $\Pi^{\text{Un}}$. Finally, the output of the PCC encoder is the sequence ${\ensuremath{\boldsymbol{v}}}=({\ensuremath{\boldsymbol{u}}},{\ensuremath{\boldsymbol{v}}}^{\text{U}},{\ensuremath{\boldsymbol{v}}}^{\text{L}})$. \[thick, scale=0.58\] at (0,3.3) [$\boldsymbol{u}$]{}; (0,3.1) – (0,2) – (5.5,2); at (5.7,1.8) [$\boldsymbol{u}$]{}; (0,2) – (0,0) – (2.5,0); (2.5,-0.5) rectangle (4.5,0.5); at (3.5,0) [$\mathcal{C}^{\text{U}}$]{}; (4.5,0) – (5.5,0); at (5.7,-0.2) [$\boldsymbol{v}^{\text{U}}$]{}; (0,0) – (0,-2); (-1,-3) rectangle (1,-2); at (0,-2.5) [$\Pi^{\text{Un}}$]{}; (1,-2.5) – (2.5,-2.5); (2.5,-3) rectangle (4.5,-2); at (3.5,-2.5) [$\mathcal{C}^{\text{L}}$]{}; (4.5,-2.5) – (5.5,-2.5); at (5.7,-2.7) [$\boldsymbol{v}^{\text{L}}$]{}; at (3,-5.6) [(a)]{}; at (7.6,4.9) [$\boldsymbol{v}^{\text{U}}$]{}; (8,4.55) circle \[radius=0.1\]; (8,4.55) – (8,4.05); (7.5,4.05) rectangle (8.5,3.05); at (8,3.55) [$N$]{}; at (9.1,3.55) [$T^{\text{U}}$]{}; (8,3.05) – (8,0); (8,0) circle \[radius=0.1\]; at (7.6,0) [$\boldsymbol{u}$]{}; (8,0) – (8,-3.05); (7.5,-4.05) rectangle (8.5,-3.05); at (8,-3.55) [$N$]{}; at (9.1,-3.55) [$T^{\text{L}}$]{}; (7.8,-1.95) – (8.2,-1.75); at (8.9,-1.8) [$\Pi^{\text{Un}}$]{}; (8,-4.05) – (8,-4.55); (8,-4.55) circle \[radius=0.1\]; at (7.6,-4.9) [$\boldsymbol{v}^{\text{L}}$]{}; at (8,-5.6) [(b)]{}; at (12.1,4.9) [$\boldsymbol{v}^{\text{U}}_t$]{}; (12.5,4.55) circle \[radius=0.1\]; (12.5,4.55) – (12.5,4.05); (12,4.05) rectangle (13,3.05); at (12.5,3.55) [$N$]{}; at (13.6,3.55) [$T^{\text{U}}_t$]{}; at (11.5,2.6) [$\Pi^{\text{U}}_t$]{}; (12.5,3.05) – (12.5,2.25); (12.3,2.5) – (12.7,2.7); (12,2.25) rectangle (13,2); (11.25,1.6) – (12.8,1.6)–(12.8,2); at (11,1.8) [$\boldsymbol{u}_{t-1,1}$]{}; (12.2,2) – (12.2,0.95); (12,0.75) rectangle (13,0.95); (12.8,0.95) – (12.8,1.35)–(13.6,1.35); at (13.8,1.5) [$\boldsymbol{u}_{t,1}$]{}; (12.5,0.75) – (12.5,0); (12.5,0) circle \[radius=0.1\]; at (12,0) [$\boldsymbol{u}_t$]{}; at (13.3,-0.5) [$\Pi_t$]{}; (12.5,0) – (12.5,-0.75); (12.3,-0.5) – (12.7,-0.3); (12,-0.75) rectangle (13,-0.95); (13.6,-1.4) – (12.8,-1.4)–(12.8,-0.95); at (13.8,-1.8) [$\tilde{\boldsymbol{u}}_{t,1}$]{}; (12.2,-0.95) – (12.2,-2); (12.8,-2) – (12.8,-1.6)–(11.25,-1.6); at (11,-1.3) [$\tilde{\boldsymbol{u}}_{t-1,1}$]{}; (12,-2.25) rectangle (13,-2); (12.5,-2.25) – (12.5,-3.05); (12.3,-2.7) – (12.7,-2.5); (12,-4.05) rectangle (13,-3.05); at (12.5,-3.55) [$N$]{}; at (13.6,-3.55) [$T^{\text{L}}_t$]{}; at (11.5,-2.7) [$\Pi^{\text{L}}_t$]{}; (12.5,-4.05) – (12.5,-4.55); (12.5,-4.55) circle \[radius=0.1\]; at (12.1,-4.9) [$\boldsymbol{v}^{\text{L}}_t$]{}; at (12.5,-5.6) [(c)]{}; \[thick, scale=0.65\] at (0.5,10) [$\boldsymbol{u}$]{}; (0.5,9.5) circle \[radius=0.1\]; (0.5,9.5) – (0.5,9); (0.5,9.5) – (1.5,9.5); (0,8) rectangle (1,9); at (0.5,8.5) [$N$]{}; at (-0.35,8.5) [$T^{\text{O}}$]{}; (0.5,8) – (0.5,7); (0.5,7) circle \[radius=0.1\]; at (0,7) [$\boldsymbol{v}^{\text{O}}$]{}; (0.5,7) – (0.5,6.5); (1.5,9.5) – (1.5,6.5); (0,6.5) rectangle (2,6.3); (1,6.3) – (1,3); (0.85,4.15) – (1.15,4.35); at (0.4,4.2) [$\Pi^{\text{Un}}$]{}; at (1.5,3.5) [$\boldsymbol{\tilde{v}}^{\text{O}}$]{}; (0.5,2) rectangle (1.5,3); at (1,2.5) [$2N$]{}; at (0.1,2.5) [$T^{\text{I}}$]{}; (1,2) – (1,1); (1,1) circle \[radius=0.1\]; at (0.5,1) [$\boldsymbol{v}^{\text{I}}$]{}; at (1,0) [(a)]{}; at (4,10) [$\boldsymbol{u}_t$]{}; (4,9.5) circle \[radius=0.1\]; (4,9.5) – (4,9); (4,9.5) – (5,9.5); (3.5,8) rectangle (4.5,9); at (4,8.5) [$N$]{}; at (3.15,8.5) [$T^{\text{O}}_t$]{}; (4,8) – (4,7); (4,7) circle \[radius=0.1\]; at (3.5,7) [$\boldsymbol{v}^{\text{O}}_t$]{}; (4,7) – (4,6.5); (5,9.5) – (5,6.5); (3.5,6.5) rectangle (5.5,6.3); (4.35,5.8) – (4.65,6); at (5.2,5.9) [$\Pi^{(1)}_t$]{}; (4.5,6.3) – (4.5,5.5); at (6,5.1) [$\tilde{\boldsymbol{v}}_{t,1}$]{}; (3.5,5.5) rectangle (5.5,5.3); (5,5.3) – (5,4.8)–(6,4.8); (4,4) – (4,5.3); (3.5,4) rectangle (5.5,3.8); (4.5,3.8) – (4.5,3); (4.35,3.2) – (4.65,3.5); at (5.2,3.4) [$\Pi^{(2)}_t$]{}; (5,4) – (5,4.4)–(3,4.4); at (2.8,4.25) [$\tilde{\boldsymbol{v}}_{t-1,1}$]{}; (4,2) rectangle (5,3); at (4.5,2.5) [$2N$]{}; at (3.6,2.5) [$T^{\text{I}}_t$]{}; (4.5,2) – (4.5,1); (4.5,1) circle \[radius=0.1\]; at (4,1) [$\boldsymbol{v}^{\text{I}}_t$]{}; at (4.5,0) [(b)]{}; at (8.2,10) [$\boldsymbol{v}^{\text{U}}$]{}; (8.2,9.5) circle \[radius=0.1\]; (8.2,9.5) – (8.2,9); (8.4,3) to \[out=70,in=0\] (8,9.5); (9.05,8.95) – (9.35,9.25); at (9.4,9.5) [$\Pi^{\text{L}}$]{}; (7.7,8) rectangle (8.7,9); at (8.2,8.5) [$N$]{}; at (7.3,8.5) [$T^{\text{U}}$]{}; (8,8) – (7.7,5.5); at (7.2,5.5) [$\boldsymbol{u}$]{}; (7.7,5.5) circle \[radius=0.1\]; (7.7,5.5) – (8,3); (7.7,2) rectangle (8.7,3); at (8.2,2.5) [$N$]{}; at (7.3,2.5) [$T^{\text{L}}$]{}; (7.7,4) – (8.05,4.25); at (7.2,4) [$\Pi$]{}; (8.2,1) – (8.2,2); (8.2,1) circle \[radius=0.1\]; (8.2,1) to \[out=0,in=290\] (8.5,8) ; (9.5,2.15) – (9.8,1.85); at (9.6,1.1) [$\Pi^{\text{U}}$]{}; at (7.7,1) [$\boldsymbol{v}^{\text{L}}$]{}; at (8.2,0) [(c)]{}; at (12.5,10) [$\boldsymbol{v}^{\text{U}}_t$]{}; (12.5,9.5) circle \[radius=0.1\]; (12.5,9.5) – (12.5,9); (12.5,9.5) to \[out=360,in=100\] (14,7); (13.2,8.95) – (13.55,9.25); at (13.6,9.5) [$\Pi^{\text{L}}_t$]{}; (12,8) rectangle (13,9); (12.8,8) – (13.3,6); at (13.8,5.9) [$\boldsymbol{v}^{\text{L}}_{t-1}$]{}; at (12.5,8.5) [$N$]{}; at (11.5,8.5) [$T^{\text{U}}_t$]{}; (12.3,8) – (12,5.5); at (11.5,5.5) [$\boldsymbol{u}_t$]{}; (12,5.5) circle \[radius=0.1\]; (12,5.5) – (12.3,3); (12.8,3) – (13.3,5); at (13.8,5.1) [$\boldsymbol{v}^{\text{U}}_{t-1}$]{}; (12,2) rectangle (13,3); at (12.5,2.5) [$N$]{}; at (11.5,2.5) [$T^{\text{L}}_t$]{}; (12,4) – (12.35,4.25); at (11.6,4) [$\Pi$]{}; (12.5,1) – (12.5,2); (12.5,1) circle \[radius=0.1\]; (12.5,1) to \[out=0,in=250\] (14,3.5) ; (13.4,2.15) – (13.7,1.85); at (13.6,1.3) [$\Pi^{\text{U}}$]{}; at (12,1) [$\boldsymbol{v}^{\text{L}}_t$]{}; at (12.5,0) [(d)]{}; at (16,10) [$\boldsymbol{u}$]{}; (16,9.5) circle \[radius=0.1\]; (16,9.5) – (16,9); (16,9.5) – (18,9.5); (16.85,9.35) – (17.15,9.65); at (17.1,10) [$\Pi^{\text{L}}$]{}; (18,9.5) – (18,9); (15.5,8) rectangle (16.5,9); at (16,8.5) [$N$]{}; at (15.1,8.5) [$T^{\text{U}}$]{}; (17.5,8) rectangle (18.5,9); at (18.9,8.5) [$T^{\text{L}}$]{}; at (18,8.5) [$N$]{}; (18,8) – (18,6.5); (16,8) – (16,6.5); (18,7) circle \[radius=0.1\]; at (17.5,7) [$\boldsymbol{v}^{\text{L}}$]{}; (16,7) circle \[radius=0.1\]; at (15.5,7) [$\boldsymbol{v}^{\text{U}}$]{}; (15.5,6.5) rectangle (18.5,6.3); (17,6.3) – (17,3); (16.85,4.15) – (17.15,4.35); at (16.4,4.2) [$\Pi^{\text{I}}$]{}; (16.5,2) rectangle (17.5,3); at (17,2.5) [$2N$]{}; at (16.1,2.5) [$T^{\text{I}}$]{}; (17,2) – (17,1); (17,1) circle \[radius=0.1\]; at (16.5,1) [$\boldsymbol{v}^{\text{I}}$]{}; at (17,0) [(e)]{}; at (20.5,10) [$\boldsymbol{u}_t$]{}; (20.5,9.5) circle \[radius=0.1\]; (20.5,9.5) – (20.5,9); (20.5,9.5) – (22.5,9.5); (21.35,9.35) – (21.65,9.65); at (21.6,10) [$\Pi^{\text{L}}_t$]{}; (20,8) rectangle (21,9); at (20.5,8.5) [$N$]{}; at (19.65,8.4) [$T^{\text{O}}_t$]{}; (20.5,8) – (20.5,7); (20.5,7) circle \[radius=0.1\]; at (19.8,7) [$\boldsymbol{v}^{\text{U}}_t$]{}; (20.5,7) – (20.5,6.5); (22.5,9.5) – (22.5,9); (22,8) rectangle (23,9); at (22.5,8.5) [$N$]{}; at (23.5,8.4) [$T^{\text{L}}_t$]{}; (22.5,8) – (22.5,6.5); (22.5,7) circle \[radius=0.1\]; at (23.3,7) [$\boldsymbol{v}^{\text{U}}_t$]{}; (20,6.5) rectangle (23,6.3); (21.35,5.8) – (21.65,6); at (22.2,5.9) [$\Pi^{(1)}_t$]{}; (21.5,6.3) – (21.5,5.5); (20,5.5) rectangle (23,5.3); (22.5,5.3) – (22.5,4.8)–(23.5,4.8); at (23.5,5.1) [$\tilde{\boldsymbol{v}}_{t,1}$]{}; (20.5,4) – (20.5,5.3); (20,4) rectangle (23,3.8); (21.5,3.8) – (21.5,3); (21.35,3.2) – (21.65,3.5); at (22.2,3.4) [$\Pi^{(2)}_t$]{}; at (21.05,4.95) [$\tilde{\boldsymbol{v}}_{t,0}$]{}; (22.5,4) – (22.5,4.4)–(19.5,4.4); at (19.5,4.25) [$\tilde{\boldsymbol{v}}_{t-1,1}$]{}; (21,2) rectangle (22,3); at (21.5,2.5) [$2N$]{}; at (20.6,2.5) [$T^{\text{I}}_t$]{}; (21.5,2) – (21.5,1); (21.5,1) circle \[radius=0.1\]; at (21,1) [$\boldsymbol{v}^{\text{I}}_t$]{}; at (21.5,0) [(f)]{}; The compact graph representation [@Moloudi_SCTC_Journal] of the PCC ensemble is depicted in Fig. \[CGPCC\](b). Each of the sequences ${\ensuremath{\boldsymbol{u}}}$, ${\ensuremath{\boldsymbol{v}}}^{\text{U}}$, and ${\ensuremath{\boldsymbol{v}}}^{\text{L}}$ is represented by a black circle, referred to as variable node. The trellises corresponding to the component encoders are shown by squares, called factor nodes, and they are labeled by the length of the trellises. The sequences ${\ensuremath{\boldsymbol{u}}}$ and ${\ensuremath{\boldsymbol{v}}}^{\text{U}}$ are connected to the upper trellis $\text{T}^{\text{U}}$. Likewise, a reordered copy of ${\ensuremath{\boldsymbol{u}}}$ and ${\ensuremath{\boldsymbol{v}}}^{\text{L}}$ are connected to the lower trellis $\text{T}^{\text{L}}$. In order to emphasize that a reordered copy of ${\ensuremath{\boldsymbol{u}}}$ is connected to $\text{T}^{\text{L}}$, the corresponding permutation is represented by a line that crosses the edge which connects ${\ensuremath{\boldsymbol{u}}}$ to $\text{T}^{\text{L}}$. Fig. \[CGPCC\](c) shows the compact graph representation of the spatially coupled PCC (SC-PCC) ensemble with coupling memory $m=1$ at time $t$. Consider a collection of PCC ensembles at time slots $t=1,\dots,L$, where $L$ is the coupling length. The SC ensemble can be obtained by dividing the information sequence at time $t$, ${\ensuremath{\boldsymbol{u}}}_t$, and its reordered copy, $\tilde{{\ensuremath{\boldsymbol{u}}}}_t$, into two subsequences, denoted by ${\ensuremath{\boldsymbol{u}}}_{t,j}$ and $\tilde{{\ensuremath{\boldsymbol{u}}}}_{t,j}$, $j=0,1$, respectively. Then these subsequences are spread over time $t$ and $t+1$. The input sequence to the upper encoder at time $t$ is the sequence $({\ensuremath{\boldsymbol{u}}}_{t,0},{\ensuremath{\boldsymbol{u}}}_{t-1,1})$, reordered by permutation $\Pi^{\text{U}}_t$.[^5] Likewise, the input sequence to the lower encoder at time $t$ is the sequence $(\tilde{{\ensuremath{\boldsymbol{u}}}}_{t,0},\tilde{{\ensuremath{\boldsymbol{u}}}}_{t-1,1})$, reordered by permutation $\Pi^{\text{L}}_t$. The information bits at time slots $t\leq 0$ are initialized by zero and the information bits at $t=L$ are chosen in such a way that ${\ensuremath{\boldsymbol{u}}}_{L,1}=0$ and $\tilde{{\ensuremath{\boldsymbol{u}}}}_{L,1}=0$ (i.e., we consider the termination of the coupled chain). Ensemble Type $\text{E}_\text{b}\text{/N}_\text{0}\text{ [dB]}_{\text{BP}}$ $\text{E}_\text{b}\text{/N}_\text{0}\text{ [dB]}_{\text{MAP}}$ $\text{E}_\text{b}\text{/N}_\text{0}\text{ [dB]}_{\text{SC}}^1$ $\text{E}_\text{b}\text{/N}_\text{0}\text{ [dB]}_{\text{SC}}^3$ $\text{E}_\text{b}\text{/N}_\text{0}\text{ [dB]}_{\text{SC}}^5$ ---------------------------------------------------------------------------- --------- --------------------------------------------------------------- ---------------------------------------------------------------- ----------------------------------------------------------------- ----------------------------------------------------------------- ----------------------------------------------------------------- ${\mathcal{C}_{\mathrm{PCC}}}$/${\mathcal{C}_{\mathrm{SC-PCC}}}$ - -0.1053 -0.3070 -0.3070 -0.3070 -0.3070 \[0.5mm\] ${\mathcal{C}_{\mathrm{SCC}}}$/${\mathcal{C}_{\mathrm{SC-SCC}}}$ - 1.4024 -0.4740 -0.1196 -0.4673 -0.4740 \[0.5mm\] ${\mathcal{C}_{\mathrm{BCC}}}$ Type-I 1.2139 -0.4723 -0.3992 -0.4573 -0.4673 \[0.5mm\] ${\mathcal{C}_{\mathrm{BCC}}}$ Type-II 1.2139 -0.4723 -0.4690 -0.4723 -0.4723 \[0.5mm\] ${\mathcal{C}_{\mathrm{HCC}}}$/${\mathcal{C}_{\mathrm{SC-HCC}}}$ Type-I 3.8846 -0.4941 1.0366 0.3038 0.0780 \[0.5mm\] ${\mathcal{C}_{\mathrm{HCC}}}$/${\mathcal{C}_{\mathrm{SC-HCC}}}$ Type-II 3.8846 -0.4941 0.2809 -0.4706 -0.4941 \[0.5mm\] \[ThresholdsAWGN\] Fig. \[CG\] shows the compact graph representation of the SCC, BCC, and HCC ensembles, and their corresponding spatially coupled ensembles. In this paper, we restrict ourselves to PCC, SCC and HCC ensembles with identical 4-state component trellises and generator matrix ${{\ensuremath{\boldsymbol{G}}}}=(1,5/7)$, in octal notation. For the BCC ensemble, we consider two identical 4-state component trellises with generator matrix $$\label{GMatrix} {\ensuremath{\boldsymbol{G}}}(D)= \left( \begin{array}{ccc}1&0&1/7\\0&1&5/7\end{array}\right) \ .$$ We also restrict ourselves to systematic TCs and SC-TCs with rate $R=1/3$. Therefore, for the SCC and HCC ensembles, we consider full puncturing of the parity sequences of the outer encoders [@Moloudi_SCTC_Journal; @MoloudiHCC]. Spatial Coupling: Waterfall Region Performance {#Waterfall} =============================================== Asymptotic Performance ---------------------- Using the decoding threshold of an ensemble computed for the BEC, it is possible to predict its decoding threshold over the AWGN channel [@LDPCPuncture; @UmarISIT2018]. This allows us to use the decoding thresholds of the TC and SC-TCs from [@Moloudi_SCTC_Journal; @MoloudiHCC] to predict the corresponding thresholds over the AWGN channel. The results are shown in Table \[ThresholdsAWGN\]. Similar to the BEC, among all the uncoupled TC ensembles, the PCC ensemble has the best BP threshold but the worst MAP threshold. Conversely, the HCC ensemble has the worst BP threshold but the best MAP threshold, which is very close to the Shannon limit. It can also be seen that for all coupled ensembles, threshold saturation occurs. In general, as the numerical results in Table \[ThresholdsAWGN\] suggest, SC-TC ensembles can achieve close-to-capacity BP thresholds. Finite Block-Length Performance ------------------------------- Fig. \[SCSCC\] shows BER simulation results for PCCs, SCCs, SC-PCCs, and spatially coupled SCC (SC-SCCs) with $R=1/3$ and input block length $K=1024$ and $K=4096$. For the coupled ensembles, we consider a coupling length $L=100$ and a sliding window decoder with window size $W=4$ [@WindowBCC]. The decoding latency is $W\cdot K$. It is well known that the PCC ensemble yields better performance than the SCC ensemble in the waterfall region [@Benedetto98Serial]; however, the SCC ensemble has a much lower error floor than the PCC ensemble. By applying spatial coupling, the performance of the PCC and SCC ensembles improves significantly for both input block lengths. This improvement is more substantial for the SCC ensemble than for the PCC ensemble. For instance, the performance of the SCC ensemble with $K=1024$ at BER$=10^{-5}$ improves more than $1$ dB with coupling. The coupling gains are in agreement with the decoding thresholds in Table \[ThresholdsAWGN\]. As it can be seen, the gap between the BP and MAP threshold of the SCC ensemble is larger than that of the PCC ensemble, hence the expected gain from coupling is bigger for the SCC ensemble. table\[row sep=crcr\][ 0 0.19251982421875\ 0.25 0.1825353515625\ 0.5 0.1690681640625\ 0.75 0.140709765625\ 1 0.084628515625\ 1.25 0.02889150390625\ 1.5 0.0045283203125\ 1.75 0.0003138671875\ 2 1.05831473309402e-05\ 2.25 5.4834799827005e-08\ ]{}; table\[row sep=crcr\][ 0 0.1930427734375\ 0.25 0.183439086914062\ 0.5 0.172591967773437\ 0.75 0.156407446289063\ 1 0.082587548828125\ 1.25 0.005832568359375\ 1.5 7.16798803979498e-05\ 1.7 1.11575837864919e-09\ 1.75 0\ 2 0\ 2.25 0\ ]{}; table\[row sep=crcr\][ 0 0.090793875933733\ 0.1 0.0842238670888573\ 0.2 0.070800686296279\ 0.3 0.0443002741301971\ 0.4 0.0171135216932222\ 0.5 0.00447514303961366\ 0.6 0.000921980399917401\ 0.7 8.87203699372661e-05\ 0.8 8.67343148630292e-06\ 0.9 4.67760382044464e-07\ 1 1.79277885886327e-08\ ]{}; table\[row sep=crcr\][ 0 0.0701610898820424\ 0.1 0.017219534475949\ 0.2 0.00116042537965636\ 0.3 4.63337261565446e-06\ 0.4 1.58213250648041e-09\ ]{}; table\[row sep=crcr\][ 0 0.05203212890625\ 0.2 0.02433720703125\ 0.4 0.00817333984375\ 0.6 0.00198681640625\ 0.8 0.00038330078125\ 1 0.000106796387894998\ 1.2 4.47352625569078e-05\ 1.4 2.37740078910576e-05\ 1.6 1.56578311002218e-05\ 1.8 1.04147357561604e-05\ 2 7.11717573897689e-06\ ]{}; table\[row sep=crcr\][ 0 0.0440502685546875\ 0.2 0.008068212890625\ 0.4 0.0004168701171875\ 0.8 1.358828163533e-05\ 1.2 5.32742538102407e-06\ 1.6 2.68697736671556e-06\ 1.8 1.92531534414703e-06\ 2 1.3320173464315e-06\ ]{}; table\[row sep=crcr\][ 0 0.0263513893235528\ 0.1 0.0160744685520261\ 0.2 0.00759753872544846\ 0.3 0.002830368920624\ 0.4 0.000939602598852041\ 0.5 0.000325916746948132\ 0.6 0.000127166994819744\ 0.7 5.95716265558421e-05\ 0.8 3.25112107623318e-05\ 0.9 2.03514312709966e-05\ 1 1.36601848259809e-05\ ]{}; table\[row sep=crcr\][ 0 0.0268895536590004\ 0.1 0.0145323003980086\ 0.2 0.00361651258879468\ 0.3 0.000552839972911081\ 0.4 0.000117946110718639\ 0.5 3.74229918134973e-05\ 0.6 1.62992615630661e-05\ 0.7 8.36527939383865e-06\ 0.8 4.92285329399379e-06\ 0.9 3.31333705357143e-06\ 1 2.33558347471541e-06\ ]{}; In Fig. \[SCSCC\], the uncoupled ensemble with $K=4096$ and the coupled ensemble with $K=1024$ have equal latency, i.e., both ensembles have a decoding latency of 4096 bits. For this latency, the SC-SCC ensemble performs better than the SCC ensemble. However, in the case of PCCs, for a latency of 4096 bits, the uncoupled ensemble performs slightly better than the corresponding coupled ensemble. Interestingly, for equal latency, the SC-SCC ensemble outperforms the SC-PCC ensemble in the waterfall region. Thus, the SC-SCC ensemble yields better performance in both the waterfall and error floor regions. In the following section, we investigate the impact of spatial coupling on the error floor performance of TCs. Spatial Coupling: Error Floor Region Performance {#Proof} ================================================ Similar to uncoupled TC ensembles, to analyze the performance of SC-TC ensembles in the error floor region, one could derive bounds based on the WEFs of the ensembles. Unfortunately, deriving the WEF for SC-TCs is cumbersome. In this section, we establish a connection between the WEF of SC-TC ensembles and that of the corresponding uncoupled ensembles. In particular, we prove that, under certain conditions, spatial coupling does not decrease the minimum distance of TCs. This allows us to use the WEF analysis of TCs to estimate the error floor performance of SC-TCs. A similar connection between LDPC and SC-LDPC codes is proved in [@TruhachevDistBoundsTBLDPCCCs; @MitchellMinDisTrapSet; @PseudocodLDPC]. Here, we restrict ourselves to SC-TCs with coupling memory $m=1$, but the proof can be generalized to higher coupling memories. Consider an uncoupled PCC, $\tilde{\mathcal{C}}$, (see Fig. \[CGPCC\](b)) with permutation $\Pi^{\text{Un}}$ and parity-check matrices ${\ensuremath{\boldsymbol{H}}}_{\text{U}}$ and ${\ensuremath{\boldsymbol{H}}}_{\text{L}}$ corresponding to the upper and lower component encoders. It is possible to unwrap the PCC to form an SC-PCC, $\mathcal{C}$ (Fig. \[CGPCC\](c)). For the SC-PCC, we assume a length-$L$ coupled chain with termination or tailbiting, and time-invariant permutations. Let us denote the permutations by $\Pi^{\text{U}}_t=\Pi^{\text{U}}$, $\Pi^{\text{L}}_t=\Pi^{\text{L}}$, and $\Pi_t=\Pi$, and assume that they satisfy $$\Pi^{\text{Un}}=(\Pi^{\text{U}})^{-1}\cdot \Pi \cdot \Pi^{\text{L}}.$$ Then, for any codeword ${\ensuremath{\boldsymbol{v}}}\in\mathcal{C}$, ${\ensuremath{\boldsymbol{v}}}=({\ensuremath{\boldsymbol{v}}}_1,\dots,{\ensuremath{\boldsymbol{v}}}_L)$, ${\ensuremath{\boldsymbol{v}}}_t=({\ensuremath{\boldsymbol{u}}}_t,{\ensuremath{\boldsymbol{v}}}_t^{\text{U}}, {\ensuremath{\boldsymbol{v}}}_t^{\text{L}})$, with Hamming weight $w_{\text{H}}({\ensuremath{\boldsymbol{v}}})$, there exists a codeword $\tilde{{\ensuremath{\boldsymbol{v}}}}\in\tilde{\mathcal{C}}$ such that $$w_{\text{H}}(\tilde{{\ensuremath{\boldsymbol{v}}}})\leq w_{\text{H}}({\ensuremath{\boldsymbol{v}}})\; .$$ We prove this theorem for tailbiting of the coupled chain, which contains termination as a special case. The result is thus valid for both cases. Any codeword ${\ensuremath{\boldsymbol{v}}}\in\mathcal{C}$ satisfies the local constraints for $t=1,\dots,L$. Therefore, at time $t$, $$\begin{aligned} \label{PCCProof1} &\big(({\ensuremath{\boldsymbol{u}}}_{t,0},{\ensuremath{\boldsymbol{u}}}_{t-1,1})\cdot\Pi^{\text{U}}\;\;\; {\ensuremath{\boldsymbol{v}}}^{\text{U}}_t\big)\cdot {\ensuremath{\boldsymbol{H}}}_{\text{U}}^{\text{T}}={\ensuremath{\boldsymbol{0}}}, \\ \label{PCCProof2} &\big(({\ensuremath{\boldsymbol{u}}}'_{t,0},{\ensuremath{\boldsymbol{u}}}'_{t-1,1})\cdot\Pi^{\text{L}}\;\;\; {\ensuremath{\boldsymbol{v}}}^{\text{L}}_t\big)\cdot {\ensuremath{\boldsymbol{H}}}_{\text{L}}^{\text{T}}={\ensuremath{\boldsymbol{0}}}, \end{aligned}$$ where ${\ensuremath{\boldsymbol{u}}}'_t={\ensuremath{\boldsymbol{u}}}_t\cdot \Pi$. The constraints are linear and time-invariant. Thus, for $t={1,\dots,L}$, any superposition of the vectors $\big(({\ensuremath{\boldsymbol{u}}}_{t-1,1},{\ensuremath{\boldsymbol{u}}}_{t,2})\cdot\Pi_1\;\;\; {\ensuremath{\boldsymbol{v}}}^{\text{U}}_t\big)$ and $ \big(({\ensuremath{\boldsymbol{u}}}'_{t-1,1},{\ensuremath{\boldsymbol{u}}}'_{t,2})\cdot\Pi_1\;\;\; {\ensuremath{\boldsymbol{v}}}^{\text{L}}_t\big)$ satisfies and , respectively. In particular, consider $$\begin{aligned} &\sum_{t=1}^{L}\big(({\ensuremath{\boldsymbol{u}}}_{t,0},{\ensuremath{\boldsymbol{u}}}_{t-1,1})\cdot\Pi^{\text{U}}\;\;\; {\ensuremath{\boldsymbol{v}}}^{\text{U}}_t\big)\nonumber \\ &=\big(\sum_{t=1}^{L}({\ensuremath{\boldsymbol{u}}}_{t,0},{\ensuremath{\boldsymbol{u}}}_{t-1,1})\cdot\Pi^{\text{U}}\;\;\; \sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{U}}_t\big)\nonumber\\ &=\big(\sum_{t=1}^{L}{\ensuremath{\boldsymbol{u}}}_{t}\cdot\Pi^{\text{U}}\;\;\; \sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{U}}_t\big)\;, \label{PCCProof3} \end{aligned}$$ and $$\begin{aligned} &\sum_{t=1}^{L}\big(({\ensuremath{\boldsymbol{u}}}'_{t,0},{\ensuremath{\boldsymbol{u}}}'_{t-1,1})\cdot\Pi^{\text{L}}\;\;\; {\ensuremath{\boldsymbol{v}}}^{\text{L}}_t\big)\nonumber \\ &=\big(\sum_{t=1}^{L}({\ensuremath{\boldsymbol{u}}}'_{t,0},{\ensuremath{\boldsymbol{u}}}'_{t-1,1})\cdot\Pi^{\text{L}}\;\;\; \sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{L}}_t\big)\nonumber\\ &=\big(\sum_{t=1}^{L}{\ensuremath{\boldsymbol{u}}}'_{t}\cdot\Pi^{\text{L}}\;\;\; \sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{L}}_t\big)\nonumber\\ &=\big(\sum_{t=1}^{L}{\ensuremath{\boldsymbol{u}}}_t\cdot\Pi\cdot\Pi^{\text{L}}\;\;\; \sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{L}}_t\big)\;. \label{PCCProof4} \end{aligned}$$ Let $$\tilde{{\ensuremath{\boldsymbol{u}}}}=\sum_{t=1}^{L}{\ensuremath{\boldsymbol{u}}}_t\cdot\Pi^{\text{U}},\;\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}}=\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}_t^{\text{U}},\;\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}}=\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}_t^{\text{L}}\;.$$ Then, the vectors obtained from and can be rewritten as $(\tilde{{\ensuremath{\boldsymbol{u}}}}\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}})$ and $(\tilde{{\ensuremath{\boldsymbol{u}}}}\cdot\Pi^{\text{Un}}\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}})$, respectively. The vectors from and satisfy and , respectively. Thus, $$\begin{aligned} &(\tilde{{\ensuremath{\boldsymbol{u}}}}\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}})\cdot {\ensuremath{\boldsymbol{H}}}_{\text{U}}^{\text{T}}={\ensuremath{\boldsymbol{0}}}, \\ &(\tilde{{\ensuremath{\boldsymbol{u}}}}\cdot\Pi^{\text{Un}}\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}})\cdot {\ensuremath{\boldsymbol{H}}}_{\text{L}}^{\text{T}}={\ensuremath{\boldsymbol{0}}}. \end{aligned}$$ Therefore, $\tilde{{\ensuremath{\boldsymbol{v}}}}=(\tilde{{\ensuremath{\boldsymbol{u}}}},\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}},\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}})$ is a codeword of the uncoupled ensemble. If all nonzero elements of ${\ensuremath{\boldsymbol{v}}}_t$, $t=1,\dots,L$, occur at different positions, then $w_H(\tilde{{\ensuremath{\boldsymbol{v}}}})=w_H({\ensuremath{\boldsymbol{v}}})$. Otherwise, the overlap of the nonzero elements reduces the weight of $\tilde{{\ensuremath{\boldsymbol{v}}}}$ and $w_{\text{H}}(\tilde{{\ensuremath{\boldsymbol{v}}}})< w_{\text{H}}({\ensuremath{\boldsymbol{v}}})$. This theorem can be extended to the other TC ensembles. Consider an uncoupled SCC (BCC/HCC), $\tilde{\mathcal{C}}$, (Fig. \[CG\]). It is possible to unwrap the SCC (BCC/HCC) to form an SC-SCC (BCC/SC-HCC), $\mathcal{C}$ (Fig. \[CG\]). For the coupled code, we assume a length-$L$ coupled chain with termination or tailbiting, and time-invariant permutations which satisfy certain conditions. Then, for any codeword ${\ensuremath{\boldsymbol{v}}}\in\mathcal{C}$, ${\ensuremath{\boldsymbol{v}}}=({\ensuremath{\boldsymbol{v}}}_1,\dots,{\ensuremath{\boldsymbol{v}}}_t,\dots,{\ensuremath{\boldsymbol{v}}}_L)$, ${\ensuremath{\boldsymbol{v}}}_t=({\ensuremath{\boldsymbol{u}}}_t,{\ensuremath{\boldsymbol{v}}}_t^{\text{U}}, {\ensuremath{\boldsymbol{v}}}_t^{\text{L}})$, there exists a codeword $\tilde{{\ensuremath{\boldsymbol{v}}}}\in\tilde{\mathcal{C}}$ such that $$w_{\text{H}}(\tilde{{\ensuremath{\boldsymbol{v}}}})\leq w_{\text{H}}({\ensuremath{\boldsymbol{v}}})\;.$$ See Appendix. The minimum distance of an SC-TC ensemble $\mathcal{C}$ is larger than or equal to the minimum distance of the underlying uncoupled TC ensemble $\tilde{\mathcal{C}}$, $$d_{\text{min}}(\mathcal{C}) \geq d_{\text{min}}(\tilde{\mathcal{C}}) \; .$$ $\square$ By the above theorems, we establish conditions on the permutations under which SC-TCs have equal or better minimum distance than their corresponding TCs. These conditions can be considered as guidelines for selecting proper permutations for SC-TCs. According to the theorems and the corollary above, the WEF analysis of uncoupled TC ensembles can be used to investigate the error floor and the minimum distance of SC-TC ensembles. Weight Enumerator Analysis {#WE} ========================== In this section, we describe how to derive upper bounds on the error rate performance of TC ensembles and bounds on the minimum distance of these ensembles based on their WEFs [@UnveilingTC; @Benedetto98Serial]. Then, we compare these bounds for different classes of TCs. For that, we first derive the average input-parity WEF (IP-WEF) of the component encoders. In particular, we describe the steps for a rate-$2/3$ recursive systematic convolutional encoder. A similar method can be used to derive the IP-WEF of any convolutional encoder with arbitrary rate $R$. Then, we use the obtained IP-WEFs to compute the average IP-WEFs of the TC ensembles. Input-Parity Weight Enumerator {#WEF} ------------------------------ Let $A(I_1,I_2,P)$ denote the IP-WEF of a rate-$2/3$ recursive systematic convolutional encoder, $$A(I_1,I_2,P)=\sum_{i_1} \sum_{i_2}\sum_{p} A_{i_1,i_2,p} \, I^{i_1} I^{i_2}P^p,$$ where the coefficient $A_{i_1,i_2,p}$ denotes the number of codewords with weight $i_1$, $i_2$, and $p$ for the first input, the second input, and the parity sequence, respectively. $A(I_1,I_2,P)$ can be computed as follows. For a trellis with $s$ states, transitions within a trellis section can be described by an $s\times s$ matrix ${\ensuremath{\boldsymbol{M}}}$. The element of ${\ensuremath{\boldsymbol{M}}}$ in the $r$th row and the $c$th column, $[{\ensuremath{\boldsymbol{M}}}]_{r,c}$, corresponds to the trellis branch from the $r$th state to the $c$th state. More precisely, $[{\ensuremath{\boldsymbol{M}}}]_{r,c}$ is a monomial $I_1^{i_1}I_2^{i_2}P^{p}$, where $i_1$, $i_2$, and $p$ are the weights corresponding to the first, second, and third outputs of the transition from the $r$th state to the $c$th state. For a trellis with $N$ sections, the overall transition matrix is ${\ensuremath{\boldsymbol{M}}}^N$. Considering that the trellis is initialized and terminated to the all-zero state, the IP-WEF is given by the element $[{\ensuremath{\boldsymbol{M}}}^{N}]_{1,1}$. Assume a terminated, rate-$2/3$ convolutional encoder with three trellis sections and generator matrix in . The transition matrix can be written as $$\label{eq:AR23} \boldsymbol{M}(I_1,I_2,P)=\left( \begin{array}{cccc}1&I_2P&I_1I_2&I_1P\\I_1&I_1I_2P&I_2&P\\I_2P&1&I_1P&I_1I_2\\I_1I_2P&I_1&P&I_2 \end{array}\right),\nonumber$$ and the IP-WEF becomes $$\begin{aligned} &A(I_1,I_2,P)= [{\ensuremath{\boldsymbol{M}}}^{3}]_{1,1}=\nonumber\\ &1+I_2^3P^2+2I_1I_2P+I_1I_2P^3+2I_1I_2^2P+\nonumber\\ &I_1I_2^2P^3+I_1^2I_2+2I_1^2I_2P^2+3I_1^2I_2^2P^2+I_1^3P+I_1^3I_2^3P \;.\nonumber\end{aligned}$$ $\triangle$ For a rate-$1/2$ convolutional encoder, we can obtain the transition matrix ${\ensuremath{\boldsymbol{M}}}$ in a similar way. Then, the IP-WEF of the encoder is given by $[{\ensuremath{\boldsymbol{M}}}^{N}]_{1,1}$ and can be written as $$A(I,P)=\sum_{i} \sum_{p}A_{i,p} \, I^iP^p,$$ where $A_{i,p}$ is the number of codewords of input weight $i$ and parity weight $p$. Consider the PCC ensemble shown in Fig. \[CG\](b). Let $A^{\text{T}_{\text{U}}}(I,P)$ and $A^{\text{T}_{\text{L}}}(I,P)$ denote the IP-WEFs corresponding to the upper and lower component encoder, respectively. The overall IP-WEF depends on the IP-WEF of the component encoders and the permutation used. Averaging over all possible permutations, the coefficients of the average IP-WEF of the PCC ensemble, $\bar{A}_{i,p}^{\text{PCC}}$, can be obtained as [@UnveilingTC] $$\bar{A}_{i,p}^{\text{PCC}}=\frac{\sum_{p_1}A^{\text{T}_{\text{U}}}_{i,p_1}\cdot A^{\text{T}_{\text{L}}}_{i,p-p_1}}{\binom{N}{i}}\;.$$ For the SCC ensemble shown in Fig. \[CG\](b), we denote the IP-WEFs of the outer and inner encoder by $A^{\text{T}_{\text{O}}}(I,P)$ and $A^{\text{T}_{\text{I}}}(I,P)$, respectively. Similar to PCCs, the average IP-WEF of the SCC ensemble, $\bar{A}_{i,p}^{\text{SCC}}$, can be computed by averaging over all possible permutations [@Benedetto98Serial]. The coefficients $\bar{A}_{i,p}^{\text{SCC}}$ can be written as $$\label{IPWESCC} \bar{A}_{i,p}^{\text{SCC}}=\sum_{p_1}\frac{A^{\text{T}_{\text{O}}}_{i,p_1}\cdot A^{\text{T}_{\text{I}}}_{i+p_1,p-p_1}}{\binom{2N}{i+p_1}}\;.$$ We denote the IP-WEFs corresponding to the upper and lower component encoders of the BCC ensemble (Fig. \[CG\](c)) by $A^{\text{T}_{\text{U}}}(I,P)$ and $A^{\text{T}_{\text{L}}}(I,P)$, respectively. The coefficients of the average IP-WEF, $\bar{A}_{i,p}^{\text{BCC}}$, can be computed as $$\label{IPWEBCC} \bar{A}_{i,p}^{\text{BCC}}=\sum_{p_1}\frac{A^{\text{T}_{\text{U}}}_{i,p_1,p-p_1}\cdot A^{\text{T}_{\text{L}}}_{i,p-p_1,p_1}}{\binom{N}{i}\binom{N}{p_1}\binom{N}{p-p_1}}\;.$$ To compute the average IP-WEF of the HCC ensemble, $\bar{A}_{i,p}^{\text{HCC}}$, it is possible to combine the methods that we used for PCCs and SCCs. First, the average IP-WEF of the parallel component is computed. Then, $\bar{A}_{i,p}^{\text{HCC}}$ can be obtained by substituting $A^{\text{T}_{\text{O}}}(I,P)$ in by the computed average IP-WEF of the parallel component [@3DTurbo], $$\label{IPWEHCC} \bar{A}_{i,p}^{\text{HCC}}=\sum_{p_1}\sum_{p_2}\frac{A^{\text{T}_{\text{U}}}_{i,p_1}\cdot A^{\text{T}_{\text{L}}}_{i,p_2}\cdot A^{\text{T}_{\text{I}}}_{p_1+p_2,p-p_1-p_2}}{\binom{N}{i}\binom{2N}{p_1+p_2}}\;.$$ It is worth mentioning that by the use of the compact graph representation, TCs can be seen as a class of protograph-based generalized LDPC (GLDPC) codes. Therefore, equivalently, it is possible to compute the average IP-WEF of TCs by the method developed for GLDPC codes in [@AbuSurrahGlobeCom07; @AbuSurra2011]. Bounds on the Error Probability ------------------------------- Consider transmission of codewords of a rate-$R$ TC ensemble over the AWGN channel. For a maximum likelihood (ML) decoder, the BER is upper bounded by $$\label{BER} P_b\leq \sum_{i=1}^{N} \sum_{p=1}^{N(1/R-1)} \frac{i}{N} \bar{A}_{i,p}\; \mathsf{Q}\left ( \sqrt{2(i+p)R\frac{E_{\mathrm b}}{N_0}}\right).$$ Likewise, the frame error rate (FER) is upper bounded by $$\label{FER} P_F\leq \sum_{i=1}^{N} \sum_{p=1}^{N(1/R-1)} \bar{A}_{i,p} \;\mathsf{Q}\left( \sqrt{2(i+p)R\frac{E_{\mathrm b}}{N_0}}\right),$$ where $Q(.)$ is the $Q$-function and ${E_{\mathrm b}}/{N_0}$ is the signal-to-noise ratio. table\[row sep=crcr\][ 0 1.08127674129687e+49\ 0.5 8.1978582039498e+37\ 1 6.70337672365652e+25\ 1.5 6401834190216.44\ 2 0.105701359658795\ 2.5 5.51539025733897e-06\ 3 2.22667984660957e-06\ 3.5 8.52945367816259e-07\ 4 3.04242737677132e-07\ 4.5 9.97706591145024e-08\ 5 2.97946570872782e-08\ 5.5 8.0490388956193e-09\ 6 1.96074384771405e-09\ 6.5 4.30954488118009e-10\ 7 8.57148726820489e-11\ 7.5 1.54296499802588e-11\ 8 2.48793008351179e-12\ 8.5 3.50328114672245e-13\ 9 4.1489315514129e-14\ 9.5 3.95930101608258e-15\ 10 2.91630066671554e-16\ ]{}; table\[row sep=crcr\][ 0 8.0601559503983e+48\ 0.5 5.06196825590139e+37\ 1 2.92925123166828e+25\ 1.5 1428167627285.4\ 2 0.00563180017896353\ 2.5 7.1726419084409e-10\ 3 2.93953470547403e-10\ 3.5 1.13727092537031e-10\ 4 4.12601938801283e-11\ 4.5 1.39472738324373e-11\ 5 4.37034079667638e-12\ 5.5 1.26699875563662e-12\ 6 3.41275332724769e-13\ 6.5 8.65669020861838e-14\ 7 2.1140547829608e-14\ 7.5 5.07123057996476e-15\ 8 1.19447226320358e-15\ 8.5 2.68570023713636e-16\ 9 5.52230355494405e-17\ 9.5 9.966986417699e-18\ 10 1.52837105335438e-18\ ]{}; table\[row sep=crcr\][ 0 1.30217371829227e+48\ 0.5 9.08895923057792e+36\ 1 5.90803775618208e+24\ 1.5 328201533373.867\ 2 0.00151217391899318\ 2.5 1.14185277283209e-07\ 3 8.79627275088619e-08\ 3.5 6.59290062332492e-08\ 4 4.79245475080335e-08\ 4.5 3.36656880657599e-08\ 5 2.27624082848239e-08\ 5.5 1.47464935538338e-08\ 6 9.10744701838694e-09\ 6.5 5.33173476636929e-09\ 7 2.93986148685546e-09\ 7.5 1.51583656558528e-09\ 8 7.25006555339804e-10\ 8.5 3.18758111981953e-10\ 9 1.27523986011843e-10\ 9.5 4.58956919878121e-11\ 10 1.46699845994459e-11\ ]{}; table\[row sep=crcr\][ 0 2.91739759742634e+33\ 0.5 1.34661005934783e+23\ 1 422050334493.007\ 1.5 0.0776733775763748\ 2 6.65040171986944e-13\ 2.5 2.89607283990355e-13\ 3 1.19660507639199e-13\ 3.5 4.70175065748866e-14\ 4 1.77365416759004e-14\ 4.5 6.52847459903677e-15\ 5 2.39004220707178e-15\ 5.5 8.8098829173984e-16\ 6 3.25419898999524e-16\ 6.5 1.17852273373363e-16\ 7 4.06311651881909e-17\ 7.5 1.29633042514568e-17\ 8 3.73478307339253e-18\ 8.5 9.50498501514067e-19\ 9 2.09133395469223e-19\ 9.5 3.88990075009214e-20\ 10 5.97111341950619e-21\ ]{}; Fig. \[BERUnion\] shows the bounds on the BER performance of the different classes of TCs for $R=1/3$ and $K=512$. The bounds are truncated at weight $w=320$, which is larger than the corresponding Gilbert-Varshamov limit. The HCC ensemble has the lowest error floor, while the BCC and PCC ensembles have the highest error floors. Surprisingly, the error floor of the BCC ensemble is not only high but also has the worst slope among all TC ensembles. On the other hand, the excellent MAP thresholds of the BCC ensemble suggest a good performance for this ensemble under MAP decoding. The contradiction between the excellent MAP threshold of the BCC ensemble and its poor bound suggests that the performance is dominated by few bad permutations. To verify this, we simulated the BCCs for two scenarios; first, a randomly selected but fixed set of permutations; second randomly chosen permutations for each simulated block. The results are shown in Fig. \[BCCFP\], together with the corresponding bounds. The figure shows that the bounds are in agreement with the simulation results for uniformly random permutations. However, it indicates a significant improvement in FER for the fixed set of permutations. For example, at $E_{\mathrm b}/N_0=2.5\;\text{dB}$, the FER improves from $9.5\cdot 10^{-5}$ to $6.8\cdot 10^{-7}$. This significant improvement caused by fixing the permutations, supports that the high floor of the BCC ensemble is caused by the poor performance of a small fraction of codes. Thus, in the next section, we compute expurgated union bounds. table\[row sep=crcr\][ 1 0.5556\ 1.5 0.0689\ 2 0.00157962286504097\ 2.5 9.41222872612207e-05\ 3 7.57213643096815e-05\ 3.5 4.84496124031008e-05\ 4 2.37334261574001e-05\ 4.5 1.60110155787182e-05\ 5 8.00313722979408e-06\ ]{}; table\[row sep=crcr\][ 1 0.0834537109375\ 1.5 0.00974296875\ 2 0.000198918328561556\ 2.5 2.27602257142089e-07\ 3 1.75623281479974e-07\ 3.5 9.46281492248062e-08\ 4 4.6354347963672e-08\ 4.5 3.12715148021839e-08\ 5 1.56311274019416e-08\ ]{}; table\[row sep=crcr\][ 1 0.5689\ 1.5 0.0725\ 2 0.00127333763291495\ 2.5 6.81625003095714e-07\ 2.7 0\ ]{}; table\[row sep=crcr\][ 1 0.0856236328125\ 1.5 0.010412109375\ 2 0.000167209540523557\ 2.5 9.76285811725631e-08\ 2.7 0\ ]{}; table\[row sep=crcr\][ 0 1.0008218632528e+41\ 0.1 1.65925966128923e+39\ 0.2 2.50582548571697e+37\ 0.3 3.44032622418754e+35\ 0.4 4.28531580689652e+33\ 0.5 4.83294766727444e+31\ 0.6 4.92481387730191e+29\ 0.7 4.5249174147433e+27\ 0.8 3.74078912127343e+25\ 0.9 2.77672141109243e+23\ 1 1.84672173021244e+21\ 1.1 1.09813616133767e+19\ 1.2 5.82623303214839e+16\ 1.3 275231599560397\ 1.4 1155328799723.91\ 1.5 4300817273.94712\ 1.6 14171221.2369379\ 1.7 41256.1924203571\ 1.8 105.941483357425\ 1.9 0.239671293503461\ 2 0.000550517905705948\ 2.1 7.14961261691498e-05\ 2.2 6.74860380463551e-05\ 2.3 6.43902437031355e-05\ 2.4 6.13800761618375e-05\ 2.5 5.84546191863483e-05\ 2.6 5.5614126069299e-05\ 2.7 5.2858748373982e-05\ 2.8 5.018853277967e-05\ 2.9 4.76034199255645e-05\ 3 4.5103243561617e-05\ 3.1 4.26877300186215e-05\ 3.2 4.03564980108887e-05\ 3.3 3.81090587835473e-05\ 3.4 3.59448166150847e-05\ 3.5 3.38630696841501e-05\ 3.6 3.18630113078904e-05\ 3.7 2.99437315571898e-05\ 3.8 2.8104219252132e-05\ 3.9 2.63433643388166e-05\ 4 2.46599606463495e-05\ 4.1 2.30527090203956e-05\ 4.2 2.15202208271634e-05\ 4.3 2.0061021819081e-05\ 4.4 1.86735563507652e-05\ 4.5 1.73561919311819e-05\ 4.6 1.610722409519e-05\ 4.7 1.49248815749623e-05\ 4.8 1.3807331749126e-05\ 4.9 1.27526863448914e-05\ 5 1.17590073659594e-05\ 5.1 1.0824313216668e-05\ 5.2 9.94658499066926e-06\ 5.3 9.12377289046608e-06\ 5.4 8.35380274240532e-06\ 5.5 7.63458257026169e-06\ 5.6 6.96400918937248e-06\ 5.7 6.3399747824325e-06\ 5.8 5.76037341755362e-06\ 5.9 5.22310746905123e-06\ 6 4.72609390166399e-06\ 6.1 4.26727037954888e-06\ 6.2 3.84460116243464e-06\ 6.3 3.4560827527596e-06\ 6.4 3.09974925946514e-06\ 6.5 2.77367744635012e-06\ 6.6 2.47599143550086e-06\ 6.7 2.20486703927131e-06\ 6.8 1.95853569757058e-06\ 6.9 1.73528800078469e-06\ 7 1.53347678247513e-06\ 7.1 1.35151977001307e-06\ 7.2 1.18790178547207e-06\ 7.3 1.04117649336126e-06\ 7.4 9.09967696074689e-07\ 7.5 7.92970182202832e-07\ 7.6 6.88950137036126e-07\ 7.7 5.96745128628244e-07\ 7.8 5.15263686618213e-07\ 7.9 4.43484494578837e-07\ 8 3.80455219911094e-07\ 8.1 3.25291008192431e-07\ 8.2 2.77172671369189e-07\ 8.3 2.3534460122561e-07\ 8.4 1.99112441137229e-07\ 8.5 1.67840550207634e-07\ 8.6 1.40949294485802e-07\ 8.7 1.17912200067898e-07\ 8.8 9.82530025128849e-08\ 8.9 8.15426261656431e-08\ 9 6.73961257130362e-08\ 9.1 5.54696206344678e-08\ 9.2 4.54572511927477e-08\ 9.3 3.70881822936694e-08\ 9.4 3.01236789781831e-08\ 9.5 2.43542745575844e-08\ 9.6 1.9597049519489e-08\ 9.7 1.56930363806752e-08\ 9.8 1.25047627010001e-08\ 9.9 9.91394155668026e-09\ 10 7.81931595335371e-09\ ]{}; table\[row sep=crcr\][ 0 1.73124693019867e+40\ 0.1 2.82436588029597e+38\ 0.2 4.1956828021137e+36\ 0.3 5.66414377915413e+34\ 0.4 6.93481071469053e+32\ 0.5 7.68443689192253e+30\ 0.6 7.69069752168299e+28\ 0.7 6.93724536906432e+26\ 0.8 5.62809550162749e+24\ 0.9 4.09798716677038e+22\ 1 2.67236018190065e+20\ 1.1 1.55746134335849e+18\ 1.2 8.09517185918063e+15\ 1.3 37447321526137.8\ 1.4 153856987861.805\ 1.5 560344207.658673\ 1.6 1805532.13078034\ 1.7 5137.85905686192\ 1.8 12.8900604152428\ 1.9 0.0284685779150781\ 2 5.54207410371086e-05\ 2.1 2.32449260849562e-07\ 2.2 1.32077340780486e-07\ 2.3 1.25849193008051e-07\ 2.4 1.19929697100271e-07\ 2.5 1.14178958144694e-07\ 2.6 1.08597197604576e-07\ 2.7 1.03184615816798e-07\ 2.8 9.79412052997195e-08\ 2.9 9.28667485616159e-08\ 3 8.79608167304609e-08\ 3.1 8.3222768820297e-08\ 3.2 7.86517516587704e-08\ 3.3 7.4246700498039e-08\ 3.4 7.00063403283588e-08\ 3.5 6.59291879104377e-08\ 3.6 6.20135545391339e-08\ 3.7 5.82575495472697e-08\ 3.8 5.4659084554252e-08\ 3.9 5.12158784598434e-08\ 4 4.79254631788606e-08\ 4.1 4.47851901077857e-08\ 4.2 4.17922373093195e-08\ 4.3 3.89436173958056e-08\ 4.4 3.62361860872561e-08\ 4.5 3.3666651414457e-08\ 4.6 3.12315835323765e-08\ 4.7 2.89274251038905e-08\ 4.8 2.67505022087345e-08\ 4.9 2.46970357276568e-08\ 5 2.27631531470211e-08\ 5.1 2.09449007246872e-08\ 5.2 1.92382559539098e-08\ 5.3 1.7639140258329e-08\ 5.4 1.6143431847929e-08\ 5.5 1.47469786631738e-08\ 5.6 1.34456113324526e-08\ 5.7 1.22351560665255e-08\ 5.8 1.11114474128994e-08\ 5.9 1.00703407930305e-08\ 6 9.10772474596058e-09\ 6.1 8.21953280347929e-09\ 6.2 7.40175492417926e-09\ 6.3 6.65044841681976e-09\ 6.4 5.96174828724651e-09\ 6.5 5.33187694769191e-09\ 6.6 4.7571532325753e-09\ 6.7 4.23400067088467e-09\ 6.8 3.7589549717927e-09\ 6.9 3.32867068726759e-09\ 7 2.93992702300214e-09\ 7.1 2.5896327769091e-09\ 7.2 2.27483039261599e-09\ 7.3 1.9926991237303e-09\ 7.4 1.7405573130212e-09\ 7.5 1.51586379895955e-09\ 7.6 1.31621847015486e-09\ 7.7 1.13936199600778e-09\ 7.8 9.83174769245422e-10\ 7.9 8.456751028147e-10\ 8 7.25016729775615e-10\ 8.1 6.19485660271951e-10\ 8.2 5.27496454282502e-10\ 8.3 4.47587972609562e-10\ 8.4 3.78418671394712e-10\ 8.5 3.18761507335779e-10\ 8.6 2.67498521700705e-10\ 8.7 2.23615171201044e-10\ 8.8 1.86194472824606e-10\ 8.9 1.54411027876639e-10\ 9 1.27524987801256e-10\ 9.1 1.04876020925093e-10\ 9.2 8.58773351715459e-11\ 9.3 7.00098071364568e-11\ 9.4 5.68162628027775e-11\ 9.5 4.5895949716867e-11\ 9.6 3.68992347693813e-11\ 9.7 2.95225559360986e-11\ 9.8 2.35036505521043e-11\ 9.9 1.86170770244034e-11\ 10 1.46700414326487e-11\ ]{}; Bound on the Minimum Distance and Expurgated Bounds --------------------------------------------------- Consider a TC ensemble consisting of $\Omega$ codes in total. The value $\Omega$ follows from the different possible combinations of permutations and depends on the type of the ensemble. Assume that all codes in the ensemble are selected with equal probability. Then, the number of codewords with weight $w$ over all possible codes in the ensemble is $\Omega\bar{A}_w$, where $\bar{A}_w$ is the average WEF of the ensemble. Therefore, given an integer value $\tilde{d}$, the total number of codewords with weight $w<\tilde{d}$ can be computed as $$\Omega_{w<\tilde{d}}=\Omega \sum^{\tilde{d}-1}_{w=1}\bar{A}_w\;.$$ By considering that these codewords are spread over different possible codes, we can obtain an upper bound on the number of codes with minimum distance $d_{\text{min}}\geq\tilde{d}$, $$\begin{aligned} \label{EX} \Omega_{w\geq\tilde{d}}<\Omega-\Omega \sum^{\tilde{d}-1}_{w=1}\bar{A}_w\;.\nonumber\end{aligned}$$ Let $\alpha$ denote the fraction of codes with $d_{\text{min}}\geq\tilde{d}$. Then, $\alpha$ is upperbounded by $$\begin{aligned} \alpha < 1-\sum^{\tilde{d}-1}_{w=1}\bar{A}_w\;.\end{aligned}$$ For a given $\alpha$ and $\bar{A}_w$, an analytical bound on the minimum distance of an ensemble can be obtained by computing the largest $\tilde{d}$ which satisfies . In fact, this bound shows that a fraction $\alpha$ of all possible codes has minimum distance $d_{\text{min}}\geq\tilde{d}$. In Fig. \[MinDis\], considering different classes of TCs with $R=1/3$, this bound is computed for $\alpha=0.5$ and different input block lengths. As it can be seen, the HCC ensemble has the best minimum distance, and the PCC ensemble the worst. As an example, for $K=300$ the values computed for HCCs, BCCs, SCCs, PCCs are $\tilde{d}=129$, $99$, $37$, and $10$, respectively. Comparing the results in Fig. \[MinDis\] and the thresholds in Table \[ThresholdsAWGN\], we can observe that the TC ensembles with good MAP threshold also have good minimum distance. According to Fig. \[MinDis\], for both the BCC and HCC ensembles, the minimum distance grows linearly with the input block length [@3DTurbo; @MoloudiISITA]. However, the bound on the minimum distance of the HCC ensemble has a higher slope and grows faster than that of the BCC ensemble. table\[row sep=crcr\][ 100 51\ 200 90\ 300 129\ 400 166\ 500 203\ 600 239\ 700 276\ 800 312\ 900 333\ ]{}; table\[row sep=crcr\][ 100 37\ 200 68\ 300 99\ 400 130\ 500 161\ 600 192\ 700 222\ 800 253\ 900 284\ 1000 315\ ]{}; table\[row sep=crcr\][ 4 7\ 24 9\ 44 9\ 64 9\ 84 10\ 104 10\ 124 10\ 144 10\ 164 10\ 184 10\ 204 10\ 224 10\ 244 10\ 264 10\ 284 10\ 304 10\ 324 10\ 344 10\ 364 10\ 384 10\ 404 10\ 424 10\ 444 10\ 464 10\ 484 10\ 504 10\ 524 10\ 544 10\ 564 10\ 584 10\ 604 10\ 624 10\ 644 10\ 664 10\ 684 10\ 704 10\ 724 10\ 744 10\ 764 10\ 784 10\ 804 10\ 824 10\ 844 10\ 864 10\ 884 10\ 904 10\ 924 10\ 944 10\ 964 10\ 984 10\ ]{}; table\[row sep=crcr\][ 100 22\ 200 31\ 300 37\ 400 43\ 500 48\ 600 53\ 700 57\ 800 62\ 900 66\ 1000 69\ ]{}; Consider excluding a fraction $(1-\alpha)$ of codes with $d_{\text{min}}<\tilde{d}$ from a TC ensemble. Then, it is possible to compute the upper bound on the performance of this expurgated ensemble. The average BER of the expurgated ensemble is upperbounded by $$P_b\leq \frac{1}{\alpha}\mathop{\sum_{i=1}^{kN}\; \sum_{w=\tilde{d}}^{nN}}\;\frac{i}{N} \bar{A}_{i,w}\; \mathsf{Q}\left ( \sqrt{2wR\frac{E_{\mathrm b}}{N_0}}\right).$$ The bounds for the expurgated TC ensembles are shown in Fig. \[EXBER\] for $\alpha=0.5$, which means that half of the codes with $d_{\text{min}}<\tilde{d}(\alpha)$ are excluded. For comparison, we also provide the corresponding union bounds in the same figure. It can be seen that for all TC ensembles except the PCC ensemble, the error floor estimated by the expurgated bound is much steeper and lower than that resulting from the union bound. In other words, expurgation improves the performance of the SCC, BCC, and HCC ensembles significantly. table\[row sep=crcr\][ 0 5.83479519485269e+33\ 0.1 6.13923485406036e+31\ 0.2 5.83908661389786e+29\ 0.3 5.01125108754083e+27\ 0.4 3.87404552987002e+25\ 0.5 2.69322011869566e+23\ 0.6 1.68101188804561e+21\ 0.7 9.40600973748269e+18\ 0.8 4.71162160334235e+16\ 0.9 211019517986770\ 1 844100668986.013\ 1.1 3013054533.82574\ 1.2 9591522.08511706\ 1.3 27219.692435055\ 1.4 68.861629285391\ 1.5 0.155346755149862\ 1.6 0.000312715158732166\ 1.7 5.62303460717369e-07\ 1.8 9.04437472937034e-10\ 1.9 1.30354720145149e-12\ 2 1.68690001553729e-15\ 2.1 1.96438783046855e-18\ 2.2 2.06324638782402e-21\ 2.3 1.95922554143366e-24\ 2.4 1.68590414311035e-27\ 2.5 1.31754028706751e-30\ 2.6 9.37128169086968e-34\ 2.7 6.07879052104575e-37\ 2.8 3.60302145164747e-40\ 2.9 1.95515809403535e-43\ 3 9.73192341646061e-47\ 3.1 4.45217197765867e-50\ 3.2 1.87580641831947e-53\ 3.3 7.29417308665927e-57\ 3.4 2.62371600784076e-60\ 3.5 8.75084274143692e-64\ 3.6 2.71319461946189e-67\ 3.7 7.84342400736308e-71\ 3.8 2.17002720632142e-74\ 3.9 1.61330912616837e-77\ 4 2.06553826053434e-79\ 4.1 3.57342160245694e-81\ 4.2 5.6663176860315e-83\ 4.3 8.1684451078017e-85\ 4.4 1.06805622576207e-86\ 4.5 1.26377665436323e-88\ 4.6 1.35005314352857e-90\ 4.7 1.29895811051223e-92\ 4.8 1.12289360949285e-94\ 4.9 8.6994566733263e-97\ 5 6.02478314057293e-99\ 5.1 3.72002993388994e-101\ 5.2 2.04239977055775e-103\ 5.3 9.94330922796276e-106\ 5.4 4.28051412043792e-108\ 5.5 1.62474920011768e-110\ 5.6 5.4215538371172e-113\ 5.7 1.58562408321707e-115\ 5.8 4.05207347843839e-118\ 5.9 9.01956371597361e-121\ 6 1.74310517795119e-123\ 6.1 2.91512803716518e-126\ 6.2 4.20455905998844e-129\ 6.3 5.21207606470582e-132\ 6.4 5.53340249436626e-135\ 6.5 5.01294602962834e-138\ 6.6 3.86104168159458e-141\ 6.7 2.51872439912962e-144\ 6.8 1.38623606885366e-147\ 6.9 6.41136644825279e-151\ 7 2.48174528965677e-154\ 7.1 8.00667648151286e-158\ 7.2 2.14382364570368e-161\ 7.3 4.74326894013434e-165\ 7.4 8.63345341852786e-169\ 7.5 1.2868579084214e-172\ 7.6 1.56347491423124e-176\ 7.7 1.54096708222797e-180\ 7.8 1.22607627463312e-184\ 7.9 7.83598228607289e-189\ 8 4.0022186543447e-193\ 8.1 1.6250525887397e-197\ 8.2 5.21756835186809e-202\ 8.3 1.3174182697501e-206\ 8.4 2.60135281704982e-211\ 8.5 3.99395683639993e-216\ 8.6 4.74008825081004e-221\ 8.7 4.32254417001507e-226\ 8.8 3.01018092257546e-231\ 8.9 1.59079376083509e-236\ 9 6.33880781040909e-242\ 9.1 1.8919640009518e-247\ 9.2 4.20147920523655e-253\ 9.3 6.89413688458471e-259\ 9.4 8.30004380157578e-265\ 9.5 7.27891589837369e-271\ 9.6 4.61560678144843e-277\ 9.7 2.10030250456712e-283\ 9.8 6.80556962841975e-290\ 9.9 1.55789086967966e-296\ 10 2.49907744585114e-303\ ]{}; table\[row sep=crcr\][ 0 5.19946636299096e+48\ 0.1 3.67142553079613e+46\ 0.2 2.36618501302388e+44\ 0.3 1.39049601765046e+42\ 0.4 7.44357856178082e+39\ 0.5 3.62648087549128e+37\ 0.6 1.60657109034683e+35\ 0.7 6.46644774608445e+32\ 0.8 2.3629069842394e+30\ 0.9 7.83307735565289e+27\ 1 2.35419374501037e+25\ 1.1 6.41102600853088e+22\ 1.2 1.58116139392232e+20\ 1.3 3.53033856593459e+17\ 1.4 713380528254042\ 1.5 1304411168908.47\ 1.6 2158122749.457\ 1.7 3231092.64665807\ 1.8 4378.73520130662\ 1.9 5.37363940066829\ 2 0.00597577489336794\ 2.1 6.0271807583264e-06\ 2.2 5.51994256935665e-09\ 2.3 4.59723137014595e-12\ 2.4 3.48816172377017e-15\ 2.5 2.41664879907572e-18\ 2.6 1.5329847081929e-21\ 2.7 8.9332920089934e-25\ 2.8 4.80150693509223e-28\ 2.9 2.39190300749717e-31\ 3 1.11087846209965e-34\ 3.1 4.84471163262622e-38\ 3.2 2.00168408171307e-41\ 3.3 7.92249008882283e-45\ 3.4 3.04659585672334e-48\ 3.5 1.15989672657066e-51\ 3.6 4.55868979992445e-55\ 3.7 5.45037222882743e-58\ 3.8 1.69785894590017e-59\ 3.9 7.42682258042216e-61\ 4 3.06104082760125e-62\ 4.1 1.17971565451355e-63\ 4.2 4.24049803961276e-65\ 4.3 1.41833254373475e-66\ 4.4 4.40446093538556e-68\ 4.5 1.26711331428014e-69\ 4.6 3.36986070927992e-71\ 4.7 8.26707133872978e-73\ 4.8 1.86681918886266e-74\ 4.9 3.87191086647739e-76\ 5 7.35992878865898e-78\ 5.1 1.27934637116072e-79\ 5.2 2.0290673181818e-81\ 5.3 2.92961648694204e-83\ 5.4 3.84172723218442e-85\ 5.5 4.5647929663537e-87\ 5.6 4.90289478870584e-89\ 5.7 4.74852597356404e-91\ 5.8 4.13672221225056e-93\ 5.9 3.23325178134048e-95\ 6 2.26141129962159e-97\ 6.1 1.41163934124018e-99\ 6.2 7.84322448602029e-102\ 6.3 3.86802433103734e-104\ 6.4 1.68841035307654e-106\ 6.5 6.50435921498645e-109\ 6.6 2.20487634976498e-111\ 6.7 6.55698146648514e-114\ 6.8 1.70537914185e-116\ 6.9 3.86689330689564e-119\ 7 7.61944894113244e-122\ 7.1 1.30037628848041e-124\ 7.2 1.91571048562996e-127\ 7.3 2.42774890030789e-130\ 7.4 2.6372664956819e-133\ 7.5 2.44685756451579e-136\ 7.6 1.93178914822759e-139\ 7.7 1.29289142348077e-142\ 7.8 7.30689252900021e-146\ 7.9 3.47336102735936e-149\ 8 1.38309536042692e-152\ 8.1 4.59448716831765e-156\ 8.2 1.26782911289723e-159\ 8.3 2.89358673667953e-163\ 8.4 5.43793536159262e-167\ 8.5 8.376801555602e-171\ 8.6 1.05280553233334e-174\ 8.7 1.07442484188567e-178\ 8.8 8.86025225829176e-183\ 8.9 5.87480397802201e-187\ 9 3.11603784873149e-191\ 9.1 1.3152474147846e-195\ 9.2 4.39428524632144e-200\ 9.3 1.15577114694776e-204\ 9.4 2.37974190945539e-209\ 9.5 3.81397120066802e-214\ 9.6 4.73011668215266e-219\ 9.7 4.51243979339962e-224\ 9.8 3.29104506936198e-229\ 9.9 1.82353828844933e-234\ 10 7.62723339431275e-240\ ]{}; table\[row sep=crcr\][ 0 1.6120311857813e+49\ 0.1 1.11548338086192e+47\ 0.2 7.04257642006771e+44\ 0.3 4.0527091666685e+42\ 0.4 2.12366348694636e+40\ 0.5 1.01239364873186e+38\ 0.6 4.38683747271313e+35\ 0.7 1.72634251502533e+33\ 0.8 6.16499310393967e+30\ 0.9 1.99640999745881e+28\ 1 5.85850245102297e+25\ 1.1 1.55697322667991e+23\ 1.2 3.74544928782908e+20\ 1.3 8.15195017551442e+17\ 1.4 1.60472294250766e+15\ 1.5 2856335249695.57\ 1.6 4596489810.9199\ 1.7 6687155.46391923\ 1.8 8796.34645077746\ 1.9 10.4645464982419\ 2 0.0112635970182631\ 2.1 1.0975669690812e-05\ 2.2 9.69005034569941e-09\ 2.3 7.7590301870022e-12\ 2.4 5.70764182303822e-15\ 2.5 3.67032505500189e-17\ 2.6 1.61974133500832e-17\ 2.7 7.83270119462798e-18\ 2.8 3.72881783978889e-18\ 2.9 1.74662266351181e-18\ 3 8.04685739933093e-19\ 3.1 3.64485398127067e-19\ 3.2 1.62249767057576e-19\ 3.3 7.09508700058834e-20\ 3.4 3.0466082351264e-20\ 3.5 1.28402107704624e-20\ 3.6 5.30922970366664e-21\ 3.7 2.15276941115872e-21\ 3.8 8.55595510211202e-22\ 3.9 3.33148786503905e-22\ 4 1.27026931725775e-22\ 4.1 4.74051452056113e-23\ 4.2 1.73063942120689e-23\ 4.3 6.17750954481503e-24\ 4.4 2.15484029495909e-24\ 4.5 7.34134079502365e-25\ 4.6 2.44147727542226e-25\ 4.7 7.92136083078206e-26\ 4.8 2.50590047975609e-26\ 4.9 7.72477820848361e-27\ 5 2.31900497211866e-27\ 5.1 6.77547834646448e-28\ 5.2 1.92541234778057e-28\ 5.3 5.31825942503594e-29\ 5.4 1.42687900679609e-29\ 5.5 3.71603501925323e-30\ 5.6 9.38736551519662e-31\ 5.7 2.29862507786088e-31\ 5.8 5.45173840384027e-32\ 5.9 1.25146557034817e-32\ 6 2.7783492091298e-33\ 6.1 5.96071044314405e-34\ 6.2 1.23482038513127e-34\ 6.3 2.46801218205785e-35\ 6.4 4.75514103362221e-36\ 6.5 8.8242627808218e-37\ 6.6 1.57583231452781e-37\ 6.7 2.70561996856849e-38\ 6.8 4.46219310000979e-39\ 6.9 7.06227930737044e-40\ 7 1.07161060193487e-40\ 7.1 1.55738330235159e-41\ 7.2 2.16561480766782e-42\ 7.3 2.87835670403569e-43\ 7.4 3.65279856920901e-44\ 7.5 4.42133877445022e-45\ 7.6 5.09855440845122e-46\ 7.7 5.59515971425709e-47\ 7.8 5.83640628426524e-48\ 7.9 5.78002573686136e-49\ 8 5.42796204728102e-50\ 8.1 4.82753897593406e-51\ 8.2 4.06109602900424e-52\ 8.3 3.22717703107475e-53\ 8.4 2.41927203530879e-54\ 8.5 1.70858176532004e-55\ 8.6 1.13519078661361e-56\ 8.7 7.0853720172004e-58\ 8.8 4.14840659056605e-59\ 8.9 2.27496179953589e-60\ 9 1.16674271649233e-61\ 9.1 5.5873101329023e-63\ 9.2 2.49436331868626e-64\ 9.3 1.03641135484552e-65\ 9.4 4.00119503540288e-67\ 9.5 1.43279783561412e-68\ 9.6 4.75064768477511e-70\ 9.7 1.45582998391759e-71\ 9.8 4.11582949094486e-73\ 9.9 1.07145325690782e-74\ 10 2.56341762458841e-76\ ]{}; table\[row sep=crcr\][ 0 2.16255348259374e+49\ 0.1 1.54123251200281e+47\ 0.2 1.00575121946955e+45\ 0.3 6.00639695678386e+42\ 0.4 3.2814535551403e+40\ 0.5 1.63957164078996e+38\ 0.6 7.49124618785127e+35\ 0.7 3.13011427831488e+33\ 0.8 1.19635460117273e+31\ 0.9 4.18473863917531e+28\ 1 1.3406753447313e+26\ 1.1 3.93831815717488e+23\ 1.2 1.06240829985288e+21\ 1.3 2.63719077753425e+18\ 1.4 6.03948987997895e+15\ 1.5 12803668380432.9\ 1.6 25236288225.2094\ 1.7 46503996.2581599\ 1.8 80696.5013460045\ 1.9 133.111365237132\ 2 0.211401432601363\ 2.1 0.000350126546946405\ 2.2 1.82693673475069e-05\ 2.3 1.48167486940589e-05\ 2.4 1.23636711307378e-05\ 2.5 1.03119420204021e-05\ 2.6 8.59241377929651e-06\ 2.7 7.15011292634068e-06\ 2.8 5.94017334581975e-06\ 2.9 4.92557922569737e-06\ 3 4.0755351188306e-06\ 3.1 3.36425498133596e-06\ 3.2 2.77004130274897e-06\ 3.3 2.27457063376989e-06\ 3.4 1.86232942201686e-06\ 3.5 1.52016158051501e-06\ 3.6 1.23690067107028e-06\ 3.7 1.00306727251744e-06\ 3.8 8.10617373980805e-07\ 3.9 6.52731312714899e-07\ 4 5.23635385733288e-07\ 4.1 4.18450140923433e-07\ 4.2 3.33060719159062e-07\ 4.3 2.6400562404114e-07\ 4.4 2.08381043224987e-07\ 4.5 1.63758406524887e-07\ 4.6 1.28113291764047e-07\ 4.7 9.97641158286835e-08\ 4.8 7.73193016756013e-08\ 4.9 5.96318110872112e-08\ 5 4.57600916168454e-08\ 5.1 3.49346145796602e-08\ 5.2 2.65292865868796e-08\ 5.3 2.0037105483476e-08\ 5.4 1.50495066634707e-08\ 5.5 1.12389105575664e-08\ 5.6 8.34403877352645e-09\ 5.7 6.15761649297486e-09\ 5.8 4.51612343188422e-09\ 5.9 3.2912957866349e-09\ 6 2.38311771987803e-09\ 6.1 1.71407353904309e-09\ 6.2 1.22446112438068e-09\ 6.3 8.68593638591565e-10\ 6.4 6.11740314428723e-10\ 6.5 4.27678358896191e-10\ 6.6 2.96746904676696e-10\ 6.7 2.04310644294952e-10\ 6.8 1.39555459406175e-10\ 6.9 9.45511640994576e-11\ 7 6.35275740357598e-11\ 7.1 4.23196482037617e-11\ 7.2 2.79455803458921e-11\ 7.3 1.82885921215516e-11\ 7.4 1.18589438175795e-11\ 7.5 7.61746784919928e-12\ 7.6 4.84587398687389e-12\ 7.7 3.05229227366276e-12\ 7.8 1.90311877182456e-12\ 7.9 1.17430868682089e-12\ 8 7.16908306724237e-13\ 8.1 4.3290812868179e-13\ 8.2 2.58500815456958e-13\ 8.3 1.52595871105821e-13\ 8.4 8.90258862909428e-14\ 8.5 5.13165834434265e-14\ 8.6 2.92172636151412e-14\ 8.7 1.64259827020403e-14\ 8.8 9.11590841600787e-15\ 8.9 4.99239709134236e-15\ 9 2.6972404454403e-15\ 9.1 1.43710766316505e-15\ 9.2 7.54870817106147e-16\ 9.3 3.9076984240923e-16\ 9.4 1.99288676827189e-16\ 9.5 1.00092640438462e-16\ 9.6 4.94903972208744e-17\ 9.7 2.40811283583621e-17\ 9.8 1.15266755173696e-17\ 9.9 5.42541816357987e-18\ 10 2.51010041225898e-18\ ]{}; table\[row sep=crcr\][ 0 1.18950384324866e-33\ 0.1 2.30615449671337e-34\ 0.2 4.30705375937924e-35\ 0.3 7.74234047735244e-36\ 0.4 1.3384017461128e-36\ 0.5 2.22298416442991e-37\ 0.6 3.5442656224782e-38\ 0.7 5.41940103864e-39\ 0.8 7.93957149787726e-40\ 0.9 1.11336979885566e-40\ 1 1.49294648667587e-41\ 1.1 1.91234842981144e-42\ 1.2 2.33750617783384e-43\ 1.3 2.72355162666594e-44\ 1.4 3.02161108905061e-45\ 1.5 3.18840301272468e-46\ 1.6 3.196240166682e-47\ 1.7 3.04035975206902e-48\ 1.8 2.74097709301399e-49\ 1.9 2.33907166404341e-50\ 2 1.88707513534634e-51\ 2.1 1.43740433385421e-52\ 2.2 1.03237082091195e-53\ 2.3 6.98181702219985e-55\ 2.4 4.43988613304187e-56\ 2.5 2.65109839298631e-57\ 2.6 1.48421063213323e-58\ 2.7 7.77912246215601e-60\ 2.8 3.81122925340261e-61\ 2.9 1.74267704212389e-62\ 3 7.42485018423727e-64\ 3.1 2.94281011932431e-65\ 3.2 1.08319893418848e-66\ 3.3 3.69637318649993e-68\ 3.4 1.16733975555512e-69\ 3.5 3.40554491528507e-71\ 3.6 9.16092095886201e-73\ 3.7 2.26793750286297e-74\ 3.8 5.15726199733905e-76\ 3.9 1.07507880609578e-77\ 4 2.05026916426625e-79\ 4.1 3.56965813270855e-81\ 4.2 5.66189727828599e-83\ 4.3 8.1633741381744e-85\ 4.4 1.06753058311844e-86\ 4.5 1.26328615367619e-88\ 4.6 1.3496421318047e-90\ 4.7 1.29864962421431e-92\ 4.8 1.12268675677297e-94\ 4.9 8.69822077946773e-97\ 5 6.02412696474684e-99\ 5.1 3.71972120596321e-101\ 5.2 2.04227141399023e-103\ 5.3 9.94283902185298e-106\ 5.4 4.28036279953993e-108\ 5.5 1.62470654876967e-110\ 5.6 5.42144887269553e-113\ 5.7 1.58560160053634e-115\ 5.8 4.05203170118498e-118\ 5.9 9.01949659256492e-121\ 6 1.74309588457536e-123\ 6.1 2.91511698805826e-126\ 6.2 4.20454781940739e-129\ 6.3 5.21206631534615e-132\ 6.4 5.53339531195507e-135\ 6.5 5.01294155234315e-138\ 6.6 3.86103932917621e-141\ 6.7 2.51872336151698e-144\ 6.8 1.38623568620333e-147\ 6.9 6.41136527335782e-151\ 7 2.48174499059022e-154\ 7.1 8.00667585315514e-158\ 7.2 2.14382353721844e-161\ 7.3 4.74326878693114e-165\ 7.4 8.63345324238662e-169\ 7.5 1.28685789201302e-172\ 7.6 1.56347490190735e-176\ 7.7 1.54096707480248e-180\ 7.8 1.22607627106234e-184\ 7.9 7.83598227244038e-189\ 8 4.00221865023487e-193\ 8.1 1.6250525877667e-197\ 8.2 5.21756835006923e-202\ 8.3 1.3174182694919e-206\ 8.4 2.60135281676376e-211\ 8.5 3.9939568361568e-216\ 8.6 4.74008825065249e-221\ 8.7 4.32254416993771e-226\ 8.8 3.01018092254687e-231\ 8.9 1.59079376082719e-236\ 9 6.33880781039288e-242\ 9.1 1.89196400094934e-247\ 9.2 4.20147920523383e-253\ 9.3 6.89413688458252e-259\ 9.4 8.30004380157451e-265\ 9.5 7.27891589837317e-271\ 9.6 4.61560678144828e-277\ 9.7 2.10030250456709e-283\ 9.8 6.80556962841971e-290\ 9.9 1.55789086967965e-296\ 10 2.49907744585113e-303\ ]{}; table\[row sep=crcr\][ 0 2.57541578402475e-25\ 0.1 6.79265259549529e-26\ 0.2 1.7380449116319e-26\ 0.3 4.31140545379372e-27\ 0.4 1.03612901288272e-27\ 0.5 2.41068629857872e-28\ 0.6 5.42610458656377e-29\ 0.7 1.18069771446604e-29\ 0.8 2.48181041459725e-30\ 0.9 5.03555831557572e-31\ 1 9.85460316550156e-32\ 1.1 1.85865963276373e-32\ 1.2 3.37582415950979e-33\ 1.3 5.89959424132829e-34\ 1.4 9.91203891092346e-35\ 1.5 1.59967360066548e-35\ 1.6 2.47770126135677e-36\ 1.7 3.6798474852713e-37\ 1.8 5.23576788880965e-38\ 1.9 7.13015570782312e-39\ 2 9.28486943518019e-40\ 2.1 1.15502730534722e-40\ 2.2 1.37126355156268e-41\ 2.3 1.55211926607345e-42\ 2.4 1.67323992589317e-43\ 2.5 1.7161830166145e-44\ 2.6 1.67291466004298e-45\ 2.7 1.54813938514716e-46\ 2.8 1.35856941285645e-47\ 2.9 1.12923943314722e-48\ 3 8.87986691741987e-50\ 3.1 6.59801155667545e-51\ 3.2 4.62659355272535e-52\ 3.3 3.05767662065644e-53\ 3.4 1.90207683622735e-54\ 3.5 1.1121891669453e-55\ 3.6 6.10427287951445e-57\ 3.7 3.14026737035771e-58\ 3.8 1.51192774481885e-59\ 3.9 6.80240590880004e-61\ 4 2.85546675627301e-62\ 4.1 1.1165312188412e-63\ 4.2 4.05994164515865e-65\ 4.3 1.3705009944069e-66\ 4.4 4.28730440745706e-68\ 4.5 1.24064781456779e-69\ 4.6 3.31485596585502e-71\ 4.7 8.16214372958866e-73\ 4.8 1.84849097318217e-74\ 4.9 3.84266507125983e-76\ 5 7.31739958716991e-78\ 5.1 1.27372359292675e-79\ 5.2 2.02232510359566e-81\ 5.3 2.92230213088542e-83\ 5.4 3.8345659704843e-85\ 5.5 4.55848131288582e-87\ 5.6 4.89789999184536e-89\ 5.7 4.74498621323765e-91\ 5.8 4.13448168826469e-93\ 5.9 3.23198861542529e-95\ 6 2.26077874572969e-97\ 6.1 1.41135878030503e-99\ 6.2 7.84212550654947e-102\ 6.3 3.86764527981583e-104\ 6.4 1.68829558065408e-106\ 6.5 6.50405508084114e-109\ 6.6 2.20480604126066e-111\ 6.7 6.55684012743264e-114\ 6.8 1.70535451587728e-116\ 6.9 3.86685624440664e-119\ 7 7.61940092515879e-122\ 7.1 1.30037095251599e-124\ 7.2 1.91570541750439e-127\ 7.3 2.42774480124513e-130\ 7.4 2.63726368324103e-133\ 7.5 2.44685593384916e-136\ 7.6 1.93178835241146e-139\ 7.7 1.29289109789196e-142\ 7.8 7.30689141691432e-146\ 7.9 3.47336071158206e-149\ 8 1.3830952862077e-152\ 8.1 4.59448702456344e-156\ 8.2 1.26782909005573e-159\ 8.3 2.89358670704389e-163\ 8.4 5.43793533034413e-167\ 8.5 8.3768015289543e-171\ 8.6 1.05280553050461e-174\ 8.7 1.07442484088085e-178\ 8.8 8.86025225389409e-183\ 8.9 5.87480397649712e-187\ 9 3.11603784831485e-191\ 9.1 1.31524741469539e-195\ 9.2 4.39428524617263e-200\ 9.3 1.15577114692853e-204\ 9.4 2.37974190943626e-209\ 9.5 3.81397120065345e-214\ 9.6 4.73011668214423e-219\ 9.7 4.51243979339593e-224\ 9.8 3.29104506936077e-229\ 9.9 1.82353828844903e-234\ 10 7.62723339431222e-240\ ]{}; table\[row sep=crcr\][ 0 2.6009146182617e-11\ 0.1 1.71657865720567e-11\ 0.2 1.12283466105641e-11\ 0.3 7.27776715942125e-12\ 0.4 4.67330912520809e-12\ 0.5 2.97239533961525e-12\ 0.6 1.87220726257135e-12\ 0.7 1.16754589604133e-12\ 0.8 7.20728690053096e-13\ 0.9 4.40301607511207e-13\ 1 2.66139746217674e-13\ 1.1 1.59128558357685e-13\ 1.2 9.40937552329903e-14\ 1.3 5.50098532225362e-14\ 1.4 3.17890244596961e-14\ 1.5 1.81534081462463e-14\ 1.6 1.02415987844464e-14\ 1.7 5.70673052567462e-15\ 1.8 3.13975215559137e-15\ 1.9 1.70516740848603e-15\ 2 9.13848906400505e-16\ 2.1 4.83154770211427e-16\ 2.2 2.51923147129994e-16\ 2.3 1.29503582231687e-16\ 2.4 6.56123894332708e-17\ 2.5 3.27517974012821e-17\ 2.6 1.61020361179059e-17\ 2.7 7.7941761310377e-18\ 2.8 3.71318794648862e-18\ 2.9 1.74040589448466e-18\ 3 8.02262768757516e-19\ 3.1 3.63560498445324e-19\ 3.2 1.61904156251609e-19\ 3.3 7.08245113651852e-20\ 3.4 3.04209049483302e-20\ 3.5 1.2824423576423e-20\ 3.6 5.30384056318301e-21\ 3.7 2.1509733417317e-21\ 3.8 8.5501143534078e-22\ 3.9 3.32963560123855e-22\ 4 1.26969682783662e-22\ 4.1 4.73879107365602e-23\ 4.2 1.73013438368904e-23\ 4.3 6.17606985822953e-24\ 4.4 2.15444131800276e-24\ 4.5 7.3402666234146e-25\ 4.6 2.44119650630341e-25\ 4.7 7.92064884418635e-26\ 4.8 2.50572544224891e-26\ 4.9 7.72436132939586e-27\ 5 2.31890885863294e-27\ 5.1 6.775263997291e-28\ 5.2 1.92536614342443e-28\ 5.3 5.31816323720156e-29\ 5.4 1.42685968364627e-29\ 5.5 3.7159975916142e-30\ 5.6 9.38729567725652e-31\ 5.7 2.29861253507576e-31\ 5.8 5.4517167412515e-32\ 5.9 1.25146197582143e-32\ 6 2.7783434840708e-33\ 6.1 5.96070169919743e-34\ 6.2 1.23481910575886e-34\ 6.3 2.46801039058125e-35\ 6.4 4.75513863532781e-36\ 6.5 8.82425971453245e-37\ 6.6 1.57583194052534e-37\ 6.7 2.7056195338511e-38\ 6.8 4.46219261903754e-39\ 6.9 7.06227880141606e-40\ 7 1.07161055139115e-40\ 7.1 1.55738325445975e-41\ 7.2 2.16561476467892e-42\ 7.3 2.87835666752662e-43\ 7.4 3.65279853991145e-44\ 7.5 4.42133875226461e-45\ 7.6 5.09855439261941e-46\ 7.7 5.59515970362534e-47\ 7.8 5.83640627755596e-48\ 7.9 5.78002573288842e-49\ 8 5.42796204507675e-50\ 8.1 4.82753897478994e-51\ 8.2 4.06109602844954e-52\ 8.3 3.22717703082395e-53\ 8.4 2.41927203520321e-54\ 8.5 1.70858176527873e-55\ 8.6 1.13519078659862e-56\ 8.7 7.08537201714997e-58\ 8.8 4.14840659055037e-59\ 8.9 2.27496179953139e-60\ 9 1.16674271649114e-61\ 9.1 5.58731013289941e-63\ 9.2 2.49436331868562e-64\ 9.3 1.03641135484539e-65\ 9.4 4.00119503540264e-67\ 9.5 1.43279783561408e-68\ 9.6 4.75064768477505e-70\ 9.7 1.45582998391758e-71\ 9.8 4.11582949094485e-73\ 9.9 1.07145325690782e-74\ 10 2.56341762458841e-76\ ]{}; table\[row sep=crcr\][ 0 0.000348335795198516\ 0.1 0.000309520520929779\ 0.2 0.000274517458942697\ 0.3 0.000243013837331099\ 0.4 0.000214715623156207\ 0.5 0.00018934712811598\ 0.6 0.000166650544830509\ 0.7 0.000146385421580615\ 0.8 0.00012832808348894\ 0.9 0.000112271008178121\ 1 9.80221638826874e-05\ 1.1 8.54043178358623e-05\ 1.2 7.42543225064377e-05\ 1.3 6.44223869331697e-05\ 1.4 5.57713400043822e-05\ 1.5 4.81758920695288e-05\ 1.6 4.15219007587266e-05\ 1.7 3.57056463375095e-05\ 1.8 3.06331213490978e-05\ 1.9 2.62193387069086e-05\ 2 2.23876618069726e-05\ 2.1 1.90691596436986e-05\ 2.2 1.6201989342471e-05\ 2.3 1.3730807977158e-05\ 2.4 1.16062150267924e-05\ 2.5 9.78422634919442e-06\ 2.6 8.22578011448883e-06\ 2.7 6.89627475134197e-06\ 2.8 5.76513861501098e-06\ 2.9 4.80543078945266e-06\ 3 3.99347218516878e-06\ 3.1 3.30850588859996e-06\ 3.2 2.73238555534659e-06\ 3.3 2.24929051527679e-06\ 3.4 1.84546616919541e-06\ 3.5 1.50898820039715e-06\ 3.6 1.22954909616284e-06\ 3.7 9.9826547010733e-07\ 3.8 8.07504692380689e-07\ 3.9 6.50729367384871e-07\ 4 5.22358244489993e-07\ 4.1 4.17642203139109e-07\ 4.2 3.32554016991878e-07\ 4.3 2.63690670021494e-07\ 4.4 2.0818706875647e-07\ 4.5 1.6364006750937e-07\ 4.6 1.28041796141075e-07\ 4.7 9.97213516427582e-08\ 4.8 7.72939847998844e-08\ 4.9 5.9616980866161e-08\ 5 4.57514981259592e-08\ 5.1 3.49296902036715e-08\ 5.2 2.65264968350277e-08\ 5.3 2.00355434732278e-08\ 5.4 1.5048642541416e-08\ 5.5 1.12384383794265e-08\ 5.6 8.34378400809923e-09\ 5.7 6.15748080522081e-09\ 5.8 4.51605211942056e-09\ 5.9 3.29125881478495e-09\ 6 2.38309881772143e-09\ 6.1 1.71406401247467e-09\ 6.2 1.22445639293286e-09\ 6.3 8.68591323712195e-10\ 6.4 6.11739199157539e-10\ 6.5 4.2767782997848e-10\ 6.6 2.96746657853767e-10\ 6.7 2.04310531001882e-10\ 6.8 1.39555408277013e-10\ 6.9 9.45511414214572e-11\ 7 6.35275641540752e-11\ 7.1 4.23196439754911e-11\ 7.2 2.7945578570021e-11\ 7.3 1.82885913897655e-11\ 7.4 1.18589435218575e-11\ 7.5 7.61746773205847e-12\ 7.6 4.84587394141155e-12\ 7.7 3.05229225638436e-12\ 7.8 1.90311876539698e-12\ 7.9 1.17430868448172e-12\ 8 7.16908305891851e-13\ 8.1 4.32908128392321e-13\ 8.2 2.58500815358635e-13\ 8.3 1.52595871073219e-13\ 8.4 8.90258862803958e-14\ 8.5 5.13165834400995e-14\ 8.6 2.92172636141185e-14\ 8.7 1.64259827017341e-14\ 8.8 9.11590841591866e-15\ 8.9 4.99239709131708e-15\ 9 2.69724044543334e-15\ 9.1 1.43710766316318e-15\ 9.2 7.54870817105663e-16\ 9.3 3.90769842409109e-16\ 9.4 1.9928867682716e-16\ 9.5 1.00092640438455e-16\ 9.6 4.94903972208727e-17\ 9.7 2.40811283583617e-17\ 9.8 1.15266755173695e-17\ 9.9 5.42541816357986e-18\ 10 2.51010041225898e-18\ ]{}; table\[row sep=crcr\][ 3 1.19660507639199e-13\ 3.5 4.70175065748866e-14\ 4 1.77365416759004e-14\ 4.5 6.52847459903677e-15\ 5 2.39004220707178e-15\ 5.5 8.8098829173984e-16\ 6 3.25419898999524e-16\ 6.5 1.17852273373363e-16\ 7 4.06311651881909e-17\ 7.5 1.29633042514568e-17\ 8 3.73478307339253e-18\ 8.5 9.50498501514067e-19\ 9 2.09133395469223e-19\ 9.5 3.88990075009214e-20\ 10 5.97111341950619e-21\ ]{}; table\[row sep=crcr\][ 3 2.93953470547403e-10\ 3.5 1.13727092537031e-10\ 4 4.12601938801283e-11\ 4.5 1.39472738324373e-11\ 5 4.37034079667638e-12\ 5.5 1.26699875563662e-12\ 6 3.41275332724769e-13\ 6.5 8.65669020861838e-14\ 7 2.1140547829608e-14\ 7.5 5.07123057996476e-15\ 8 1.19447226320358e-15\ 8.5 2.68570023713636e-16\ 9 5.52230355494405e-17\ 9.5 9.966986417699e-18\ 10 1.52837105335438e-18\ ]{}; table\[row sep=crcr\][ 3 9.00738329690746e-08\ 3.5 6.75113023828472e-08\ 4 4.90747366482262e-08\ 4.5 3.44736645793382e-08\ 5 2.33087060836597e-08\ 5.5 1.51004093991258e-08\ 6 9.32602574682822e-09\ 6.5 5.45969640076215e-09\ 7 3.01041816253999e-09\ 7.5 1.55221664315933e-09\ 8 7.42406712667959e-10\ 8.5 3.26408306669519e-10\ 9 1.30584561676127e-10\ 9.5 4.69971885955196e-11\ 10 1.50220642298326e-11\ ]{}; table\[row sep=crcr\][ 3 2.28997965256643e-06\ 3.5 8.7723490805736e-07\ 4 3.13022560386828e-07\ 4.5 1.02719015394744e-07\ 5 3.07065281936111e-08\ 5.5 8.30744199352229e-09\ 6 2.02766522230175e-09\ 6.5 4.46769952987495e-10\ 7 8.91130630934014e-11\ 7.5 1.60867171341276e-11\ 8 2.60000773307729e-12\ 8.5 3.66708746363477e-13\ 9 4.34702279292935e-14\ 9.5 4.1502466372239e-15\ 10 3.05753288238676e-16\ ]{}; (axis cs:1,4.6) ellipse \[x radius=10, y radius=140\]; (axis cs:5,5.4) ellipse \[x radius=10, y radius=50\]; at (rel axis cs:0.01,0.5) [Expurgated Bounds]{}; at (rel axis cs:0.615,0.83) [Union Bounds]{}; For the BCC and HCC ensembles, the gap between the expurgated bound and the union bound is very large and notable. To investigate the influence of expurgation on the performance of these ensembles, in Fig. \[EXBERBCCHCC\] we provide the expurgated bound on the BER of the BCC and HCC ensembles for $\alpha=0.5$ and $\alpha=0.99$. Note that for $\alpha=0.99$, the expurgated bounds are computed by excluding only $1\%$ of the possible codes, and these bounds are still significantly lower and steeper than the corresponding union bounds. For the BCC ensemble, the gap between the expurgated bounds for $\alpha=0.5$ and $\alpha=0.99$ is much smaller than that of the HCC ensemble. Therefore, for $\alpha=0.99$, the BCC ensemble has slightly steeper and lower error floor than the HCC ensemble. The fact that changing $\alpha$ has a little impact on the expurgation of the BCC ensemble suggests that only a small fraction of the codes have poor distance properties. This means that for a BCC with randomly selected but fixed permutations, with high probability the error floor is as steep and low as the corresponding expurgated bound for an ML decoder. table\[row sep=crcr\][ 0 2.59973318149547e+48\ 0.1 1.83571276539807e+46\ 0.2 1.18309250651194e+44\ 0.3 6.95248008825227e+41\ 0.4 3.72178928089043e+39\ 0.5 1.81324043774564e+37\ 0.6 8.03285545173413e+34\ 0.7 3.23322387304222e+32\ 0.8 1.1814534921197e+30\ 0.9 3.91653867782643e+27\ 1 1.17709687250519e+25\ 1.1 3.20551300426543e+22\ 1.2 7.90580696961158e+19\ 1.3 1.76516928296729e+17\ 1.4 356690264127020\ 1.5 652205584454.239\ 1.6 1079061374.72849\ 1.7 1615546.3233292\ 1.8 2189.36760081126\ 1.9 2.686819851531\ 2 0.00298803205798424\ 2.1 3.15178109309571e-06\ 2.2 1.34696253817189e-07\ 2.3 1.25851306426461e-07\ 2.4 1.19929698630987e-07\ 2.5 1.14178958145684e-07\ 2.6 1.08597197604576e-07\ 2.7 1.03184615816798e-07\ 2.8 9.79412052997195e-08\ 2.9 9.28667485616159e-08\ 3 8.79608167304609e-08\ 3.1 8.3222768820297e-08\ 3.2 7.86517516587704e-08\ 3.3 7.4246700498039e-08\ 3.4 7.00063403283588e-08\ 3.5 6.59291879104377e-08\ 3.6 6.20135545391339e-08\ 3.7 5.82575495472697e-08\ 3.8 5.4659084554252e-08\ 3.9 5.12158784598434e-08\ 4 4.79254631788606e-08\ 4.1 4.47851901077857e-08\ 4.2 4.17922373093195e-08\ 4.3 3.89436173958056e-08\ 4.4 3.62361860872561e-08\ 4.5 3.3666651414457e-08\ 4.6 3.12315835323765e-08\ 4.7 2.89274251038905e-08\ 4.8 2.67505022087345e-08\ 4.9 2.46970357276568e-08\ 5 2.27631531470211e-08\ 5.1 2.09449007246872e-08\ 5.2 1.92382559539098e-08\ 5.3 1.7639140258329e-08\ 5.4 1.6143431847929e-08\ 5.5 1.47469786631738e-08\ 5.6 1.34456113324526e-08\ 5.7 1.22351560665255e-08\ 5.8 1.11114474128994e-08\ 5.9 1.00703407930305e-08\ 6 9.10772474596058e-09\ 6.1 8.21953280347929e-09\ 6.2 7.40175492417926e-09\ 6.3 6.65044841681976e-09\ 6.4 5.96174828724651e-09\ 6.5 5.33187694769191e-09\ 6.6 4.7571532325753e-09\ 6.7 4.23400067088467e-09\ 6.8 3.7589549717927e-09\ 6.9 3.32867068726759e-09\ 7 2.93992702300214e-09\ 7.1 2.5896327769091e-09\ 7.2 2.27483039261599e-09\ 7.3 1.9926991237303e-09\ 7.4 1.7405573130212e-09\ 7.5 1.51586379895955e-09\ 7.6 1.31621847015486e-09\ 7.7 1.13936199600778e-09\ 7.8 9.83174769245422e-10\ 7.9 8.456751028147e-10\ 8 7.25016729775615e-10\ 8.1 6.19485660271951e-10\ 8.2 5.27496454282502e-10\ 8.3 4.47587972609562e-10\ 8.4 3.78418671394712e-10\ 8.5 3.18761507335779e-10\ 8.6 2.67498521700705e-10\ 8.7 2.23615171201044e-10\ 8.8 1.86194472824606e-10\ 8.9 1.54411027876639e-10\ 9 1.27524987801256e-10\ 9.1 1.04876020925093e-10\ 9.2 8.58773351715459e-11\ 9.3 7.00098071364568e-11\ 9.4 5.68162628027775e-11\ 9.5 4.5895949716867e-11\ 9.6 3.68992347693813e-11\ 9.7 2.95225559360986e-11\ 9.8 2.35036505521043e-11\ 9.9 1.86170770244034e-11\ 10 1.46700414326487e-11\ ]{}; table\[row sep=crcr\][ 0 5.83479519485269e+33\ 0.5 2.69322011869566e+23\ 1 844100668986.013\ 1.5 0.155346755149862\ 2 1.68690001553729e-15\ 2.5 1.31754028706751e-30\ 3 9.73192341646061e-47\ 3.5 8.75084274143692e-64\ 4 2.06553826053434e-79\ 4.5 1.26377665436323e-88\ 5 6.02478314057293e-99\ 5.5 1.62474920011768e-110\ 6 1.74310517795119e-123\ 6.5 5.01294602962834e-138\ 7 2.48174528965677e-154\ 7.5 1.2868579084214e-172\ 8 4.0022186543447e-193\ 8.5 3.99395683639993e-216\ 9 6.33880781040909e-242\ 9.5 7.27891589837369e-271\ 10 2.49907744585114e-303\ ]{}; table\[row sep=crcr\][ 0 2.91739759742634e+33\ 0.5 1.34661005934783e+23\ 1 422050334493.007\ 1.5 0.0776733775763748\ 2 6.65040171986944e-13\ 2.5 2.89607283990355e-13\ 3 1.19660507639199e-13\ 3.5 4.70175065748866e-14\ 4 1.77365416759004e-14\ 4.5 6.52847459903677e-15\ 5 2.39004220707178e-15\ 5.5 8.8098829173984e-16\ 6 3.25419898999524e-16\ 6.5 1.17852273373363e-16\ 7 4.06311651881909e-17\ 7.5 1.29633042514568e-17\ 8 3.73478307339253e-18\ 8.5 9.50498501514067e-19\ 9 2.09133395469223e-19\ 9.5 3.88990075009214e-20\ 10 5.97111341950619e-21\ ]{}; table\[row sep=crcr\][ 0 2.91739759742634e+33\ 0.5 1.34661005934783e+23\ 1 422050334493.007\ 1.5 0.0776733775763748\ 2 6.65040171986944e-13\ 2.5 2.89607283990355e-13\ 3 1.19660507639199e-13\ 3.5 4.70175065748866e-14\ 4 1.77365416759004e-14\ 4.5 6.52847459903677e-15\ 5 2.39004220707178e-15\ 5.5 8.8098829173984e-16\ 6 3.25419898999524e-16\ 6.5 1.17852273373363e-16\ 7 4.06311651881909e-17\ 7.5 1.29633042514568e-17\ 8 3.73478307339253e-18\ 8.5 9.50498501514067e-19\ 9 2.09133395469223e-19\ 9.5 3.88990075009214e-20\ 10 5.97111341950619e-21\ ]{}; table\[row sep=crcr\][ 0 5.19946636299096e+48\ 0.1 3.67142553079613e+46\ 0.2 2.36618501302388e+44\ 0.3 1.39049601765046e+42\ 0.4 7.44357856178082e+39\ 0.5 3.62648087549128e+37\ 0.6 1.60657109034683e+35\ 0.7 6.46644774608445e+32\ 0.8 2.3629069842394e+30\ 0.9 7.83307735565289e+27\ 1 2.35419374501037e+25\ 1.1 6.41102600853088e+22\ 1.2 1.58116139392232e+20\ 1.3 3.53033856593459e+17\ 1.4 713380528254042\ 1.5 1304411168908.47\ 1.6 2158122749.457\ 1.7 3231092.64665807\ 1.8 4378.73520130662\ 1.9 5.37363940066829\ 2 0.00597577489336794\ 2.1 6.0271807583264e-06\ 2.2 5.51994256935665e-09\ 2.3 4.59723137014595e-12\ 2.4 3.48816172377017e-15\ 2.5 2.41664879907572e-18\ 2.6 1.5329847081929e-21\ 2.7 8.9332920089934e-25\ 2.8 4.80150693509223e-28\ 2.9 2.39190300749717e-31\ 3 1.11087846209965e-34\ 3.1 4.84471163262622e-38\ 3.2 2.00168408171307e-41\ 3.3 7.92249008882283e-45\ 3.4 3.04659585672334e-48\ 3.5 1.15989672657066e-51\ 3.6 4.55868979992445e-55\ 3.7 5.45037222882743e-58\ 3.8 1.69785894590017e-59\ 3.9 7.42682258042216e-61\ 4 3.06104082760125e-62\ 4.1 1.17971565451355e-63\ 4.2 4.24049803961276e-65\ 4.3 1.41833254373475e-66\ 4.4 4.40446093538556e-68\ 4.5 1.26711331428014e-69\ 4.6 3.36986070927992e-71\ 4.7 8.26707133872978e-73\ 4.8 1.86681918886266e-74\ 4.9 3.87191086647739e-76\ 5 7.35992878865898e-78\ 5.1 1.27934637116072e-79\ 5.2 2.0290673181818e-81\ 5.3 2.92961648694204e-83\ 5.4 3.84172723218442e-85\ 5.5 4.5647929663537e-87\ 5.6 4.90289478870584e-89\ 5.7 4.74852597356404e-91\ 5.8 4.13672221225056e-93\ 5.9 3.23325178134048e-95\ 6 2.26141129962159e-97\ 6.1 1.41163934124018e-99\ 6.2 7.84322448602029e-102\ 6.3 3.86802433103734e-104\ 6.4 1.68841035307654e-106\ 6.5 6.50435921498645e-109\ 6.6 2.20487634976498e-111\ 6.7 6.55698146648514e-114\ 6.8 1.70537914185e-116\ 6.9 3.86689330689564e-119\ 7 7.61944894113244e-122\ 7.1 1.30037628848041e-124\ 7.2 1.91571048562996e-127\ 7.3 2.42774890030789e-130\ 7.4 2.6372664956819e-133\ 7.5 2.44685756451579e-136\ 7.6 1.93178914822759e-139\ 7.7 1.29289142348077e-142\ 7.8 7.30689252900021e-146\ 7.9 3.47336102735936e-149\ 8 1.38309536042692e-152\ 8.1 4.59448716831765e-156\ 8.2 1.26782911289723e-159\ 8.3 2.89358673667953e-163\ 8.4 5.43793536159262e-167\ 8.5 8.376801555602e-171\ 8.6 1.05280553233334e-174\ 8.7 1.07442484188567e-178\ 8.8 8.86025225829176e-183\ 8.9 5.87480397802201e-187\ 9 3.11603784873149e-191\ 9.1 1.3152474147846e-195\ 9.2 4.39428524632144e-200\ 9.3 1.15577114694776e-204\ 9.4 2.37974190945539e-209\ 9.5 3.81397120066802e-214\ 9.6 4.73011668215266e-219\ 9.7 4.51243979339962e-224\ 9.8 3.29104506936198e-229\ 9.9 1.82353828844933e-234\ 10 7.62723339431275e-240\ ]{}; table\[row sep=crcr\][ 0 2.94686626002661e+33\ 0.5 1.36021218115943e+23\ 1 426313469184.855\ 1.5 0.0784579571463949\ 2 8.51969704816813e-16\ 2.5 6.65424387407849e-31\ 3 4.95595542986756e-47\ 3.5 3.00669711153673e-54\ 4 5.34922201174505e-60\ 4.5 1.93287072526881e-66\ 5 1.16703618844879e-73\ 5.5 9.45867937507813e-82\ 6 8.04827532774767e-91\ 6.5 5.45638239597988e-101\ 7 2.16265438225738e-112\ 7.5 3.54047943875019e-125\ 8 1.62108027911558e-139\ 8.5 1.34033440199667e-155\ 9 1.2249032575657e-173\ 9.5 7.13321663833151e-194\ 10 1.42699254658577e-216\ ]{}; table\[row sep=crcr\][ 0 2.6259931126217e+48\ 0.1 1.8542553185839e+46\ 0.2 1.19504293587064e+44\ 0.3 7.02270715985079e+41\ 0.4 3.75938311201052e+39\ 0.5 1.83155599772287e+37\ 0.6 8.11399540579206e+34\ 0.7 3.26588270004265e+32\ 0.8 1.19338736577748e+30\ 0.9 3.95609967457217e+27\ 1 1.18898673990423e+25\ 1.1 3.23789192350044e+22\ 1.2 7.98566360566827e+19\ 1.3 1.78299927572454e+17\ 1.4 360293196087900\ 1.5 658793519650.743\ 1.6 1089960984.57424\ 1.7 1631864.97305963\ 1.8 2211.48242490233\ 1.9 2.71395929326682\ 2 0.0030180681279636\ 2.1 3.04403068602344e-06\ 2.2 2.78784978250336e-09\ 2.3 2.32183402532624e-12\ 2.4 1.76169784028796e-15\ 2.5 1.22052969650289e-18\ 2.6 7.74234701107527e-22\ 2.7 4.51176364090576e-25\ 2.8 2.42500350257183e-28\ 2.9 1.20803182196827e-31\ 3 5.61049728333157e-35\ 3.1 2.44682405688265e-38\ 3.2 1.01095155647934e-41\ 3.3 4.00125766548556e-45\ 3.4 1.53868801706693e-48\ 3.5 5.86026987996207e-52\ 3.6 2.44359553725945e-55\ 3.7 1.12492092002161e-57\ 3.8 5.6534897814387e-59\ 3.9 2.91138557959629e-60\ 4 1.40934726140697e-61\ 4.1 6.39391744260495e-63\ 4.2 2.7128139239879e-64\ 4.3 1.07420582946718e-65\ 4.4 3.96185275265119e-67\ 4.5 1.35828165653249e-68\ 4.6 4.32017530957386e-70\ 4.7 1.27224158341331e-71\ 4.8 3.46198290681578e-73\ 4.9 8.68738079966121e-75\ 5 2.0061886103237e-76\ 5.1 4.25472058156585e-78\ 5.2 8.26931753192832e-80\ 5.3 1.46972071113974e-81\ 5.4 2.38350856553098e-83\ 5.5 3.51922973637856e-85\ 5.6 4.71998701925826e-87\ 5.7 5.73706414878253e-89\ 5.8 6.30473819824547e-91\ 5.9 6.24919513811073e-93\ 6 5.57300888296419e-95\ 6.1 4.4603599182391e-97\ 6.2 3.19555802130836e-99\ 6.3 2.04398515613227e-101\ 6.4 1.16411353997294e-103\ 6.5 5.88716067115364e-106\ 6.6 2.63626493971832e-108\ 6.7 1.0423104206004e-110\ 6.8 3.62788946415887e-113\ 6.9 1.1082943786413e-115\ 7 2.9625511393477e-118\ 7.1 6.90750701923695e-121\ 7.2 1.40031422074024e-123\ 7.3 2.46008616373596e-126\ 7.4 3.73280244111642e-129\ 7.5 4.87511886732901e-132\ 7.6 5.46100019367182e-135\ 7.7 5.22795680094225e-138\ 7.8 4.26152030746426e-141\ 7.9 2.94667959478473e-144\ 8 1.72171879059483e-147\ 8.1 8.46718203043003e-151\ 8.2 3.49067727938952e-154\ 8.3 1.20138101070172e-157\ 8.4 3.43729370070095e-161\ 8.5 8.14027961192663e-165\ 8.6 1.58864630414366e-168\ 8.7 2.54339686605876e-172\ 8.8 3.32497374567648e-176\ 8.9 3.53257104401947e-180\ 9 3.03539894247202e-184\ 9.1 2.09897605345261e-188\ 9.2 1.16214565879542e-192\ 9.3 5.12528819194477e-197\ 9.4 1.79090051767007e-201\ 9.5 4.9312518068823e-206\ 9.6 1.06403911350375e-210\ 9.7 1.78895170309922e-215\ 9.8 2.32996082629877e-220\ 9.9 2.33677601749756e-225\ 10 1.79371419822568e-230\ ]{}; at (rel axis cs:0.615,0.48) [$\alpha=0.99$]{}; (70,19) – (70,13); (70,19) – (85,19); (70,43) – (70,35); (70,43) – (80,25); at (rel axis cs:0.615,0.25) [$\alpha=0.5$]{}; (axis cs:0.712656,0.875312) ellipse \[x radius=0.0197095, y radius=0.0548628\]; at (rel axis cs:0.55,0.83) [Union Bounds]{}; Discussion and Conclusion {#con} ========================= We investigated the performance of SC-TC ensembles with finite block length in both waterfall and error floor regions. The two primary results can be summarized as follows. First, spatial coupling not only improves the asymptotic decoding threshold of an ensemble but also, for finite length and given latency, it improves the performance of the ensemble in the waterfall region. Second, considering certain conditions, spatial coupling either improves or preserves the minimum distance of the ensemble. Therefore, the coupled ensembles cannot have worse error floor than the corresponding uncoupled TC ensembles. Using this fact, instead of performing the cumbersome WEF analysis for the coupled ensemble, we derived the WEF for uncoupled ensembles. Then, based on the WEF, we computed bounds on BER performance and the minimum distance of TCs. As the coupled ensembles have equal or larger minimum distance than the uncoupled ensembles, the computed bounds for TCs can be used to approximate the error floor of SC-TCs. The results from the WEF analysis of TCs demonstrate very low error floors for SCC, BCC, and HCC esnembles. Moreover, for the BCC and HCC ensembles, the minimum distance grows linearly with block length. Comparing the thresholds of SC-TC ensembles (see Table \[ThresholdsAWGN\]) and the results from the WEF analysis, we observe that the ensembles with better MAP thresholds also have larger minimum distance and lower error floor. However, so far, only PCCs have been widely used in various standards—such as UMTS and LTE— because of their good BP thresholds and good performance in the waterfall region. Other TC ensembles have received much less attention for commercial use, although they have better MAP threshold and distance properties than PCCs. Our results confirm that the BP thresholds of these ensembles can be significantly improved by applying coupling. Also, regarding the finite length regime, while their error floor stays at very low error probabilities, their waterfall performance gets much closer to capacity. This brings us to the conclusion that by coupling a given ensemble with close to capacity MAP threshold and low error floor, such as SCCs, BCCs, and HCCs, the resulting ensemble is very promising and can perform close-to-capacity, yet achieving low error floor, with a low complexity iterative decoder. Finally, we should remark that the considered bounds estimate the error floor of an ML decoder. To obtain bounds on the performance of the BP decoder, more investigations on the corresponding absorbing sets [@Mitchell_AS2014] and pseudo-codewords [@RosnesPseudoweight] need to be done. Proof of [*Theorem 2*]{} ======================== We prove the theorem for the general case of tailbiting. Serially Concatenated Codes --------------------------- Consider the SCC and SC-SCC ensembles in Fig. \[CG\](a) and (b), and assume that $$\Pi^{\text{Un}}=\Pi^{(1)}\cdot\Pi^{(2)}.$$ Any codeword ${\ensuremath{\boldsymbol{v}}}\in\mathcal{C}$ satisfies the local constraints for $t=1,\dots,L$. Therefore, at time $t$, $$\begin{aligned} \label{SCCProof1} &\big({\ensuremath{\boldsymbol{u}}}_t\;\; {\ensuremath{\boldsymbol{v}}}_t^{\text{O}}\big)\cdot {\ensuremath{\boldsymbol{H}}}_{\text{O}}^T={\ensuremath{\boldsymbol{0}}} \; ,\\ \label{SCCProof2} &\big((\tilde{{\ensuremath{\boldsymbol{v}}}}'_{t,0}\;\;\; \tilde{{\ensuremath{\boldsymbol{v}}}}'_{t-1,1}) \cdot \Pi^{(2)} \;\;{\ensuremath{\boldsymbol{v}}}_t^{\text{I}}\big)\cdot {\ensuremath{\boldsymbol{H}}}_{\text{I}}^T={\ensuremath{\boldsymbol{0}}} \; , \end{aligned}$$ where $\tilde{{\ensuremath{\boldsymbol{v}}}}'_t=({\ensuremath{\boldsymbol{u}}}_t,{\ensuremath{\boldsymbol{v}}}_t^{\text{O}})\cdot \Pi^{(1)}$. The constraints are linear and time-invariant. Therefore, any superposition of the vectors $\big({\ensuremath{\boldsymbol{u}}}_t\;\; {\ensuremath{\boldsymbol{v}}}_t^{\text{O}}\big)$ and $\big((\tilde{{\ensuremath{\boldsymbol{v}}}}'_{t-1,1}\;\;\; \tilde{{\ensuremath{\boldsymbol{v}}}}'_{t,2}) \cdot \Pi^{(2)} \;\;{\ensuremath{\boldsymbol{v}}}_t^{\text{I}}\big)$ from different time slots $t=1,\dots,L$, satisfy and , respectively. In particular, we can consider $$\begin{aligned} \label{SCCProof3} \sum_{t=1}^{L}\big({\ensuremath{\boldsymbol{u}}}_t\;\; {\ensuremath{\boldsymbol{v}}}_t^{\text{O}}\big)=\big(\sum_{t=1}^{L}{\ensuremath{\boldsymbol{u}}}_t\;\; \sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}_t^{\text{O}}\big), \end{aligned}$$ and $$\begin{aligned} &\sum_{t=1}^{L}\big((\tilde{{\ensuremath{\boldsymbol{v}}}}'_{t,0}\;\;\; \tilde{{\ensuremath{\boldsymbol{v}}}}'_{t-1,1}) \cdot \Pi^{(2)} \;\;{\ensuremath{\boldsymbol{v}}}_t^{\text{I}}\big)\nonumber\\ &=\big(\sum_{t=1}^{L}(\tilde{{\ensuremath{\boldsymbol{v}}}}'_{t,0}\;\;\; \tilde{{\ensuremath{\boldsymbol{v}}}}'_{t-1,1}) \cdot \Pi^{(2)} \;\;\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}_t^{\text{I}}\big)\nonumber\\ &=\big(\sum_{t=1}^{L}\tilde{{\ensuremath{\boldsymbol{v}}}}'_{t} \cdot \Pi^{(2)} \;\;\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}_t^{\text{I}}\big)\nonumber\\ \label{SCCProof4} &=\big(\sum_{t=1}^{L}({\ensuremath{\boldsymbol{u}}}_t\;\;{\ensuremath{\boldsymbol{v}}}_t^{\text{O}})\cdot\Pi^{(1)}\cdot\Pi^{(2)} \;\;\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}_t^{\text{I}}\big). \end{aligned}$$ Let $$\tilde{{\ensuremath{\boldsymbol{u}}}}=\sum_{t=1}^{L}{\ensuremath{\boldsymbol{u}}},\;\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{O}}=\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{O}},\;\;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{I}}=\sum_{t=1}^{L}{\ensuremath{\boldsymbol{v}}}^{\text{I}},$$ and substitute and into and , respectively. Then $$\begin{aligned} &\big(\tilde{{\ensuremath{\boldsymbol{u}}}}\;\; \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{O}}\big)\cdot {\ensuremath{\boldsymbol{H}}}_{\text{O}}^T={\ensuremath{\boldsymbol{0}}} \; , \\ &\big((\tilde{{\ensuremath{\boldsymbol{u}}}}\;\;\; \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{O}}) \cdot \Pi^{\text{Un}} \;\;\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{I}}\big)\cdot {\ensuremath{\boldsymbol{H}}}_{\text{I}}^T={\ensuremath{\boldsymbol{0}}} \; . \end{aligned}$$ Therefore, $\tilde{{\ensuremath{\boldsymbol{v}}}}=(\tilde{{\ensuremath{\boldsymbol{u}}}},\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{O}},\tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{I}})$ is a codeword of the uncoupled code. If all nonzero elements of ${\ensuremath{\boldsymbol{v}}}_t$, $t=1,\dots,L$, occur at different positions, then $w_{\text H}(\tilde{{\ensuremath{\boldsymbol{v}}}})=w_{\text H}({\ensuremath{\boldsymbol{v}}})$. Otherwise, the overlap of the non zero elements reduces the weight of $\tilde{{\ensuremath{\boldsymbol{v}}}}$ and $w_{\text{H}}(\tilde{{\ensuremath{\boldsymbol{v}}}})< w_{\text{H}}({\ensuremath{\boldsymbol{v}}})$. Braided Convolutional Codes --------------------------- Consider the SCC and SC-SCC ensembles in Fig. \[CG\](c) and (d), and assume that $\Pi_t=\Pi$, $\Pi_t^{\text{U}}=\Pi^{\text{U}}$ and $\Pi_t^{\text{L}}=\Pi^{\text{L}}$. A valid code sequence of $\mathcal{C}$ has to satisfy the local constraints $$\begin{aligned} \begin{pmatrix} {\ensuremath{\boldsymbol{u}}}_t & {\ensuremath{\boldsymbol{v}}}_{t-1}^{\text{L}} \cdot \Pi_t^{\text{U}} & {\ensuremath{\boldsymbol{v}}}_{t}^{\text{U}} \end{pmatrix} \cdot {\ensuremath{\boldsymbol{H}}}_{\text{U}}^T & = {\ensuremath{\boldsymbol{0}}} \label{eq:coupledUpper} \; , \\ \begin{pmatrix} {\ensuremath{\boldsymbol{u}}}_t \cdot \Pi_t & {\ensuremath{\boldsymbol{v}}}_{t-1}^{\text{U}} \cdot \Pi_t^{\text{L}} & {\ensuremath{\boldsymbol{v}}}_{t}^{\text{L}} \end{pmatrix} \cdot {\ensuremath{\boldsymbol{H}}}_{\text{L}}^T & = {\ensuremath{\boldsymbol{0}}} \label{eq:coupledLower} \end{aligned}$$ for all $t=1,\dots,L$, where ${\ensuremath{\boldsymbol{H}}}_{\text{U}}$ and ${\ensuremath{\boldsymbol{H}}}_{\text{L}}$ are the parity-check matrices that represent the constraints imposed by the trellises of the upper and lower component encoders, respectively. Since these constraints are linear and time-invariant, it follows that any superposition of vectors ${\ensuremath{\boldsymbol{v}}}_t=({\ensuremath{\boldsymbol{u}}}_t,{\ensuremath{\boldsymbol{v}}}_t^{\text{U}},{\ensuremath{\boldsymbol{v}}}_t^{\text{U}})$ from different time instants $t \in \{1,\dots,L\}$ will also satisfy and . In particular, if we let $$\tilde{{\ensuremath{\boldsymbol{u}}}}=\sum_{t=1}^{L} {\ensuremath{\boldsymbol{u}}}_t \ , \quad \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}}=\sum_{t=1}^{L} {\ensuremath{\boldsymbol{v}}}_t^{\text{L}} \ , \quad \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}}=\sum_{t=1}^{L} {\ensuremath{\boldsymbol{v}}}_t^{\text{U}}$$ then $$\begin{aligned} \begin{pmatrix} \tilde{{\ensuremath{\boldsymbol{u}}}} & \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}} \cdot \Pi^{\text{U}} & \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}} \end{pmatrix} \cdot {\ensuremath{\boldsymbol{H}}}_{\text{U}}^T & = {\ensuremath{\boldsymbol{0}}} \label{eq:uncoupledUpper} \; , \\ \begin{pmatrix} \tilde{{\ensuremath{\boldsymbol{u}}}} \cdot \Pi & \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}} \cdot \Pi^{\text{L}} & \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}} \end{pmatrix} \cdot {\ensuremath{\boldsymbol{H}}}_{\text{L}}^T & = {\ensuremath{\boldsymbol{0}}} \label{eq:uncoupledLower} \enspace . \end{aligned}$$ Here we have implicitly made use of the fact that ${\ensuremath{\boldsymbol{v}}}_t={\ensuremath{\boldsymbol{0}}}$ for $t<1$ and $t>L$. But now it follows from and that $\tilde{{\ensuremath{\boldsymbol{v}}}}=(\tilde{{\ensuremath{\boldsymbol{u}}}}, \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{U}}, \tilde{{\ensuremath{\boldsymbol{v}}}}^{\text{L}}) \in \tilde{\mathcal{C}}$, i.e., we obtain a codeword of the uncoupled code. If all nonzero symbols within ${\ensuremath{\boldsymbol{v}}}_t$ occur at different positions for $t=1,\dots,L$, then $w_{\text{H}}(\tilde{{\ensuremath{\boldsymbol{v}}}}) = w_{\text{H}}({{\ensuremath{\boldsymbol{v}}}})$. If, on the other hand, the support of nonzero symbols overlaps, the weight of $\tilde{{\ensuremath{\boldsymbol{v}}}}$ is reduced accordingly and $w_{\text{H}}(\tilde{{\ensuremath{\boldsymbol{v}}}}) < w_{\text{H}}({{\ensuremath{\boldsymbol{v}}}})$. The same result can be proved for HCCs by combining the proofs for PCCs and SCCs. [10]{} C. Berrou, A. Glavieux, and P. Thitimajshima, “Near [Shannon]{} limit error-correcting coding and decoding: turbo-codes (1),” in [*Proc.  [IEEE]{} Int. Conf. Commun. (ICC)*]{}, Geneva, Switzerland, 1993. R. Gallager, , MIT Press, Cambridge, MA, 1963. and [K.Sh. Zigangirov]{}, “Periodic time-varying convolutional codes with low-density parity-check matrices,” , vol. 45, no. 5, pp. 2181–2190, Sep. 1999. M. Lentmaier, A. Sridharan, D. J. Costello, and K. S. Zigangirov, “Iterative decoding threshold analysis for [LDPC]{} convolutional codes,” , vol. 56, no. 10, pp. 5274–5289, Oct. 2010. S. Kudekar, T.J. Richardson, and R.L. Urbanke, “Threshold saturation via spatial coupling: [W]{}hy convolutional [LDPC]{} ensembles perform so well over the [BEC]{},” , vol. 57, no. 2, pp. 803 –834, Feb. 2011. A. Yedla, Yung-Yih Jian, P.S. Nguyen, and H.D. Pfister, “A simple proof of threshold saturation for coupled scalar recursions,” in [*Proc. Int. Symp. Turbo Codes and Iterative Inf. Processing (ISTC)*]{}, Gothenburg, Sweden, 2012. A. Yedla, Yung-Yih Jian, P.S. Nguyen, and H.D. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” in [*Proc.  Inf. Theory Work. (ITW)*]{}, Lausanne, Switzerland, 2012. A. Yedla, Yung-Yih Jian, P.S. Nguyen, and H.D. Pfister, “A simple proof of [Maxwell]{} saturation for coupled scalar recursions,” , vol. 60, no. 11, pp. 6943–6965, Nov. 2014. S. Moloudi, M. Lentmaier, and A. Graell i Amat, “Spatially coupled turbo codes,” in [*Proc. Int. Symp. Turbo Codes and Iterative Inf. Processing (ISTC)*]{}, Bremen, Germany, 2014. S. Moloudi, M. Lentmaier, and A. Graell i Amat, “Spatially coupled turbo-like codes,” , vol. 63, no. 10, pp. 6199–6215, Oct. 2017. S. Moloudi, M. Lentmaier, and A. Graell i Amat, “Spatially coupled hybrid concatenated codes,” in [*Proc. 11th Int. ITG Conf. Systems, Commun. and Coding (SCC)*]{}, Hamburg, Germany, 2017. S. Benedetto, D. Divsalar, G. Montorsi, and F. Pollara, “Serial concatenation of interleaved codes: performance analysis, design, and iterative decoding,” , vol. 44, no. 3, pp. 909–926, May 1998. A. Graell i Amat, G. Montorsi, and F. Vatta, “Design and performance analysis of a new class of rate compatible serially concatenated convolutional codes,” , vol. 57, no. 8, pp. 2280–2289, Aug. 2009. , [M. Lentmaier]{}, [K.Sh. Zigangirov]{}, and [D.J. Costello, Jr.]{}, “Braided convolutional codes: a new class of turbo-like codes,” , vol. 56, no. 1, pp. 316–331, Jan. 2010. M. Zhu, D. G. M. Mitchell, M. Lentmaier, D. J. Costello, and B. Bai, “Braided convolutional codes with sliding window decoding,” , vol. 65, no. 9, pp. 3645–3658, Sep. 2017. E. Rosnes and A. Graell i Amat, “Performance analysis of 3-[D]{} turbo codes,” , vol. 57, no. 6, pp. 3707–3720, June 2011. C. Koller, A. Graell i Amat, J. Kliewer, F. Vatta, and D. J. Costello, “Hybrid concatenated codes with asymptotically good distance growth,” in [*Proc. 5th Int. Symp. Turbo Codes and Related Topics (ISTC)*]{}, 2008. D. G. M. Mitchell, M. Lentmaier, A. E. Pusane, and D. J. Costello, “Randomly punctured [LDPC]{} codes,” , vol. 34, no. 2, pp. 408–421, Feb. 2016. M. U. Farooq, S. Moloudi, and M. Lentmaier, “[Thresholds of Braided Convolutional Codes on the AWGN Channel]{},” submitted to [*IEEE Int. Sym. Inf. Theory (ISIT)*]{}, available online at <https://arxiv.org/abs/1802.10540>, 2018. S. Benedetto and G. Montorsi, “Unveiling turbo codes: some results on parallel concatenated coding schemes,” , vol. 42, no. 2, pp. 409–428, Mar. 1996. S. Abu-Surra, W.E. Ryan, and D. Divsalar, “Ensemble enumerators for protograph-based generalized [LDPC]{} codes,” , pp. 1492–1497, 2007. S. Abu-Surra, D. Divsalar, and W. E. Ryan, “Enumerators for protograph-based ensembles of [LDPC]{} and generalized [LDPC]{} codes,” , vol. 57, no. 2, pp. 858–886, Feb. 2011. D. Truhachev, K. S. Zigangirov, and D. J. Costello, “Distance bounds for periodically time-varying and tail-biting [LDPC]{} convolutional codes,” , vol. 56, no. 9, pp. 4301–4308, Sep. 2010. D. G. M. Mitchell, A. E. Pusane, and D. J. Costello, “Minimum distance and trapping set analysis of protograph-based [LDPC]{} convolutional codes,” , vol. 59, no. 1, pp. 254–281, Jan. 2013. R. Smarandache, A. E. Pusane, P. O. Vontobel, and D. J. Costello, “Pseudocodeword performance analysis for [LDPC]{} convolutional codes,” , vol. 55, no. 6, pp. 2577–2598, Jun. 2009. S. Moloudi, M. Lentmaier, and A. Graell i Amat, “Finite length weight enumerator analysis of braided convolutional codes,” in [*Proc. Int. Symp. Inf. Theory and Its Applications (ISITA)*]{}, Oct. 2016. D. G. M. Mitchell, L. Dolecek, and D. J. Costello, “Absorbing set characterization of array-based spatially coupled [LDPC]{} codes,” in [*IEEE Int. Sym. Inf. Theory (ISIT)*]{}, Honolulu, HI, USA, 2014. E. Rosnes, M. Helmling, and A. Graell i Amat, “Minimum pseudoweight analysis of 3-dimensional turbo codes,” , vol. 62, no. 7, pp. 2170–2182, Jul. 2014. [^1]: Parts of this paper have been presented at the IEEE International Symposium on Information Theory and Its Applications (ISITA), 2016. [^2]: This work was supported in part by the Swedish Research Council (VR) under grant \#621-2013-5477. [^3]: S. Moloudi and M. Lentmaier are with the Department of Electrical and Information Technology, Lund University, Lund, Sweden (e-mail: {saeedeh.moloudi,michael.lentmaier}@eit.lth.se). [^4]: A. Graell i Amat is with the Department of Electrical Engineering, Chalmers University of Technology, SE-41296 Gothenburg, Sweden (e-mail: [email protected]). [^5]: The multiplexer is represented by a rectangular in the compact graph representation.

ArXiv

--- abstract: 'Quantum state transfer in optical microcavities plays an important role in quantum information processing, and is essential in many optical devices, such as optical frequency converter and diode. Existing schemes are effective and realized by tuning the coupling strengths between modes. However, such approaches are severely restricted due to the small amount of strength that can be tuned and the difficulty to perform the tuning in some situations, such as on-chip microcavity system. Here, we propose a novel approach that realizes the state transfer between different modes in optical microcavities by tuning the frequency of an intermediate mode. We showed that for typical functions of frequency-tuning, such as linear and periodic functions, the state transfer can be realized successfully with different features. To optimize the process, we use gradient descent technique to find an optimal tuning function for the fast and perfect state transfer. We also showed that our approach has significant nonreciprocity with appropriate tuning variables, where one can unidirectionally transfer a state from one mode to another, but the inverse direction transfer is forbidden. This work provides an effective method for controlling the multimode interactions in on-chip optical microcavities via simple operations and it has practical applications in all-optical devices.' address: | $^1$State Key Laboratory of Low-Dimensional Quantum Physics, Department of Physics, Tsinghua University, Beijing 100084, China\ $^2$Beijing Information Science and Technology National Research Center, Beijing 100084, China\ $^3$Beijing Academy of Quantum Information Sciences, Beijing 100193, China author: - 'Xu-sheng Xu$^{1*}$, Hao Zhang$^{1}$[^1], Xiang-Yu Kong$^{1}$, Min Wang$^{1}$, and Gui-Lu Long$^{1,2,3}$[^2]' title: 'Frequency-tuning induced state transfer in optical microcavities' --- Introduction ============ As an important fundamental task, state transfer is widely studied in atomic, optical physics and quantum information for its indispensable role in building optical and quantum devices, such as optical transistor [@transistor1; @transistor2], frequency conversion [@conversion1] and quantum interface [@interface1; @interface2; @JICiracPRL1997; @markprx]. In atomic system, the typical approaches for realizing state transfer are the rapid adiabatic passage [@NVVitanovARPC2001] for two-level, the stimulated Raman adiabatic passage [@KBergmannRMP1998] for excited state assisted three-level $\Lambda$ quantum systems and their optimized shortcuts to adiabaticity technique[@DGORMPSTA; @MVBerry2009; @XChenPRL20102; @ABaksicPRL2016; @YHChenpra; @YLiangPRA2015; @XKSongNJP2016]. Optical microcavity, which can effectively enhance the interaction between light and matters [@Vahala], is a good platform for studying optical physics and useful applications. For instance, some interesting physics, such as parity-time-symmetry [@pt1; @pt2; @pt3], chaos [@chao1; @chao2] and nonreciprocity [@optomnonre1; @optomnonre2], have been demonstrated in microcavities. In applications, microcavities show significant functions for sensing [@sensing; @cavitysensing1; @cavitysensing2; @cavitysensing3; @cavitysensing4] and processing quantum information [@CQED; @qip1; @qip2; @qip3; @qip4; @qip5; @qip6; @qip7]. Photon can be confined in microcavity and can also be transferred to another one via evanescent waves coupling or other interactions. Realizing state transfer between microcavities is important for making microcavity a good physical system for quantum information processing and optical devices. In quantum computing, all optical microcavity coupling lattice structure can be used for performing boson sampling [@boson] and microcavity can also be considered as a quantum bus to connect solid qubits for building quantum computer. To make all-optical device, such as transistor [@transistor1; @transistor2] and router [@router], the target is achieved by performing the state transfer between microcavities successfully. Some effective protocols for state transfer between optical modes are reported with adiabatic methods [@LTianPRL2012; @YDWangPRL2012; @YDWangNJP], nonadiabatic approaches [@YDWangNJP] and shortcuts to adiabaticity technique. [@HZhangOE; @XZhouLPL]. By using optomechanical interactions [@MAspelmeyerRMP2014; @Long1; @Long2; @Long3; @CHDongscience; @Kuzykpra17; @optomaqip; @optomcooling], the protocols are completed successfully by tuning coupling strengths very well to satisfy technique constraints. When we consider the situation of state transfer in on-chip all optical microcavity system, the coupling strength tuning becomes difficult. To solve this problem, in this paper, we proposed an approach to realize the state transfer task between separated modes in optical microcavities via frequency-tuning. In our protocol, we assume that all the coupling strengths are constant and tune the frequency of intermediate microcavity to control the interactions. With linear and periodic tuning, one can transfer the state from the initial cavity mode to the target successfully. To achieve the faster frequency-tuning induced state transfer (FIST) with high fidelity, we use gradient descent to optimize result and acquire an optimal tuning function. Our protocol also shows an significant nonreciprocity in appropriate area of parameters. Good experimental feasibility and interesting features of our work provides potential applications in quantum computing and optical devices. This article is organized as follows: We describe the basic model of multimode interactions in optical microcavities in Sec. \[sec2\]. In Sec. \[sec3\], we show how to realize the FIST via numerically calculation. We discuss the linear, periodic and optimized tuning in Sec. \[sec31\], B and C, respectively. In Sec. \[sec4\], we investigate the nonreciprocity in our model. Discussion and conclusion are given in Sec. \[sec5\]. ![Schematic diagram for the model of multimode interactions in optical microcavities. All the modes have very narrow linewidth. A mode in one cavity couples to two different optical modes, (a) in the same cavity, (b) in two different cavities separately. (c) Resonance frequency tuning of intermediate cavity to induce state transfer. The tuning domain is divided into three parts labelled with I, II and III.[]{data-label="basicmodel"}](fig1.pdf){width="8.5cm"} Basic model for multimode interactions in optical microcavities {#sec2} =============================================================== We consider a model of multimode interactions in coupled optical microcavities shown in Fig. \[basicmodel\]. The model is universal for situations which two optical modes coupled to the intermediate one and all the modes have very narrow linewidth. Fig. \[basicmodel\] (a) gives a setup with nearest neighbour couplings in three optical cavities. We assume that the coupling strengths between corresponding modes are constant and the system can be controlled by tuning the resonance frequency of intermediate mode. Under the rotating frame with unitary transformation $U=\exp[i\omega_{1}t(a_{1}^{\dag}a_{1}+a_{t}^{\dag}a_{t}+a_{2}^{\dag}a_{2})]$, the Hamiltonian is given by $$\begin{aligned} \label{eq1} H_{1}=\delta a_{2}^{\dag} a_{2}+\Delta(t) a_{t}^{\dag}a_{t}+g_{1}a_{1}^{\dag}a_{t}+g_{2}a_{2}^{\dag}a_{t}+\text{H.c.},\end{aligned}$$ where $a_{i}$ $(a^{\dag}_{i})$ $(i=1,2,t)$ are the annihilation (creation) operators for the corresponding $i$-th mode of the cavity and the corresponding frequency is $\omega_{i}$, respectively. The detunings are $\delta=\omega_{2}-\omega_{1}$ and $\Delta(t)=\omega_{t}-\omega_{1}$. $g_{i}$ $(i=1,2)$ is the coupling strength between modes $a_{t}$ and $a_{i}$. The Hamiltonian can be expressed in Heisenberg equations with $id\vec{a}(t)/dt=M(t)\vec{a}(t)$ and the matrix $M(t)$ is given by $$\begin{aligned} \label{eqM} M(t)= \left[ \begin{array}{ccc} 0&g_{1}&0\\ g_{1}&\Delta(t)&g_{2}\\ 0&g_{2}&\delta\\ \end{array} \right].\end{aligned}$$ Here the vector is $\vec{a}(t)=[\hat{a}_{1}(t),\hat{a}_{t}(t),\hat{a}_{2}(t)]^{T}$. To show the results more clearly, we omit the dissipation and noise terms in our calculation due to the very narrow linewidth. In general, the coupling strength between cavities is difficult to be modulated for on-chip sample. Therefore, we keep the coupling strengths constant here and tuning the resonance frequency of intermediate cavity to control the evolution path of system. FIST between separated modes {#sec3} ============================ The frequency-tuning can be realized with different functions. Here we perform the FIST task with two common typical envelopes, i.e. linear and periodic functions, and use the gradient descent technique to optimize the process. ![The result of FIST between $a_1$ and $a_2$ by linearly tuning the resonance frequency of $a_t$. The speed is chosen with $0.08\delta^2$. The inset is the plot of tuning function and the unit of time $t$ is $\delta^{-1}$.[]{data-label="simulationlinear"}](fig2.pdf){width="8.5cm"} FIST with linear function {#sec31} ------------------------- We assume that the envelope of frequency-tuning is chosen with $\Delta(t)=vt+\Delta_0$. Here $v$ and $\Delta_0$ are tuning speed and the initial value of detuning, respectively. At the initial time, the mode $a_{1}$ is stimulated and the other modes are kept in their ground state, i.e. $a_{1}(0)=1$ and $a_{t}(0)=a_{2}(0)=0$. Without loss of generality, we choose the frequency-detuning of mode $a_{1}$ and $a_{2}$ with $\delta>0$ and sweep the frequency of mode $a_{t}$ from left to right in Fig. \[basicmodel\] (c). As the frequency of intermediate mode $a_t$ is swept, the state is transferred from the initial mode to other modes. We numerically simulate the process and show the result in Fig. \[simulationlinear\]. The resonance frequency of intermediate mode is swept from $-6\delta$ to $7\delta$ with a constant speed $0.08\delta^2$, i.e. $\Delta_0=-6\delta$ and $v=0.08\delta^2$. The coupling strength between modes are chosen with $g_{1}=0.6$ and $g_{2}=0.2$ in our simulations. Fig. \[simulationlinear\] shows that the population $P$ of mode $a_1$ is transferred to $a_t$ when the frequency of $a_t$ is swept to $a_1$. When the frequency of intermediate mode is moving in domain I of Fig. \[basicmodel\] (c), the mode $a_2$ has no effective exchange of population with $a_t$ and almost keeps its initial state due to the large detuning with $a_t$. When the frequency of $a_t$ is swept into domain II (between $a_1$ and $a_2$), $a_t$ exchanges the population with $a_1$ and $a_2$ simultaneously. As the frequency of $a_t$ arrived at domain III, i.e. the right of $a_2$, the mode $a_1$ returns to its ground state and keep the state to the end due to the large detuning with $a_t$. The population of $a_t$ also keep transferring to $a_2$ until the $a_t$ evolves to its ground state. When the detuning $\Delta(t)$ is larger than about $4\delta(t)$, the system evolves to our target state, i.e. $a_{2}(T)\approx1$ and $a_{1}(T)\approx a_{t}(T)\approx 0$. In whole process, one only need to keep a stable tuning speed to get the perfect state transfer after the $a_t$ becomes larger than $4\delta(t)$. The maximal population of $a_2$ is about $P=0.9536$. ![Simulation of final population of mode $a_2$ effected by tuning variables. The units of $d$ and $v$ are chosen with $\delta$ and $\delta^2$, respectively. (a) Population $P$ vs tuning range $d$ with $v=0.27\delta^2$. All the $\Delta_0$ are chosen with $\Delta_0=(\delta-d)/2$. (b) Population $P$ vs tuning speed $v$ with $d=2.65\delta$ and $\Delta_0=-0.825\delta$. (c)Population $P$ vs tuning range $d$ and tuning speed $v$. The dashed line is all the points of evolution time with $10\delta^{-1}$. []{data-label="popuvd"}](fig3.pdf){width="8.5cm"} To investigate how FIST is effected by different tuning range $d$ and tuning speed $v$, we show the final population of mode $a_2$ with respective to $d$ and $v$ in Fig. \[popuvd\]. The Fig. \[popuvd\] (a) and (b) are the 2-dimensional cross-section analysis of two dot-dashed line in Fig. \[popuvd\] (c). In Fig. \[popuvd\] (a), the tuning speed is fixed with $v=0.27\delta^2$. The population is proportional to the tuning range $d$ within a short range less than about $3.0\delta$ and oscillates to a stable value as the range becomes larger. The reason is that system has more time to transfer the state from $a_1$ to $a_2$ as tuning range becomes larger in the domain about $(0,3.0\delta)$. When the tuning range becomes very long, the population has little change. Because the large detuning makes little contribution to state transfer. As shown in Fig. \[popuvd\] (b), on the contrary, when the tuning range is kept with $2.65\delta$, the population is inversely proportional to the tuning speed along with a little oscillation, since faster speed makes shorter interaction time. Fig. \[popuvd\] (c) indicates the clearer conclusion that large tuning range and slow tuning speed are the best tuning manner. FIST with periodic function {#sec32} --------------------------- Beside the linear tuning, we consider the situation with periodic function. For simplicity, we choose sine function here. The frequency-detuning is described as $\Delta(t)=A[\sin(\Omega t)+c_0]/2$. With the parameters chosen with $A=9.6$, $\Omega=0.95$ and $c_0=0.5$, the populations of each modes are performed in Fig. \[popusin\]. The envelope curve of populations of $a_1$ and $a_2$ are evolved as a sine function. The modes $a_1$ and $a_t$ exchange their populations with Rabi frequency $\Omega$, which is the frequency of tuning function. The maximal value of population $P$ of mode $a_2$ in this periodic tuning is $0.9365$. Compared with linear tuning, the population transferred here is realized successfully by controlling the evolution time of tuning function. For instance, in Fig. \[popusin\], the evolution should be ended in around $160\delta^{-1}$. In principle, if the end time is missed, one can wait for the neat periodic time, but the long time will decrease the fidelity of protocol by bring more decoherence. ![The populations change with respective to evolution time via sine tuning function. Lines labelled with $a_1$, $a_2$ and $a_3$ are the populations of corresponding modes, respectively.[]{data-label="popusin"}](fig4.pdf){width="8.5cm"} Optimizing the FIST via gradient descent {#sec33} ---------------------------------------- To achieve a perfect and fast state transfer from $a_1$ to $a_3$, we use the gradient descent technique to optimize the FIST and pick an optimal function for frequency tuning. The gradient descent technique is usually used for finding a optimal evolution path for system along the opposite direction of gradient descent. In our model, the evolution equations are given by $$\begin{aligned} \label{eqGD1} a(t)=U(t,0)a(0),\end{aligned}$$ where the evolution operator is a Dyson series with expression given by $U(t,0)=\exp[\intop^{t}_{0}M(t')dt']$. The time domain is chosen with $t=[0,t_1]$ and the frequency of intermediate mode, considered as parameter to be optimized, is divided into $n$ discrete constant variables, i.e. $\Delta_I=[\Delta_1,\cdots,\Delta_n]$. Therefore, the operator can be rewritten with $$\begin{aligned} \label{eqGD2} a(t_1)=U(\Delta_1,\cdots,\Delta_n;t_1,0)a(0).\end{aligned}$$ The target function is chosen with $\mathcal{L}(a,t)=|a_2(t)|^2$ and can be described as $$\begin{aligned} \label{eqGD3} \mathcal{L}(a,t_1)=\mathcal{L}(\Delta_1,\cdots,\Delta_n;t_1,0).\end{aligned}$$ Our goal is to optimize above function and get the maximal value. So the gradient of the target function is calculated as $\nabla\mathcal{L}=\frac{\partial\mathcal{L}}{\partial\omega}$. In numerical calculation, the gradient is written approximatively by $$\begin{aligned} \label{eqGD4} \nabla\mathcal{L}=(\nabla\mathcal{L}_1,\cdots,\nabla\mathcal{L}_n),\end{aligned}$$ where $\nabla\mathcal{L}_i$ is $$\begin{aligned} \label{eqGD5} \nabla\mathcal{L}_i\!=\!\frac{\mathcal{L}(\cdots,\Delta_i+\delta\Delta,\cdots;t_1,0)\!-\!\mathcal{L}(\Delta_1,\cdots,\Delta_n;t_1,0)}{\delta\Delta}.\end{aligned}$$ ![Simulation of fast FIST from $a_1$ to $a_2$ by using gradient descent technique. The parameters are the cross point of Fig. \[popuvd\] (c) with ($d=2.65\delta$,$v=0.27\delta^2$). (a) The result of optimized population transfer process. (b) The corresponding optimal tuning function of intermediate mode. The unit of time here is $\delta^{-1}$.[]{data-label="optimize"}](fig5.pdf){width="8.5cm"} The simulation of optimized fast FIST is shown in Fig. \[optimize\]. We plot the transfer results and the tuning functions in Fig. \[optimize\] (a) and (b), respectively. In our simulations, we first choose the $\Delta_I(t=0)=0$ and cut the tuning function $\Delta_I(t)$ into discrete $100$ segments which are mutually independent to be optimized via gradient descent algorithm. The results and optimal tuning envelope are plotted respectively in Fig. \[optimize\] (a) and (b) with dotted curves. During the evolution time, the population of mode $a_2$ monotonically increases to its maximal value $0.9923$. However, the curves of modes $a_1$ and $a_t$ decrease to its ground state with an oscillating process. To make the scheme conveniently controlled, according to the envelope of dotted line in Fig. \[optimize\] (b), we reasonably give the tuning function with expression as $$\begin{aligned} \label{optimalfunc} \Delta_{II}(t)=A\sin[(\Omega t+\theta)+C_1]*[e^{-\gamma t}+C_2],\end{aligned}$$ where the parameters $A$, $\Omega$, $\theta$, $\gamma$, $C_1$ and $C_2$ are undetermined coefficients to be optimized. By calculating the state transfer task with the Eq. (\[optimalfunc\]), the population of mode $a_2$ can be optimized to $0.9826$ and the parameters are given with $A=-5.555$, $\Omega=1.276$, $\theta=0.564$, $\gamma=1.467$, $C_1=-0.797$ and $C_2=0.119$. The evolutions of populations and tuning functions are plotted in Fig. \[optimize\]. The envelope of Eq. (\[optimalfunc\]) and the evolution of populations are similar with the dotted curves in each figures, respectively. In linear tuning shown in Fig. \[popuvd\] (c), the maximal population $P$ of mode $a_2$ with evolution time less then $10\delta^{-1}$ is $P=0.618$ labelled with a cross point. With the evolution time $10\delta^{-1}$, the optimized protocol can achieve the population with $P>0.98$. Compared with linear tuning, the state transfer under this protocol can be optimized with a faster evolution path and higher fidelity. ![Nonreciprocal state transfer between modes $a_1$ and $a_2$. (a) The populations of modes $a_1$ and $a_2$ vs tuning speed. The tuning range is $d=14\delta$. (b)The populations of modes $a_1$ and $a_2$ vs tuning range. The tuning speed is $v=0.1\delta^2$.[]{data-label="Nonreciprocity"}](fig6.pdf){width="8.5cm"} Nonreciprocity in multimode interactions {#sec4} ======================================== Our model shows the significant nonreciprocity in state transfer between $a_1$ and $a_2$. For example, when the frequency of intermediate mode is swept from $a_1$ to $a_2$, the state can be transferred from $a_1$ to $a_2$, but it completely failed for $a_2$ to $a_1$. The results of the nonreciprocal state transfer are plotted in Fig. \[Nonreciprocity\] with linear tuning from $a_1$ to $a_2$, i.e. from left to right in Fig. \[basicmodel\] (c). We fix the tuning range with $d=14\delta$ and change the tuning speed in Fig. \[Nonreciprocity\] (a). The top (bottom) green (blue) line is the final population of $a_2$ ($a_1$) with the initial state is prepared in $a_1$ ($a_2$). As the speed becomes slower, the nonreciprocity is more clearly. At very slow speed, the population of $a_2$ transferred from $a_1$ is almost $1$ but the population of inverse transfer process is less than $10^{-4}$. In Fig. \[Nonreciprocity\] (b), we investigate the nonreciprocity with respective to tuning range with constant tuning speed $v=0.1\delta^2$. The small populations of both $a_1$ and $a_2$ in different directions in small tuning range indicate that the nonreciprocity is not clearly. When we increase the tuning range $d$, the green and blue line converge towards unit and $10^{-5}$, respectively. ![All-optical on-chip microcavities structures. (a) One-dimensional microcavity array. (b)Two-dimensional optical microcavity lattice.[]{data-label="network"}](fig7.pdf){width="8.5cm"} Discussion and Conclusion {#sec5} ========================= All the results shown above are considered in all optical cavities system. Actually, our model is a universal approach for multimodes interaction system, such as all mechanical phonon modes or the photon-phonon interactions. For example, the direct interaction between phonons is difficult. So one can transfer the state from one mechanical resonator to another one via an intermediate optical cavity mode [@Kuzykpra17; @HZhangOE]. Beside the coupling strength tuning way, one can use frequency tuning approach described here to control the interactions. To perform more interesting applications, our model can be extended to design one-dimensional microcavity array and two-dimensional microcavity lattice. The detailed schematic diagrams are shown in Fig. \[network\]. An array is built by coupling $n+1$ microcavities with $n$ tuning cavities in Fig. \[network\](a). The structure can be used for realizing all-optical transistor with more abundant physical tuning. Fig. \[network\] (b) shows the two-dimensional lattice structure which designed by connecting one tuning cavity with three storage cavities. Every unit of this structure is a simple optical router. The structure has function for building all-optical on-chip quantum network [@JICiracPRL1997; @markprx]. In our model, we always keep the coupling strengths constant with the assumption that the distances between cavities are fixed. The typical corresponding physical system is the on-chip optical microcavity sample. Because the distances between each cavity are difficult to change after complete the fabrication. Our frequency tuning manner is possible. This is because the frequency of microcavity is sensitive to its shape which can be modulated by some operations, such as temperature [@temp1; @temp2; @temp3], external forces [@mech1; @mech2; @mech3; @mech4; @mech5; @mech6; @mech7] and etc. The above frequency tuning approaches have been realized with high resolution in experiments, but the tuning speed is limited. So the fast tuning way is needed for improving the feasibility of the practical applications. In conclusion, we have proposed an approach to realize the state transfer between two separated modes in optical microcavities. Our proposal are valid for both two and three microcavities. The FIST can be realized with high fidelity via different tuning manner, i.e. linear and periodic function, of resonance frequency of intermediate mode. To optimize the tuning function, a fast and perfect evolution process is performed by using gradient descent technique. Our proposal also shows the significant nonreciprocity. The state can be transferred successfully in the same direction with frequency tuning and it is fail in the opposite direction. Our work provides an effective approach for controlling the optical mode in on-chip microcavities and has important applications in all-optical devices. ACKNOWLEDGMENT {#acknowledgment .unnumbered} ============== The authors thank Guo-Qing Qin for helpful discussions. This work was supported by the National Natural Science Foundation of China (20171311628); National Key Research and Development Program of China (2017YFA0303700); Beijing Advanced Innovation Center for Future Chip (ICFC). H.Z. acknowledges the China Postdoctoral Science Foundation under Grant No.2019M650620. [1]{} D. A. B. Miller, Are optical transistors the logical next step?, Nat. Photonics **4**, 3 (2010). W. Chen, K. M. Beck, R. Bücker, M. Gullans, M. D. Lukin, H. Tanji-Suzuki, and V. Vuletić, All-optical switch and transistor gated by one stored photon, Science **341**, 768 (2013). X. Guo, C.-L. Zou, H. Jung, and H. X. Tang, On-Chip Strong Coupling and Efficient Frequency Conversion between Telecom and Visible Optical Modes, Phys. Rev. Lett. **117**, 123902 (2016). K. Hammerer, A. S. S[ø]{}rensen, and E. S. Polzik, Rev. Mod. Phys. **82**, 1041 (2010). T. Wilk, S. C. Webster, A. Kuhn, and G. Rempe, Single-atom single-photon quantum interface, Science **317**, 488 (2007). J. I. Cirac, P. Zoller, H. J. Kimble, and H. Mabuchi, Quantum state transfer and entanglement distribution among distant nodes in a quantum network, Phys. Rev. Lett. **78**, 3221 (1997). M. C. Kuzyk and H. Wang, Scaling Phononic Quantum Networks of Solid-State Spins with Closed Mechanical Subsystems, Phys. Rev. X **8**, 041027 (2018). N. V. Vitanov, T. Halfmann, B. W. Shore, and K. Bergmann, Laser-induced population transfer by adiabatic passage techniques, Annu. Rev. Phys. Chem. **52**, 763 (2001). K. Bergmann, H. Theuer, and B. Shore, Coherent population transfer among quantum states of atoms and molecules, Rev. Mod. Phys. **70**, 1003 (1998). D. Guéry-Odelin, A. Ruschhaupt, A. Kiely, E. Torrontegui, S. Martinez-Garaot, and J. G. Muga, Shortcuts to adiabaticity: Concepts, methods, and applications, Rev. Mod. Phys. **91**, 045001 (2019). M. V. Berry, Transitionless quantum driving, J. Phys. A: Math. Theor. **42**, 365303 (2009). X. Chen, I. Lizuain, A. Ruschhaupt, D. Guéry-Odelin, and J. G. Muga, Shortcut to adiabatic passage in two- and three-Level atoms, Phys. Rev. Lett. **105**, 123003 (2010). A. Baksic, H. Ribeiro, and A. A. Clerk, Speeding up adiabatic quantum state transfer by using dressed states, Phys. Rev. Lett. **116**, 230503 (2016) Y. H. Chen, Y. Xia, Q. Q. Chen, and J. Song, Efficient shortcuts to adiabatic passage for fast population transfer in multiparticle systems, Phys. Rev. A **89**, 033856 (2014). Y. Liang, Q. C. Wu, S. L. Su, X. Ji, and S. Zhang, Shortcuts to adiabatic passage for multiqubit controlled-phase gate, Phys. Rev. A **91**, 032304 (2015). X. K. Song, H. Zhang, Q. Ai, J. Qiu, and F. G. Deng, Shortcuts to adiabatic holonomic quantum computation in decoherence-free subspace with transitionless quantum driving algorithm, New J. Phys. **18**, 023001 (2016). K. J. Vahala, Optical microcavities, Nature (London) **424**, 839 (2003). H. Jing, S. K. Özdemir, X.-Y. Lü, J. Zhang, L. Yang, and F. Nori, PT-symmetric phonon laser, Phys. Rev. Lett. **113**, 053604 (2014). B. Peng, S. K. Özdemir, F. Lei, F. Monifi, M. Gianfreda, G.-L. Long, S. Fan, F. Nori, C. M Bender, and L. Yang, Parity–time-symmetric whispering-gallery microcavities, Nat. Phys. **10**, 394 (2014). L. Chang, X. Jiang, S. Hua, C. Yang, J. Wen, L. Jiang, G. Li, G. Wang, and M. Xiao, Parity–time symmetry and variable optical isolation in active–passive-coupled microresonators Nat. photonics **8**, 524 (2014). X. Jiang, L. Shao, S.-X. Zhang, X. Yi, J. Wiersig, L. Wang, Q. Gong, M. Lončar, L. Yang, and Y.-F. Xiao, Chaos-assisted broadband momentum transformation in optical microresonators, Science **358**, 344-347 (2017). X.-Y. Lü, H. Jing, J. Y. Ma, and Y. Wu, PT-Symmetry-Breaking Chaos in Optomechanics, Phys. Rev. Lett. **114**, 253601 (2015). Z. Shen, Y.-L. Zhang, Y. Chen, C.-L. Zou, Y.-F. Xiao, X.-B. Zou, F.-W. Sun, G.-C. Guo, and C.-H. Dong, Experimental realization of optomechanically induced non-reciprocity, Nat. Photonics **10**, 657 (2016). C.-H. Dong, Z. Shen, C.-L. Zou, Y.-L. Zhang, W. Fu, and G.-C. Guo, Brillouin-scattering-induced transparency and non-reciprocal light storage, Nat. Commun. **6**, 6193 (2015). C. L. Degen, F. Reinhard, and P. Cappellaro, Quantum sensing, Rev. Mod. Phys. **89**, 035002 (2017). W. Chen, S. K. Özdemir, G. Zhao, J. Wiersig, and L. Yang, Exceptional points enhance sensing in an optical microcavity, Nature (London) **548**, 192 (2017). N. Zhang, Z. Gu, S. Liu, Y. Wang, S. Wang, Z. Duan, W. Sun, Y.-F. Xiao, S. Xiao, and Q. Song, Far-field single nanoparticle detection and sizing, Optica **4**, 1151 (2018). J. M. Ward, Y. Yang, F. Lei, X.-C. Yu, Y.-F. Xiao, and S. N. Chormaic, Nanoparticle sensing beyond evanescent field interaction with a quasi-droplet microcavity, Optica **5**, 674 (2018). G.-Q. Qin, M. Wang, J.-W. Wen, D. Ruan, and G.-L. Long, Brillouin cavity optomechanics sensing with enhanced dynamical backaction, Photon. Res. **7**, 1440 (2019). C. J. Hood, T. W. Lynn, A. C. Doherty, A. S. Parkins, and H. J. Kimble, The atom-cavity microscope: Single atoms bound in orbit by single photons, Science **287**, 1447 (2000). T. J. Wang and C. Wang, Universal hybrid three-qubit quantum gates assisted by a nitrogen-vacancy center coupled with a whispering-gallery-mode microresonator, Phys. Rev. A **90**, 052310 (2014). X.-F. Liu, T. J. Wang, and C. Wang, Optothermal control of gains in erbium-doped whispering-gallery microresonators, Opt. lett. **43**, 326-329 (2018). H.-R. Wei and F.-G. Deng, Compact quantum gates on electron-spin qubits assisted by diamond nitrogen-vacancy centers inside cavities, Phys. Rev. A **88**, 042323 (2013). B. C. Ren and F. G. Deng, Robust hyperparallel photonic quantum entangling gate with cavity QED, Opt. Express **25**, 10863 (2017). G.-Y. Wang, T. Li, Q. Ai, A. Alsaedi, T. Hayat, and F.-G. Deng, Faithful Entanglement Purification for High-Capacity Quantum Communication with Two-Photon Four-Qubit Systems, Phys. Rev. Appl. **10**, 054058 (2018). M. Li and M. Zhang, Robust universal photonic quantum gates operable with imperfect processes involved in diamond nitrogen-vacancy centers inside low-Q single-sided cavities, Opt. Express **26**, 33129 (2018). S.-S. Chen, H. Zhang, Q. Ai, and G.-J. Yang, Phononic entanglement concentration via optomechanical interactions, Phys. Rev. A **100**, 052306 (2019). J. B. Spring, B. J. Metcalf, P. C. Humphreys, W. S. Kolthammer, X.-M. Jin, M. Barbieri, A. Datta, N. Thomas-Peter, N. K. Langford, D. Kundys, J. C. Gates, B. J. Smith, P. G. R. Smith, and I. A. Walmsley, Boson Sampling on a Photonic Chip, Science **339**, 798 (2013). K. Xia and J. Twamley, All-optical switching and router via the direct quantum control of coupling between cavity modes, Phys. Rev. X **3**, 031013 (2013). Y. D. Wang and A. A. Clerk, Using interference for high fidelity quantum state transfer in optomechanics, Phys. Rev. Lett. **108**, 153603 (2012). L. Tian, Adiabatic state conversion and pulse transmission in optomechanical systems, Phys. Rev. Lett. **108**, 153604 (2012). Y.-D. Wang and A. A. Clerk, Using dark modes for high-fidelity optomechanical quantum state transfer, New J. Phys. **14**, 105010 (2012). H. Zhang, X.-K. Song, Q. Ai, H. Wang, G.-J. Yang, and F.-G. Deng, Fast and robust quantum control for multimode interactions using shortcuts to adiabaticity, Opt. Express **27**, 7384 (2019). X. Zhou, B. J. Liu, L. B. Shao, X. D. Zhang, and Z. Y. Xue, Quantum state conversion in opto-electro-mechanical systems via shortcut to adiabaticity, Laser Phys. Lett. **14**, 095202 (2017). M. Aspelmeyer, T. J. Kippenberg, and F. Marquardt, Cavity optomechanics, Rev. Mod. Phys. **86**, 1391 (2014). M. Gao, F. C. Lei, C. G. Du, and G. L. Long, Self-sustained oscillation and dynamical multistability of optomechanical systems in the extremely-large-amplitude regime, Phys. Rev. A **91**, 013833 (2015). F. C. Lei, M. Gao, C. G. Du, Q. L. Jing, and G. L. Long, Three-pathway electromagnetically induced transparency in coupled-cavity optomechanical system, Opt. Express **23**, 11508–11517 (2015). X. Jiang, M. Wang, M. C. Kuzyk, T. Oo, G. L. Long, and H. Wang, Chip-based silica microspheres for cavity optomechanics, Opt. Express **23**, 27260–27265 (2015). C. Dong, V. Fiore, M. C. Kuzyk, and H. Wang, Optomechanical dark mode, Science **338**, 1609–1613 (2012). M. C. Kuzyk and H. Wang, Controlling multimode optomechanical interactions via interference, Phys. Rev. A **96**, 023860 (2017). J.-Q. Liao and L. Tian, Macroscopic quantum superposition in cavity optomechanics, Phys. Rev. Lett. **116**, 163602 (2016). Y.-C. Liu, Y.-F. Xiao, X. Luan, and C. W. Wong, Dynamic Dissipative Cooling of a Mechanical Resonator in Strong Coupling Optomechanics, Phys. Rev. Lett. **110**, 153606 (2013). Y. S. Park, A. K. Cook and H. Wang, Cavity QED with diamond nanocrystals and silica microspheres, Nano Lett. **6**, 2075-2079 (2006). B. Peng, S. K. Özdemir, W. Chen, F. Nori and L. Yang, What is and what is not electromagnetically induced transparency in whispering-gallery microcavities, Nat. Commun. **5**, 5082 (2014). T. Wang, Y.-Q. Hu, C.-G. Du and G.-L. Long, Multiple EIT and EIA in optical microresonators, Opt. Express **27**, 7344-7353 (2019). V. S. Ilchenko, P. S. Volikov, V. L. Velichansky , F. Treussart, V. Lefevre-Seguin, J. M. Raimond and S. Haroche, Strain-tunable high-Q optical microsphere resonator, Opt. Commun. **145**, 86-90 (1998). W. von Klitzing, R. Long, V. S. Ilchenko, J. Hare and V. Lefevre-Seguin, Frequency tuning of the whispering-gallery modes of silica microspheres for cavity quantum electrodynamics and spectroscopy, Opt. Lett. **26**, 166-168 (2001). K. N. Dinyari, R. J. Barbour, D. A. Golter and H. Wang, Mechanical tuning of whispering gallery modes over a 0.5 THz tuning range with MHz resolution in a silica microsphere at cryogenic temperatures, Opt. Express **19**, 17966-17972 (2011). R. Henze, T. Seifert, J. Ward and O. Benson, Tuning whispering gallery modes using internal aerostatic pressure, Opt. Lett. **36**, 4536-4538 (2011). R. Henze, J. M. Ward and O. Benson, Temperature independent tuning of whispering gallery modes in a cryogenic environment, Opt. Express **21**, 675-680 (2013). Z. H. Zhou, C. L. Zou, Y. Chen, Z. Shen, G. C. Guo and C. H. Dong, Broadband tuning of the optical and mechanical modes in hollow bottle-like microresonators, Opt. Express **25**, 4046-4050 (2017). X.-S. Xu, H. Zhang, M. Wang, D. Ruan, G.-L. Long, Arbitrary function resonance tuner of the optical microcavity with sub-MHz resolution, Opt. Lett. **44**, 3250-3253 (2019). [^1]: These two authors contributed equally to this work. [^2]: [email protected]

ArXiv

--- abstract: 'Based on the concept and properties of $C^{*}$-algebras, the paper introduces a concept of $C_{*}$-class functions. Then by using these functions in $C^{*}$-algebra-valued modular metric spaces of moeini et al. [@MACP], some common fixed point theorems for self-mappings are established. Also, to support of our results an application is provided for existence and uniqueness of solution for a system of integral equations.' --- [**[Common fixed points for $C^{*}$-algebra-valued modular metric spaces via $C_{*}$-class functions with application]{}**]{} Bahman Moeini$^{*,a}$ and Arsalan Hojat Ansari$^{b}$  \ $^{a}$Department of Mathematics, Hidaj Branch, Islamic Azad University, Hidaj, Iran $^{b}$Department of Mathematics, Karaj Branch, Islamic Azad University, Karaj, Iran [^1] [^2] [^3] [^4] [^5] [^6] Introduction ============ As is well known, the Banach contraction mapping principle is a very useful, simple and classical tool in modern analysis, and it has many applications in applied mathematics. In particular, it is an important tool for solving existence problems in many branches of mathematics and physics. In order to generalize this principle, many authors have introduced various types of contraction inequalities (see [@5; @7; @8; @9; @13; @14]). In 2014, Ansari [@16] introduced the concept of $C$-class functions which cover a large class of contractive conditions. Afterwards, Ansari et al.[@ANSEGR] defined and used concept of complex $C$-class functions involving $C$-class functions in complex valued $G_{b}$-metric spaces to obtain some fixed point results. One of the main directions in obtaining possible generalizations of fixed point results is introducing new types of spaces. In 2010 Chistyakov [@CHIS2] defined the notion of modular on an arbitrary set and develop the theory of metric spaces generated by modular such that called the modular metric spaces. Recently, Mongkolkeha et al. [@MONG1; @MONG2] have introduced some notions and established some fixed point results in modular metric spaces. In [@ZHLJ2], Ma et al. introduced the concept of $C^{*}$-algebra-valued metric spaces. The main idea consists in using the set of all positive elements of a unital $C^{*}$-algebra instead of the set of real numbers. This line of research was continued in [@ZKSR; @KPG; @ZHLJ; @DSTK; @ASZ], where several other fixed point results were obtained in the framework of $C^{*}$-algebra valued metric, as well as (more general) $C^{*}$-algebra-valued $b$-metric spaces. Recently, Moeini et al. [@MACP] introduced the concept of $C^{*}$-algebra-valued modular metric spaces which is a generalization of modular metric spaces and next proved some fixed point theorems for self-mappings with contractive conditions on such spaces. In this paper, we introduce a concept of $C_{*}$-class functions on a set of unital $C^{*}$-algebra and via these functions some common fixed point results are proved for self-mappings with contractive conditions in $C^{*}$-algebra-valued modular metric spaces. Also, some examples to elaborate and illustrate of our results are constructed. Finally, as application, existence and uniqueness of solution for a type of system of nonlinear integral equations is discussed. Basic notions ============= Let $X$ be a non empty set, $\lambda\in (0, \infty)$ and due to the disparity of the arguments, function $\omega :(0, \infty)\times X\times X \rightarrow [0, \infty]$ will be written as $\omega_{\lambda}(x, y) = \omega(\lambda,x, y)$ for all $\lambda > 0$ and $x,y \in X$. \[d2.1\][@CHIS1] Let $X$ be a non empty set. a function $\omega :(0, \infty)\times X\times X \rightarrow [0, \infty]$ is said to be a modular metric on $X$ if it satisfies the following three axioms: - given $x,y \in X$, $\omega_{\lambda}(x, y) = 0$ for all $\lambda > 0$ if and only if $x=y$; - $\omega_{\lambda}(x, y) =\omega_{\lambda}(y, x)$ for all $\lambda > 0$ and $x,y \in X$; - $\omega_{\lambda+\mu}(x, y) \leq\omega_{\lambda}(x,z)+\omega_{\mu}(z, y)$ for all $\lambda > 0$ and $x,y,z\in X$, and $(X,\omega)$ is called a modular metric space.\ Recall that a Banach algebra $\mathbb{A}$ (over the field $\mathbb{C}$ of complex numbers) is said to be a $C^{*}$-algebra if there is an involution $*$ in $\mathbb{A}$ (i.e., a mapping $*:\mathbb{A}\rightarrow \mathbb{A}$ satisfying $a^{**}=a$ for each $a\in \mathbb{A}$) such that, for all $a,b\in \mathbb{A}$ and $\lambda, \mu \in \mathbb{C}$, the following holds: - $(\lambda a+\mu b)^{*}=\bar \lambda a^{*}+\bar \mu b^{*}$; - $(ab)^{*}=b^{*}a^{*}$; - $\Vert a^{*}a \Vert=\Vert a \Vert^{2}$. Note that, form $(iii)$, it easy follows that $\Vert a \Vert= \Vert a^{*} \Vert$ for each $a\in \mathbb{A}$. Moreover, the pair $(\mathbb{A},*)$ is called a unital $*$-algebra if $\mathbb{A}$ contains the identity element $1_{\mathbb{A}}$. A positive element of $\mathbb{A}$ is an element $a\in \mathbb{A}$ such that $a^{*}=a$ and its spectrum $\sigma(a)\subset \mathbb{R_{+}}$, where $\sigma(a)=\{\lambda\in \mathbb{R}: \lambda 1_{\mathbb{A}}-a \ \text{is noninvertible}\}$. The set of all positive elements will be denoted by $\mathbb{A_{+}}$. Such elements allow us to define a partial ordering ’$\succeq$’ on the elements of $\mathbb{A}$. That is, $$b\succeq a \ \ \text{if and only if} \ \ b-a\in\mathbb{A_{+}}.$$ If $a\in \mathbb{A}$ is positive, then we write $a\succeq \theta$, where $\theta$ is the zero element of $\mathbb{A}$. Each positive element $a$ of a $C^{*}$-algebra $\mathbb{A}$ has a unique positive square root. From now on, by $\mathbb{A}$ we mean a unital $C^{*}$-algebra with identity element $1_{\mathbb{A}}$. Further, $\mathbb{A_{+}}=\{a\in \mathbb{A}: a\succeq \theta\}$ and $(a^{*}a)^{\frac{1}{2}}=\vert a\vert$. \[l2.2\] [@DRG] Suppose that $\mathbb{A}$ is a unital $C^{*}$-algebra with a unit $1_{\mathbb{A}}$. - For any $x\in \mathbb{A_{+}}$, we have $x \preceq 1_{\mathbb{A}} \Leftrightarrow \Vert x\Vert \leq 1$. - If $a \in \mathbb{A_{+}}$ with $\Vert a\Vert < \frac{1}{2}$, then $1_{\mathbb{A}}- a $ is invertible and $\Vert a(1_{\mathbb{A}}-a)^{-1}\Vert< 1$. - Suppose that $a, b \in \mathbb{A}$ with $a, b \succeq \theta$ and $ab = ba$, then $ab \succeq \theta$. - By $\mathbb{A'}$ we denote the set $\{a\in \mathbb{A}: ab=ba, \forall b\in \mathbb{A}\}$. Let $a \in \mathbb{A'}$ if $b, c \in \mathbb {A}$ with $b \succeq c \succeq \theta$, and $1_{\mathbb{A}}- a \in \mathbb{A'}$ is an invertible operator, then $$(1_{\mathbb{A}}- a)^{-1}b \succeq (1_{\mathbb{A}}- a)^{-1}c.$$ Notice that in a $C^{*}$-algebra, if $\theta \preceq a, b$, one cannot conclude that $\theta \preceq ab$. For example, consider the $C^{*}$-algebra $\mathbb{M}_{2}(\mathbb{C})$ and set $a = \left ( \begin{array}{l} 3\ \ \ 2\\ 2\ \ \ 3 \end{array} \right )$, $b = \left ( \begin{array}{l} 1\ \ \ -2\\ -2\ \ \ 4 \end{array} \right ),$ then $ab = \left ( \begin{array}{l} -1\ \ \ 2\\ -4\ \ \ 8 \end{array} \right )$. Clearly $a,b \in \mathbb{M}_{2}(\mathbb{C})_{+}$, while $ab$ is not.\ \[d2.3\][@MACP] Let $X$ be a non empty set. a function $\omega :(0, \infty)\times X\times X \rightarrow \mathbb{A}_{+}$ is said to be a $C^{*}$-algebra-valued modular metric (briefly, $C^{*}$.m.m) on $X$ if it satisfies the following three axioms: - given $x,y \in X$, $\omega_{\lambda}(x, y) = \theta$ for all $\lambda > 0$ if and only if $x=y$; - $\omega_{\lambda}(x, y) =\omega_{\lambda}(y, x)$ for all $\lambda > 0$ and $x,y \in X$; - $\omega_{\lambda+\mu}(x, y) \preceq \omega_{\lambda}(x,z)+\omega_{\mu}(z, y)$ for all $\lambda, \mu > 0$ and $x,y,z\in X$. The truple $(X,\mathbb{A}, \omega)$ is called a $C^{*}$.m.m space. If instead of $(i)$, we have the condition\ $(i')$ $\omega_{\lambda}(x, x) = \theta$ for all $\lambda > 0$ and $x\in X$, then $\omega$ is said to be a $C^{*}$-algebra-valued pseudo modular metric (briefly, $C^{*}$.p.m.m) on $X$ and if $\omega$ satisfies $(i')$, $(iii)$ and\ $(i'')$ given $x, y \in X$, if there exists a number $\lambda> 0$, possibly depending on $x$ and $y$, such that $\omega_{\lambda}(x, y) = \theta$, then $x = y$, then $\omega$ is called a $C^{*}$-algebra-valued strict modular metric (briefly, $C^{*}$.s.m.m) on $X$.\ A $C^{*}$.m.m (or $C^{*}$.p.m.m, $C^{*}$.s.m.m) $\omega$ on $X$ is said to be convex if, instead of $(iii)$, we replace the following condition: - $\omega_{\lambda+\mu}(x, y) \preceq \frac{\lambda}{\lambda + \mu}\omega_{\lambda}(x,z)+\frac{\mu}{\lambda + \mu}\omega_{\mu}(z, y)$ for all $\lambda, \mu > 0$ and $x,y,z\in X$. Clearly, if $\omega$ is a $C^{*}$.s.m.m, then $\omega$ is a $C^{*}$.m.m, which in turn implies $\omega$ is a $C^{*}$.p.m.m on $X$, and similar implications hold for convex $\omega$. The essential property of a $C^{*}$.m.m $\omega$ on a set $X$ is a following given $x,y\in X$, the function $0 < \lambda \rightarrow \omega_{\lambda}(x, y) \in \mathbb{A}$ is non increasing on $(0,\infty)$. In fact, if $0 < \mu < \lambda$, then we have $$\label{eq2.1} \omega_{\lambda}(x, y)\preceq \omega_{\lambda-\mu}(x, x) +\omega_{\mu}(x, y)=\omega_{\mu}(x, y).$$ It follows that at each point $\lambda> 0$ the right limit $\omega_{\lambda+0}(x, y):=\lim_{\varepsilon \rightarrow +0}\omega_{\lambda+\varepsilon}(x, y)$ and the left limit $\omega_{\lambda-0}(x, y):=\lim_{\varepsilon \rightarrow +0}\omega_{\lambda-\varepsilon}(x, y)$ exist in $\mathbb{A}$ and the following two inequalities hold: $$\label{eq2.2} \omega_{\lambda+0}(x, y)\preceq \omega_{\lambda}(x, y) \preceq\omega_{\lambda-0}(x, y).$$ It can be check that if $x_{0}\in X$, the set $$X_{\omega}=\{ x\in X: \lim_{\lambda \rightarrow \infty}\omega_{\lambda}(x, x_{0})=\theta\},$$ is a $C^{*}$-algebra-valued metric space, called a $C^{*}$-algebra-valued modular space, whose $d_{\omega}^{0}:X_{\omega}\times X_{\omega}\rightarrow \mathbb{A}$ is given by $$d^{0}_{\omega}=\inf \{\lambda >0: \Vert \omega_{\lambda}(x, y)\Vert \leq \lambda \}\ \text{ for all}\ x,y \in X_{\omega}.$$ Moreover, if $\omega$ is convex, the set $X_{\omega}$ is equal to $$X_{\omega}^{*}=\{x\in X: \exists \ \lambda=\lambda(x)>0\ \text{such that}\ \Vert \omega_{\lambda}(x, x_{0})\Vert <\infty \},$$ and $d_{\omega}^{*}:X_{\omega}^{*}\times X_{\omega}^{*}\rightarrow \mathbb{A}$ is given by $$d^{*}_{\omega}=\inf \{\lambda >0: \Vert \omega_{\lambda}(x, y)\Vert \leq 1 \}\ \text{ for all}\ x,y \in X_{\omega}^{*}.$$ It is easy to see that if $X$ is a real linear space, $\rho: X\rightarrow\mathbb{A}$ and $$\label{eq2.3} \omega_{\lambda}(x, y)=\rho(\frac{x-y}{\lambda})\ \text{ for all} \ \lambda>0 \ \text{and}\ x,y\in X,$$ then $\rho$ is $C^{*}$-algebra valued modular (convex $C^{*}$-algebra-valued modular) on $X$ if and only if $\omega$ is $C^{*}$.m.m (convex $C^{*}$.m.m, respectively) on $X$. On the other hand, if $\omega$ satisfy the following two conditions: - $\omega_{\lambda}(\mu x,0)=\omega_{\frac{\lambda}{\mu}}(x, 0)$ for all $\lambda, \mu >0$ and $x\in X$; - $\omega_{\lambda}( x+z,y+z)=\omega_{\lambda}(x, y)$ for all $\lambda>0$ and $x,y,z\in X$. If we set $\rho(x)= \omega_{1}( x,0)$ with (\[eq2.3\]) holds, where $x\in X$, then \(a) $X_{\rho}=X_{\omega}$ is a linear subspace of $X$ and the functional $\Vert x\Vert_{\rho}= d_{\omega}^{0}(x,0), \ x\in X_{\rho}$ is a $F$-norm on $X_{\rho}$; \(b) If $\omega$ is convex, $X_{\rho}^{*}\equiv X_{\omega}^{*}=X_{\rho}$ is a linear subspace of $X$ and the functional $\Vert x\Vert_{\rho}= d_{\omega}^{*}(x,0), \ x\in X_{\rho}^{*}$ is a norm on $X_{\rho}^{*}$. Similar assertions hold if replace $C^{*}$.m.m by $C^{*}$.p.m.m. If $\omega$ is $C^{*}$.m.m in $X$, we called the set $X_{\omega}$ is $C^{*}$.m.m space.\ By the idea of property in $C^{*}$-algebra-valued metric spaces and $C^{*}$-algebra-valued modular spaces, we defined the following: \[d2.4\][@MACP] Let $X_{\omega}$ be a $C^{*}$.m.m space. - The sequence $(x_{n})_{n\in \mathbb{N}}$ in $X_{\omega}$ is said to be $\omega$-convergent to $x\in X_{\omega}$ with respect to $\mathbb{A}$ if $\omega_{\lambda}(x_{n},x)\rightarrow \theta$ as $n\rightarrow \infty$ for all $\lambda>0$. - The sequence $(x_{n})_{n\in \mathbb{N}}$ in $X_{\omega}$ is said to be $\omega$-Cauchy with respect to $\mathbb{A}$ if $\omega_{\lambda}(x_{m},x_{n})\rightarrow \theta$ as $m,n\rightarrow \infty$ for all $\lambda>0$. - A subset $C$ of $X_{\omega}$ is said to be $\omega$-closed with respect to $\mathbb{A}$ if the limit of the $\omega$-convergent sequence of $C$ always belong to $C$. - $X_{\omega}$ is said to be $\omega$-complete if any $\omega$-Cauchy sequence with respect to $\mathbb{A}$ is $\omega$-convergent. - A subset $C$ of $X_{\omega}$ is said to be $\omega$-bounded with respect to $\mathbb{A}$ if for all $\lambda>0$ $\delta_{\omega}(C)=\sup\{\Vert \omega_{\lambda}(x,y)\Vert; \ x,y \in C\}<\infty$. \[d2.5\][@MACP] Let $X_{\omega}$ be a $C^{*}$.m.m space. Let $f, g$ self-mappings of $X_{\omega}$. a point $x$ in $X_{\omega}$ is called a coincidence point of $f$ and $g$ iff $fx = gx$. We shall call $w = fx = gx$ a point of coincidence of $f$ and $g$. \[d2.6\][@MACP] Let $X_{\omega}$ be a $C^{*}$.m.m space. Two maps $f$ and $g$ of $X_{\omega}$ are said to be weakly compatible if they commute at coincidence points. \[d2.7\][@MACP] Let $X_{\omega}$ be a $C^{*}$.m.m space. Two self-mappings $f$ and $g$ of $X_{\omega}$ are occasionally weakly compatible (owc) iff there is a point x in $X_{\omega}$ which is a coincidence point of $f$ and $g$ at which $f$ and $g$ commute. \[l2.8\][@GJBR] Let $X_{\omega}$ be a $C^{*}$.m.m space and $f, g$ owc self-mappings of $X_{\omega}$. If $f$ and $g$ have a unique point of coincidence, $w= fx = gx$, then $w$ is a unique common fixed point of $f$ and $g$. In 2017, Ansari et al. [@ANSEGR] introduced the concept of complex $C$-class functions as follows: \[com C-class\] Suppose $S=\{z\in \mathbb{C}: z\succeq 0\}$, then a continuous function $F:S^{2}\rightarrow \mathbb{C}$ is called a complex $C$-class function if for any $s,t\in S$, the following conditions hold: \(1) $F(s,t)\preceq s $; \(2) $F(s,t)=s$ implies that either $s=0$ or $t=0$. An extra condition on $F$ that $F(0,0)=0$ could be imposed in some cases if required. For examples of these functions see [@ANSEGR]. Main results ============ In this section, we introduce a $C_{*}$-class function. The main idea consists in using the set of elements of a unital $C^{*}$-algebra instead of the set of complex numbers. ($C_{*}$-class function) \[C\*-class\] Suppose $\mathbb{A}$ is a unital $C^{*}$-algebra, then a continuous function $F:\mathbb{A}_{+}\times \mathbb{A}_{+}\rightarrow \mathbb{A}$ is called $C_{*}$-class function if for any $A,B\in\mathbb{A}_{+}$, the following conditions hold: \(1) $F(A,B)\preceq A $; \(2) $F(A,B)=A$ implies that either $A=\theta$ or $B=\theta$. An extra condition on $F$ that $F(\theta,\theta)=\theta$ could be imposed in some cases if required. The letter $\mathcal{C_{*}}$ will denote the class of all $C_{*}$-class functions. The class $\mathcal{C_{*}}$ includes the set of complex $C$-class functions. It is suffitient to take $\mathbb{A}=\mathbb{C}$ in Definition \[C\*-class\]. The following examples show that the class $\mathcal{C_{*}}$ is nonempty: \[ex3.2\] Let $\mathbb{A} = M_{2}(\mathbb{R})$, of all $2 \times 2$ matrices with the usual operation of addition, scalar multiplication, and matrix multiplication. Define norm on $\mathbb{A}$ by $\Vert A\Vert =\Big(\sum^{2}_{i,j=1}\vert a_{ij}\vert^{2}\Big)^{\frac{1}{2}}$, and $*: \mathbb{A} \rightarrow \mathbb{A}$, given by $A^{*}=A$, for all $A\in \mathbb{A}$, defines a convolution on $ \mathbb{A}$. Thus $\mathbb{A}$ becomes a $C^{*}$-algebra. For $$A = \left ( \begin{array}{l} a_{11}\ \ \ a_{12}\\ a_{21}\ \ \ a_{22} \end{array} \right ), B = \left ( \begin{array}{l} b_{11}\ \ \ b_{12}\\ b_{21}\ \ \ b_{22} \end{array} \right ) \in \mathbb{A}= M_{2}(\mathbb{R}) ,$$ we denote $A\preceq B$ if and only if $(a_{ij}-b_{ij})\leq 0$, for all $i,j=1,2$.\ (1)   Define $F_{*}:\mathbb{A}_{+} \times \mathbb{A}_{+}\rightarrow \mathbb{A}$ by $$\ \ F_{*}\Big(\left ( \begin{array}{l} a_{11}\ \ \ a_{12}\\ a_{21}\ \ \ a_{22} \end{array} \right ), \left ( \begin{array}{l} b_{11}\ \ \ b_{12}\\ b_{21}\ \ \ b_{22} \end{array} \right )\ \Big) =\left ( \begin{array}{l} a_{11}-b_{11}\ \ \ a_{12}-b_{12}\\ a_{21}-b_{21}\ \ \ a_{22}-b_{22} \end{array} \right )$$ for all $a_{i,j},b_{i,j}\in \mathbb{R_{+}}$, $(i,j\in \{1,2\})$. Then $F_{*}$ is a $C_{*}$-class function.(2)   Define $F_{*}:\mathbb{A}_{+} \times \mathbb{A}_{+} \rightarrow\mathbb{A}$ by $$\ \ F_{*}\Big(\left ( \begin{array}{l} a_{11}\ \ \ a_{12}\\ a_{21}\ \ \ a_{22} \end{array} \right ), \left ( \begin{array}{l} b_{11}\ \ \ b_{12}\\ b_{21}\ \ \ b_{22} \end{array} \right )\ \Big) =m\left ( \begin{array}{l} a_{11}\ \ \ a_{12}\\ a_{21}\ \ \ a_{22} \end{array} \right )$$ for all $a_{i,j},b_{i,j}\in \mathbb{R_{+}}$, $(i,j\in \{1,2\})$, where, $m\in(0,1)$. Then $F_{*}$ is a $C_{*}$-class function. \[ex4.3\] Let $X=L^{\infty}(E)$ and $H=L^{2}(E)$, where $E$ is a lebesgue measurable set. By $B(H)$ we denote the set of bounded linear operator on Hilbert space $H$. Clearly, $B(H)$ is a $C^{*}$-algebra with the usual operator norm.\ Define $F_{*}:B(H)_{+}\times B(H)_{+}\rightarrow B(H)$ by $$F_{*}(U,V)=U-\varphi(U),$$ where $\varphi:B(H)_{+}\rightarrow B(H)_{+}$ is is a continuous function such that $\varphi(U)=\theta$ if and only if $U=\theta$ ($\theta=0_{B(H)}$). Then $F_{*}$ is a $C_{*}$-class function. Let $\Phi _{u}$ denote the class of the functions $\varphi :\mathbb{A}_{+}\rightarrow \mathbb{A}_{+}$ which satisfy the following conditions: 1. $\varphi$ is continuous and non-decreasing; 2. $\varphi (T)\succ \theta , T\succ \theta$  and $\varphi (\theta)\succeq \theta$. Let $\Psi $ be a set of all continuous functions $\psi :\mathbb{A}_{+} \rightarrow \mathbb{A}_{+}$ satisfying the following conditions: - $\psi$ is continuous and non-decreasing; - $\psi (T)=\theta$ if and only if $T=\theta$. A tripled $(\psi ,\varphi ,F_{*})$ where $\psi \in \Psi ,$ $\varphi \in \Phi _{u} $ and $F_{*}\in \mathcal{C_{*}}$ is said to be monotone if for any $A,B\in \mathbb{A}_{+}$ $$A\preceq B\Longrightarrow F_{*}(\psi (A),\varphi (A))\preceq F_{*}(\psi (B),\varphi (B)).$$ We now give detailed proofs of main results of this paper. \[t3.1\] Let $X_{\omega}$ be a $C^{*}$.m.m space and $I, J, R, S, T, U : X_{\omega} \rightarrow X_{\omega}$ be self-mappings of $X_{\omega}$ such that the pairs $(SR, I)$ and $(TU, J)$ are occasionally weakly compatible. Suppose there exist $a,b,c\in \mathbb{A}$ with $0<\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}\leq1$ such that the following assertion for all $x,y\in X_{\omega}$ and $\lambda >0$ hold: - $\psi(\omega_{\lambda}(SRx,TUy))\preceq F_{*}\Big( \psi(M(x,y)),\varphi(M(x,y)\Big)$, where $$M(x,y)=a^{*}\omega_{\lambda}(Ix,Jy)a+b^{*}\omega_{\lambda}(SRx,Jy)b+c^{*}\omega_{2\lambda}(TUy,Ix)c,$$ $\psi \in \Psi, \varphi\in \Phi_{u}$ and $F_{*}\in \mathcal{C_{*}}$ such that $(\psi, \varphi, F_{*})$ is monotone; - $\Vert \omega_{\lambda}(SRx,TUy)\Vert< \infty$. Then $SR, TU, I$ and $J$ have a common fixed point in $X_{\omega}$. Furthermore if the pairs $(S,R), (S,I),(R,I),(T,J),(T,U), (U,J)$ are commuting pairs of mappings then $I, J, R, S, T$ and $U$ have a unique common fixed point in $X_{\omega}$. Since the pair $(SR, I)$ and $(TU, J)$ are occasionally weakly compatible then there exists $u,v\in X_{\omega}: SRu=Iu$ and $TUv=Jv$. Moreover; $SR(Iu) =I(SRu)$ and $TU(Jv) =J(TUv)$. Now we can assert that $SRu=TUv$. By $(3.1.1)$, we have $$\label{eq3.4} \psi(\omega_{\lambda}(SRu,TUv))\preceq F_{*}\Big( \psi(M(u,v)),\varphi(M(u,v)\Big),$$ where $$\begin{aligned} M(u,v)&= a^{*}\omega_{\lambda}(Iu,Jv)a+b^{*}\omega_{\lambda}(SRu,Jv)b+c^{*}\omega_{2\lambda}(TUv,Iu)c\notag \\ &=a^{*}\omega_{\lambda}(Iu,Jv)a+b^{*}\omega_{\lambda}(Iu,Jv)b+c^{*}\omega_{2\lambda}(Jv,Iu)c. \label{11}\end{aligned}$$ By definition of $C^{*}$.m.m space and inequalites (\[eq2.1\]), (\[eq3.4\]) and (\[11\]), we get $$\label{eq3.5} \begin{array}{rl} &\psi(\omega_{\lambda}(Iu,Jv))=\psi(\omega_{\lambda}(SRu,TUv))\\ &\preceq F_{*}\Big(\psi(a^{*}\omega_{\lambda}(Iu,Jv)a+b^{*}\omega_{\lambda}(Iu,Jv)b+c^{*}(\omega_{\lambda}(Iu,Iu)+\omega_{\lambda}(Iu,Jv))c),\\ &\varphi(a^{*}\omega_{\lambda}(Iu,Jv)a+b^{*}\omega_{\lambda}(Iu,Jv)b+c^{*}(\omega_{\lambda}(Iu,Iu)+\omega_{\lambda}(Iu,Jv))c)\Big)\\ &=F_{*}\Big(\psi(a^{*}\omega_{\lambda}(Iu,Jv)a+b^{*}\omega_{\lambda}(Iu,Jv)b+c^{*}\omega_{\lambda}(Iu,Jv)c),\\ &\varphi(a^{*}\omega_{\lambda}(Iu,Jv)a+b^{*}\omega_{\lambda}(Iu,Jv)b+c^{*}\omega_{\lambda}(Iu,Jv)c)\Big)\\ &=F_{*}\Big(\psi(a^{*}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}a+b^{*}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}b\\ &+c^{*}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}c),\\ &\varphi(a^{*}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}a+b^{*}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}b\\ &+c^{*}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}c)\Big)\\ &=F_{*}\Big(\psi((a(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})^{*}(a(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})+(b(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})^{*}(b(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})\\ &+(c(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})^{*}(c(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})),\\ &\varphi((a(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})^{*}(a(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})+(b(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})^{*}(b(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})\\ &+(c(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}})^{*}(c(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}))\Big)\\ &\preceq F_{*}\Big(\psi(\Vert a(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}\Vert^{2}1_{\mathbb{A}}+\Vert b(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}\Vert^{2}1_{\mathbb{A}}+\Vert c(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}\Vert^{2}1_{\mathbb{A}}),\\ &\varphi(\Vert a(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}\Vert^{2}1_{\mathbb{A}}+\Vert b(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}\Vert^{2}1_{\mathbb{A}}+\Vert c(\omega_{\lambda}(Iu,Jv))^{\frac{1}{2}}\Vert^{2}1_{\mathbb{A}})\Big)\\ &=F_{*}\Big(\psi(\Vert \omega_{\lambda}(Iu,Jv)\Vert (\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2})1_{\mathbb{A}}),\\ &\varphi(\Vert \omega_{\lambda}(Iu,Jv)\Vert (\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2})1_{\mathbb{A}})\Big). \end{array}$$ So, $$\label{eq3.6} \begin{array}{rl} &\psi(\Vert \omega_{\lambda}(Iu,Jv) \Vert 1_{\mathbb{A}})\\ &\leq F_{*}\Big(\psi(\Vert \omega_{\lambda}(Iu,Jv)\Vert (\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2})1_{\mathbb{A}}),\varphi(\Vert \omega_{\lambda}(Iu,Jv)\Vert (\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2})1_{\mathbb{A}})\Big)\\ &\preceq F_{*}\Big(\psi(\Vert \omega_{\lambda}(Iu,Jv)\Vert 1_{\mathbb{A}}),\varphi(\Vert \omega_{\lambda}(Iu,Jv)\Vert 1_{\mathbb{A}})\Big). \end{array}$$ Thus, $\psi(\Vert \omega_{\lambda}(Iu,Jv) \Vert 1_{\mathbb{A}})=\theta $ or $\varphi(\Vert \omega_{\lambda}(Iu,Jv) \Vert 1_{\mathbb{A}})=\theta $, which means $Iu=Jv$. Hence $SRu=TUv$ and thus $$\label{eq3.7} SRu=Iu=TUv=Jv.$$ Moreover, if there is another point $z$ such that $SRz=Iz$, and using condition $(3.1.1)$ $$\label{eq3.8} \begin{array}{rl} &\psi(\omega_{\lambda}(SRz,TUv))\preceq F_{*}\Big(\psi(a^{*}\omega_{\lambda}(Iz,Jv)a+b^{*}\omega_{\lambda}(SRz,Jv)b+c^{*}\omega_{2\lambda}(TUv,Iz)c),\\ &\varphi(a^{*}\omega_{\lambda}(Iz,Jv)a+b^{*}\omega_{\lambda}(SRz,Jv)b+c^{*}\omega_{2\lambda}(TUv,Iz)c)\Big) \\ &=F_{*}\Big(\psi(a^{*}\omega_{\lambda}(SRz,TUv)a+b^{*}\omega_{\lambda}(SRz,TUv)b+c^{*}\omega_{2\lambda}(SRz,TUv)c),\\ &\varphi(a^{*}\omega_{\lambda}(SRz,TUv)a+b^{*}\omega_{\lambda}(SRz,TUv)b+c^{*}\omega_{2\lambda}(SRz,TUv)c)\Big). \end{array}$$ By above similar way, we conclude that $$\begin{aligned} &\psi(\Vert \omega_{\lambda}(SRz,TUv)\Vert 1_{\mathbb{A}})\\ &\preceq F_{*}\Big(\psi(\Vert \omega_{\lambda}(SRz,TUv)\Vert (\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}))1_{\mathbb{A}},\\ &\varphi(\Vert \omega_{\lambda}(SRz,TUv)\Vert (\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2})1_{\mathbb{A}})\Big)\\ &\preceq F_{*}\Big(\psi(\Vert \omega_{\lambda}(SRz,TUv)\Vert 1_{\mathbb{A}},\varphi(\Vert \omega_{\lambda}(SRz,TUv)\Vert 1_{\mathbb{A}})\Big).\end{aligned}$$ Therefore, $\psi(\Vert \omega_{\lambda}(SRz,TUv)\Vert 1_{\mathbb{A}})=\theta$ or $\varphi(\Vert \omega_{\lambda}(SRz,TUv)\Vert 1_{\mathbb{A}})=\theta$, which means $SRz=TUv$, and so, $$\label{eq3.9} SRu=Iu=TUv=Jv.$$ Thus from equation (\[eq3.8\]) and (\[eq3.9\]) it follows that $SRu=SRz$. Hence, $w=SRu=Iu$ for some $w\in X_{\omega}$ is the unique point of coincidence of $SR$ and $I$. Then by Lemma \[l2.8\], $w$ is a unique common fixed point of $SR$ and $I$. So, $SRw=Iw=w$.\ Similarly, there is another common fixed point $w'\in X_{\omega}: TUw'=Jw'=w'$.\ For the uniqueness, by $(3.1.1)$ we have $$\label{eq3.10} \begin{array}{rl} &\psi(\omega_{\lambda}(SRw,TUw'))=\psi(\omega_{\lambda}(w,w') )\\ &\preceq F_{*}\Big(\psi(a^{*}\omega_{\lambda}(Iw,Jw')a+b^{*}\omega_{\lambda}(SRw,Jw')b+c^{*}\omega_{2\lambda}(TUw,Iw')c),\\ &\varphi(a^{*}\omega_{\lambda}(Iw,Jw')a+b^{*}\omega_{\lambda}(SRw,Jw')b+c^{*}\omega_{2\lambda}(TUw,Iw')c)\Big)\\ &=F_{*}\Big(\psi(a^{*}\omega_{\lambda}(w,w')a+b^{*}\omega_{\lambda}(w,w')b+c^{*}\omega_{2\lambda}(w,w')c,\\ &\varphi(a^{*}\omega_{\lambda}(w,w')a+b^{*}\omega_{\lambda}(w,w')b+c^{*}\omega_{2\lambda}(w,w')c)\Big). \end{array}$$ Thus, $$\begin{aligned} &\psi(\Vert \omega_{\lambda}(w,w')\Vert 1_{\mathbb{A}})\\ &\preceq F_{*}\Big(\psi(\Vert \omega_{\lambda}(w,w')\Vert(\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}) 1_{\mathbb{A}}), \\ &\varphi(\Vert \omega_{\lambda}(w,w')\Vert(\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}) 1_{\mathbb{A}})\Big)\\ &\preceq F_{*}\Big(\psi(\Vert \omega_{\lambda}(w,w')\Vert 1_{\mathbb{A}}), \varphi(\Vert \omega_{\lambda}(w,w')\Vert 1_{\mathbb{A}})\Big).\end{aligned}$$ So, $\psi(\Vert \omega_{\lambda}(w,w')\Vert 1_{\mathbb{A}})=\theta$ or $\varphi(\Vert \omega_{\lambda}(w,w')\Vert 1_{\mathbb{A}})=\theta$. Hence $w=w'$. Therefore, $w$ is a unique common fixed point of $SR, TU, I$ and $J$.\ Furthermore, if we take pairs $(S, R),\ (S, I),\ (R, I),\ (T, J),\ (T, U),\ (U, J)$ are commuting pairs then $$\begin{array}{rl} & Sw = S(SRw) = S(RS)w = SR(Sw)\\ &Sw = S(Iw) = S(RS)w = I(Sw)\\ &Rw = R(SRw) = RS(Rw) = SR(Rw)\\ &Rw = R(Iw) = (Rw), \end{array}$$ this shows that $Sw$ and $Rw$ is common fixed point of $(SR, I)$ and this gives $SRw = Sw = Rw = Iw = w$. Similarly, we have $TUw = Tw = Uw = Jw = w$. Hence, $w$ is a unique common fixed point of $S,\ R,\ I,\ J,\ T,\ U$. \[c3.2\] Let $X_{\omega}$ be a $C^{*}$.m.m space and $I, J, S, T : X_{\omega} \rightarrow X_{\omega}$ be self-mappings of $X_{\omega}$ such that the pairs $(S, I)$ and $(T, J)$ are occasionally weakly compatible. Suppose there exist $a,b,c\in \mathbb{A}$ with $0<\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}\leq1$ such that the following assertion for all $x,y\in X_{\omega}$ and $\lambda >0$ hold: - $\psi(\omega_{\lambda}(Sx,Ty))\preceq F_{*}\Big(\psi(N(x,y)),\varphi(N(x,y))\Big)$ where, $$\begin{aligned} N(x,y)=a^{*}\omega_{\lambda}(Ix,Jy)a+b^{*}\omega_{\lambda}(Sx,Jy)b+c^{*}\omega_{2\lambda}(Ty,Ix)c\end{aligned}$$ $\psi\in \Psi, \ \varphi\in \Phi_{u}$ and $F_{*}\in \mathcal{C_{*}}$ such that $(\psi, \varphi, F_{*})$ is monotone; - $\Vert \omega_{\lambda}(Sx,Ty)\Vert< \infty$. Then $S, T, I$ and $J$ have a unique common fixed point in $X_{\omega}$. If we put $R=U :=Ix_{\omega}$ where $Ix_{\omega}$ is an identity mapping on $X_{\omega}$, the result follows from Theorem \[t3.1\]. \[c3.3\] Let $X_{\omega}$ be a $C^{*}$.m.m space and $ S, T : X_{\omega} \rightarrow X_{\omega}$ be self-mappings of $X_{\omega}$ such that $S$ and $T$ are occasionally weakly compatible. Suppose there exist $a,b,c\in \mathbb{A}$ with $0<\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}\leq1$ such that the following assertion for all $x,y\in X_{\omega}$ and $\lambda >0$ hold: - $\psi(\omega_{\lambda}(Tx,Ty))\preceq F_{*}\Big(\psi(O(x,y)),\varphi(O(x,y))\Big)$ where, $$\begin{aligned} O(x,y)=a^{*}\omega_{\lambda}(Sx,Sy)a+b^{*}\omega_{\lambda}(Tx,Sy)b+c^{*}\omega_{2\lambda}(Ty,Sx)c\end{aligned}$$ $\psi\in \Psi, \ \varphi\in \Phi_{u}$ and $F_{*}\in \mathcal{C_{*}}$ such that $(\psi, \varphi, F_{*})$ is monotone; - $\Vert \omega_{\lambda}(Tx,Ty)\Vert< \infty$. Then $S$ and $T$ have a unique common fixed point in $X_{\omega}$. If we put $I=J:=S$, and $S:=T$ in $(3.2.1)$ and $(3.2.2)$ the result follows from Theorem \[t3.1\]. \[c3.4\] Let $X_{\omega}$ be a $C^{*}$.m.m space and $ S, T : X_{\omega} \rightarrow X_{\omega}$ be self-mappings of $X_{\omega}$ such that $S$ and $T$ are occasionally weakly compatible. Suppose there exist $a\in \mathbb{A}$ with $0<\Vert a\Vert\leq1$ such that the following assertion for all $x,y\in X_{\omega}$ and $\lambda >0$ hold: - $\psi(\omega_{\lambda}(Tx,Ty))\preceq F_{*}\Big(\psi(a^{*}\omega_{\lambda}(Sx,Sy)a), \varphi(a^{*}\omega_{\lambda}(Sx,Sy)a)\Big)$, where, $\psi\in \Psi, \ \varphi\in \Phi_{u}$ and $F_{*}\in \mathcal{C_{*}}$ such that $(\psi, \varphi, F_{*})$ is monotone; - $\Vert \omega_{\lambda}(Tx,Ty)\Vert< \infty$. Then $S$ and $T$ have a unique common fixed point in $X_{\omega}$. If we put $b=c:=\theta$, in $(3.3.1)$ the result follows from Corollary \[c3.3\]. Examples ======== In this section we furnish some nontrivial examples in favour of our results. \[ex4.1\] Let $X=\mathbb{R}$ and consider, $\mathbb{A} = M_{2}(\mathbb{R})$ as in Example \[ex3.2\]. Define $\omega: (0,\infty)\times X \times X\rightarrow \mathbb{A_{+}}$ by $$\omega_{\lambda}(x,y) = \left ( \begin{array}{l} \vert \frac{x-y}{\lambda}\vert\ \ \ \ \ \ \ \ 0\\ \ \ 0\ \ \ \ \ \ \ \ \vert \frac{x-y}{\lambda}\vert \end{array} \right ),$$ for all $x,y\in X$ and $\lambda>0$. It is easy to check that $\omega$ satisfies all the conditions of Definition \[d2.3\]. So, $(X, \mathbb{A}, \omega)$ is a $C^{*}$.m.m space. \[ex4.2\] Let $X = \{ \frac{1}{c^{n}}: n=1,2, \cdots \}$ where $0 < c < 1$ and $\mathbb{A} = M_{2}(\mathbb{R})$. Define $\omega: (0,\infty)\times X \times X\rightarrow \mathbb{A_{+}}$ by $$\omega_{\lambda}(x,y) = \left ( \begin{array}{l} \Vert \frac{x-y}{\lambda}\Vert\ \ \ \ \ \ \ \ 0\\ \ \ 0\ \ \ \ \ \ \ \ \alpha \Vert \frac{x-y}{\lambda}\Vert \end{array} \right ),$$ for all $x,y\in X$, $\alpha\geq 0$ and $\lambda>0$. Then it is easy to check that $\omega$ is a $C^{*}$.m.m. \[ex4.3\] Let $X=L^{\infty}(E)$ and $H=L^{2}(E)$, where $E$ is a Lebesgue measurable set. By $B(H)$ we denote the set of bounded linear operator on Hilbert space $H$. Clearly, $B(H)$ is a $C^{*}$-algebra with the usual operator norm.\ Define $\omega:(0,\infty)\times X\times X\rightarrow B(H)_{+}$ by $$\omega_{\lambda}(f,g)=\pi_{\vert \frac{f-g}{\lambda}\vert},\ \ \ (\forall f,g\in X),$$ where $\pi_{h}:H\rightarrow H$ is the multiplication operator defined by $$\pi_{h}(\phi)=h.\phi,$$ for $\phi \in H$. Then $\omega$ is a $C^{*}$.m.m and $(X_{\omega},B(H),\omega)$ is a $\omega$-complete $C^{*}$.m.m space. It suffices to verify the completeness of $X_{\omega}$. For this, let $\{f_{n}\}$ be a $\omega$-Cauchy sequence with respect to $B(H)$, that is for an arbitrary $\varepsilon> 0$, there is $N\in \mathbb{N}$ such that for all $m, n \geq N$, $\Vert \omega_{\lambda}(f_{m},f_{n})\Vert =\Vert \pi_{\vert \frac{f_{m}-f_{n}}{\lambda}\vert}\Vert =\Vert \frac{f_{m}-f_{n}}{\lambda}\Vert_{\infty}\leq \varepsilon$, so $\{f_{n}\}$ is a Cauchy sequence in Banach space $X$. Hence, there is a function $f \in X$ and $N_{1} \in \mathbb{N}$ such that $\Vert \frac{f_{n}- f}{\lambda}\Vert_{\infty} \leq \varepsilon,\ \ (n \geq N_{1})$. It implies that $\Vert \omega_{\lambda}(f_{n}, f)\Vert= \Vert \pi_{\vert \frac{f_{n}-f}{\lambda}\vert}\Vert= \Vert \frac{f_{n}- f}{\lambda}\Vert _{\infty} \leq \varepsilon,\ \ (n \geq N_{1})$. Consequently, the sequence $\{f_{n}\}$ is a $\omega$-convergent sequence in $X_{\omega}$ and so $X_{\omega}$ is a $\omega$-complete $C^{*}$.m.m space. \[ex4.4\] Let $(X,\mathbb{A}, \omega)$ is $C^{*}$.m.m space defined as in Example \[ex4.1\]. Define $S,T,I,J: X_{\omega}\rightarrow X_{\omega}$ by $$Sx=Tx=2,\ \ Jx=4-x,\ \ Ix= \left \{ \begin{array}{l} \frac{2x}{3}\ \ \ \ \ \ \ \ \ if\ x\in (-\infty,2),\\ 2\ \ \ \ \ \ \ \ \ \ if\ x=2,\\ 0\ \ \ \ \ \ \ \ \ \ if\ x\in (2,\infty). \end{array} \right.$$ Suppose, $$\left \{ \begin{array}{l} \psi:\mathbb{A_{+}}\rightarrow \mathbb{A_{+}}\\ \psi(A)=2A, \end{array} \right. \ \ \left \{ \begin{array}{l} \varphi:\mathbb{A_{+}}\rightarrow \mathbb{A_{+}}\\ \varphi(A)=A, \end{array} \right. \ \ \left \{ \begin{array}{l} F_{*}:\mathbb{A_{+}}\times \mathbb{A_{+}}\rightarrow \mathbb{A}\\ F_{*}(A,B)=A-B. \end{array} \right.$$ Then, $(\psi,\varphi, F_{*})$ is monotone. For all $x,y\in X_{\omega}=\mathbb{R}$ and $\lambda >0$, we have $0=\Big\Vert \left ( \begin{array}{l} 0\ \ \ \ \ \ 0\\ 0\ \ \ \ \ \ 0 \end{array} \right )\Big\Vert=\Vert \omega_{\lambda}(Sx,Ty)\Vert< \infty$. For every $a,b,c\in \mathbb{A}$ with $0<\Vert a\Vert^{2}+\Vert b\Vert^{2}+\Vert c\Vert^{2}\leq1$, we get $\left ( \begin{array}{l} 0\ \ \ \ \ \ 0\\ 0\ \ \ \ \ \ 0 \end{array} \right )=\psi(\omega_{\lambda}(Sx,Ty))\preceq F_{*}\Big(\psi(M(x,y)),\varphi(M(x,y))\Big)$, for all $x,y\in X_{\omega}$ and $\lambda >0$. Also clearly, the pairs $(S, I)$ and $(T, J)$ are occasionally weakly compatible. So all the conditions of the Corollary \[c3.2\] are satisfied and $x=2$ is a unique common fixed point of $S,\ T,\ I$ and $J$. Application =========== Remind that if for $\lambda>0$ and $x,y\in L^{\infty}(E)$, define $\omega: (0,\infty)\times L^{\infty}(E)\times L^{\infty}(E) \rightarrow B(H)_{+}$ by $$\omega_{\lambda}(x,y)=\pi_{\vert \frac{x-y}{\lambda}\vert},$$ where, $\pi_{h}:H\rightarrow H$ be defined as in Example \[ex4.3\], then $( L^{\infty}(E)_{\omega}, B(H), \omega)$ is a $\omega$-complete $C^{*}$.m.m space.\ Let $E$ be a Lebesgue measurable set, $X=L^{\infty}(E)$ and $H=L^{2}(E)$ be the Hilbert space. Consider the following system of nonlinear integral equations: $$\label{eq5.11} x(t)=w(t)+k_{i}(t,x(t))+\mu\int_{E}n(t,s)h_{j}(s,x(s))ds,$$ for all $t\in E$, where $w\in L^{\infty}(E)_{\omega}$ is known, $k_{i}(t,x(t)),\ n(t,s),\ h_{j}(s,x(s))$, $i,j=1,2$ and $i\neq j$ are real or complex valued functions that are measurable both in $t$ and $s$ on $E$ and $\mu$ is real or complex number, and assume the following conditions:\ - $sup_{s\in E}\int_{E}|n(t,s)|dt=M_{1}<+\infty$, - $k_{i}(s,x(s))\in L^{\infty}(E)_{\omega}$ for all $x\in L^{\infty}(E)_{\omega},$ and there exists $L_{1}>1$ such that for all $s\in E$, $$\frac{|k_{1}(s,x(s))-k_{2}(s,y(s))|}{\sqrt{2}}\geq L_{1}|x(s)-y(s)|\ \ \text{for all}\ x,y\in L^{\infty}(E)_{\omega},$$ - $h_{i}(s,x(s))\in L^{\infty}(E)_{\omega}$ for all $x\in L^{\infty}(E)_{\omega}$, and there exists $L_{2}>0$ such that for all $s\in E$, $$|h_{1}(s,x(s))-h_{2}(s,y(s))|\leq L_{2}|x(s)-y(s)|\ \ \text{for all}\ x,y\in L^{\infty}(E)_{\omega},$$ - there exists $x(t)\in L^{\infty}(E)_{\omega}$ such that $x(t)-w(t)-\mu\int_{E}n(t,s)h_{1}(s,x(s))ds=k_{1}(t,x(t))$, implies $$\begin{array}{rl} &k_{1}(t,x(t))-w(t)-\mu\int_{E}n(t,s)h_{1}(s,k_{1}(s,x(s)))ds\\ &= k_{1}(t,x(t)-w(t)-\mu\int_{E}n(t,s)h_{1}(s,x(s))ds). \end{array}$$ - there exists $y(t)\in L^{\infty}(E)_{\omega}$ such that $y(t)-w(t)-\mu\int_{E}n(t,s)h_{2}(s,y(s))ds=k_{2}(t,y(t))$, implies $$\begin{array}{rl} &k_{2}(t,y(t))-w(t)-\mu\int_{E}n(t,s)h_{i}(s,k_{2}(s,y(s)))ds\\ &= k_{2}(t,y(t)-w(t)-\mu\int_{E}n(t,s)h_{2}(s,y(s))ds). \end{array}$$ \[t5.1\] With the assumptions ()-(e), the system of nonlinear integral equations (\[eq5.11\]) has a unique solution $x^{*}$ in $L^{\infty}(E)_{\omega}$ for each real or complex number $\mu$ with $\frac{1+|\mu|L_{2}M_{1}}{L_{1}}\leq1$. Define $$Sx(t)=x(t)-w(t)-\mu\int_{E}n(t,s)h_{1}(s,x(s))ds,$$ $$Tx(t)=x(t)-w(t)-\mu\int_{E}n(t,s)h_{2}(s,x(s))ds,$$ $$Ix(t)=k_{1}(t,x(t)),\ Jx(t)=k_{2}(t,x(t)).$$ Set $a=\sqrt{\frac{1+\vert \mu\vert M_{1}L_{2}}{L_{1}}}.1_{B(H)}$, $b=c=\theta=0_{B(H)}$, then $a\in B(H)_{+}$ and $0 <\Vert a \Vert^{2}+\Vert b \Vert^{2}+\Vert c \Vert^{2}=\frac{1+\vert \mu\vert M_{1}L_{2}}{L_{1}}\leq1$.Define $$\left \{ \begin{array}{l} \psi:B(H)_{+}\rightarrow B(H)_{+}\\ \psi(B)=\frac{1}{2}B, \end{array} \right. \ \ \left \{ \begin{array}{l} \varphi:B(H)_{+}\rightarrow B(H)_{+}\\ \varphi(B)=\frac{1}{4}B, \end{array} \right. \ \ \left \{ \begin{array}{l} F_{*}:B(H)_{+}\times B(H)_{+}\rightarrow B(H)\\ F_{*}(A,B)=\frac{1}{\sqrt{2}}A. \end{array} \right.$$ Then, $(\psi,\varphi, F_{*})$ is monotone.For any $h\in H$, we have $$\begin{array}{rl} &\Vert \psi(\omega_{\lambda}(Sx,Ty))\Vert =\frac{1}{2}\sup_{\Vert h \Vert=1}(\pi_{\vert \frac{Sx-Ty}{\lambda} \vert}h, h)\\\\ &=\sup_{\Vert h \Vert=1}\int_{E}\Big[ \frac{1}{2\lambda}\Big\vert (x-y)+\mu\int_{E}n(t,s)(h_{2}(s,y(s)-h_{1}(s,x(s))ds \Big\vert\Big]h(t)\overline{h(t)}dt\\\\ &\leq \sup_{\Vert h \Vert=1}\int_{E}\Big[ \frac{1}{2\lambda}\Big\vert (x-y)+\mu\int_{E}n(t,s)(h_{2}(s,y(s)-h_{1}(s,x(s))ds \Big\vert\Big]\vert h(t)\vert^{2}dt\\\\ &\leq \frac{1}{2\lambda}\sup_{\Vert h \Vert=1}\int_{E} \vert h(t)\vert^{2}dt\Big[\Vert x-y\Vert_{\infty}+ \vert \mu \vert M_{1}L_{2}\Vert x-y\Vert_{\infty} \Big]\\\\ &\leq(\frac{1+\vert \mu\vert M_{1}L_{2}}{2\lambda})\Vert x-y\Vert_{\infty}\\\\ &\leq\frac{1}{\sqrt{2}}(\frac{1+\vert \mu\vert M_{1}L_{2}}{2L_{1}})\Vert \frac{k_{1}(t,x(t))-k_{2}(t,y(t))}{\lambda}\Vert_{\infty}\\\\ &=\frac{1}{\sqrt{2}}.\frac{1}{2}(\frac{1+\vert \mu\vert M_{1}L_{2}}{L_{1}})\Vert \omega_{\lambda}(Ix, Jy) \Vert\\\\ &=\frac{1}{\sqrt{2}}.\frac{1}{2}\Vert a\Vert^{2}\Vert \omega_{\lambda}(Ix, Jy) \Vert\\\\ &=\frac{1}{\sqrt{2}}.\frac{1}{2}\Big(\Vert a\Vert^{2}\Vert \omega_{\lambda}(Ix, Jy) \Vert+ \Vert b\Vert^{2}\Vert \omega_{\lambda}(Sx, Jy) \Vert+\Vert c\Vert^{2}\Vert \omega_{2\lambda}(Ty, Ix) \Vert \Big)\\\\ &=\Vert F_{*}\Big(\psi(N(x,y)), \varphi(N(x,y))\Big)\Vert \end{array}$$ Then, $\Vert \psi(\omega_{\lambda}(Sx,Ty))\Vert \leq \Vert F_{*}\Big(\psi(N(x,y)), \varphi(N(x,y))\Big)\Vert $, for all $x,y\in L^{\infty}(E)_{\omega}$ and $\lambda>0$. Also by conditions $(d)$ and $(e)$ the pairs $(S,I)$ and $(T,J)$ are occasionally weakly compatible. Therefore, by the Corollary \[c3.2\], there exists a unique common fixed point $x^{*}\in L^{\infty}(E)_{\omega}$ such that $x^{*}=Sx^{*}=Tx^{*}=Ix^{*}=Jx^{*}$, which proves the existence of unique solution of (\[eq5.11\]) in $L^{\infty}(E)_{\omega}$. This completes the proof. [99]{} A.H. Ansari, Note on  $\varphi $-$\psi $ -contractive type mappings and related fixed point, The 2nd Regional Conference on Mathematics And Applications, Payame Noor University, 2014, pages 377-380. A.H. Ansari, O. Ege and S. Randenović, Some fixed point results on complex valued $G_{b}$-metric spaces, RACSAM (2017). doi:10.1007/s13398-017-0391-x. A. Branciari, A fixed point theorem for mappings satisfying a general contractive condition of integral type, Hindawi Publishing Corpration, Inter. J. Math. Math. Sci., 29 (2002), 531-536. V.V. Chistyakov, , , 14 (2008), 3-25. V.V. Chistyakov, , , 72 (2010), 1-14. S. Dhompongsa, H. Yingtaweesittikul, Fixed point for multivalued mappings and the metric completeness, Fixed Point Theory and Applications, 2009, 15 pages, Article ID 972395. R. Douglas, , (1998). G. Jungck and B.E. Rhoades , , 7(2) (2006), 287-296. Z. Kadelburg and S. Radenovi$\Acute{c}$, , , 2016, Article ID 53 (2016). T. Kamran, M. Postolache, A. Ghiura, S. Batul and R. Ali, , , 2016, Article ID 10 (2016). M. Kikkawa, T. Suzuki, Three fixed point theorems for generalized contractions with constants in complete metric spaces, Nonlinear Analysis, 69 (2008), 2942-2949. Z. Ma and L. Jiang, , , 2015, Article ID 222 (2015). Z. Ma, L. Jiang and H. Sun, , , 2014, Article ID 206 (2014). B. Moeini, A.H. Ansari and C. Park, $C^{*}$-algebra-valued modular metric spaces and related fixed point results, submitted. C. Mongkolkeha, W. Sintunavarat and P. Kumam, , , 2011(2011), Article ID 93. C. Mongkolkeha, W. Sintunavarat and P. Kumam, , , 2012(2012), Article ID 103. G. Mot, A. Petruşel, Fixed point theory for a new type of contractive multivalued operators, Nonlinear Analysis, 70 (2009), 3371-3377. D. Shehwar and T. Kamran, , , 2015, Article ID 304 (2015). T. Suzuki, A new type of fixed point theorem in metric spaces, Nonlinear Analysis, 71 (2009), 5313-5317. T. Suzuki, A generalized Banach contraction principle that characterizes metric completeness, Proc. Amer. Math. Soc., 136 (2008), 1861-1869. A. Zada, S. Saifullah and Z. Ma, , , Vol. 11, no. 1 (2016), 23-27. [^1]: $^*$ Corresponding author [^2]: E-mail:[email protected] [^3]: E-mail:[email protected] [^4]: *2010 Mathematics Subject Classification:* 47H10, 46L07, 46A80. [^5]: *Keywords:* $C_{*}$-class function, $C^{*}$-algebra-valued modular metric space, common fixed point, occasionally weakly compatible, integral equation. [^6]: **

ArXiv

--- author: - 'A. Finoguenov, R.P. Kudritzki, C. Jones' date: 'Received October 5 2001; accepted March 20 2002' title: 'Probing the Intracluster star-light with Chandra' --- Introduction ============ The integrated light and the mass-to-light ratio in galaxy clusters plays an important role in several aspects of observational cosmology. Carlberg (1996), for example, use the mean M/L ratio determined for the CNOC cluster sample to derive an estimate of the matter density of the universe, $\Omega_{\rm o}$. Mass-to-light ratios are important diagnostics of biased/non-biased star-formation in clusters (Bahcall 2000), a key quantity for the explanation of the heavy element enrichment of the ICM (Arnaud 1992; Renzini 1993), and important for describing the statistical properties of cluster surveys (Borgani & Guzzo 2001). The classical approach to estimate the optical luminosity of the cluster was to sum the light from the individual member galaxies. A detailed review on the luminosity function of galaxies can be found in Binggeli, Sandage, Tammann (1988). Recent studies provide more refined integrated light estimates, adding in dwarf galaxies and an estimate for the amount of diffuse stellar light (also called the intergalactic stellar population, IGSP), as was recently done for the Coma cluster (Gregg & West 1998). The result of galaxy merging and galaxy harassment as proposed by Moore (1996) will leave stellar debris that accumulates in the cluster potential. Therefore the observation of an intracluster stellar component would place interesting constraints on the history of galaxy collisions in clusters. The efficiency of tidal stripping is controlled by the probability of a galaxy crossing the cluster core and by the duration of the impact. As simulations show, the two effects compensate each other and the tidal radius is similar among clusters and equals $100h^{-1}$ kpc (Klypin 1999). Observationally, the determination of the level of such diffuse emission is very difficult due to its low optical surface brightness. It should be observed best in nearby clusters, but there it can easily be confused with background and foreground emission. The present sensitivity limit for optical searches for IGSP corresponds to 20% of the total light of the cluster (Melnick & Sargent 1977). This value may still be considered uncertain, as values up to 80% are still cited ( Feldmeier 1998). One of the solutions proposed to solve this problem is the observation of individual bright objects, that represent the light density of the IGSP. The most important sources are supernovae, red giants and planetary nebulae. For the Virgo Cluster, the detection of intra-cluster red giants (Ferguson 1998) and PN (Arnaboldi 1996, Mendez 1997) provided evidence for the existence of an intracluster stellar population. However recently Kudritzki (2000) showed that many of the intra-cluster PN candidates are Lyman-alpha emitting background galaxies, although Freeman (2000) demonstrated that a significant fraction must be real PN belonging to an intracluster stellar population. Mass estimates for this population suggest that it contains roughly 10% to 20% of the stellar mass in the Virgo cluster galaxies. Durrell (2002) using independent method based on the red giant stars confirmed this conclusion. With the advent of high-angular resolution, X-ray astronomy could play a leading role in this seemingly purely optical field, via studies of the intracluster X-ray binaries. The method ========== Chandra observations reveal a numerous population of galactic X-ray binaries (XRB’s) in most elliptical galaxies (Sarazin, Irwin, Bregman 2000; Finoguenov & Jones 2001). With luminosities of $10^{37-39}$ ergs/s, these sources are easily detected even at the distance of Virgo cluster galaxies. The total emitted flux is proportional to the optical luminosity of stars (Matsushita 1998) and within individual galaxies, the number density of the XRB’s follows the distribution of the stellar light (Finoguenov & Jones 2002; Sarazin, Irwin, Bregman 2001). Finding one XRB with a luminosity[^1] exceeding $10^{37}$ ergs/s corresponds to revealing $4.3\times10^8 L_{\odot}$ light in the B band. Although, the sensitivity of the PN method is higher, $0.5\times10^8 L_{\odot}$ for each detected PN, for bright \[OIII\] PN the limit $5\times10^8 L_{\odot}$ is used (Durrell 2002). Moreover, the first results for the Virgo PN survey indicate that it is limited by contamination by background objects, rather than the lack of detections. Therefore, we propose to use X-ray binaries, as another indicator for the level of the intracluster stellar population. Given the large field of view of contemporary X-ray telescopes, finding of 1 XRB in a single pointing corresponds to probing the average surface brightness of intracluster star-light on the $\mu_{B}=30$ mag arcsec$^2$ level. Typical surface brightness of the intracluster star-light is $\mu_{B}\sim27$ mag arcsec$^2$. At the moment, only a few Chandra studies of the X-ray binary population in early-type galaxies have been published. Relations between the optical and hard X-ray luminosity, based on ASCA studies alone have a scatter of a factor of 1.5 (Matsushita 2001) for most of the galaxies, while higher X-ray luminosities are met, they are attributed to AGN activity, in the form of bright central point source or hard diffuse X-ray emission. White (2001) demonstrated that the deviation from the $L_{X}-L_{B}$ relation correlates with the specific frequency of globular clusters. This dependency could easily be eliminated by comparison with the optical image. Our normalization corresponds to the case of nearly (90%) clean sample of non-globular XRB. Theoretical modeling of the XRB’s luminosity function predicts its gradual evolution with time since star-formation (Wu 2001). Most drastic changes are expected between the disk and the bulge population of XRB’s, as revealed by observation of M81 (Tennant 2001). Therefore, having determined the luminosity function of Virgo intracluster XRB’s, we will be able to differentiate between the scenarios for diffuse light production, which is not possible in any other methods. {width="3.2in"} The feasibility of the XRB detection in Virgo is illustrated in Fig.\[fig:igsp\]. We assumed that the distribution of diffuse light has a core radius equal to the cluster tidal radius and a total amount of light in the IGSP equal to 20% of the total light in Virgo galaxies, according to the revised estimate using observations of planetary nebulae (Freeman 2000). Using the model presented in Fig.\[fig:igsp\], detection of 1 XRB (with luminosity exceeding $10^{37}$ ergs/s) corresponds to an IGSP contribution of 0.2% to the total light in the Virgo cluster. The gradient in the source number density is expected to be sharp (see. Fig.\[fig:igsp\]). The amplitude of the CXB correlation function on such angular scales (10) is also $\sim10$% (Vikhlinin & Forman 1995; Giacconi 2001). This is therefore the principal limitation of the method, unless optical follow-up or X-ray spectral information is used. The resulting sensitivity is $\sim1$% of the total light in the Virgo cluster. Emission from the hot X-ray gas in the Virgo cluster, acting as a background, puts stringent requirements on the angular resolution. For regions in the Virgo cluster close to M87 (less than half a degree), even for Chandra, variation of the PSF (point spread function) over the field of view (Weisskopf 1996) results in the strongly decreasing sensitivity. Study of the spectral characteristics of the XRB in M84 have revealed that they are spectrally different from the constituents of the CXB (Finoguenov & Jones 2002). Of particular interest are the binaries with a soft component, which exhibit very similar spectra, characterized by a multicolor disk black-body model (Makishima 1986; [*diskbb*]{} model in XSPEC) with a central black-body temperature of 0.5 keV. For the luminosity range of interest ($10^{37}-10^{39}$ ergs/s), over 50% of the XRB in M84 exhibit such a spectrum. By limiting the sources to only those, we can reduce substantially the contamination by CXB to the sample and thus increase the sensitivity of this method to the limit implied by the quality of the hardness ratio determination (typically 30 source counts are required). For an ACIS-I exposure of 40 ksec (or ACIS-S exposure of 25 ksec), such determination will be possible for one third of the sample, with a resulting sensitivity is 1% of the total light in the Virgo cluster. Most of the background objects will be type-1 QSO, for which optical counterpart can readily be found in USNO A2.0 catalogs. Type-2 QSO do not overlap in X-ray colors with XRBs due to characteristic strong absorption. This allows for an independent check of the X-ray color selection. Detection of PNs was also reported for the field far off the Virgo center. Due to lower flux from the thermal gas the requirements on the quality of imaging are reduced, so routine detection of XRBs in fore-coming XMM shallow surveys, (covering large areas, typical for X-ray surveys, challenging for optics) will be feasible. The proposed X-ray method for estimating the amount of intracluster star-light provides a useful alternative to optical methods and is less affected by systematics in the search of candidates, which was recently realized to be a problem for the PN (Planetary Nebulae) method (Kudritzki 2000). Therefore the use of X-ray observations provide an important test of the results obtained optically. In addition, the shape of the luminosity function of the X-ray binaries can shed light on the age of the intracluster stellar population. The authors thank the referee, Jimmy Irwin, for useful suggestions on the manuscript. This work was supported by NASA grants GO0-1045X and AG5-3064 and the Smithsonian Institution. AF has benefited from discussions with Hans Boehringer. AF acknowledges receiving the Max-Plank-Gesellschaft Fellowship. Arnaboldi, M., Freeman, K.C., Mendez, R.H., 1996, ApJ 472, 145 Arnaud, M., Rothenflug, R., Boulade, O., Vigroux, L., and Vangioni-Flam, E. 1992, A&A, 254, 49 Bahcall, N.A., Cen, R., Davé, R., Ostriker, J. P., Yu, Q. 2000, ApJ, 541, 1 Binggeli, B., Sandage, A., Tammann, G.A. 1988, ARAA, 26, 509 Borgani, S., & Guzzo, L. 2001, Nature, 409, 39 Carlberg, R.G., Yee, H.K.C., Ellingson, E., 1996, ApJ, 462, 32 Durrell, P.R., Ciardullo, R., Feldmeier, J.J., Jacoby, G.I., Sigurdsson, S. 2002, ApJ, in press (astro-ph/0201219) Feldmeier, J.J., Ciarullo R., Jacoby G.H. 1998, ApJ, 503, 109 Ferguson, H.C., Tanvir, N.R., von Hippel, T. 1998, Nature 391, 461 Finoguenov, A., and Jones, C. 2001, ApJ, 547, 107 Finoguenov, A., and Jones, C. 2002, ApJ, submitted Freeman, K. C., Arnaboldi, M., Capaccioli, M., 2000, ASP Conf. Proc. 197, 389 Giacconi, R., Rosati, P., Tozzi, P., 2001, ApJ, 551, 624 Gregg, M.D., & West M.J. 1998, Nature, 396, 359 Klypin, A., Gottloeber S., Kravtsov A., Khokhlov A. 1999, ApJ, 516,530 Kudritzki, R.P., Méndez, R. H., Feldmeier, J.J., 2000, ApJ 536, 19 Makishima, K., Maejima, Y., Mitsuda, K., et al. 1986, ApJ, 308, 635 Matsushita, K. 1998, PhD thesis, The University of Tokyo Melnick, K., Sargent, W.L.W. 1977, ApJ, 215, 401 Mendez, R.H., Guerrero, M.A., Freeman, K.C., et al. 1997, ApJ 491, L23 Moore, B., Katz, N., Lake, G., Dressler, A., Oemler, A. Jr. 1996, Nature, 379, 613 Renzini, A., Ciotti, L., D’Ercole, A., Pellegrini, S. 1993, ApJ, 419, 52 Sarazin, C.L., Irwin, J.A., Bregman, J.N. 2000, ApJ, 544L, 101 Sarazin, C.L., Irwin, J.A., Bregman, J.N. 2001, ApJ, 556, 533 Schindler, S., Binggeli, B., Böhringer, H. 1999, 343, 420 Tennant, A.F., Wu, K., Ghosh, K.K., Kolodziejczak, J.J., Swartz, D.A. 2001, ApJ, 549L, 43 Vikhlinin, A., Forman, W. 1995, 455L, 109 Weisskopf, M.C., O’dell, S.L., van Speybroeck, L.P. 1996, SPIE, 2805, 2 White, R.E. III 2001, preprint astro-ph/0111293 [^1]: Spectral characteristics of XRB’s are not uniform. Thus to compare different measurements, one should specify the spectral model assumed. We use the disk blackbody model of Makishima (1986) with the characteristic temperature of 0.5 keV and cite the bolometric luminosity. This spectral model corresponds to the soft excess in the XRB, which is well matched to the energy window of Chandra. If we use the mean spectral index of 0.4 instead, the luminosity in the 0.4–10 keV is twice as high. Spectrally harder sources are detected with Chandra at lower countrate.

ArXiv

--- abstract: 'Ideal monolayers of common semiconducting transition metal dichalcogenides (TMDCs) such as MoS$_2$, WS$_2$, MoSe$_2$, and WSe$_2$ possess many similar electronic properties. As it is the case for all semiconductors, however, the physical response of these systems is strongly determined by defects in a way specific to each individual compound. Here we investigate the ability of exfoliated monolayers of these TMDCs to support high-quality, well-balanced ambipolar conduction, which has been demonstrated for WS$_2$, MoSe$_2$, and WSe$_2$, but not for MoS$_2$. Using ionic-liquid gated transistors we show that, contrary to WS$_2$, MoSe$_2$, and WSe$_2$, hole transport in exfoliated MoS$_2$ monolayers is systematically anomalous, exhibiting a maximum in conductivity at negative gate voltage (V$_G$) followed by a suppression of up to 100 times upon further increasing V$_G$. To understand the origin of this difference we have performed a series of experiments including the comparison of hole transport in MoS$_2$ monolayers and thicker multilayers, in exfoliated and CVD-grown monolayers, as well as gate-dependent optical measurements (Raman and photoluminescence) and scanning tunneling imaging and spectroscopy. In agreement with existing [*ab-initio*]{} calculations, the results of all these experiments are consistently explained in terms of defects associated to chalcogen vacancies that only in MoS$_2$ monolayers – but not in thicker MoS$_2$ multilayers nor in monolayers of the other common semiconducting TMDCs – create in-gap states near the top of the valence band that act as strong hole traps. Our results demonstrate the importance of studying systematically how defects determine the properties of 2D semiconducting materials and of developing methods to control them.' author: - Evgeniy Ponomarev - Árpád Pásztor - Adrien Waelchli - Alessandro Scarfato - Nicolas Ubrig - Christoph Renner - 'Alberto F. Morpurgo' title: 'Hole Transport in Exfoliated Monolayer MoS$_{2}$' --- Extensive studies of monolayers (MLs) of group VI semiconducting transition metal dichalcogenides (TMDCs) have demonstrated that all these materials share many electronic properties. For example, they all have a direct band gap at the K and K‘ points,[@Splendiani2010; @Mak2010; @Zhao2013] a finite Berry curvature in the K and K‘ valley [@Zeng2012; @Mak2012; @Cao2012; @Jones2013; @Mak2014] responsible for the occurrence of the valley Hall effect[@Mak2014; @Ubrig2017], an extremely strong spin-orbit coupling (as large as a few hundreds meV in the valence band),[@Zhu2011; @Xiao2012; @Komider2013] very large exciton binding energies due to the reduced screening characteristic of 2D systems,[@Peimyoo2013; @He2014; @Chernikov2014] stable trion excitations[@Mak2013; @Shang2015], and more. Differences are also present, such as the relative sign of the spin orientation at the conduction band minimum (CBM) and valence band maximum (VBM) – the same for Mo-based compounds and opposite for W-based ones – that leads to a different temperature dependence of the measured photoluminescence.[@Ye2014; @Zhang2015] Although important for specific physical phenomena, these differences mostly concern more subtle aspects of the electronic properties. Having access to a broad class of semiconducting 2D materials with many similar properties is very attractive because – for instance – it facilitates the realization of van der Waals heterostructures obtained by stacking two or more monolayers of different TMDCs on top of each other. It should be realized, however, that these considerations do not take into account that real materials unavoidably contain defects that are specific to each individual compound, and that drastically affect their electronic response. This is certainly the case for the systems considered here, since in semiconductors defects generally determine crucial characteristics such as the work function,[@Addou2015] the position of electrochemical potential,[@McDonnell2014; @Mahjouri-Samani2016] the transport properties (*e.g.*, the carrier mobility [@Qiu2013; @Hong2015]), the rate of non-radiative electron-hole recombination, [@Wang2015] *etc*. That is why an increasing research effort is currently being devoted to the investigation of defects present in all types of semiconducting 2D materials, whose identification, understanding and control will be necessary if these systems will eventually be employed in technological applications (not to mention the possibility to exploit new functionalities that are sometimes offered by defects in 2D materials, such as – for instance – their ability to act as single photon emitters[@Koperski2015; @Srivastava2015; @He2015; @Chakraborty2015; @Bourrellier2016; @Grosso2017]). One important aspect that is seemingly common to semiconducting TMDC monolayers is their ability to support well-balanced ambipolar transport. Measurements done on suitable field-effect transistor (FET) devices upon sweeping the gate voltage show that an equally good conductivity is found in a same monolayer irrespective of whether the chemical potential is in the conduction or in the valence band. This feature is particularly relevant for the realization of opto-electronic devices, since it is the ability to transport simultaneously electrons and holes that allows the controlled generation of electroluminescence from electron-hole recombination or the conversion of light into an electrical signal. Surprisingly, however, if we look at experiments reported on exfoliated monolayers of common group VI semiconducting TMDCs, well-balanced ambipolar transport has been observed in MoSe$_{2}$,[@Onga2016] WSe$_{2}$,[@Allain2014] and WS$_{2}$,[@Jo2014] but not in MoS$_{2}$. This is unexpected both because exfoliated MoS$_{2}$ monolayers are probably the most studied among these compounds (which is why they are often used to benchmark other 2D semiconductors) and because for thick exfoliated MoS$_{2}$ multilayers excellent ambipolar conduction is routinely observed. [@Zhang2012; @Zhang2013] Motivated by these considerations, here we investigate ambipolar conduction in exfoliated MoS$_{2}$ monolayers using ionic liquid-gated FETs and demonstrate a systematic and reproducible anomalous behavior of transport upon hole accumulation. Specifically, after the onset of hole conduction a virtually complete suppression of source-drain current is observed in all devices as the gate voltage is biased to shift the chemical potential deeper into the valence band. In contrast, experiments on identical devices realized on exfoliated bi, tri, and tetralayer MoS$_{2}$ show excellent ambipolar conduction, demonstrating that the anomalous hole transport is an inherent characteristic of monolayer MoS$_{2}$ devices. By combining transistor measurements, gate-dependent optical studies (photo-luminescence and Raman spectroscopy), scanning tunneling imaging and spectroscopy, and a thorough analysis of existing studies based on *ab-initio* calculations, we establish the origin of this phenomenon as due to the presence of atomic-scale defects that induce states inside the band-gap of MoS$_{2}$ monolayers, approximately 300-400 meV above the top of the valence band. These defects – whose manifestations are consistent with what is expected from sulfur vacancies [@Yuan2014; @Vancso2016] – act as traps for holes and prevent hole conduction in exfoliated MoS$_2$ monolayers. The phenomenon is specific to MoS$_2$ monolayers – and not to monolayers of other semiconducting TMDCs or to MoS$_2$ bilayers/thicker multilayers – because only in MoS$_2$ monolayers the states created by chalcogen vacancies near the top of the valence band appear to be inside the band gap. Results ans discussion {#results-ans-discussion .unnumbered} ====================== The most effective way to investigate ambipolar transport in TMDC monolayers is by integrating them into a FET employing an ionic liquid top gate, as illustrated schematically in Figure 1a. The very large capacitance of the ionic gate allows the electrochemical potential to be shifted over a large range – from deep in the conduction band to deep in the valence band – as demonstrated experimentally by the ambipolar conduction observed in WS$_{2}$[@Jo2014], WSe$_{2}$ [@Allain2014] and MoSe$_{2}$ [@Onga2016; @Chen2017] monolayers. For FETs using exfoliated monolayers of MoS$_{2}$, however, a similar observation has not been reported, and our measurements show that an unexpected behavior indeed occurs when the gate is biased to shift the electrochemical potential into the valence band. Specifically, Figure 1b shows transfer curves (source-drain current I$_{SD}$ as a function of gate voltage V$_{G}$ at a finite applied source-drain bias $V_{SD}$) measured on two of our devices representative of the behavior of eight nominally identical devices that we investigated. Upon increasing V$_{G}$ above the electron threshold voltage ($V_{th}^{e}$) carriers are accumulated in the conduction band and the source-drain current increases steeply (the estimated value of electron density $n _{e}$ at the largest V$_G$ value reached approximately $n _{e} = C_{*}(V_{G}-V_{th}^{e})/e=7\cdot10^{12}$ cm$^{-2}$ where $C_{*}= 7 \mu Fcm^{-2}$ is the capacitance per unit of the ionic liquid [@Braga2012; @Jo2014]; values as large as $n _{e} = 2-3 \cdot10^{14}$ cm$^{-2}$ can be obtained at larger positive V$_G$ values [@Costanzo2016]). This is the behavior that is commonly expected in a field-effect transistor. As V$_{G}$ is swept past the onset of hole transport, however, we observe that after an initial increase, I$_{SD}$ exhibits a maximum followed by a steep drop to a virtually vanishing current, with the conductivity $\sigma_{\square}$ dropping by approximately two orders of magnitude (see the inset in Figure 1b). Depending on the specific device either one or two peaks in I$_{SD}$ are observed, as shown by the red and blue curves in Figure 1b. With any further increase of gate voltage to more negative values $I_{SD}$ stays low, and remains low as the gate voltage is swept back to $V_G=$ 0V. The phenomenon is reversible upon re-cycling $V_G$ from zero to large negative values and back, implying that the observed behavior is not due to a permanent degradation of the device. This can be concluded from Figure 1c, which shows transfer curves acquired successively one after the other, by sweeping $V_G$, for increasingly large V$_{SD}$ values. The phenomenon was observed in FET devices with source and drain contacts made of two different materials (Au and Pt), indicating that the current suppression – and the corresponding anomalous behavior of transport upon hole accumulation – is not due to a contact effect (FET transfer curves measured on devices with Pt contacts are shown in figure S1). The anomalous behavior of hole conduction in MoS$_{2}$ monolayers is also clearly visible in the FET output characteristics, *i.e.*, when measuring the source-drain current I$_{SD}$ as a function of source-drain bias V$_{SD}$ at a fixed gate voltage $V_G$. At low positive V$_{SD}$ values, with the gate biased to have electron accumulation, the usual transistor behavior is observed, with $I_{SD}$ increasing linearly until the onset of the saturation regime, occurring at V$_{SD} \simeq$ V$_{G}$-V$_{th}^{e}$, past which $I_{SD}$ stays constant (Figure 1d). As $V_{SD}$ is increased up to much higher values, $I_{SD}$ starts to grow again rapidly (Figure 1e), because the applied source-drain bias causes the channel potential to reverse its polarity and to exceed the threshold for hole accumulation near the drain contact. This is the so-called ambipolar injection regime in which the steep increase in I$_{SD}$ is due to holes injected from the drain contact. In the experiment we see that, sweeping $V_{SD}$ past the point when $I_{SD}$ starts to increase (orange shaded area in Figure 1e) causes a rapid drop of current, and the current stays low as $V_{SD}$ is swept back. We conclude that, irrespectively of the way in which holes are accumulated in the MoS$_{2}$ monolayer channel – either by sweeping the gate voltage or the source-drain bias – an unexpected anomalous suppression of I$_{SD}$ is observed in all cases. The observed phenomenon is inherent to monolayers, and is absent in ionic liquid gated FETs realized with MoS$_2$ bilayers or thicker multilayers. Figure 2 shows the transfer curves of devices fabricated on bi, tri, and tetra layer MoS$_{2}$, in which the measured $I_{SD}$ is excellently balanced (*i.e.*, the magnitude of the measured current is approximately the same upon electron and hole accumulation), with no indication of any anomalous behavior. These observations explain why earlier experiments on ionic liquid gated FETs realized using thick (essentially bulk-like) exfoliated MoS$_{2}$ crystals[@Zhang2012; @Zhang2013] have reported high-quality ambipolar transport, despite the absence of hole current in monolayers. Note that, as the thickness of the MoS$_2$ layer is increased, the range of gate voltages in which the current vanishes (*i.e.*, in which the electrochemical potential is located inside the band gap) decreases. This is because ionic liquid gating used on systems with a small density of states, such as in the gap of a semiconductor, has spectroscopic capabilities[@Braga2012; @Jo2014; @Lezama2014] and the difference in $V_G$ between the threshold voltage for electron and hole conduction is a direct measure of the band gap (to increase the precision with which the gap is determined, data should be plotted as a function of a reference potential – which avoids effects caused by a possible voltage drop at the gate-liquid interface – but in high quality devices the reference potential and the gate voltage nearly coincide). We conclude from these measurements that the band gap is $\Delta$V$_{gap }^{2L}$ = 1.6 eV for MoS$_2$ bilayers, $\Delta$V$_{gap }^{3L}$ = 1.4 eV for trilayers, and $\Delta$V$_{gap }^{4L}$ = 1.25 eV for tetralayers, in all cases with a precision of approximately 10%. Importantly, since the threshold voltage determines the energy of accumulation of individual charge carriers in the respective band, the values extracted are the actual band gaps of the different multilayers – as measured for instance from scanning tunneling spectroscopy [@Huang2015a]– and not the exciton energies which are commonly measured in optical experiments. Irrespective of these considerations, we re-iterate that these experiments unambiguously show that the suppression of conductivity under hole accumulation is a distinct property of exfoliated MoS$_{2}$ monolayers. The systematic anomalous hole transport behavior observed in MoS$_2$ monolayer devices is unexpected and an explanation is called for. In view of the very large negative gate voltage required to populate the valence band, one may wonder whether the correspondingly large electric field (estimated by dividing the maximum applied gate voltage by the thickness of the ionic liquid double layer, and reaching up to several tens of MV/cm) can actually affect the crystalline structure of the material. Indeed, such a drastic structural effect has been reported previously for MoTe$_{2}$ [@Li2016; @Wang2017], and its occurrence in monolayer MoS$_2$ cannot be excluded a priori. To address this question we performed Raman and photoluminescence spectroscopy of ionic-liquid gated MoS$_{2}$ monolayers *in-situ* , in the presence of a large negative gate voltage (see schematics in Figure 3a). Selected Raman spectra collected at different $V_G$ values, from 0 V up to -3.3 V (see Figure 3b), show no change in the characteristic E$_{2g}$ and A$_{1g}$ modes of 2H MoS$_{2}$ throughout the $V_G$ range investigated (the same is true for other parts of the spectrum that we looked at). Results from photoluminescence (PL) spectroscopy are shown in Figure 3c, with a clear peak centered at about 670nm that originates from the A-exciton recombination in MoS$_{2}$ and represents a characteristic signature of MoS$_2$ monolayers. Remarkably, the peak (broadened by large potential fluctuations generated by the ionic liquid gate) remains unchanged, irrespective of the applied gate voltage: the complete photoluminescence map together with the corresponding FET transfer curve are represented in figure S2 (in these measurements the V$_G$-range of the FET transfer curve is shifted relative to the one shown in Figure 1b due to bias stress and the anomalous hole transport is observed at somewhat more negative values of $V_G$). The observed insensitivity of Raman and PL measurements to the gate voltage indicates the absence of significant structural changes in the MoS$_2$ monolayers, and allows us to exclude that the reentrance of the insulating state observed at large negative $V_G$ has a structural origin. What makes the suppression of hole current in exfoliated MoS$_{2}$ monolayers even more surprising is that FETs realized on large-area monolayers grown by chemical vapor deposition (CVD) do exhibit good ambipolar characteristics.[@Ponomarev2015] The top panel of figure 4 shows the conductivity ($\sigma_{\square}$) of a device realized on such a CVD monolayer (blue curve) that, in contrast to what is measured on a device realized using an exfoliated monolayer (green curve), exhibits rather conventional ambipolar conduction. From the data, we first extract the threshold voltage for electrons (which coincides for CVD-grown and exfoliated monolayers) and holes (which can only be determined for the device realized on the CVD-grown monolayer); then we compare transport in the electron and hole subthreshold regimes by looking at the source-drain current $I_{SD}$ on a logarithmic scale (Figure 4, bottom panel). In the electron sub-threshold regime (light-blue shaded region in Figure 4) the conductivity is higher and the subthreshold slope is less steep for the device realized on the CVD-grown monolayer. This behavior is due to a broader tail of disorder-induced localized states that in CVD-grown monolayers is present inside the band gap near the bottom of the conduction band. It is expected, because CVD-grown monolayers are generally more disordered than exfoliated ones, due to a larger density of grain boundaries. A very different behavior is observed in the hole subthreshold regime (light yellow shaded region in Figure 4): the current enhancement extends much more deeply into the gap and it is much larger (by nearly two orders of magnitude) for exfoliated monolayers as compared to CVD-grown material. Furthermore, the very pronounced broad peaks seen in the source-drain current measured on exfoliated monolayers – also present at comparable energy in CVD-grown material, albeit much less pronouncedly – seem to indicate the presence of discrete energy states. These observations indicate that the nature of disorder affecting states near the top of the valence band is qualitatively different from that influencing states near the bottom of the conduction band. Specifically, finding that the effect of disorder near the top of the valence band is much stronger in exfoliated MoS$_2$ monolayers despite their superior structural quality indicates that the origin of the effect is not due to the presence of grain boundaries, but rather to other kind of defects. These observations also strongly suggest that disorder at energies near the top of the valence band has the same origin in exfoliated and CVD-grown monolayers, since the sub-gap states are present in the same range of energies in the two cases. To investigate the nature of the discrete in-gap energy states near the top of the valence band we employ scanning tunneling microscopy (STM) and scanning tunneling spectroscopy (STS). These techniques provide detailed information on the local crystalline and electronic structure of a material down to the atomic scale. STM measurements were performed on a CVD-grown monolayer deposited on a dielectric substrate (SiO$_{2}$/Si), made conductive by electrostatic accumulation of charge carriers (*i.e.*, by integrating it into a FET device; the set-up scheme and FET transfer characteristics are shown in figure 5a and 5b respectively). Under these conditions, with a large positive gate voltage applied to the Si back gate, we can successfully establish stable tunneling conditions, as it is needed to take hiqh-quality images and perform tunneling spectroscopy. Figure 5c demonstrates that atomic resolution can indeed be achieved through the observation of the triangular atomic lattice of MoS$_2$. Atomic resolution is further highlighted by the intense six-fold symmetric peak structure in the image Fourier-transform. These results illustrate how STM measurements performed under an applied gate voltage allow the local properties of the material to be investigated. We probe the spatial distribution of the local density of states (LDOS) by acquiring STS spectra over a dense grid. A map obtained in this way is presented in Figure 5d, where we plot the differential tunneling conductance dI/dV measured at a fixed dc bias (V$_{tip}$ = -1.5 V). Three regions can be identified. The first region occupies the majority of the scanned area and is composed of semiconducting spectra (such as spectrum 1, on the left side of the figure) with a band gap value of 2.1 eV, or just slightly larger (which is why the differential conductance measured at V$_{tip}$ = -1.5 V vanishes). This value of the gap confirms that the material studied is indeed a monolayer. In the two other regions – delimited by the green (region 2) and red (region 3) dashed lines – additional states are visible, and manifest themselves as an enhanced tunneling conductance inside the gap of monolayer MoS$_2$. Interestingly, in both cases the image shows that the states have sub-nanometer dimension, indicating that they originate from atomic defects (no more specific fingerprints enabling a direct identification of the defects could be obtained on gated CVD-grown MoS$_2$ layers on SiO$_2$, which prevents the direct identification of the defects by STM imaging alone). The states have energy close to the top of the valence band, as illustrated by spectra 2 and 3 on the left side of Figure 5. Specifically, in region 2 the in-gap states lead to a fully developed, broad peak 400 meV above the top of the valence band, and in region 3 in a shoulder at a comparable energy (again 300-400 meV above the top of the valence band). The energy of these features (i.e., their distance from the top of the valence band) coincide with what we observed in measurements of the source-drain current as a function of gate/reference potential (see Figure 1b and Figure 4). From this we conclude the in-gap states created by atomic scale defects are the same states acting as hole traps, which are responsible for the observed anomalous transport properties discussed above. The combination of all the measurements presented so far (transport experiments on exfoliated and CVD-grown MoS$_2$ monolayers, their comparison, the comparison with the normal ambipolar behavior observed in other semiconducting TDMCs, photoluminescence and Raman measurements, and the STM and STS measurements just discussed), in conjunction with theoretical investigations based on [*ab-initio*]{} calculations reported in the literature[@Noh2014; @Yuan2014; @Huang2015; @Mahjouri-Samani2016; @Vancso2016], point to a realistic and fully consistent scenario accounting for the experimental observations. Specifically, we claim that the atomic scale defects that we observe and that are responsible for the anomalous hole transport in exfoliated MoS$_2$ monolayers are individual or clustered sulfur vacancies, which are known to be present in relative high density in MoS$_2$ bulk crystals, and therefore also in exfoliated monolayers. Spectroscopy measurements show that they create in-gap states near the top of the valence band acting as traps and strongly affecting hole transport, which is in agreement with expectations based on [*ab-initio*]{} calculations [@Yuan2014; @Huang2015; @Vancso2016]. More specifically, it is known in the context of semiconducting TMDC monolayers that the detailed analysis of the electronic states generated by sulfur and selenium vacancies is complex, as these defects create different types of localized states. However, if we confine our attention exclusively to defect-induced states near the top of the valence band (*i.e.*, those states that can act as effective traps for holes), the results of virtually all existing [*ab-initio*]{} calculations[@Noh2014; @Yuan2014; @Huang2015; @Mahjouri-Samani2016; @Vancso2016] indicate that only in the case of MoS$_2$ monolayers these states are inside the band gap. In other semiconducting TMDC monolayers (WS$_2$, MoSe$_2$, and WSe$_2$) the energy of states near the top of the valence band generated by sulfur or selenium vacancies is actually in the band. As a result, in these other monolayers the defect states do not act as hole traps, because they can hybridize with delocalized states at the same energy. That is: sulfur vacancies explain why only in MoS$_2$ monolayers hole transport is anomalous. For MoS$_2$ bilayers or thicker multilayers the absence of any anomaly in hole transport is explained in a similar way: the decrease of the band gap with increasing the number of layers (by at least 0.5 eV for bilayers as compared to monolayers, and more for thicker multilayers) makes the states induced by sulfur vacancies “fall” into the valence band, preventing them from acting as hole traps. Sulfur vacancies also accounts for other aspects of our observations. In particular, as we have remarked above, the defects affecting hole transport are the same in exfoliated monolayers and CVD-grown ones, albeit in CVD-grown monolayers their effect is less intense. Sulfur vacancies are compatible with this conclusion, as they are expected to be present in both systems. This is not the case for most other types of atomic defects: indeed, mineral MoS$_2$ crystals are known to contain a number of different impurities [@Addou2015a], but these same impurities are neither necessarily present in the Mo and S source material used for CVD growth, nor would they be effectively transferred to the monolayer during the growth process if present. That is why finding that the defects responsible for hole trapping are the same in exfoliated and CVD-grown monolayers strongly constraints the number of possibilities, and sulfur vacancies is one of the few. Finally, sulfur vacancies also naturally explain why the effect of disorder on hole transport is much less pronounced in devices fabricated on CVD-grown material, simply because under the conditions at which CVD growth is done a large excess of sulfur is present, which leads to largely reduced density of sulfur vacancies as compared to that present in exfoliated flakes. Conclusion {#conclusion .unnumbered} ========== Simple theoretical considerations suggest that monolayers of group VI semiconducting transition metal dichalcogenides all possess very similar electronic properties and functionality. Inasmuch as this may be the case for ideal defect-free systems, actual materials contain defects whose influence on the physical properties can (and do) strongly depend on the compound considered. Here we have shown that this is clearly the case for hole transport, which in monolayer MoS$_2$ – but not in monolayers of other common TDMCs – is strongly affected by defects naturally present. Our results very strongly support the conclusion that these naturally present defects are sulfur vacancies, and explain why these defects have a different effect in MoS$_2$ monolayers as compared to monolayers of other similar compounds, *i.e.* the fact that for MoS$_2$ monolayers sulfur vacancies cause localized defect states at energies inside the band gap. They also illustrate a possible solution to minimize the influence of these defects on the performance of devices realized using monolayer MoS$_2$, namely the use of material grown in a large excess of sulfur.\ There is a broad consensus that 2D semiconducting materials have an important potential for technology in the long term. It is clear that exploiting this potential cannot only rely on properties of the idealized systems, but has to take into account the effect of defects unavoidably present, understand it in detail, and find strategies to minimize it. Our results illustrate how this can be done in a concrete specific case. Methods {#methods .unnumbered} ======= MoS$_{2}$ flakes are exfoliated from commercially available crystals (SPI supplies and 2D Semiconductors) by conventional scotch tape exfoliation technique. CVD monolayers of MoS$_{2}$ are synthesized through chemical reaction of a molybdenum containing precursor (MoO$_{3}$) and sulfur, according to a protocol reported in ref.45. In short, crucibles with the solid precursors are loaded into a tube furnace so that the MoO$_{3}$ powder is positioned in the middle of the furnace and the sulfur approximately 20 cm upstream. The substrate (either sapphire or SiO$_{2}$) is mounted ’face-down’ above the MoO$_{3}$-containing crucible. The synthesis is done under constant Ar flow of 75 sccm. Throughout the growth the temperatures of MoO$_{3}$ and sulfur are maintained at 700 and 250$^{\circ}$C respectively. Electronic transport measurements are done using FETs with ionic liquid top gate. Electric contacts are fabricated with electron-beam lithography, metal evaporation, and lift-off. Devices are further annealed at 200$^{\circ}$C in an inert atmosphere of Ar for 2 hours in order to decrease the contact resistance. A drop of ionic liquid (P14-FAP 1-butyl-1-methylpyrrolidiniumtris(pentafluoroethyl)trifluorophosphate) is put on top of the devices right before loading them into a vacuum chamber (p$\thickapprox$10$^{-6}$ mbar). Prior to measurements devices are kept in the chamber overnight to pump out humidity and oxygen. The STM experiments were done in ultra-high vacuum (base pressure 5$\cdot$10$^{-12}$ mbar) at liquid nitrogen temperature using tips electrochemically etched from an annealed tungsten wire. Topographic images were recorded in constant current mode. Tunneling spectroscopic I(V) curves were acquired with open feedback loop while $\mathbf{d}I/\mathbf{d}V$ curves were obtained by numerical derivation of the measured I(V) curves. Positive and negative bias voltages correspond to empty and filled states of the sample, respectively. Photoluminescence and Raman measurements were performed both in back-scattering geometry (*i.e.* collecting the emitted light with the same microscope used to couple the laser beam onto the device). The light collected from the sample was sent to a Czerny-Turner monochromator and detected with a Si CCD-array (Andor). For both type of measurements the excitation wavelength of the laser was set to 514.5 nm and the power was kept below 20 $\mu W$. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to A. Ferreira, R. Villarreal, and G. Manfrini for fruitful discussions and technical help. AFM and CR gratefully acknowledge financial support from the Swiss National Science Foundation (Division 2 and Sinergia program). AFM also gratefully acknowledges financial support from the EU Graphene Flagship project. NU acknowledges funding from an Ambizione grant of the Swiss National Science Foundation. Associated content {#associated-content .unnumbered} ================== Supporting Information Available: FET transfer curves with Pt and Au contacts, photoluminescence map, FET transfer curve at reduced temperature. This material is available free of charge via the Internet at http://pubs.acs.org. Competing financial interests {#competing-financial-interests .unnumbered} ============================= The authors declare no competing financial interests. @ifundefined [50]{} Splendiani, A.; Sun, L.; Zhang, Y.; Li, T.; Kim, J.; Chim, C. Y.; Galli, G.; Wang, F. [Emerging Photoluminescence in Monolayer MoS$_2$]{}. *Nano Lett.* **2010**, *10*, 1271–1275 Mak, K. F.; Lee, C.; Hone, J.; Shan, J.; Heinz, T. F. [Atomically Thin MoS$_2$: A New Direct-Gap Semiconductor]{}. *Phys. Rev. Lett.* **2010**, *105*, 2–5 Zhao, W.; Ghorannevis, Z.; Chu, L.; Toh, M.; Kloc, C.; Tan, P. H.; Eda, G. [Evolution of Electronic Structure in Atomically Thin Sheets of WS$_2$ and WSe$_2$]{}. *ACS Nano* **2013**, *7*, 791–797 Zeng, H.; Dai, J.; Yao, W.; Xiao, D.; Cui, X. [Valley Polarization in MoS$_2$ Monolayers by Optical Pumping]{}. *Nat. Nanotechnol.* **2012**, *7*, 490–493 Mak, K. F.; He, K.; Shan, J.; Heinz, T. F. [Control of Valley Polarization in Monolayer MoS$_2$ by Optical Helicity]{}. *Nat. Nanotechnol.* **2012**, *7*, 494–498 Cao, T.; Wang, G.; Han, W.; Ye, H.; Zhu, C.; Shi, J.; Niu, Q.; Tan, P.; Wang, E.; Liu, B.; Feng, J. [Valley-Selective Circular Dichroism of Monolayer Molybdenum Disulphide]{}. *Nat. Commun.* **2012**, *3*, 885–887 Jones, A. M.; Yu, H.; Ghimire, N. J.; Wu, S.; Aivazian, G.; Ross, J. S.; Zhao, B.; Yan, J.; Mandrus, D. G.; Xiao, D.; Yao, W.; Xu, X. [Optical Generation of Excitonic Valley Coherence in Monolayer WSe$_2$]{}. *Nat. Nanotechnol.* **2013**, *8*, 634–638 Mak, K. F.; McGill, K. L.; Park, J.; McEuen, P. L. [The Valley Hall Effect in MoS$_2$ Transistors]{}. *Science (80-. ).* **2014**, *344*, 1489–1492 Ubrig, N.; Jo, S.; Philippi, M.; Costanzo, D.; Berger, H.; Kuzmenko, A. B.; Morpurgo, A. F. [Microscopic Origin of the Valley Hall Effect in Transition Metal Dichalcogenides Revealed by Wavelength-Dependent Mapping]{}. *Nano Lett.* **2017**, *17*, 5719–5725 Zhu, Z. Y.; Cheng, Y. C.; Schwingenschl[ö]{}gl, U. [Giant Spin-Orbit-Induced Spin Splitting in Two-Dimensional Transition-Metal Dichalcogenide Semiconductors]{}. *Phys. Rev. B - Condens. Matter Mater. Phys.* **2011**, *84*, 1–5 Xiao, D.; Liu, G. B.; Feng, W.; Xu, X.; Yao, W. [Coupled Spin and Valley Physics in Monolayers of MoS$_2$ and Other Group-VI Dichalcogenides]{}. *Phys. Rev. Lett.* **2012**, *108*, 1–5 Komider, K.; Gonz[á]{}lez, J. W.; Fern[á]{}ndez-Rossier, J. [Large Spin Splitting in the Conduction Band of Transition Metal Dichalcogenide Monolayers]{}. *Phys. Rev. B - Condens. Matter Mater. Phys.* **2013**, *88*, 1–7 Peimyoo, N.; Shang, J.; Cong, C.; Shen, X.; Wu, X.; Yeow, E. K. L.; Yu, T. [Nonblinking, Intense Two-Dimensional Light Emitter: Monolayer WS$_2$ Triangles]{}. *ACS Nano* **2013**, *7*, 10985–10994 He, K.; Kumar, N.; Zhao, L.; Wang, Z.; Mak, K. F.; Zhao, H.; Shan, J. [Tightly Bound Excitons in Monolayer WSe$_2$]{}. *Phys. Rev. Lett.* **2014**, *113*, 1–5 Chernikov, A.; Berkelbach, T. C.; Hill, H. M.; Rigosi, A.; Li, Y.; Aslan, O. B.; Reichman, D. R.; Hybertsen, M. S.; Heinz, T. F. [Exciton Binding Energy and Nonhydrogenic Rydberg Series in Monolayer WS$_2$]{}. *Phys. Rev. Lett.* **2014**, *113*, 1–5 Mak, K. F.; He, K.; Lee, C.; Lee, G. H.; Hone, J.; Heinz, T. F.; Shan, J. [Tightly Bound Trions in Monolayer MoS$_2$]{}. *Nat. Mater.* **2013**, *12*, 207–211 Shang, J.; Shen, X.; Cong, C.; Peimyoo, N.; Cao, B.; Eginligil, M.; Yu, T. [Observation of Excitonic Fine Structure in a 2D Transition-Metal Dichalcogenide Semiconductor]{}. *ACS Nano* **2015**, *9*, 647–655 Ye, Z.; Cao, T.; O’Brien, K.; Zhu, H.; Yin, X.; Wang, Y.; Louie, S. G.; Zhang, X. [Probing Excitonic Dark States in Single-Layer Tungsten Disulphide]{}. *Nature* **2014**, *513*, 214–218 Zhang, X. X.; You, Y.; Zhao, S. Y. F.; Heinz, T. F. [Experimental Evidence for Dark Excitons in Monolayer WSe$_2$]{}. *Phys. Rev. Lett.* **2015**, *115*, 1–6 Addou, R.; McDonnell, S.; Barrera, D.; Guo, Z.; Azcatl, A.; Wang, J.; Zhu, H.; Hinkle, C. L.; Quevedo-Lopez, M.; Alshareef, H. N.; Colombo, L.; Hsu, J. W. P.; Wallace, R. M. [Impurities and Electronic Property Variations of Natural MoS$_2$ Crystal Surfaces]{}. *ACS Nano* **2015**, *9*, 9124–9133 McDonnell, S.; Addou, R.; Buie, C.; Wallace, R. M.; Hinkle, C. L. [Defect-Dominated Doping and Contact Resistance in MoS$_2$]{}. *ACS Nano* **2014**, *8*, 2880–2888 Mahjouri-Samani, M.; Liang, L.; Oyedele, A.; Kim, Y. S.; Tian, M.; Cross, N.; Wang, K.; Lin, M. W.; Boulesbaa, A.; Rouleau, C. M.; Puretzky, A. A.; Xiao, K.; Yoon, M.; Eres, G.; Duscher, G.; Sumpter, B. G.; Geohegan, D. B. [Tailoring Vacancies Far Beyond Intrinsic Levels Changes the Carrier Type and Optical Response in Monolayer MoSe$_2-x$ Crystals]{}. *Nano Lett.* **2016**, *16*, 5213–5220 Qiu, H.; Xu, T.; Wang, Z.; Ren, W.; Nan, H.; Ni, Z.; Chen, Q.; Yuan, S.; Miao, F.; Song, F.; Long, G.; Shi, Y.; Wang, J.; Wang, X. [Hopping Transport Through Defect-Induced Localized States in Molybdenum Disulphide]{}. *Nat. Commun.* **2013**, *4*, 1–6 Hong, J.; Hu, Z.; Probert, M.; Li, K.; Lv, D.; Yang, X.; Gu, L.; Mao, N.; Feng, Q.; Xie, L.; Zhang, J.; Wu, D.; Zhang, Z.; Jin, C.; Ji, W.; Zhang, X.; Yuan, J.; Zhang, Z. [Exploring Atomic Defects in Molybdenum Disulphide Monolayers]{}. *Nat. Commun.* **2015**, *6*, 1–8 Wang, H.; Zhang, C.; Rana, F. [Ultrafast Dynamics of Defect-Assisted Electron-Hole Recombination in Monolayer MoS$_2$]{}. *Nano Lett.* **2015**, *15*, 339–345 Koperski, M.; Nogajewski, K.; Arora, A.; Cherkez, V.; Mallet, P.; Veuillen, J. Y.; Marcus, J.; Kossacki, P.; Potemski, M. [Single Photon Emitters in Exfoliated WSe$_2$ Structures]{}. *Nat. Nanotechnol.* **2015**, *10*, 503–506 Srivastava, A.; Sidler, M.; Allain, A. V.; Lembke, D. S.; Kis, A.; Imamoglu, A. [Optically Active Quantum Dots in Monolayer WSe$_2$]{}. *Nat. Nanotechnol.* **2015**, *10*, 491–496 He, Y. M.; Clark, G.; Schaibley, J. R.; He, Y.; Chen, M. C.; Wei, Y. J.; Ding, X.; Zhang, Q.; Yao, W.; Xu, X.; Lu, C. Y.; Pan, J. W. [Single Quantum Emitters in Monolayer Semiconductors]{}. *Nat. Nanotechnol.* **2015**, *10*, 497–502 Chakraborty, C.; Kinnischtzke, L.; Goodfellow, K. M.; Beams, R.; Vamivakas, A. N. [Voltage-Controlled Quantum Light from an Atomically Thin Semiconductor]{}. *Nat. Nanotechnol.* **2015**, *10*, 507–511 Bourrellier, R.; Meuret, S.; Tararan, A.; St[é]{}phan, O.; Kociak, M.; Tizei, L. H.; Zobelli, A. [Bright UV Single Photon Emission at Point Defects in h-BN]{}. *Nano Lett.* **2016**, *16*, 4317–4321 Grosso, G.; Moon, H.; Lienhard, B.; Ali, S.; Efetov, D. K.; Furchi, M. M.; Jarillo-Herrero, P.; Ford, M. J.; Aharonovich, I.; Englund, D. [Tunable and High-Purity Room Temperature Single-Photon Emission from Atomic Defects in Hexagonal Boron Nitride]{}. *Nat. Commun.* **2017**, *8*, 1–8 Onga, M.; Zhang, Y.; Suzuki, R.; Iwasa, Y. [High Circular Polarization in Electroluminescence from MoSe$_2$]{}. **2016**, *073107*, 2–6 Allain, A.; Kis, A. [Electron and Hole Mobilities in Single-Layer WSe$_2$]{}. *ACS Nano* **2014**, *8*, 7180–7185 Jo, S.; Ubrig, N.; Berger, H.; Kuzmenko, A. B.; Morpurgo, A. F. [Mono- and Bilayer WS$_2$ Light-Emitting Transistors]{}. *Nano Lett.* **2014**, *14*, 2019–2025 Zhang, Y.; Ye, J.; Matsuhashi, Y.; Iwasa, Y. [Ambipolar MoS$_2$ Thin Flake Transistors]{}. *Nano Lett.* **2012**, *12*, 1136–1140 Zhang, Y. J.; Ye, J. T.; Yomogida, Y.; Takenobu, T.; Iwasa, Y. [Formation of a Stable p-n Junction in a Liquid-Gated MoS$_2$ Ambipolar Transistor]{}. *Nano Lett.* **2013**, *13*, 3023–3028 Yuan, S.; Rold[á]{}n, R.; Katsnelson, M. I.; Guinea, F. [Effect of Point Defects on the Optical and Transport Properties of MoS$_2$ and WS$_2$]{}. *Phys. Rev. B - Condens. Matter Mater. Phys.* **2014**, *90*, 1–5 Chen, M.-W.; Ovchinnikov, D.; Lazar, S.; Pizzochero, M.; Whitwick, M. B.; Surrente, A.; Baranowski, M.; Sanchez, O. L.; Gillet, P.; Plochocka, P.; Yazyev, O. V.; Kis, A. [Highly Oriented Atomically Thin Ambipolar MoSe$_2$ Grown by Molecular Beam Epitaxy]{}. *ACS Nano* **2017**, *11*, 6355–6361 Braga, D.; [Guti[é]{}rrez Lezama]{}, I.; Berger, H.; Morpurgo, A. F. [Quantitative Determination of the Band Gap of WS$_2$ with Ambipolar Ionic Liquid-Gated Transistors]{}. *Nano Lett.* **2012**, *12*, 5218–5223 Costanzo, D.; Jo, S.; Berger, H.; Morpurgo, A. F. [Gate-Induced Superconductivity in Atomically Thin MoS$_2$ Crystals]{}. *Nat. Nanotechnol.* **2016**, *11*, 339–344 Lezama, I. G.; Ubaldini, A.; Longobardi, M.; Giannini, E.; Renner, C.; Kuzmenko, A. B.; Morpurgo, A. F. [Surface Transport and Band Gap Structure of Exfoliated 2H-MoTe$_2$ crystals]{}. *2D Mater.* **2014**, *1*, 021002 Huang, Y. L.; Chen, Y.; Zhang, W.; Quek, S. Y.; Chen, C.-H.; Li, L.-J.; Hsu, W.-T.; Chang, W.-H.; Zheng, Y. J.; Chen, W.; Wee, A. T. S. [Bandgap Tunability at Single-Layer Molybdenum Disulphide Grain Boundaries.]{} *Nat. Commun.* **2015**, *6*, 6298 Li, Y.; Duerloo, K.-A. N.; Wauson, K.; Reed, E. J. [Structural Semiconductor-to-Semimetal Phase Transition in Two-Dimensional Materials Induced by Electrostatic Gating]{}. *Nat. Commun.* **2016**, *7*, 10671 Wang, Y.; Xiao, J.; Zhu, H.; Li, Y.; Alsaid, Y.; Fong, K. Y.; Zhou, Y.; Wang, S.; Shi, W.; Wang, Y.; Zettl, A.; Reed, E. J.; Zhang, X. [Structural Phase Transition in Monolayer MoTe$_2$ Driven by Electrostatic Doping]{}. *Nature* **2017**, *550*, 487–491 Ponomarev, E.; Guti[é]{}rrez-Lezama, I.; Ubrig, N.; Morpurgo, A. F. [Ambipolar Light-Emitting Transistors on Chemical Vapor Deposited Monolayer MoS$_2$]{}. *Nano Lett.* **2015**, *15*, 8289–8294 Noh, J. Y.; Kim, H.; Kim, Y. S. [Stability and Electronic Structures of Native Defects in Single-Layer MoS$_2$]{}. *Phys. Rev. B - Condens. Matter Mater. Phys.* **2014**, *89*, 1–12 Huang, B.; Yoon, M.; Sumpter, B. G.; Wei, S. H.; Liu, F. [Alloy Engineering of Defect Properties in Semiconductors: Suppression of Deep Levels in Transition-Metal Dichalcogenides]{}. *Phys. Rev. Lett.* **2015**, *115*, 1–5 Vancs[ó]{}, P.; Magda, G. Z.; Pető, J.; Noh, J.-Y.; Kim, Y.-S.; Hwang, C.; Bir[ó]{}, L. P.; Tapaszt[ó]{}, L. [The Intrinsic Defect Structure of Exfoliated MoS$_2$ Single Layers Revealed by Scanning Tunneling Microscopy]{}. *Sci. Rep.* **2016**, *6*, 29726 Addou, R.; Colombo, L.; Wallace, R. M. [Surface Defects on Natural MoS$_2$]{}. *ACS Appl. Mater. Interfaces* **2015**, *7*, 11921–11929

ArXiv

--- abstract: 'We formulate and solve a two-state model for the elasticity of nicked, double-stranded DNA that borrows features from both the Worm Like Chain and the Bragg–Zimm model. Our model is computationally simple, and gives an excellent fit to recent experimental data through the entire overstretching transition. The fit gives the first value for the bending stiffness of the overstretched state as about $10\,\mathrm{nm}\cdot\kbt$, a value quite different from either B-form or single-stranded DNA.' author: - 'C. Storm[^1]' - 'P. C. Nelson' title: 'The bend stiffness of S-DNA' --- Introduction and Summary ======================== When double-stranded DNA is subjected to longitudinal forces greater than about 65$\,\pNunit$ it undergoes a radical conformational change, marked by a sudden, almost twofold increase in contour length[@Cluzel; @Smith]. The structural characterization of the resulting overstretched or (“S-state”) DNA is complicated by the fact that techniques such as X-ray crystallography are not applicable to single molecules. In this Letter, we instead characterize overstretched DNA by examining its elastic constants, and to this end formulate and solve a model that synthesizes features of both the Worm Like Chain (WLC) and the Bragg–Zimm model of the helix–coil transition in peptides. Thus we model DNA as consisting of two different, coexisting conformations, each with its own elastic constants. We solve this model and show that it gives a good fit to recent data on the overstretching transition in nicked, double-stranded DNA. From these fits, we conclude that the bend stiffness of S-DNA is intermediate between the known values for single stranded and double stranded DNA. Our result supports the work of Léger [*et al.*]{} [@lege99a; @lege99b], who argued that S-DNA has a definite helical pitch and hence is a new duplex conformation of DNA. Our model and solution method differ from those offered by Marko [@mark98a], who assumes the bend stiffnesses of the two conformational states to be identical; our analysis will show that on the contrary the stiffnesses are markedly different. The analysis of Viovy and Cizeau [@Acize97a] is essentially a mean-field approximation to the model we study here; in addition, the authors did not quote any value for the S-DNA bend stiffness, presumably because the experimental data available at that time did not permit such a determination. The model studied here is a continuum limit of a more general class of discrete persistent-chain models. Such models give better fits to the stretching curves of *single-*stranded DNA than either the continuum WLC or the freely jointed chain models. Details will appear elsewhere [@promise]. Our model and method are also of some general interest beyond DNA. For example, both can be adapted to the study of the stretching of polypeptides with a helix-coil transition. Model ===== We begin by formulating a discretized form of the WLC, which we call the the “Discrete Persistent Chain.” Later we will introduce an Ising-like variable on each chain link describing a cooperative transition from B- to S-form. The DPC models the polymer as a chain of $N$ segments of length $b$, whose conformation is fully described by the collection of orientation vectors $\{\hat t_i\}$ for each segment. Thus the relaxed total contour length is $L_{\text{tot}}\equiv Nb$. Bend resistance is taken into account by including an energy penalty at each link proportional to the square of the angle $\Theta_{i,i+1}=\arccos(\hat t_i \cdot \hat t_{i+1})$ between two adjacent links. The energy functional describing this model is thus given by $$\frac{{\cal E}[\{\hat t_i\}]}{\kb T}=-\sum_{i=1}^{N}\frac{fb}{\kb T} \, \hat t_i \cdot \hat z\, +\sum_{i=1}^{N-1}\frac{A}{2b}(\Theta_{i,i+1})^2 \, .$$ The partition function for this energy functional is $ {\cal Z}=\left[ \prod_{i=1}^{N} \int_{{\mathbb S}^2} \! {\rm d}^2\hat t_i\!\right] e^{{\cal E}[\{\hat t_i\}]/{\kb T}}$, where ${\mathbb S}^2$ is the two-dimensional unit sphere. To compute ${\cal Z}$ we use the transfer matrix technique[@KramersWannier], interpreting each integral above as a generalized matrix product among matrices with continuous indices: $ {\cal Z}=\vec v \cdot {\sf T}^{N-1}\cdot \vec w \, $. In this formula $\vec v$ and $\vec w$ are vectors indexed by $\hat t$, or in other words functions $v(\that\,),w(\that\,)$. The matrix product $\Tmat\cdot\vec v$ is a new vector, defined by the convolution: $$\label{dp1} ({\sf T}\cdot \vec v\,)(\hat t_i)=\int_{{\mathbb S}^2} \! {\rm d}^2\hat t_j \, {\mathsf T}(\hat t_i,\hat t_j) v(\hat t_j) \, .$$ Here the matrix elements of $\Tmat$ are given by $ {\mathsf T}(\hat t_i,\hat t_j)=e^{-{\cal E}_i(\hat t_i,\hat t_j)/\kb T}\, $; we will not need the explicit forms of $\vec v$ and $\vec w$ below. The force-extension relation can be obtained from ${\cal Z}$ by differentiating with respect to the force. It is here that the transfer matrix formulation can be used to greatly simplify the calculation of the force-extension relation, since all that is needed to compute the logarithmic derivative of ${\cal Z}$ in the limit of long chains is the largest eigenvalue of ${\sf T}$, which we will call $\lambda_{\text{max}}$: $$\label{zldef} \langle \frac{z}{L_{\text{tot}}} \rangle\stackrel{\mathrm{large }N}{\longrightarrow}\left(\frac{\kb T}{L_{\text{tot}}}\right) \frac{\rm d}{{\rm d} f} \ln (\lambda_{\max})^{N} =\left(\frac{\kb T}{b}\right) \frac{\rm d}{{\rm d} f} \ln {\lambda_{\max}}\, .$$ Analogously to ref. [@Aodij95a], it is straightforward to add an intrinsic stretch modulus to the calculation outlined above, obtaining an “Extensible DPC” model. To study overstretching, we now extend the extensible DPC by giving each link a discrete variable $\sigma$, which takes the values $\pm 1$. We will take $\sigma=+1$ to mean the segment is in the B-state and $\sigma=-1$ for the S-state. The factor by which a segment elongates when going from B to S will be called $\zeta$, [*i.e.*]{} $b^{(\mathrm{S})}=\zeta b$ (with $\zeta>1$). We assign a bend stiffness parameter $A$ to B-DNA, and a different $A^{(\mathrm{S})}\equiv\beta \zeta A$ to S-DNA; $\beta$ is a dimensionless parameter with $\beta\zeta<1$. Similarly we assign a bend stiffness $\Nxterm A$ to a hinge joining a B and an S segment. The full energy functional for the Ising–DPC model is thus: $$\begin{aligned} \label{idpcen} \frac{{\cal E}[\{\hat t_i,\sigma_i\}]}{\kb T}&=&-\sum_{i=1}^{N-1} \biggl\{ \frac{\alpha_0}{2}(\sigma_i\!+\!\sigma_{i+\!1})+\gamma(\sigma_i \sigma_{i+\!1}\!-\!1)\nonumber+ \\&&\hspace{-1.5cm}+\frac{fb}{2\kb T}\left[\Bigl(\frac{1\!+\!\sigma_i}{2}\!+\!\frac{1\!-\!\sigma_i}{2}\zeta\Bigr)\hat t_i \cdot \hat z\!+\!\Bigl(\frac{1\!+\!\sigma_{i+\!1}}{2}\!+\!\frac{1\!-\!\sigma_{i+\!1}}{2}\zeta\Bigr) \hat t_{i+\!1}\cdot \hat z\right]\!- \nonumber \\ & &\hspace{-.6cm} -\frac{A}{2b}\left[ \frac{(1\!-\!\sigma_i)(1\!-\!\sigma_{i+\!1})}{4}\beta+|\sigma_i\!-\!\sigma_{i+\! 1}|\Nxterm+ \frac{(1\!+\!\sigma_i)(1\!+\!\sigma_{i+\!1})}{4}\right] (\Theta_{i,i+\!1})^2 \biggr\} \, .\end{aligned}$$ The first line is the pure-Ising part, with $2 \alpha_0 \kb T$ the intrinsic free energy cost of converting a single segment from B to S and $2\gamma \kb T$ the energy cost of creating a B$\to$S interface. Note that we ignore a contribution to the energy functional from the first and last segments. In the long-chain limit this does not affect the outcome of our calculation. The partition function for the energy functional (\[idpcen\]) is given by $$\label{2dpcZ} {\cal Z}=\left[\prod_{i=1}^{N-1}\sum_{\sigma_i=\pm 1} \! \int_{{\mathbb S}^2} \! {\rm d}^2\hat t_i\right] \! \prod_{i=1}^{N-1}e^{-{\cal E}_i(\hat t_i,\sigma_i,\hat t_{i+1},\sigma_{i+1})/\kb T}\, ,$$ where now ${\cal E}[\{ \hat t_i,\sigma_i \}]=\sum_{i=1}^{N-1}{\cal E}_i(\hat t_i,\sigma_i,\hat t_{i+1},\sigma_{i+1})$. We again calculate ${\cal Z}$ with the aid of the transfer matrix technique, writing as $ {\cal Z}=\vec v \cdot {\sf T}^{N-1}\cdot \vec w\, $, with ${\sf T}$ now the transfer matrix for our Ising-DPC model, which carries an additional 2-by-2 structure due to the Ising variables. The dot products are thus defined as $$({\sf T}\cdot \vec v)_{\sigma_i}(\hat t_i)=\sum_{\sigma_j=\pm 1} \! \int_{{\mathbb S}^2} \! {\rm d}^2\hat t_j \, {\mathsf T}_{\sigma_i \sigma_j}(\hat t_i,\hat t_j) v_{\sigma_j}(\hat t_j) \, .$$ The individual matrix elements ${\mathsf T}_{\sigma_i \sigma_j}$ are given explicitly by $$\begin{aligned} \label{Telts} {\mathsf T}_{-1,-1}(\hat t_i,\hat t_{i+1})& = & \exp\left[\frac{1}{2} \zeta \tilde f(\hat t_i\!+\!\hat t_{i+1})\cdot \hat z\!-\!\frac{\beta A}{b}(1\!-\!\hat t_i \cdot \hat t_{i+1})\!-\!\alpha_0\right]\, ,\end{aligned}$$ and related expressions for ${\mathsf T}_{1,1}$, ${\mathsf T}_{1,-1}$, and ${\mathsf T}_{-1,1}$, where $\tilde f\equiv\frac{fb}{\kb T}$. We approximate the largest eigenvalue of the transfer matrix ${\sf T}$ using a variational approach[@MarkoSiggia]. We choose a three-parameter family of trial eigenfunctions with azimuthal symmetry, peaked in the direction of the force $\hat z$: $$\label{eomphi} v_{\omega_{1},\omega_{-1},\varphi}(\hat t\,)= \left(\begin{array}{c } \left(\frac{\omega_{1}}{\sinh(2 \omega_1)}\right)^{\frac12} e^{\omega_1 \hat t \cdot \hat z} \cos \varphi \\ \left(\frac{\omega_{-1}}{\sinh(2 \omega_{-1})}\right)^{\frac12}e^{\omega_{-1} \hat t \cdot \hat z}\sin \varphi \end{array}\right)\, .$$ These trial functions were chosen such that their squared norm is independent of all parameters: $ \| \vec v_{\omega_{1},\omega_{-1},\varphi}\|^2=2 \pi$. shows that the $\omega$’s give the degree of alignment of the monomers (how forward-peaked their probability distribution is), whereas $\varphi$ describes the relative probability of a monomer to be in the two states. The variational estimate for the maximal eigenvalue is thus $$\label{varesti} \lambda^*_{\text{max}}\equiv\max_{\omega_{1},\omega_{-1},\varphi} y(\omega_{1},\omega_{-1},\varphi) \equiv\max_{\omega_{1},\omega_{-1},\varphi}\, \frac{\vec v_{\omega_{1},\omega_{-1},\varphi} \cdot {\sf T} \cdot \vec v_{\omega_{1},\omega_{-1},\varphi} }{{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}} \, ,$$ The maximization over $\varphi$ can be done analytically: defining the $2\times 2$ matrix $\tilde {\sf T}(\omega_1,\omega_{-1})$ by $$\vec v_{\omega_{1},\omega_{-1},\varphi} \cdot {\sf T} \cdot \vec v_{\omega_{1},\omega_{-1},\varphi}=(\cos \varphi,\sin \varphi) \cdot \tilde {\sf T}(\omega_1,\omega_{-1}) \cdot \left(\begin{array}{c } \cos \varphi \\ \, \sin \varphi \end{array}\right) \, ,$$ gives that $$\label{e:lstmax} \lambda^*_{\text{max}}=\max_{\omega_1,\omega_{-1}}\, \frac{\tilde y(\omega_1,\omega_{-1})}{{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}}\, ,$$ where $\tilde y(\omega_1,\omega_{-1})$ is the maximal eigenvalue of the $2\times2$ matrix $\tilde{\sf T}(\omega_1,\omega_{-1})$. The following section will calculate this eigenvalue in a continuum approximation to $\tilde{\sf T}(\omega_1,\omega_{-1})$, illustrating the procedure by considering in some detail the matrix element $\tilde{\mathsf T}_{1,1}(\omega_1,\omega_{-1})$. The other matrix elements can be obtained analogously. Writing out the integrals explicitly, we have $$\tilde{\mathsf T}_{1,1}(\omega_1)=\frac{\omega_{1} e^{\alpha_0-\frac{A}{b}}}{\sinh(2 \omega_1)}\int_{{\mathbb S}^2} \! {\rm d}^2\hat t_i e^{\hat a \hat t_i \cdot \hat z}\int_{{\mathbb S}^2} \! {\rm d}^2\hat t_{i+1} \left[e^{(\hat a \hat z+\frac{A}{b}\hat t_i)\cdot \hat t_{i+1}}\right]\, ,$$ where we have introduced $\hat a \equiv \omega_1+\frac{\tilde f}{2}$. Condensing notation even further we define $\mu^2=\hat a^2+(\tfrac{A}{b})^2+2\hat a\tfrac{A}{b}\hat t_i \cdot \hat z$, which allows us to write $$\label{eqT} \tilde{\mathsf T}_{1,1}(\omega_1)\!=\!(2 \pi)^2\frac{\omega_{1} e^{\alpha_0-{A}/{b}}}{\sinh(2 \omega_1)}\int_{|\tfrac{A}{b}\!-\!\hat a|}^{\tfrac{A}{b}\!+\hat a}\!\! \ \frac{ b\,{\rm d}\mu}{\hat a A} e^{{b}(\mu^2\!-\!\hat a^2\!-\!(\frac{A}{b})^2)/{(2 A)}}\left[ e^\mu\!-\!e^{-\mu}\right]\, .$$ Continuum Limit\[s:cl\] ======================= We could now proceed to evaluate the force-extension relation of the Ising-DPC model, by evaluating numerically and using . To simplify the calculations, however, we first pass to a continuum limit. To justify this step, we note that the continuum (WLC) approximation gives a good account of single-stranded DNA stretching out to forces beyond those probed in overstretching experiments (about $90\,\pNunit$) [@Clausen]. As mentioned earlier, the continuum approximation is also quite good for double-stranded DNA, because the latter’s persistence length is so much longer than its monomer size. In the continuum limit $b$ is sent to zero holding $L\tot$ fixed; hence $N\to\infty$. The bookkeeping is more manageable after a shift in $\mu$: $ x\equiv \mu-({A}/{b})$. then reduces to $$\begin{aligned} \tilde{\mathsf T}_{1,1}(\omega_1)\!&=&\!\frac{\omega_{1} e^{\alpha_0}}{\sinh(2 \omega_1)} \frac{(2 \pi)^2 b}{\hat a A}\int_{-\hat a}^{+\hat a} \!\! \, {\rm d}x\, \exp \left[ \frac{b}{2 A} x^2 +2x-\frac{\hat a^2 b}{2A}\right] \nonumber \\ &\approx& \!\frac{\omega_{1} e^{\alpha_0}}{\sinh(2 \omega_1)} \frac{(2 \pi)^2 b}{\hat a A}\int_{-\hat a}^{+\hat a} \!\! \, {\rm d}x\, e^{2x}(1+\frac{x^2 b}{2 A}) e^{-\frac{\hat a^2 b}{2 A}}\, .\end{aligned}$$ The last integral can be worked out exactly, and expanding the result to second order in $b$ we end up with $$\frac{A}{2\pi b} \frac{1}{{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}}\tilde{\mathsf T}_{1,1}(\omega_1)\!=\!e^{\alpha_0}\left[ 1+{b}\left(\frac{f}{\kb T}-\frac{\omega_1}{2A}\right)\left(\coth(2\omega_1)-\frac{1}{2\omega_1}\right)\right]\, .$$ In similar fashion, we can obtain the following expressions for the other matrix elements. $$\begin{aligned} \frac{A}{2\pi b} \frac{1}{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}\tilde{\mathsf T}_{-1,-1}(\omega_{-1})\!&=&\!\beta{^{\raise.15ex\hbox{${\scriptscriptstyle -}$}\kern-.05em 1}}e^{-\alpha_0}\left[ 1+{b}\left(\frac{\zeta f}{\kb T}-\frac{\omega_{-1}}{2 \beta A}\right)\left(\coth(2\omega_{-1})-\frac{1}{2\omega_{-1}}\right)\right] \nonumber \\ \frac{A}{2\pi b} \frac{1}{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}\tilde{\mathsf T}_{1,-1}(\omega_1,\omega_{-1})\!&=&\!\frac{e^{-2\gamma}}{\Nxterm} \left( \frac{\omega_1 \omega_{-1}}{\sinh(2\omega_1)\sinh(2\omega_{-1})}\right)^{\frac12} \left( \frac{2 \sinh(\omega_1+\omega_{-1})}{\omega_1+\omega_{-1}}\right)\,.\end{aligned}$$ To obtain a nontrivial continuum limit we must now specify how the parameters $A$, $\alpha_0$, and $\gamma$ depend on $b$ as $b \to 0$. The choices $$\alpha_0=-\frac12\ln\beta+b\bar\alpha\,,\qquad \gamma=-\frac12\ln(\bar g b)$$ give a well-defined limit, where we hold $A$, $\bar\alpha$, $\beta$ and $\bar g$ fixed as $b\to0$. With these choices, the matrix $\frac1{{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}}\tilde{\sf T}(\omega_{1},\omega_{-1})$ takes the form $$\label{tmatcons} \frac{1}{\|\vec v_{\omega_{1},\omega_{-1},\varphi}\|^2}\tilde{\sf T}(\omega_{1},\omega_{-1})=\frac{2 \pi b}{A \sqrt{\beta}}\left({\mathsf 1} +b \begin{pmatrix} {\cal P} & {\cal Q} \\ {\cal Q} & {\cal R} \end{pmatrix}\right)\, ,$$ with $$\begin{aligned} {\cal P}&=&\bar\alpha + \left(\frac{f}{\kb T}-\frac{\omega_1}{2A}\right)\left(\coth(2\omega_1)-\frac{1}{2\omega_1}\right)\nonumber\, ,\\ {\cal R}&=&-\bar\alpha + \left(\frac{\zeta f}{\kb T}-\frac{\omega_{-1}}{2A\beta }\right)\left(\coth(2\omega_{-1})-\frac{1}{2\omega_{-1}}\right)\nonumber \, ,\\ {\cal Q}&=&\frac{\bar g \sqrt{\beta}}{\Nxterm}\left( \frac{\omega_1 \omega_{-1}}{\sinh(2\omega_1)\sinh(2\omega_{-1})}\right)^{\frac12} \left( \frac{2 \sinh(\omega_1+\omega_{-1})}{\omega_1+\omega_{-1}}\right)\, .\end{aligned}$$ Note that the prefactor $\frac{2\pi b}{A \sqrt\beta}$ in does not contribute to the force-extension result , because it does not depend on the force. In terms of the individual matrix entries, the quantity to be maximized now reads (see ): $$\label{lamom} \ln \tilde y(\omega_{1},\omega_{-1})=\frac{b}{2}\left({\cal P}+{\cal R}+\sqrt{({\cal P}-{\cal R})^2+4{\cal Q}^2}\right)\, .$$ Writing $\Omega\equiv b{^{\raise.15ex\hbox{${\scriptscriptstyle -}$}\kern-.05em 1}}\ln\lambda^*_{\text{max}}= b{^{\raise.15ex\hbox{${\scriptscriptstyle -}$}\kern-.05em 1}}\times\max \ln \tilde y(\omega_{1},\omega_{-1})$, the force-extension in the continuum limit is finally given by $$\langle \frac{z}{L_{\rm tot,b}}\rangle=\kb T\frac{{\rm d}\Omega}{{\rm d} f} \, .$$ We evaluate $\Omega$ by numerically maximizing . So far, we have not included stretch moduli for the B- and S-DNA. This is easily implemented to first order in $f/\Nestiff$ by replacing $f$ with $f(1+\frac{f}{2 \Nestiff^{(\mathrm{S,B})}})$ in the matrix elements for the two states respectively (). This procedure yields theoretical force-extension curves like the one plotted in Fig. (\[smithfit\]). In summary, our model contains the following seven parameters. $2 \bar \alpha \kb T$ is the free energy per unit length required to flip B-DNA into the S-state, and is measured in \[J/nm\]. ${\cal Q}$ measures the cooperativity of the transition and has units \[1/nm\]. $A$ is the bend stiffness parameter of B-DNA, with units \[nm\]. The dimensionless parameter $\beta$ is the ratio of the B- and S-DNA bend stiffnesses. $\Nestiff^{(\mathrm{B})}$ and $\Nestiff^{(\mathrm{S})}$ are the stretch stiffnesses of B and S-DNA, and are measured in pN. Finally, $\zeta$ is the dimensionless elongation factor associated with the B$\to$S transition. Discussion of fit\[fitdisc\] ============================ Fig. (\[smithfit\]) shows a fit to some recent experimental data (similar data appear in [@bust01a]). Our model reproduces the experimental data rather well, but with so many fit parameters one may ask whether it actually makes any falsifiable predictions. To answer this question we note that the data below the transition suffice to fix $A$ and $\Nestiff^{(\mathrm{B})}$ as usual, roughly speaking from the curvature and slope of the curve below the transition. Similarly, the data above the transition fix $A^{(\mathrm{S})}=\zeta\beta A$ and $\Nestiff^{(\mathrm{S})}$. The vertical jump in the curve at the transition fixes $\zeta$. The horizontal location of the jump fixes $\bar\alpha$, and the steepness of the jump fixes the cooperativity ${\cal Q}$.[^2] Thus all of the model’s parameters are fixed by specific features of the data. Two additional, independent features of the data now remain, namely the rounding of the curve at the start and end of the transition. Our model predicts these features fairly succesfully. The fit recovers the known values for the effective persistence length of B-DNA of around $50\,$nm and its stretch modulus of about $1000\,$pN. Our first result is that the bend stiffness of S-DNA from our fit as $A^{(\mathrm{S})}=\beta\zeta A=12.32\,$nm. Similar results were obtained using the older data of Cluzel [*et al.*]{} [@Cluzel; @promise]. If S-DNA consisted of two unbound, single strands, we might have expected $A^{(\mathrm{S})}$ to be twice as large as the value $A^{\rm ss}\approx0.75\,$nm appropriate to single-stranded DNA (as obtained from stretching experiments on ssDNA, restricted to forces above those required to pull out secondary structure [^3] [@Smith; @Clausen; @promise]). On the contrary, we find that the bend stiffness of S-DNA is intermediate between that of B-DNA and that of two single strands. Finally, our fit gives the stretch modulus of S-DNA is substantially higher than that of B-DNA. This conclusion is consistent with the idea that the contour length of S-DNA is determined by its covalently bonded sugar-phosphate backbones, which are much straighter than in B-DNA; the contour length of B-DNA is instead determined by weaker, base-stacking interactions. Relation to prior work ====================== Several authors have studied the entropic elasticity of two-state chains. As soon as the overstretching transition was discovered, Cluzel proposed a pure Ising model by analogy to the helix-coil transition [@cluz96a]. Others then introduced entropic elasticity, but required that both states have the same bending stiffness as B-DNA [@mark98a; @Ahsan] or took one of the two states to be infinitely stiff [@tama01], or to be a FJC [@RouzinaBloomfield1]. Also several earlier works made a mean-field approximation instead of diagonalizing the full transfer matrix. We believe our Ising-DPC model to be the first consistent formulation incorporating the coexistence of two different states with arbitrary elastic constants. Our approach also is calculationally more straightforward than some, and minimal in the sense that no unknown potential function needs to be chosen. We thank T. Burkhardt, D. Chatenay, A. Grosberg, R. Kamien, J. Marko and M. Rief for valuable discussions, and C. Bustamante, D. Chatenay, J.-F. Léger, J. Marko, M. Rief, and S. Smith for sending us experimental data. CS acknowledges support from NIH grant R01 HL67286 and from NSF grant DMR00-79909. PN acknowledges partial support from NSF grant DMR98-07156. ![Least-squares fit of the Ising-DPC model to an overstretching dataset (48.5kbp $\lambda$ DNA construct; buffer 500mM NaCl, 20mM Tris, pH 8). Data kindly supplied by C. Bustamante and S. Smith. Fit parameters: $A=43.75\,\nmunit$, $\bar\alpha =5.45\, \nmunit^{-1}$, $\beta=0.16$, ${\cal Q}=0.13\, \nmunit^{-1}$, $\zeta=1.76$, $\Nestiff^{(\mathrm B)}=1.2\cdot 10^3\,\pNunit$ and $\Nestiff^{(\mathrm S)}=1.0\cdot10^4\,\pNunit$. $\chi^2=9.22$ at $N=825$; points with $1.11<\langle\frac{z}{L}\rangle < 1.55$ were excluded from the fit.\[smithfit\]](smithfit.eps){width="4truein"} [0]{} . . . . . . . . . . . . . . . . . . [^1]: E-mail: [^2]: The fit value of $\bar\alpha$ should be regarded as an average of the two different costs to convert AT or GC pairs. The fit value of ${\cal Q}$ has no direct microscopic significance, as the apparent cooperativity of the transition will be reduced by the sequence disorder. [^3]: Hagerman’s result that the persistence length of a single strand of poly(dT) DNA is between 2 nm and 3 nm [@Hagerman] does not come from a stretching experiment and should not be compared directly.

ArXiv

--- abstract: 'Metal wire networks rely on percolation paths for electrical conduction, and by suitably introducing break-make junctions on a flexible platform, a network can be made to serve as a resistive strain sensor. Several experimental designs have been proposed using networks made of silver nanowires, carbon nanotubes and metal meshes with high sensitivities. However, there is limited theoretical understanding; the reported studies have taken the numerical approach and only consider rearrangement of nanowires with strain, while the critical break-make property of the sensor observed experimentally has largely been ignored. Herein, we propose a generic geometrical based model and study distortion, including the break-make aspect, and change in electrical percolation of the network on applying strain. The result shows that when a given strain is applied, wire segments below a critical angle with respect to the applied strain direction end up breaking, leading to increased resistance of the network. The percolation shows interesting attributes; the calculated resistance increases linearly in the beginning and at a higher rate for higher strains, consistent with the experimental findings. In a real scenario, the strain direction need not necessarily be in the direction of measurement, and therefore, strain value and its direction both are incorporated into the treatment. The study reveals interesting anisotropic conduction features; strain sensitivity is higher parallel to the strain, while strain range is wider for perpendicular measurement. The percolation is also investigated on direct microscopic images of metal networks to obtain resistance-strain characteristics and identification of current percolation pathways. The findings will be important for electrical percolation in general, particularly in predicting characteristics and improvising metal network-based strain sensors.' author: - Ankush Kumar bibliography: - 'references.bib' title: Electrical percolation in metal wire network based strain sensors --- Introduction ============ Electrical transport in a metal wire network has been an area of keen interest for their fascinating percolation properties [@balberg1984percolation; @vzevzelj2012percolating; @lebovka2018anisotropy] and extensive range of promising transparent conducting electrode applications. [@ye2014metal; @kumar2016evaluating; @mutiso2013integrating] A strain sensor is an emerging application of metal wire networks with broader strain working range, extraordinary response and high transmittance, desired for robotics and health monitoring systems applications. [@yao2014wearable; @yamada2011stretchable; @trung2016flexible] The metal network coupled with flexible substrate forms resistive-type strain sensor, its unknown value of applied strain can be determined, using the measured value of resistance. Various efforts have been devoted in the direction of fabricating metal network based strain sensors using silver nanowires, carbon nanotubes and template based metal meshes  [@xu2012highly; @amjadi2014highly] and their characteristics have been studied in detail. The studies exhibit that the resistance increases linearly for a lower strain, while non-linearly for higher strain values, resistance variation in the parallel and perpendicular direction of strain are different [@kim2015highly; @gupta2018cosmetically] and sensitivity is higher by using binary width distribution. [@duan2018highly] It is yet a challenge to improve their sensitivity and strain range significantly, for diverse applications. A comprehensive theoretical model is a need of the hour to explain their mechanism and discuss the pathway for their further improvisation. The electrical properties of metal networks are primarily investigated using the numerical method and have successfully calculated the sheet resistance with various parameters such as nanowire density, [@vzevzelj2012percolating] length distribution [@mutiso2013integrating], anisotropy [@lebovka2018anisotropy] etc. Besides, effective medium theory [@o2016effective; @he2018conductivity], block matrix approach [@kim2018systematic], excluded volume percolation model [@mutiso2012simulations] have also been discussed for deeper insights. In this direction, we attempted to model it based on geometrical consideration and successfully obtained a relationship of sheet resistance and current carrying region with geometrical parameters of the network, [@kumar2016evaluating; @kumar2017current] which is found to be in experimental [@darmakkolla2019morphology] and numerical agreement. [@kim2018systematic] However, modeling of metal network based strain sensor is intensely involved, as it comprises distortion of the network with a given strain and then analyzing electrical transport of the distorted networks with the strain. Shengbo et al. [@shengbo2018highly] have presented schematic representation for strain effect; however, theoretical or numerical calculations were not performed in the study. Amjadi et al. [@amjadi2014highly] and Kim et al. [@kim2015highly] developed 3-dimensional models of silver nanowire network and examined its resistance variation numerically during elongation. Yao et al. [@yao2018characterizing] have very recently determined percolative coefficients, which are quantifiers of electrical transport properties, using the strain sensor characteristics, exhibiting the interesting utility of strain studies for electrical percolation understanding. All these existing studies account for elongation and rearrangement of nanowires, however, do not include breakdown of wire segments into the treatment. On the other hand, recent microscopic study by Gupta et al. [@gupta2018cosmetically] reveals that wire segments, particularly for metal mesh, break during the strain and being half embedded in the elastic substrate, the broken wire segments reconnect at the same places on restoring strain, leading to the recovery of conductance for several such cycles. The present work pertains to the modeling of metal-network based strain sensor as a 2-dimensional graph; the application of strain distorts the network and breaks a few wire segments which in turn, modifies its electrical percolation significantly. Moreover, we calculate resistance and gauge factor as a function of strain and compare it for different strain directions and discuss some approaches for their improvement. Results and discussion ====================== To describe the mechanism of such sensors, consider a random conducting network as a 2D graph [@kumar2016evaluating] of size $a \times b$, where wire segments signify the edges and junctions denote nodes of the graph. We assume that the network elongates along the direction of strain and increase in resistance is primarily due to the breakdown of specific edges (wire segments). A wire segment in a network breaks, if the experienced strain exceeds the critical strain for breakdown ($\lambda$&gt;$\lambda_c$), here, critical strain, $\lambda_c$ depends on the strength of the wire (Fig.  \[strain\_sensor1\]a). Further, we understand the role of strain in a network (Fig.  \[strain\_sensor1\]b) with $\lambda$ being elongation along the horizontal direction. Fig.  \[strain\_sensor1\]c-d represents network with length $\lambda $ = 1.004 and 1.04 times $\lambda_c$. Since all the edges have different angles w.r.t direction of strain and hence experience unequal strains. The strain experienced ($\lambda_E$) by an edge having angle $\theta_i$ w.r.t direction of strain is $\lambda_{E}= \lambda cos\theta_i $. The wire segments break if $\lambda_E$&gt;$\lambda_c$, in other words, the wire segments with an angle lesser than critical angle end up breaking ($\theta_i$ &lt;$\theta_C$). The critical angle can be defined as $$\theta_C = \cos^{-1} \frac{\lambda_c}{\lambda} \label{thetab}$$ As an example, with the strain values 1.004 $\lambda_c$ and 1.04 $\lambda_c$, edges with angle less than 5$^0$ and 16$^0$ respectively end up breaking. To model its electrical resistance , lets begin with, $\alpha =0$ i.e. resistance is measured along the direction of the strain as $R^{\|}$ (see Fig. \[strain\_sensor1\]e left). As, edge length $<<$ network size, one can assume potential drops uniformly between the verticle electrodes, and equipotential lines are perpendicular to the electric field. Let $L_i$ be the length of wire segment (edge) placed in the electric field $E$ at an angle $\theta_i$. The potential difference across the edge, $V_i$, therefore depends on the orientation of edge as $$\label{Vr} V_{i} = E L_{i} cos \theta_{i}$$ As wire segments with an angle lesser than $\theta_C$ are broken with the starin, thus the average potential of all unbroken wire segments can be calculated as $$\label{Vr_avg} V_{am} = \frac{ \int_{\theta_C}^{\frac{\pi}{2}} EL_i cos\theta_i d\theta_i} {\int_{\theta_C}^{\frac{\pi}{2}} d\theta_i} = EL \frac{1 - sin\theta_C}{\frac{\pi}{2}-\theta_C}$$ ![ (a) Comparision of variation in resistance parallel to strain, perpendicular to strain and biaxial strain with respect to the density of broken wire segments in upper axis and the critical angle of the breakdown in the upper axis. (b) Variation of resistance with respect to strain values with scale $\frac{\lambda_c}{\lambda}$ in upper axis and typical values for Au network in the lower axis. (c) Variation of gauge factor for Au network for the increasing strain. (d) Variation in resistance, $R$ with respect to strain angle at different strain values[]{data-label="strain_sensor3"}](figure2.png){width="50.00000%"} Say, $\rho$ is the resistivity, $w$ is the width, and $t$ is the thickness of the wire segment (edge), then the mean resistance of an individual edge is $ \frac{\rho L_{am}}{w t}$. Hence, mean current, $I_{am}$ passing through an individual edge can be written as $$\label{ir2} I_{am}= \frac{E wt} {\rho} \frac{1-sin\theta_C}{\frac{\pi}{2}-\theta_C}$$ Consider, $N$, $N_b$ and $N_u$ as the edge density of total, broken and unbroken wire segments respectively. The current passing across an equipotential line, $I_{eq}$, depends on the current carried by a single edge and a total number of unbroken edges on the equipotential line which equals $\sqrt{N_u} \, b $ by symmetry arguments. $$\label{i} I_{eq}= \frac{E wt} {\rho} \frac{1-sin\theta_C}{\frac{\pi}{2}-\theta_C} \sqrt{N_u}b$$ Using, $E = \frac{V}{a} $ and applying Ohm’s law, resistance, $R^{\|}$ can be written as $$\label{R11} R^{\|} = \frac { \rho} {wt\sqrt{N_u}} \frac{\frac{\pi}{2}-\theta_C} {1-sin\theta_C} \frac{a}{b}$$ As the wire segemnts with angles up-to angle $\theta_C$ are broken, hence $ \sqrt{N_u} = \frac{\frac{\pi}{2} -\theta_C}{ \pi /2} \sqrt{N} $. If $R^0$ is the resistance, in the unstreched state. $$\label{R311} R^{\|} = \frac{R^{0}} {1-sin\theta_C}$$ Using $\theta_C$ = $ \cos^{-1} \frac{\lambda_c}{\lambda}$ from Eq.  \[thetab\], above can written as $$\label{R51} R^{\|} = \frac{R^{0}} {1-sin(cos^{-1} \frac{\lambda_c}{\lambda})} =\frac{R^{0}} {1- \sqrt{ 1- \big(\frac{\lambda_c}{\lambda}}\big)^2}$$ Using the first term of series expension, for lower strain values, the resistance shows linear increase as $$\label{R5178} R^{\|} \approx 2 R^{0} \frac{\lambda}{\lambda_c}$$ Now, if one is measuring resistance perpendicular to the direction of strain, the average potential difference can be written as $V_{am}^{\bot} = \frac{ \int_{0}^{\frac{\pi}{2}-\theta_C} EL_i cos\theta_i d\theta_i} {\int_{0}^{\frac{\pi}{2}-\theta_C} d\theta_i}$ and the resistance,$R^{\bot}$ can be calculated as $$\label{R311p} R^{\bot} = \frac{R^{0}} {cos\theta_C} = {R^{0}} \frac{\lambda}{\lambda_c}$$ Further, consider a generic case having an angle of strain, $\alpha$ with respect to the resitance measurement direction (Fig.  \[strain\_sensor1\]e right). As derived in supporting information, $R$ depends on value of $\alpha +\theta_C$. For $\alpha +\theta_C <\frac{\pi}{2}$, based on Eq. S8, $$\label{Rx1} R = \frac{R^{0}} { 1-cos\alpha \, sin\theta_C}$$ and for $\alpha +\theta_C > \frac{\pi}{2}$, based on Eq. S13, $$R = \frac{R^{0}} { sin\alpha \, cos\theta_C} \label{Rx11}$$ In this way, one can calculate resistance for any given value of strain and its direction, which is necessary in most of the biological and robotics applications, where the strain direction need not necessarily be in the direction of measurement. Similarly, any other experimentally interesting case, such as biaxial strain [@ryu2015extremely] can also be studied by the treatment. For biaxial strain, with $\lambda$ being starin along both the direction, the resistance can be expressed as $$\label{Rb} R^{B} = \frac{R^{0}} {cos\theta_C - sin\theta_C} = \frac{R^{0}} { \frac{\lambda_c}{\lambda} - \sqrt{ 1- \big(\frac{\lambda_c}{\lambda}}\big)^2}$$ Based on the above calculations, Fig.  \[strain\_sensor3\]a represents variation in resistance with an increase in critical angle in lower axis and edge density of broken wires, in the upper axis as $\frac{\theta_C}{\pi/2} = \sqrt \frac{N_b}{N}$. The increase in resistance is gradual in the beginning and rises abruptly for larger critical values or broken edge densities. Fig.  \[strain\_sensor3\]b shows the results in terms of critical strain ($\lambda_c$), and for typical strain value for Au. Here, $\lambda_c$ for Au considered is approximately $1.6 \times 10^{-3}$ calculated using $\lambda_c= \frac{G}{Y}$ by substituting typical Yield strength $(G)$ of 50 to 200 $MPa$ and Young Modulus $(Y)$ 8 GP for Au. [@wu2005mechanical] It is clear, for resistance measurement along the direction of strain (black curve), the resistance increases linearly for lower strain, while it increases drastically for higher strain values. Further, the working range of the strain sensor in all cases is different. For parallel strain (black curve), resistance shoots up at relatively lower strains as compared to the perpendicular strain (red curve); thus higher strain measurement may be more appropriate to measure in the perpendicular direction. For biaxial strain, resistance shoots up for relatively smaller strains and thus such sensors may not be suitable to measure higher biaxial strains. The phenomena can be explained by considering the importance of the current carried by the broken wire segments. For perpendicular strain, broken wire segments are nearly on equipotential lines carrying negligible current; thus, there is very less effect on resistance after their breakdown. On the other hand, for parallel strain, broken wire segments are nearly in the direction of the electric field carrying maximum current; thus their breakdown leads to more increase in overall resistance. For biaxial strain, both kinds of wire segments are broken, leading to their maximum increase in resistance. Gauge factor defined as $\frac{\bigtriangleup R}{\lambda}$ can also be derived easily and shown in Fig. \[strain\_sensor3\]c. In this way, one can choose and understand the limitations of measurement mode based on sensitivity and working strain range. The decrease in resistance with increasing strain angle from parallel to perpendicular is also clear in (Fig.  \[strain\_sensor3\]d) and can be used to measure resistance change for arbitrary strain direction. Further, we demonstrate the pertinence of the model on a microscopic image of a metal wire network as an example (Fig.  \[image\_strain\]). The analysis involves converting a metal network image into 2- dimensional graph using image analysis, distorting the graph for various strain values and numerically calculating their percolating path and corresponding sheet resistance values. Only the current carrying backbone is shown in the initial image for clarity (Fig.  \[image\_strain\]a); with the applied strain, edges with projection angle lower than the critical angle, break down (white), and some edges become dangling with no current (blue), and current flows in the rest of current carrying backbone region (red). Interestingly, the current carrying backbone (red) gradually decreases for lower strain values, $\lambda = 1.004\lambda_c$ (Fig.  \[image\_strain\]b) while for higher strain values, $\lambda = 1.04\lambda_c$ (Fig.  \[image\_strain\]c), the current carrying backbone forms bottleneck (Fig.  \[image\_strain\]d), in the end, at $\lambda = 1.15\lambda_c$, there is no current carrying path. The resistance is determined for all the networks, using the two-point resistance method, based on solving Kirchhoff’s law in the network.[@kumar2017current] Resistance is found to increase linearly for lower strain (see inset of Fig.  \[image\_strain\]f) and increases with the higher rate for higher strain values before the breakdown of percolation pathway (see Fig.  \[image\_strain\]f). In this way, interesting percolation scenareos can be seen to work in different strain regimes. The model is generic and can be applied to many special cases of metal networks and different strain modes. If the deposited wire networks have good adhesion with the substrate, the nanowires cannot individually move, and network as a whole elongates, as discussed in the model. The technique can compare various microscopic images of metal networks, and help in the selection of the optimum one, for strain sensor applications. Additionally, the model can be utilized in understanding the robustness of electrical percolation in the networks for stretchable electronics applications [@guo2014highly], where a minimum variation of resistance with a strain is desired. The study brings out important features of anisotropic conduction and its implications in strain sensor characteristics. The present day strain sensor devices only measure strain values, the model may be an important step towards realizing strain sensor devices, which could measure strain value and its direction simultaneously. Note that the analytical treatment is based on applying effective medium theory, which may not hold good at very sparse network in case of very high strains; image analysis based numerical approach discussed here, should be applied on bigger images in such contexts. For a more accurate and detailed picture, future work should improvise the model by calculating dangling regions with strain, introducing corrections for a sparse network, and introducing self-healing effects [@kumar2019self] and quantum conductance of interconnects. The present model estabilishes important modeling ideas, which will be helpful in approaching much complex scenareos. Conclusion ========== In conclusion, a geometrical model is presented to understand the impact of strain on the electrical percolation properties of metal wire networks. The analysis explains that strain applied to the network leads to a preferential breakdown of certain wire segments based on their angle with respect to the strain direction, which increases their resistance. The calculated resistance increases linearly for the lower strain values and abruptly for higher strains, consistent with experiments. The electrical percolation manifests interesting anisotropic effects; the variation in resistance with strain is higher along the direction of strain, while the strain range is higher if the measurement is performed perpendicular to the strain. The study devised a numerical approach to obtain the percolating pathways and strain sensor characteristics from microscopic images of metal networks. The analysis suggests that the current carrying region decreases gradually for lower strains, while forms a bottleneck pathway at higher strains and finally collapse completely at very high strain values. The proposed model and its interesting findings will be useful in addressing various electrical percolation problems and improvising strain sensor devices. See supplementary material for detailed calculations of resistance variation for strain in an arbitrary direction. Acknowledgements {#acknowledgements .unnumbered} ================ The author acknowledges his Ph.D. mentor Professor Giridhar U. Kulkarni for valuable suggestions, encouragement, and proposing the problem based on lab experimental observations and acknowledges the Department of Science and Technology, India for financial support.

ArXiv